COURSE NOTES

Traffic Engineering and

Intelligent Transport Systems
Gaetano Fusco

Volume 1: Traffic Flow Theory

Table of Contents
CHAPTER 1 MACROSCOPIC MODELS OF TRAFFIC FLOW ....................................................................... 4
1.1 FUNDAMENTALS OF TRAFFIC FLOW............................................................................................................ 4
1.1.1 Conservation Equation ............................................................................................................. 4
1.1.2 State Equation .......................................................................................................................... 5
1.2 CONGESTION FUNCTIONS ........................................................................................................................ 7
1.2.1 Elementary Spacing Model ....................................................................................................... 7
1.2.2 Greenshields Model ................................................................................................................ 11
1.2.3 Greenberg Model.................................................................................................................... 13
1.2.4 Comparison of Different Models ............................................................................................ 16
1.3 DYNAMIC FIRST-ORDER MODEL ............................................................................................................. 17
1.3.1 Theory of Kinematic Waves .................................................................................................... 17
1.3.2 Departure Model .................................................................................................................... 19
1.3.3 Shock Waves ........................................................................................................................... 24
1.4 SECOND-ORDER TRAFFIC MODELS .......................................................................................................... 31
1.4.1 The Density Gradient Model ................................................................................................... 31
1.4.2 Payne’s Dynamic Model ......................................................................................................... 33
1.5 BIBLIOGRAPHY ..................................................................................................................................... 34
CHAPTER 2 CAR-FOLLOWING MODELS ............................................................................................... 35
2.1 FORMULATION OF THE CAR-FOLLOWING MODELS ...................................................................................... 35
2.1.1 Formulation of the Stimulus-Response Model ........................................................................ 35
2.1.2 Collision Avoidance Models .................................................................................................... 36
2.1.3 Perception Threshold Models ................................................................................................. 37
2.1.4 Lower Order Newell’s model .................................................................................................. 41
2.2 STABILITY ANALYSIS .............................................................................................................................. 44
2.2.1 Local Stability ......................................................................................................................... 45
2.2.2 Asymptotic Stability ................................................................................................................ 49
2.3 ANALYSIS OF THE STATIONARY STATE ....................................................................................................... 53
2.3.1 Linear Model ........................................................................................................................... 53
2.3.2 Non-Linear Models ................................................................................................................. 55
2.4 EXPERIMENTAL RESULTS ........................................................................................................................ 58
2.4.1 General Motors Experiments .................................................................................................. 58
2.4.2 Experiments with Differential GPS .......................................................................................... 61
2.5 BIBLIOGRAPHY ..................................................................................................................................... 65
CHAPTER 3 TRAFFIC MODELS AT INTERSECTIONS............................................................................... 67
3.1 GENERAL ............................................................................................................................................ 67
3.2 GAP ACCEPTANCE MODEL ..................................................................................................................... 69
3.3 PROBABILISTIC ARRIVAL MODELS ............................................................................................................ 71
3.3.1 Negative Exponential Distribution .......................................................................................... 72
3.3.2 Translated Negative Exponential Distribution ........................................................................ 73
3.3.3 Dichotomized Distributions .................................................................................................... 74
3.4 QUEUEING MODEL ............................................................................................................................... 76
3.4.1 Fundamental Relations and Variables .................................................................................... 77
3.4.2 Hypotheses on the Queue Model at an Intersection .............................................................. 79
3.4.3 Stationary Queue Model M|M|1 ........................................................................................... 79
3.4.4 Stationary Queue Model M|G|1 ............................................................................................ 82
3.4.5 Time-dependent queue model ................................................................................................ 86
3.5 TRAFFIC MODEL AT UNSIGNALIZED INTERSECTIONS..................................................................................... 89

2
 3.5.1 Hierarchy between maneuvers ............................................................................................... 89
3.5.2 Capacity of a Rank 2 Traffic Stream ....................................................................................... 91
3.5.3 Delay Model for Unsignalized Intersections ........................................................................... 93
3.5.4 Queue length .......................................................................................................................... 94
3.6 TRAFFIC MODEL AT SIGNALIZED INTERSECTIONS ......................................................................................... 95
3.6.1 Signal Settings at Road Intersections ..................................................................................... 95
3.6.2 Departure Model at a Traffic Signal ....................................................................................... 96
3.6.3 Delay model .......................................................................................................................... 100
3.7 BIBLIOGRAPHY ................................................................................................................................... 104
Chapter 1 Macroscopic Models of Traffic Flow
1.1 Fundamentals of Traffic Flow
The realization of transport implies by its nature the movement of people or things. Depending on
the technological systems used, the movement can take place using vehicles (i.e., excluding the
pedestrian mode in the movement of people and the mode of transport by pipeline in the case of goods)
and either linear or nodal transport infrastructures (for example, naval and air transport do not use linear
infrastructures). The motion of the individual vehicle depends on the mechanical characteristics and its
interactions with the road and with the external environment. The study of the laws that describe the
motion of the isolated vehicle concern the mechanics of locomotion. However, while driving on the
street, the vehicles condition each other. The modes of movement of a stream of vehicles depend on the
physical and behavioral laws that determine its qualitative and quantitative characteristics: travel times,
operating costs, level of service, level of safety, and impacts on the external environment. The discipline
that studies them is called traffic flow theory.
This paragraph introduces the fundamental relationships of physical nature that are common to the
circulation of different modes of transport, while the next paragraphs will deal with specific details of
the individual modes.
Fundamentals of traffic flow theory, which are summarized in this text, assimilate the movements
of vehicles along the way to a fluid stream: therefore, the representation of the motion of each unit
(vehicle or pedestrian) in the traffic stream is ignored; however, it is described at a macroscopic level
through the characteristic variables that are also used in fluid mechanics:
• flow (generally indicated with the symbol q): quantity (i.e.: number of vehicles or people) traveling
in the unit of time through a given section of the facility;
• speed (generally indicated with the symbol v): average speed of the traffic stream in a given instant
(i.e.: average speed of the vehicles that at a given instant are in a unit of space of the facility);
• density (generally indicated with the symbol k): quantity (i.e.: number of vehicles or people) that in
a given time interval are in a unit of space of the facility.
In the general case, the traffic stream moves with a non-uniform motion: therefore, the values of the
characteristic variables of the flow vary in time and space. Under steady-state driving conditions, it can
be assumed that the speeds of vehicles (or people) of a homogeneous traffic stream (i.e.: of vehicles of
the same type) vary insignificantly for the purpose of studying the overall behavior of the traffic stream.
This state of flow is called stationary state and is characterized by constant values in time and space.

1.1.1 Conservation Equation

A stream of moving vehicles (or people) on a transport facility must comply with the physical
condition of conservation of the mass (i.e.: the number of vehicles or people traveling through a finite
element of space). Referring to an element of a linear infrastructure, for example, a road or railway
section or a section of an escalator, this condition requires that the variation in the number of vehicles
that are present on the section in a given time interval is equal to the difference between the number of
vehicles entered and the number of vehicles exited in the same time interval. Figure 1-I provides a
graphical representation of the conservation (or continuity) condition.

4
 Traffic Engineering and Intelligent Transportation Systems

x + Dx 1 2 3 4
6
4
5
3
4
3 2
2 1
x 1
1 2

t t+Dt

Figure 1–I. Graphical representation of the conservation equation.

In formal terms, referring to a road segment of length Δx> 0 with origin at a generic abscissa x and
to a generic time interval Δt> 0, without entering or exiting vehicles, we can write the conservation
equation using the characteristic variables of traffic flow by imposing that the difference between the
number of vehicles (or number of people or quantity of freight) at the end (t + Δt) and at the beginning
(t) of the time interval is equal to the difference between the entering flow at the initial section x and the
exiting flow from the final section x + Δx:

k (x, t + Dt )Dx - k (x, t )Dx = q(x, t )Dt - q(x + Dx, t )Dt (1.1)
Transferring the increments Δx and Δt to the denominator and changing the sign to the second
member, we obtain the increment ratios of the functions k (x, t) and q (x, t):
k ( x, t + Dt ) - k ( x, t ) q( x + Dx, t ) - q( x, t )
=- (1.2)
Dt Dx
Introducing the limit for Δx → 0 and Δt → 0, we obtain the classic expression of the conservation
equation, which also holds for any fluid:
¶k ( x, t ) ¶q( x, t )
+ =0 (1.3)
¶t ¶x

1.1.2 State Equation

At steady state the three characteristic quantities of traffic flow are not independent of each other,
but are linked by a relationship, called the state equation, which allows one variable to be expressed as
a function of the other two:
q = kv (1.4)
The state equation can be effectively illustrated through a simple graphical demonstration. Let us
consider a section of a transport infrastructure of length Δx, traveled by vehicles (or people) running at
an approximately constant speed, not being necessarily the same for all vehicles, for an observation
interval Δt. Vehicles traveling at different speeds, however, do not interact with each other. This
condition corresponds to the motion of vehicles of different categories (fast and slow) on a multi-lane
highway, or to a moving walkway on which a fraction of users walk and another remains stationary.
Figure 1–II illustrates the case of two vehicular classes, denoted with the indices 1 and 2.

5
 Macroscopic Models of Traffic Flow

h1 v1 h2
s1 v2
x + Dx
s2

t t+Dt
Figure 1–II. Graphical representation of the state equation.

It is straightforward to verify from the geometry of the figure, or from elementary physics
considerations, that the spacing s between two vehicles of the same class is given by the product of the
headway h for the speed v, that is: sl = vlhl.
It is to note now that the flow q of vehicles through any given section of the infrastructure is equal
to the inverse of the average spacing in time, h. In fact, provided that Δt is quite large compared to h,
the number n of vehicles traveling in the interval Δt through the given section is:

n n 1
q= ≅ n =
Δt h (1.5)
∑ hi
i=1

We can apply the same reasoning to the density, which is equal to the inverse of the average spacing
s:

n n 1
k= ≅ n =
Δx s (1.6)
∑ si
i=1

Substituting the flow and the density respectively for the headway and the spacing, we obtain, for a
generic class l:

q l = vl k l (1.7)
Summing up the vehicles of the m different classes, we get the total flow and the total density:
m m
q= åq ; k = åk
l =1
l
l =1
l (1.8)

Substituting equation (1.7) in the first expression of (1.8) and dividing numerator and denominator
by the total density k, we finally have:
m m

åv k å
vl k l
q= l l =k = kv s (1.9)
k
l =1 l =1

It is worth noting that the fractional term under the summation represents the average speed of the
traffic stream, computed, according to the definition provided in the previous subsection, as the average

6
 Traffic Engineering and Intelligent Transportation Systems

value of the speeds of the vehicles that are present at a given moment on an infrastructure segment. The
subscript s highlights that the average has been calculated with respect to space: indeed, this is not the
only possible definition of the average speed. It is possible to define an average speed over time, defined
as the average of the speeds of vehicles passing during a given time interval through a given section of
the infrastructure. It is useful to specify that the two averages generally provide different values. In
particular, the average over time generally gives values greater than the average in space1.

1.2 Congestion Functions

The equations introduced in the two previous paragraphs are of physical nature and are valid to
describe the flow of any traffic stream, being it a fluid, a particle, a vehicle or a person.
Furthermore, in traffic streams of vehicles or people, the motion of the individual unit is governed
by the driving behavior (or gait in the case of pedestrians), which derives from safety or comfort
considerations and induces drivers (or pedestrians) to adjust their speed and trajectory according to the
position and the movement of the other vehicles (or pedestrians).
The implementation of safety and comfort conditions is generally assisted by a regulation system,
which acts in ways that vary from one mode of transport to another and, in the same mode, that depend
on the road element. The regulation system consists of a set of traffic rules (road code, navigation code,
rail traffic regulation).
Furthermore, the mechanisms that regulate the flow are generally different for nodes (intersections
between different road elements and stops of public transport systems) and for links (road elements
between one node and another).
Safety conditions on links essentially consist in keeping the safety distance from the preceding
vehicle (or pedestrian) along the trajectory followed. Safety conditions on nodes, where different traffic
streams share the use of a common element of the road, concern the drivers’ decision of entering the
common area or not.
In road traffic, implementation of safety conditions on links is entirely left to the driver. This driving
mode is called free-density. However, there are modes of transport in which the spacing between
vehicles is fixed, such as cableway installations, or is controlled by a signaling system, such as railways.
These modes are called controlled-density.
Different models have been developed to describe vehicular congestion on road links, starting from
theoretical considerations –that is, following a deductive approach– or starting from experimental
considerations –following an empirical or inductive approach. The elementary spacing model, reported
in paragraph 1.2.1, was derived from considerations on driving safety for an individual vehicle and is,
therefore, an example of a deductive model. The Greenshields and Greenberg models, reported in the
following paragraphs, were obtained assuming functional relations between the macroscopic flow
variables and then calibrating them with observed data. Therefore, they belong to models of empirical
nature.

1.2.1 Elementary Spacing Model

In the case of straight-ahead moving vehicles without the possibility of overtaking, it is natural to
think that the safety condition is linked to the braking distance sf.

1
The reason can be easily understood if we consider the extreme case of a 1 km long segment with a queue of 99 standing vehicles
and one single vehicle that, in 1-hour interval of observation, travels on the fast lane at speed of 100km/h. The space average speed
is calculated over all vehicles in the segment and is therefore 1 km/h; the time average speed is calculated on the only vehicle that
passes through the initial section of the segment (the others are standing and are not counted by an observer placed at the section)
and is therefore 100 km/h.

7
 Macroscopic Models of Traffic Flow

v2
sf = + vt (1.10)
2g
which is supposed to consist of a first term at a constant speed during the driver's reaction time τ and a
second term uniformly decelerated with rate γ.
If every vehicle traveled in such a way as to remain at the braking distance from the tail of the
previous vehicle, the average spacing would be sf+l, being l the average length of the vehicle. This is
actually extremely cautionary, since it must be considered that during the braking phase also the
previous vehicle that, if the traffic stream were perfectly homogeneous, would still be equal to sf.
For safety, it is necessary to take into account that for a number of factors (reaction time, efficiency
of the braking system, ability to adjust the effort at the brake pedal, uneven conditions of the road surface
or tires) the braking distance of the previous vehicle can be shorter; then it can be assumed that the
safety condition consists in keeping a spacing s equal to a fraction a of the braking distance sf, increased
by the average length of the vehicle to take into account the definition of spacing between the front
bumper of the leading vehicle and the front bumper of the following vehicle.

æ v2 ö
s =aç + vt ÷ + l (1.11)
ç 2g ÷
è ø
To be used in the macroscopic model, the previous relationship, obtained starting from behavioral
hypotheses on the link between spacing and speed of single vehicles, must be expressed through the
fundamental flow variables. Considering that the average density k is equal to the inverse of the average
spacing between the vehicles s, we obtain the density-speed function k-v:

1
k=
æv 2
ö (1.12)
a çç + vt ÷÷ + l
è 2g ø
The function is more easily interpreted in terms of the outflow of the traffic stream if it is expressed
in the speed, inverting the equation (1.12):
2γ & 1 )
v 2 + 2γτ v − ( − l+ = 0 (1.13)
α 'k *
and solving the second-degree equation (1.13):

€ v = -gt + (gt )2 + 2 g æ1 ö
ç -l÷ (1.14)
a èk ø
Note how this model derives from theoretical considerations so that in order to be operational it
requires that the values of the quantities a, g, t and l, be determined by experimental observations
(calibration of the model).
Figure 1–III graphically illustrates this function, in correspondence with the following values of the
parameters2: α=0,30; γ=2,9m/s2; τ=1,5s; l=7m.

2
With reference to the values of the parameters, one can observe that the engine-braking on a flat road produces a deceleration of
about 1m/s2, while a relatively sharp deceleration is around 4÷5m/s2. The length l between standing vehicles must take into account
both the distance between standing vehicles (about 1m) and the possible presence of longer vehicles (in the case of heterogeneous
traffic stream). The reaction time based on experimental observations varies from 0.5 to 2s, depending on the drivers, while the α
coefficient, being an exclusively psychological parameter, has been calibrated so as to obtain a capacity equal to 0.605veh/s at a
critical speed of 11.5 m/s.

8
 Traffic Engineering and Intelligent Transportation Systems

60

50

40
v (m/s)

30

20

10

0
0 0,02 0,04 0,06 0,08 0,1 0,12 0,14 0,16
k (veic/m)

Figure 1–III. Speed-Density function according to the elementary spacing model.

The v (k) function describes the relationship between the average speed of the traffic stream and the
number of vehicles on the same road segment. Therefore, it represents the well-known phenomenon of
congestion, or, more generally, of vehicle conditioning: as the average density increases, the traffic
speed reduces, which means that a dense traffic stream (with very close vehicles) is less fast than a fluid
traffic stream, with vehicles are very far apart. However, (14) is not acceptable as is throughout the
whole range of density, because:
• speed tends to infinity when density tends to zero;
• for densities higher than a given value, the speed assumes a negative value.

Therefore, it is necessary to introduce additional constraints, necessary to make the representation

of the (discrete) physical phenomenon acceptable in terms of continuous flow quantities:
• the first constraint requires the existence of a limitation of the vehicle speed due to factors other
than the mutual conditioning between vehicles described by (1.14), such as the existence of a
maximum vehicle speed or the conditioning due to the type of road (safety when skidding or
overturning): the traffic speed must not exceed the speed that the isolated vehicle (free of
conditioning) keeps on a given type of road, which is called the free-flow speed (denoted here as vl):
v ≤ vl
• the second constraint requires that the traffic speed does not take negative values and therefore limits
the density to a maximum value, generally called in the technical literature jam density (and denoted
as kJ). It is to note that the need to introduce this constraint stems from the presence of the term l,
which represents the average length of the vehicle, in (1.14): in fact, setting l = 0, v tends to zero for
k which tends to infinity.
Taking into account the two previous conditions, the elementary spacing model is finally written:
-1
é æ vl 2 ö ù
v = vl ; 0 < k £ êa çç + vlt ÷÷ + l ú (1.15)
ëê è 2g ø ûú

Remark. The state equation (1.4) and the arc conditioning model (1.15) allow you to fully describe
the state of outflow of steady-state traffic flow. In fact, by replacing (1.12) in (1.4) it is possible to
express the flow as a function of speed:

9
 Macroscopic Models of Traffic Flow

v
q= ; 0 £ v £ vl
æ v2 ö (1.16)
a çç + vt ÷÷ + l
è 2g ø
The q(v) function is very useful for the operational analysis of a road section since it expresses the
traffic conditions directly through the speed, which describes the infrastructure performance in terms of
users’ travel time, and the flow, which measures the effectiveness of the infrastructure in terms of
transported users: that is, of the satisfied demand for transport. To highlight the variation of
performances as the flow of vehicles changes, the flow-speed relationship is generally represented by
reporting the speed in the ordinate axis. Analytically, the inversion of (1.16) is not immediate; the curve
v = v (q), illustrated in Figure 1–IV, was therefore obtained numerically.

60

50

40
v (m/s)

30

20

10

0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7
q (veic/s)

Figure 1–IV. Speed-Flow function according to the elementary spacing model.

Unlike the diagram v-k, the function v-q does not define a biunivocal correspondence: two values
of speed can correspond to the same value of flow: in fact, the speed is first decreasing as the flow
increases (also, the condition v≤vl prevents it from going to infinity for q approaching zero); then, after
the flow has reached its maximum value (a condition that involves heavy traffic), the speed begins to
decrease together with the flow, until it reaches the condition of blocked traffic (jam): zero speed and
zero flow. It is clear that the traffic state corresponding to the maximum value of the flow acquires
particular importance: on the one hand, it identifies the road capacity (how many vehicles it can
discharge in a unit of time); on the other hand, it identifies the speed below which the flow starts to
decrease (i.e.: traffic is congested); this value, clearly fundamental for traffic modeling, is called critical
speed and is denoted here as vc.

10
 Traffic Engineering and Intelligent Transportation Systems

0,7

0,6

0,5
q (veic/s)

0,4

0,3

0,2

0,1
vl vc
0
0 0,02 0,04 0,06 0,08 0,1 0,12 0,14 0,16
k (veic/m)

Figure 1–V. Flow-Density function (Fundamental Diagram) according to the elementary spacing model.

The q-k diagram (often denoted as fundamental diagram in the literature) provides an easier
interpretation of the inversion of the flow curve as the density changes. Consider a ring-shaped road
section, on which the vehicles run for a long enough time until reaching a running condition with a
constant speed and a constant distance (stationary state or state steady-state). The density is naturally
given by the ratio between the number n of vehicles on the road and its length L. The flow (that is, the
number of vehicles that travel through a generic section in the unit of time), for the same number of
vehicles on the segment is proportional to the speed (q = k·v). Now, imagine to progressively increase
n, starting from the free-flow state (k = 0; q = 0).
Keeping in mind the equation of state q = k·v and the link v = v(k), i.e. q = k·v(k), the following
considerations can be made:
• as long as the density is low, that is, the vehicles are very distant from each other and do not
condition much each other, an increase in the number k of vehicles on the link leads to a zero
or a modest reduction in speed and, consequently, to an increase in flow almost proportional to
the increase in density, i.e. to the increase in the number of vehicles placed on the ring;
• as the number n of vehicles increases, i.e. the density increases, the conditioning between the
vehicles v(k) becomes relevant, so that the vehicles travel more slowly than the free-speed and,
consequently, the flow increases less than proportionally with the density;
• when the density exceeds a certain value, denoted as critical density, the vehicles are so close
and the conditioning is so strong that a further increase in n involves a reduction in speed that
is prevalent compared to the increase in density that produced it: the result is a reduction in
flow.

1.2.2 Greenshields Model

Following the opposite approach to the deductive one used for the spacing model, numerous
empirical models have been developed since the 1930s, starting from the interpretation of experimental
observations. The first and simplest of these models, due to Greenshields (1934), assumed a decreasing
linear relationship between speed and density.

11
 Macroscopic Models of Traffic Flow

æ k ö
v = vl çç1 - ÷÷ (1.17)
è kJ ø
where vl is the free-flow speed (the maximum speed held on average by an isolated vehicle in the
roadway in safe driving condition) and kJ the maximum density (obtained for a column of stationary
vehicles).
Applying the state equation (1.4), we get that the relationship between flow and velocity (as well as
that between flow and density) is parabolic.

æ k ö
q = vl k çç1 - ÷÷ (1.18)
è kJ ø

æ v ö
q = k J vçç1 - ÷÷ (1.19)
è vl ø
In Figure 1–VI and Figure 1–VII, the experimental points and the corresponding regression curves
obtained by Greenshields are shown.

Figure 1–VI. Experimental data and v-k function of the Greenshields model (1935).

12
 Traffic Engineering and Intelligent Transportation Systems

Figure 1–VII. Experimental data and v-q function of the Greenshields model (1935).

Undoubtedly, this model has the advantage of a great simplicity, but it does not adequately
reproduce the modes of vehicular outflow. In fact:
• the linear shape implies that conditioning also occurs for very low densities, when vehicles are farther
than the distance of visibility;
• the parabolic function between flow and speed implies that the maximum flow be obtained at the
average value of the velocity; that is: vc = vl /2;
• the same condition applies to the parabolic shape of the flow with respect to density, so: kc = kJ /2; it
follows that the capacity would be:
k J vl
qmax = (1.20)
4

This expression of capacity, obtained in a range of values where detected data are lacking, does not
comply with subsequent experimental observations.

1.2.3 Greenberg Model

A more suitable model from both the theoretical and the experimental point of view was proposed
by Greenberg (1959) who hypothesized for the stationary state a logarithmic function between speed
and density:

æk ö
v = vc logç J ÷ (1.21)
è k ø
In this model, the critical speed is an independent parameter and its value can be estimated, like the
maximum density, based on experimental observations. Thus, the model can fit the observed data in a
more satisfactory way. In Figure 1–VIII, the experimental data collected in the Holland Tunnel in New
York and the corresponding regression curve obtained by Greenberg (1959) are reported.

13
 Macroscopic Models of Traffic Flow

Figure 1–VIII. Experimental data and the corresponding v-k function of the Greenberg model.

At first glance, one may regard a deficiency of the model that, at zero density, the speed tends to
infinity. However, it should be considered that the speed-density relationship represents the conditioning
between vehicles and, therefore, is not active for dislocations greater than the safe distance at free speed.
Thus, the equation (1.21) must then be completed with the specification of the domain of the function:

ìvl , 0 £ k < k (vl )

ï
v=í æk ö (1.22)
vc logç J ÷, k (vl ) £ k < k J
ï
î è k ø
The (1.22) can be easily reversed to express k as a function of v:

k = kJ e −v / v c 0 ≤ v < vl (1.23)
By applying the state equation (1.4) to (1.22) and (1.23), the relationships between flow and speed
and between flow and speed are also derived.
€ ìkvl , 0 £ k < k (vl )
ï
q=í æk ö (1.24)
vc k logç J ÷, k (vl ) £ k < k J
ïî è k ø

q = kJ ve −v / v c 0 ≤ v < vl (1.25)

The three diagrams of the Greenberg model v-k, v-q, q-k are shown below, obtained by assuming
the following parameter€
values: vc=12m/s; kJ=143veh/m.

14
 Traffic Engineering and Intelligent Transportation Systems

60

50

40
v (m/s)

30

20

10

0
0 0,02 0,04 0,06 0,08 0,1 0,12 0,14 0,16
k (veic/m)

Figure 1–IX. Speed-density function according to the Greenberg model.

60

50

40
v (m/s)

30

20

10

0
0 0,1 0,2 0,3 0,4 0,5 0,6 0,7
q (veic/s)

Figure 1–X. Speed-flow function according to the Greenberg model.

15
 Macroscopic Models of Traffic Flow

0,7

0,6

0,5
q (veic/s)

0,4

0,3

0,2

0,1
vl vc
0
0 0,02 0,04 0,06 0,08 0,1 0,12 0,14 0,16
k (veic/m)

Figure 1–XI. Flow-density function according to the Greenberg model.

1.2.4 Comparison of Different Models

The following figure shows a comparison of the flow-density curves obtained using the three
patterns described above. It can be observed how the elementary spacing model, having 4 parameters,
can be easily adapted to approximate the Greenberg model. It is also worth to be noted that the
Greenshields model provides a slightly higher capacity value than the other two models, even for very
low free speed values (v=18m/s corresponding to about 65=km/h), while adopting the same maximum
density value used for the other models: kJ=0.143veh/m.

0.8

0.7

0.6

0.5
q (veic/s)

0.4

0.3

0.2

0.1

0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

k (veic/m)

Greenshields Greenberg Distanziamento

Figure 1–XII. Comparison of different link congestion models.

In addition to the models described in this paragraph, over the years various scholars have proposed
various other congestion models, assuming other functional forms of the v-k function: negative

16
 Traffic Engineering and Intelligent Transportation Systems

exponential (Underwood, 1961), discontinuous Drake et al. (1967) or “bi-regime” (Edie, 1961),
polynomial generalization of the Greenshields model (Sanwal et al., 1996), up to more complex
functional forms in the three-dimensional v-k-q plane (Persaud and Hall, 1983), based on the catastrophe
theory.
Recently, a radical criticism of the progressive complication of flow models has led Daganzo and
Muñoz (2000) to use a simple function with a triangular form for flow-density relationship.

1.3 Dynamic First-Order Model

1.3.1 Theory of Kinematic Waves

The models introduced in the previous section describe the vehicular flow in the stationary state
(alternatively called steady-state) when the speed, density, and flow do not vary in time and space. The
analysis for stationary states is appropriate in the transport systems planning and design, which are
concerned with the sizing of the components of the system, given the expected demand for mobility.
The study and design of traffic control and regulation systems, however, require representing traffic
dynamics, which describe how the characteristic variables of the traffic flow (v, k, and q) change in time
and space. This means to specify functions of the type: v=v(x,t) or, equivalently, since v=v(k), of the
type: k=k(x,t).
Such a function describes the time variation of the speed (or density) on a given road facility or,
alternatively, the space variation of the speed (or density) along a road segment at a given instant.
It is natural that to obtain such a dynamic function we must start from the conservation equation
(1.3), which relates the temporal variation of the density to the spatial variation of the flow.
Assuming that, even in dynamic conditions, the state equation (1.4) still holds locally, the value of
the flow in a generic point (x,t) is equal to the product of the velocity in this point v(x,t) for the density
at the same point k(x,t) and the equation (1.4) can be specified in the x-t plane as follows:

q( x,t ) = k ( x,t )⋅ v ( x,t ) (1.26)

Assuming also that the v=v(k) relationship observed in the stationary state still holds even when k
and v vary in time and space, it is possible to write:
€ q = q[k ( x, t )] (1.27)
It should be noted that the resulting hypothesis that the flow is a point function of the only density,
perfectly consistent with the initial assumption to assimilate the traffic stream to a stream of a fluid,
means that the speed at a given time in a given section depends only on the value of the traffic density
in the same section at that same time. In other words, by interpreting the condition on state variables in
terms of vehicle behavior, we are assuming that vehicles instantly adapt their speed to any changes in
density.
The dependency of the flow q from dynamics of density k(x,t) allows us to express the conservation
equation (1.3) as a function of the only density k:

¶k ( x, t ) ¶q[k ( x, t )]
+ =0 (1.28)
¶t ¶x
Applying the chain rule for the derivation of composite functions3, we can write:

3
The derivative with respect to the variable x of a composite function y=f[g(x)] is: y’=f’(g)g’= df/dgּ dg/dx.

17
 Macroscopic Models of Traffic Flow

¶k ( x, t ) dq ¶[k ( x, t )]
+ =0 (1.29)
¶t dk ¶x
Now, we pose:
dq
=w (1.30)
dk
Considering small variations in state variables, we can assume that w is locally constant4.
Under these assumptions, the conservation equation is a quasi-linear homogeneous partial derivative
differential equation:
𝜕𝑘(𝑥, 𝑡) 𝜕[𝑘(𝑥, 𝑡)]
+w =0 (1.31)
𝜕𝑡 𝜕𝑥
whose solution is represented by a general integral5 of the form:

k = k ( x - wt ) (1.32)
As it is easy to verify by replacing the solution (1.32) in (1.28) and differentiating it.
The solution (1.32) implies that all points of the space-time plane placed on a straight line of
equation
u = x - wt (1.33)
have the same density value (and therefore also the same speed and flow).
This equation represents how a traffic state with density k propagates in the plane (x, t). Unlike the
dynamic waves observed in fluids, which have an oscillatory feature, the propagation of perturbations
along traffic streams occurs according to straight lines, which are called kinematic waves6, in analogy
to dynamic waves. This is true under the hypothesis that w is constant and does not change over time7.
This condition can be used to determine the state of the traffic stream (for example, its density)
anywhere along the straight line u, starting with a known value detected at any point (x, t) belonging to
the same line. Since the speed the kinematic waves along a traffic stream moving at speed v is given by
the tangent to the fundamental diagram q(k) at the point k(v), we can observe that:
• when k>kc is w<0: dense traffic states propagate backward along the traffic stream;
• when k<kc is w>0: fluid propagate forward along the traffic stream;
• when k=kc is w=0, the critical traffic state (where is q=qmax) does not propagate along the traffic
stream8.
Figure 1–XIII illustrates the propagation of traffic states in a vehicular stream in non-stationary
conditions. The graph on the left depicts the propagation of the kinematic waves for two generic traffic
states; in the right chart, the two traffic states are identified over the fundamental diagram. Although the
graphs are purely qualitative, the respective scales have been chosen so that the angles are equal in the
two floors. Note that, in the q-k plane, speeds are represented by the trigonometric tangents of inclination
angles to the horizontal axis: the angle of the radius vector from the origin corresponds to the traffic
speed; angles of the geometric tangents to the fundamental diagram identify the speed of kinematic
waves.

4
It is obvious that, in general, the derivative of the flow q(k) varies with x and t if k varies with x and t.
5
Cfr. For example: Elsgolts L. (1973). Differential Equations and the Calculus of Variations. Mir Publishers.
6
Kinematic waves are often referred to as even Characteristics, because they are obtained from the characteristic equation of the
differential equation.
7
If we remove this hypothesis, the general integral of the equation (3”) would be a non-linear function and the states at constant
density k(x, t) would not propagate in the plane x-t along straight lines but along curves.
8
This property means that if an observer were to place himself at the section where the first vehicle reaches the critical speed (called
the critical section), it would see all vehicles travel at that same speed.

18
 Traffic Engineering and Intelligent Transportation Systems

In this way, a graphical method to construct the kinematic waves consists of: measure a traffic speed
in the x-t plane, draw it onto the q-k plane to find the corresponding traffic state, and plot the tangent to
the fundamental diagram at this point to determine the speed of the kinematic wave.
A particular application is described in the following subsection.

0,7
x wA wA A
A’
0,6
A
0,5

q (veic/s)
0,4
vA vA
0,3
B
B
vB B’ 0,2
vB B”
0,1
vA vB wB
wB
0
t 0 k (veic/m)

Figure 1–XIII. Graphical representation of the traffic state propagation according to the kinematic wave theory.

Remark. The linear dynamic model consisting of equations (1.26) and (1.29), due to Lighthill and
Whitham (1955), is a complete model: that is, it is sufficient to describe the dynamics of vehicular traffic
without the need to introduce other hypotheses or equations, except the necessary boundary conditions.
The solution is a family of curves u[k(x–wt)], characteristic curves or kinematic waves, consisting of the
points of the space-time plane having the same density k. Note, as can also be deduced from Figure 1–
VII, that the solution of the Lighthill and Whitham model is not necessarily unique: nothing excludes,
in fact, that two kinematic waves intersect at a point in the plane (x,t). We will see in the following
paragraphs that this condition occurs only in particular circumstances.

1.3.2 Departure Model

A particularly significant example of application of the kinematic wave theory is the discharge
process of a queue of standing vehicles. It is clear that the solution of the conservation equation, being
a differential equation, is a general integral that provides a family of curves (with our assumptions, a
family of straight lines), the parameters of which must be determined by imposing the suitable boundary
conditions that characterize the specific problem. In our case, these require:
a) that all vehicles are initially in rest; so that, the speed of the traffic stream is everywhere zero for
t=0:
v = v ( x,t = 0); x ∈ [ −∞,0)
b) that the first vehicle (which in this case is the element of perturbance for the system, which initially
in rest) moves with a known trajectory:
x1 = x1 ( t ); v1 = v1 ( t ); t ∈ [0,∞)
€
The trajectory of the first vehicle represents, in the context of the macroscopic representation of the
flow, the boundary of the traffic stream. Therefore, in the section x1(t) that the first vehicle reaches at
time t, the speed of the traffic stream v corresponds to the speed of the first vehicle v1(t). According to
€
the theory of kinematic waves, this state of flow characterized by the speed v1 propagates on the plane

19
 Macroscopic Models of Traffic Flow

(x, t) along a straight line of equation u=x–wt, whose angular coefficient w is equal to the derivative of
the function q(k) at the point corresponding to the speed v1.
The traffic stream dynamics can then be constructed using a graphical procedure:
1. for any given time instant, determine the section x1(ti) and the speed v1(ti) reached by the first
vehicle by the known trajectory;
2. calculate the speed w(v1) of propagation of the state v1 as the derivative of the function q(k) over
k at the point k(v1);
3. draw the kinematic wave corresponding to the speed w(v1) from the point x1(ti), whose equation
is:

dq ⎡⎣ k ( v1 ) ⎤⎦
u ( v1 ) = x ( ti ) + (t − t )
i
(1.34)
dk
Figure 1–XIV illustrates the graphical construction procedure. On the top side of the figure, the
trajectory of the first vehicle is depicted, which starts from the rest and accelerates until it reaches the
free speed. The graph at the bottom side of the figure depicts the traffic states corresponding to the speed
reached by the first vehicle: even in the q-k plane, the speed ranges from zero to the free-flow value, in
the counterclockwise direction indicated by the arrow shown in the q-k plane. At every speed value, the
tangent to the q-k curve is plotted, which detects the speed of the kinematic wave that propagates that
traffic speed along the traffic stream.
Point J represents the jam condition of standing vehicles in a queue: the kinematic wave in J (which
corresponds to the speed v=0) determines the starting instant of the vehicles following the first. It
naturally has a negative slope. It should be noted that, as the speed increases from zero to the free speed,
the kinematic waves expand fan-like.
As v increases, in fact, the speed of the kinematic waves decreases in absolute value and becomes
zero for v=vc: this implies that in the section xc, where the first vehicle reaches the speed vc, all the
following vehicles arrive at the same speed and are equally interspaced over the time by the quantity
h=1/qc. It should be noted, however, that in a given section x<xc the following vehicles arrive at faster
speeds than that of the first vehicle and that those speeds increase as the traffic stream flows out.
For speeds greater than vc, the traffic states propagate forward, so the following vehicles will arrive
at a generic section x>xc at speeds lower than that of the first vehicle and gradually decreasing as the
traffic stream flows out.
The picture at the top of the figure shows how graphically construct the time variation of the traffic
speed at a given section x: the intersections of the horizontal line with the kinematic waves
corresponding to the different speeds v1, v2,... identify the time instants in which the traffic speeds at the
section x are precisely equal to v1, v2,...
The points with white background in the picture exemplify the application of the construction
procedure for the stop line section x=0.

20
 Traffic Engineering and Intelligent Transportation Systems

v= vL
x

v= v4

xc v= vc

v= v3

x=0
t
v= v2

v=v1
v=0

0,7
C
P4 P3
0,6
L P2
0,5
q (veic/s)

0,4

0,3
P1
0,2

0,1
vL v1 J
0
0 k (veic/m)

Figure 1–XIV. Graphical method for constructing the departure model of a traffic stream starting form of a
queue of standing vehicles.

The same result can be obtained mathematically, analytically or numerically, using the speed v1 of
the first vehicle as a parameter and inverting the trajectory of the first vehicle, so as to have the abscissae
x1(v1) and the time instants t1(v1) at which the first vehicle reaches the speed v1.
Assuming that the first vehicle moves from standing with a uniformly accelerated motion (v1=at;
x1=1/2at2), the inverse of the trajectory is: t1=v1/a; x=v2/2a.
The equation of the kinematic waves

" $ v12 & v)

( ) ( ) ( )
u v1 = x v1 + w v1 #t − t1 v1 % = − w v1 ( t − 1 +
2a ' a*
( ) ( ) (1.35)

as the value of the parameter v1 changes, it describes a family of curves, locus of the points having a
constant speed v1 in the x-t plane.

21
 Macroscopic Models of Traffic Flow

Remark. If we adopt the Greenberg model to describe the traffic flow, we get an extremely simple
and effective expression of the speed of the kinematic waves. To derive it, we have just to apply the
well-known chain rule for composite functions:

dq d[ k⋅ v ( k )] dv ( k ) dk dv ( k ) (1.36)
w= = =k +v =v+k
dk dk dk dk dk
It is also:

dq d[ k⋅ v ( k )] dv ( k ) dk dv ( k ) (1.37)
€ w= = =k +v =v+k
dk dk dk dk dk

dv ( k ) d ( " k %+ d
dk
= *v c ln$ J ' - = v c
dk ) # k &, dk
[
ln( k J ) − ln( k ) ] (1.38)
€
dv ( k ) d 1
= −v c ln( k ) = −v c (1.39)
dk dk k
€
Substituting the (1.39) and (1.40)in the equation (1.36), we finally get:

dv(k ) æ v ö
€ w=v+k = v + k ç - c ÷ = v - vc (1.40)
dk è k ø
Note that the equation (1.40) of the kinematic waves speed obtained from the Greenberg model is
the simplest expression that complies with the two necessary conditions of kinematic waves:
• The speed w of the perturbation must be always lower than or equal to the speed of the traffic stream
v9;
• The following relations must hold: for v<vc, it must be w<0; for v>vc, it must be w>0; for v=vc, w=0.
Then, the departure model can be expressed as a function of v, t, and u. The kinematic waves of the
become:

" $ v2 & v)
() ( ) ()
u = x1 v + w v1 #t − t1 v % = − v − vc ( t − +
2a ' a*
( ) (1.41)

By fixing the value of the abscissa u, we can get the expression of the traffic speed in the section u
as a function of the time by solving the (1.41) with respect to the speed v.

v2 " %
2a
(
− v − vc )$#t − av '& − u = 0 (1.42)

Sorting by v, we get the following second order equation:

1 2
2
( ) (
v − at + vc v + a vc t + u = 0 ) (1.43)

It has two solutions:

2
v = ( at + vc ) ± ( at + vc ) − 2a( vc t + u) (1.44)

9
€
This condition, evident in the model of Greenberg, where it is w=v-vc, has a general validity, as it can be seen from the equation (29),
where the traffic speed v is always not increasing with the density k; it follows dv/dk≤0.

22
 Traffic Engineering and Intelligent Transportation Systems

defined for t ³ 2u/a , of which only the one with the negative sign is possible10.
In Figure 1–XV, the temporal variation of the traffic speed is depicted with reference to several
different sections of the road: the starting section, which in the case of a queue of vehicles stopped at a
traffic light coincides with the stop line (u=0, red-colored), in sections 10m and 20m downstream the
stop line and upstream the critical section (green-colored curves), and in sections 100m, 200m, 300m,
and 400m away, downstream the critical section (blue-colored curves). In this model we assumed:
vc=10m/s; a=2m/s2.

20

u=100m u=200m u=300m u=400m

15
v (m/s)

10

u=20m
5
u=10m
u=0

0
0 20 40
60 80 100
t (s)
Figure 1–XV. Time variation of the traffic for a vehicular stream initially stopped at several sections
downstream of the starting line, assumed as the origin of the space axis (u=0).

It should be noted that the traffic speed in all sections before the critical one (and, of course, also at
the starting section) increases continuously and tends to the critical speed from below, while in the
sections after the critical one the speed decreases and still tends asymptotically towards critical speed
but in this case from above. This different trend is a direct consequence of the slope inversion of the
flow trend at the critical state in the q-k diagram.
By applying the state equation q=k·v and the relationship k=k(v), which in the Greenberg model is
k=kJ exp(–v/vc), we can determine also the function that expresses the space-time variation of the
vehicular flow:
at
" 2 % −v
#
( ) (at + v )
q = $ at + vc ± c ( )
− 2a vc t + u ' k J e c
&
(1.45)

As shown in the figure, for each value of v, the function q=q(x,t,u) increases with t and
asymptotically tends to q(vc).

10
To realize this, it is sufficient to impose that x0 the speed is zero at time t0 at the stopline section.

23
 Macroscopic Models of Traffic Flow

0.6
u=10m

0.5
u=100m
0.4
u=200m
q (veic/s)

0.3 u=0 u=300m

0.2
u=400m

0.1

0
0 20 40 60 80 100
t (s)
Figure 1–XVI. Time variation of the flow of a traffic stream initially standing for several sections located at
different distances from the stop line, assumed as the origin of the spatial axis (u=0).

The time variation of the flow in a section sufficiently far from the starting line, shown in Figure 1–
XVI, reproduces a phenomenon known and easily experienced on the fastest arteries, whereby the
vehicles forming to a platoon that starts after the red at a traffic signal tend to be more spaced at the
front of the platoon and closer at the tail (platoon dispersion) because the first vehicles arrive at the free
speed while the following vehicles, conditioned by the previous ones, arrive at speeds gradually lower
tending to the critical speed for an infinitely long platoon.

1.3.3 Shock Waves

Consider now the case of a traffic stream near a bottleneck that limits the speed below the critical
value and the flow to a value lower than that upstream of the bottleneck (that is, lower than the demand
flow). The traffic states upstream and downstream the bottleneck and of it are represented on the left
side of Figure 1–XVII respectively by the points A and B.

24
 Traffic Engineering and Intelligent Transportation Systems

0,7
x wA A
B 0,6
vB B’
0,5
vB B”

q (veic/s)
0,4

0,3
B
0,2
A A’ wA
0,1
vA vB wB
vA vA wB
0
t 0 k (veic/m)

Figure 1–XVII. Intersection of kinematic waves in the transition from light traffic upstream (A) to dense traffic
downstream (B).

Unlike the departure of a stream of vehicles at a traffic light, in which the boundary condition is an
hourly law of increasing speed, in this case, the bottleneck causes the first vehicle of the platoon to
reduce its speed. Applying the Lighthill and Whitham method, you can see how the speed of cinematic
waves this time is progressively decreasing. Whereas in the starting model the cinematic waves open in
fan, in this case, they close on themselves. In other words, there are points where the cinematic waves
intersect. This implies that at a given point in space the traffic speed should have two different values at
the same time, which is not physically acceptable. As a result, from the point of intersection, these two
cinematic waves must cease to exist, and a single wave must depart.
This condition is represented in Figure 1–XVIII, which still illustrates the case of a stationary state
of fluid traffic (A), which subsequently joins a denser stationary state (B). This case corresponds, for
example, to a lightly congested highway with a big amount of traffic entering from an on-ramp, which
increases the downstream flow, without nevertheless exceeding the capacity. In this case, both speeds
are greater than the critical one. In the picture, continuous lines represent vehicle trajectories, while
continuous lines represent the kinematic waves of each state. In every stationary zone, they are parallel
to each other because the traffic speed is constant: parallel kinematic waves indicate that the same traffic
state propagates along the road and all vehicles travel at the same speed. Thus, they translate a condition
of constant speed (over time) to a condition of uniform speed (over space).
However, these stationary conditions hold until the kinematic waves of the lighter traffic state,
which are faster than the denser one, intersect those of such denser state. One can observe from the
figure that the two families of waves representing states A and B intersect along a straight line, which
represents the discontinuity transition from state A to state B and has a slope equal to the chord of the
fundamental diagram that connects the two corresponding points on the q-k curve.

25
 Macroscopic Models of Traffic Flow

x wB 0,7
wB
B
0,6
B
0,5
vB wA A

q (veic/s)
0,4

0,3
wA
0,2
A
0,1
vA vB
0
t 0 k (veic/m)

Figure 1–XVIII. Graphical representation of shock waves in the transition from a denser traffic upstream (A) to
lighter traffic downstream (B).

This phenomenon also occurs in fluid dynamics: when a faster traffic reaches denser traffic, a shock
wave is observed, which is an element of discontinuity of the flow, consisting in an abrupt transition
from one stationary state to another.
The shock wave propagates along a traffic stream with a speed that can be determined by solving
the dynamic problem (1.31) by imposing a discrete variation from one state to another.
To do an easy graphical demonstration of this, let us consider a speed wave w, which represents the
transition from an upstream traffic state (A) to a downstream traffic state (B), denser than the previous
one, and apply the condition of conservation of the number of vehicles to the road segment traveled by
the wave in the time interval Dt, having a length equal to wDt.
At the initial time instant, the traffic density on the whole segment is kB; at the end of the interval is
kA. By the conservation law, the difference of the number of vehicles on the road segment at the
beginning and at the end of the time interval must be equal to the difference between the number of
vehicles entering the initial section (x=0) and the number of vehicles are exiting the final section
(x=wΔt) in the same range (Fig.1-XIX).

k A wDt - k B wDt = q A Dt - q B Dt (1.46)

By deleting the Δt, which is an arbitrary nonzero quantity, and by solving the equality with respect
to w, we get the following expression of the propagation speed of the shock wave:

q A - qB
w= (1.47)
k A - kB

26
 Traffic Engineering and Intelligent Transportation Systems

x 0,7
0,7
A
B wB
wAB vB
0,6
0,6
wAB
0,5
0,5
B A

q (veic/s)
wΔt 0,4
0,4 B
wΔt

q(veh/s)
A 0,3
0,3
wA
0,2
0,2

vA 0,1
0,1
vvBA vB
00
0 Δt t k (veh/m)
0 k(veic/m)

Figure 1–XIX. Application of the principle of conservation to the number of vehicles in a road segment having a
length equal to the distance traveled by the shock wave that represents the transition from a fluid traffic upstream
(A) to denser traffic downstream (B).

In other words, in the transition from state B (holding at the beginning of the observation time
interval) to state A (holding at the end of the observation time interval) a densification of traffic occurs
(it is, in fact, kB>kA), which must be compensated by a higher downstream flow than upstream (i.e.:
qB>qA). It is clear that, for the number of vehicles to conserve, the wave that constitutes the separation
between the two traffic states must propagate forward to allow the more fluid state to advance.
Therefore, the shock wave arises from a discontinuity of density, in which the conservation equation in
its continuous form (1.3) is no longer valid, but the conservation principle of vehicles is still valid, which
requires that the discontinuity move with a speed given by equation (1.47).

An even easier demonstration can be provided in the density-space plane. In this case, we consider
a growing queue at a bottleneck (Figure 1–XX). The speed w at which the tail of the queue spills back
is given by the shock wave that represents the interface between uncongested upstream traffic
(characterized by a flow qA and a density kA) and the queueing state (characterized by a flow qB and a
density kB). After a time interval Δt, the queue has spilled back by the length wΔt and has grown by a
number of vehicles (kB–kA)wΔt. By the conservation law, the increase of the number of vehicles in the
queue must be equal to the difference of the number of vehicles that joined the queue (qAΔt) and the
number of vehicles that left the queue (qAΔt) in the same time interval. The equations (1.46) and (1.47)
follow from this consideration.

27
 Macroscopic Models of Traffic Flow

k k
kB kB
Congested Congested
flow flow
kA kA
Uncongested flow Δx=wΔt Uncongested flow
Bottleneck

Bottleneck
xQ(t) x xQ(t) xQ(t+Δt) x
Direction of flow

Figure 1–XX. Representation of a queue of vehicles in the plane space (x)-density (k) at two successive time
instants t (on the left) and t+Δt (on the right).

Remark. The results obtained in this and in the previous subsection show that the Lighthill and
Whitham’s model admits a single solution when the boundary condition of the differential equation is
represented by a perturbation with an increasing speed (dv1/dt(t)>0); however, the solution is not unique
when the boundary condition is represented by a perturbation with a decreasing speed (dv1/dt(t)<0). In
the case of stationary traffic (dv1/dt(t)=0), all the kinematic waves have the same speed: it is the case of
a non-disrupted traffic stream that travels at a constant speed and which is then crossed by parallel
cinematic waves.

Microscopic interpretation of the shock waves. The expression of shock waves (1.47) can also
be derived by using the trajectories of individual vehicles and then transforming microscopic variables
(distance, headway, and individual speed) into the descriptive macroscopic quantities of the traffic
stream (density, flow, and average speed).
We assume (as the macroscopic first-order theory does) that the transition of each vehicle from the
speed vA to the speed vB occurs instantaneously when the vehicle is at the distance sB from the previous
one that the driver judges to be safe for the speed vB that he is going to assume.
This state transition is illustrated in Fig.1.XXI, that represents the following driving actions in the
space-time plane:
• At the time tP, the leading vehicle I reduces its speed from vA to vB;
• the following vehicle II adapts its speed to the new one vB and starts traveling at the safe distance
for that speed s(vB)=1/kB
•
the adjustment takes place after a time interval h, at which the following vehicle, initially at a
distance sA from the leading one, has covered the distance xH needed to place itself at the
distance sB from the vehicle I, which in the meantime has traveled the space xP.
These kinematic considerations, graphically illustrated in Figure 1–XXI, correspond to the
following equality:
1 1 1 1
s A + xP = xH + sB Þ + vB h = v A h + Þ h(v B - v A ) = - (1.48)
kA kB kB k A
from which:
1 1
-
k k A k A - kB 1 (1.49)
h= B =
vB - v A vB - v A k B k A

28
 Traffic Engineering and Intelligent Transportation Systems

sB
xP = vB h H
x vB

P w

I II

sA xH =v h
A
s=1/kB
vB H vB h

P w wh
h V
A
t

s A =1/kA

I II
H

w wh
P
V V
A A h

I II

Figure 1–XXI. Graphical method to determine the speed of the shock wave from the simplified trajectories of
two vehicles.

The straight line passing for the points P and H of the plane x-t represents the locus of the points of
that plane at which the velocity adjustment occurs; that is, with reference to the vehicular stream, the
transition from state A to state B. The angular coefficient of this line, w, represents the rate at which the
transition from one traffic state to another propagates. From previous considerations, we find again the
equation of the well-known shock wave speed, as it follows.
1
vB h -
xB - sB kB 1 (1.50)
w= = = vB -
h h kBh

1 1 kBk A
w = vB - = vB - (v B - v A ) (1.51)
kBh kB k A - kB

1 1 kBk A
w = vB - = vB - (v B - v A ) (1.52)
kBh kB k A - kB

vB k A - vB k B - k AvB + k Av A - vB k B + k Av A q A - qB
w= = = (1.53)
k A - kB k A - kB k A - kB

Approximate resolution of the dynamic problem. Shock waves can be seen as a discrete
approximation of kinematic waves, in which the chord that joins the two traffic states of initial and final
flow replaces the gradual passage for tangents to the curve.

29
 Macroscopic Models of Traffic Flow

Shock waves do not provide a detailed representation of the transition from one state to another
through every individual traffic state; however, they approximate such a transition through a
discontinuity in which the speed changes abruptly in value.
The resolution of the dynamic problem of vehicular traffic in the presence of shock waves is easy
when switching from one stationary state to another, like in the cases of a bottleneck that temporarily
reduces the speed or of a slow vehicle entering the traffic stream from a side entrance.
On the other hand, it is burdensome if you want to represent the transient period in which the traffic
stream changes in a greater detail. The solution of this problem can be done numerically or graphically.
For each traffic event, the procedure requires to identify the intersections between two kinematic waves.
From every intersection, a new shock wave departs that represent the transition from one state to another,
the speed of which is then given by the chord that joins the two corresponding traffic states in the q-k
plane (Sasaki et al., 1984). An example, taken from Kunhe and Michalopoulos (2000) is reported in
Figure 1–XXII.

Figure 1–XXII. Representation of traffic flow between two signalized intersections by applying the finite
difference method for a discrete approximation of kinematic waves.

Remark. The first-order model is a complete and consistent theoretical scheme that is a very useful
tool to analyze road traffic in stationary or near-stationary states. However, it introduces some
approximate hypotheses that are unrealistic if one wants to get a more detailed representation of dynamic
traffic variations as it can be obtained by microscopic models. Specifically:
• the analogy of the vehicular stream to a fluid stream prevents the possibility of representing the
variability of performances among vehicles and does not allow to describe the probabilistic
distribution of the desired speeds among different drivers, nor to represent important aspects of the
phenomenon such as overtaking or lane changing.
• the assumption that the flow is a function of the density only at the point assumes that vehicles
instantaneously adapt their speed to density changes.

30
 Traffic Engineering and Intelligent Transportation Systems

• From the above hypothesis, for boundary conditions of the type dv1/dt(t)<0, the non-uniqueness of
the solution and therefore the discontinuity of the outflow represented by a shock wave11.
• The solution of the first-order model does not allow to reproduce phenomena of traffic stream
instability, often observed on the highway for sudden increases in density (“stop-and-go” conditions),
nor the so-called hysteresis of traffic, often observed in the discharge process of queues on highways.
The first observation is relevant only in low traffic conditions on roads with allowed overtaking. In
the case of roads with not allowed overtaking, the first-order model provides a good representation of
traffic, apart from the detailed description of the motion of a single vehicle and the representation of
stop-and-go.
Higher-order models and additional variable dynamics were introduced in an attempt to overcome
the remaining inadequacies. The following paragraph describes two of the best-known of these models,
both due to Payne (1971).

1.4 Second-Order Traffic Models

Second-order models aim to improve the linear model by introducing additional features that take
into account the impossibility for drivers to instantaneously update speed to traffic density. In the context
of the analogy with fluid dynamics, these behavioral assumptions result in the introduction of the
concepts of relaxation and traffic diffusion, which result in a lag of speed adjustment in space and time,
respectively.

1.4.1 The Density Gradient Model

The density gradient model is based on the assumption that the traffic speed v at a point x depends
on the density k at a farther point x+Dx, so:

v ( x,t ) = V ⎡⎣ k ( x + Δx,t ) ⎤⎦ , Δx > 0 (1.54)

This expression replaces the classic linear model expression:

v ( x, t ) = V !"k ( x, t )#$ (1.55)

keeping the same shape for the v(k) function.
To transform the model defined at an incremented point x+Δx to a function at the same point (x, t),
we apply a Taylor’s series expansion with respect to the increment of space Δx>0

∂V
v ( x,t ) = V ⎡⎣ k ( x,t ) ⎤⎦ + ⋅ Δx (1.56)
∂ x ( x,t )

and then the chain rule for composite functions:

∂V ∂ k
v ( x, t ) = V !"k ( x, t )#$ + ⋅ Δx (1.57)
∂ k ∂ x ( x,t )
Assuming also that the space shift ∆x of the speed is a function of the perceived safe distance s, and
remembering that the spacing equals the inverse of the density k:

11
Note that the existence of shock waves is an undesirable property mathematically, but it is not in contrast to experimental
applications, of which it is a linear approximation.

31
 Macroscopic Models of Traffic Flow

α
Δx = α s = (1.58)
k
and replacing the (46) and (47) in the (45), we get a function in only the density k:

α ∂V ∂ k ( x,t )
v ( x,t ) =V !"k ( x,t )#$ + (1.59)
k ∂k ∂x
Then, assuming v-k function be linear and introducing the positive constant µ:

∂V
= −µ (1.60)
∂k
we obtain a congestion model, often called first Payne’s model, that relates the traffic speed to both the
density in the same point (as the stationary congestion model) and the spatial gradient of the density:

µ ∂k ( x,t ) (1.61)
[ ]
v ( x,t ) = V k ( x,t ) −
k ∂x
The Payne’s model takes into account that drivers adjust their speed to a variation in density that
occurs not at the same point but some a distance ahead; consequently, it reduces the speed value given
by the steady-state model
€ V[k(x,t)] by an anticipation term (or diffusion term), represented by the spatial
gradient of the density.
The equation (1.61) can be expressed in terms of flow by applying the state equation q=kv:

∂k ( x,t )
[ ]
q ( x,t ) = Q k ( x,t ) − µ (1.62)
∂x
A direct consequence of equation (1.61) is that the flow at the critical state (corresponding to the
critical density) may be higher or lower than the capacity, depending on the downstream density.
Finally, replacing
€ the (1.61) in the continuity equation (1.3) we obtain a second-order partial
differential equation, from which the naming of these traffic flow models:

¶k ( x, t ) dq ¶[k ( x, t )] ¶ 2 k ( x, t )
+ -µ =0 (1.63)
¶t dk ¶x ¶x 2

Remark. An important mathematical property of the density gradient model is that it provides a
single solution without introducing discontinuities. The flow changes are also more smoothed than in
the linear model.
This formulation also lends to radical criticisms, vigorously argued by Daganzo (1995) and
previously refuted by Del Castillo et al. (1993). Indeed, the addition of the term diffusion, which was
introduced to compensate for an obvious inadequacy of the macroscopic model and has its own
correspondence in the dynamics of isotropic fluids, ends up producing unacceptable results in traffic
streams, which are anisotropic.
In fact, the equation (1.61) that provides the value of speed depending on the density gradient does
not distinguishing the direction of travel and so neglect the fundamental peculiarity of traffic that vehicle
conditioning only occurs in opposite direction of the travel; that is, from downstream to upstream. This
negligence would imply that drivers would be affected even by traffic incoming from upstream, which
is unrealistic. Moreover, the structure of (1.63) does not exclude that, for sharp increases in density, the
model can predict negative flows.

32
 Traffic Engineering and Intelligent Transportation Systems

Nevertheless, this unsatisfactory theoretical structure does not preclude us using this model because
we can simply replace the equation (1.63) by the following expression

⎧⎪
( )
q = max ⎨0,Q ⎡⎣ k x,t ⎤⎦ − µ
∂k x,t ( ) ⎫⎪ (1.64)
⎬
⎪⎩ ∂x ⎪⎭
To react to the increasing complexity of the traffic models introduced in the recent years, even with
weak adherence to the actual phenomenon, Newell (2002) presented a simplified model which he called
lower-order model, starting from basic considerations on car-following mechanism. It will be presented
in the next Chapter, which deals with car following theory.

1.4.2 Payne’s Dynamic Model

Payne’s second model is based on an interpretation of the drivers’ behavior in dynamic rather than
static conditions and introduces a delay in the speed adaptation due to a variation in density. Specifically,
it assumes that the speed v in the point x is adapted at the instant t+t depending on the density k at a
farther point x+Dx in the previous instant t12:

[
v ( x,t + T ) = V k ( x +∆x,t ) ] (1.65)

To get the differential equation that describes the dynamics of flow we apply a Taylor expansion to
the first member of the equation (1.65) with respect to a small time interval T, provided that the
expansion of the second term was already obtained in equation (1.61).
€

( ) ( )
v x,t + T = v x,t +
( )T
dv x,t
(1.66)
dt
Then, we apply the chain rule to the total derivative of the traffic speed v with respect to the time t
and we get an expression corresponding to the sum of a local term and a convection term:

dv ( x,t ) ∂v ∂t ∂v ∂ x ∂v ∂v
= + = +v (1.67)
dt ∂t ∂t ∂x ∂ t ∂t ∂x
Replacing the equations (1.66), (1.67) and (1.61) in the equation (1.65), we get the following
congestion model:
€
∂v ∂v µ ∂k ( x,t ) (1.68)
v ( x,t ) + T
∂t
+ vT
∂x
= V k ( x,t ) − [
k ∂x
]
Reordering and taking the derivative term with respect to time on the left side we obtain a dynamic
equation of the traffic speed:
€ $ '
∂v ( x,t ) ∂v 1 µ ∂k ( x,t )
= −v + &V ( k ) − v ( x,t ) − ) (1.69)
∂t ∂x T &% k ∂x )(
The equation (1.69) has many similarities to fluid dynamics, as it expresses the acceleration of the
traffic stream by the sum of three terms, which represent, respectively:
• the term v·∂v/∂x is a convection factor;
€
• the term [V(k) - v] /t is a relaxation term;

12
It is worth noticing the close similarity between the dynamic model of Payne and microscopic car following models described in
Chapter 2.

33
 Macroscopic Models of Traffic Flow

• the term (µ/k)·∂k/∂x is an anticipation (or diffusion) term.

Remarks. The speed gradient equation (1.69) and the conservation equation (1.3) constitute a
system of two differential equations to partial derivatives in v and k, which form a complete traffic flow
model, whose solution provides values of speed in time and space, given boundary constraints.
In principle, this model should allow for a more realistic representation of vehicular outflow. On the
other hand, this increased complexity is offset by the fact that there is no exact method of solution.
Moreover, the dynamic model does not eliminate the disadvantages of the density gradient model. Even
in this dynamic congestion model, in fact, traffic would be affected by the conditioning that comes from
behind, unless we introduce the non-null constraint like in equation (1.64).

1.5 Bibliography
Castillo, J. M., P. Pintado, and F. G. Benitez (1993). A Formulation for the Reaction Time of Traffic Flow
Models. In C. Daganzo (ed). Theory of Transportation and Traffic Flow. Elsevier Science.
Daganzo C. (1995). Requiem for Second-Order Fluid Approximations of Traffic Flow. Transportation
Research-B, Vol.29B, No.4, pp. 277-286.
Daganzo C., Muñoz J.C. (2000). Experimental Characterization of Multi-lane Freeway Traffic Upstream of an
Off-ramp Bottleneck. California PATH Working Paper UCB-ITS-PWP-2000-13.
Drake, J.S., Schofer, J.L., and May, A.D. (1967). A Statistical Analysis of Speed Density Hypotheses, Highway
Research Record. 154, 53-87.
Edie, L.C. 1961. Car following and steady-state theory for non-congested traffic, Operations Research, 9, 66-76.
Greenberg, H. (1959). An Analysis of Traffic Flow. Operations Research, Vol 7, pp. 78-85.
Greenshields, B. D. (1935). A Study in Highway Capacity. Highway Research Board, Proceedings, Vol. 14, p.
458.
Kunhe e Michalopoulos (2000). Continuum Flow Models. Traffic Flow Theory, in Traffic Flow Theory, in
Nathan H. Gartner, Carroll Messer, Ajay K. Rathi (Eds.). Transportation Research Board.
http://www.tfhrc.gov/its/tft/tft.htm.
Lighthill, M. J. and G. B. Whitham, (1955). On Kinematic Waves: II. A Theory of Traffic Flow on Long
Crowded Roads. Proceedings of the Royal Society: A229, pp. 317-347, London.
Michalopoulos, P. G., G. Stephanopoulos, and V. B. Pisharody (1980). Modelling of Traffic Flow at Signalized
Links. Transportation Science, Vol. 14, No. 1, pp. 9-41.
Newell, G. F. (2002). A simplified car-following theory: a lower order model. Transportation Research Part B:
Methodological, 36(3), 195-205.
Papageorgiou, M., J. M. Blosseville, and H. Hadj-Salem (1990). Modelling and Real-Time Control of Traffic
Flow on the Southern Part of the Boulevard Peripherique in Paris Part I: Modelling, Part II: Coordinated
Ramp Metering. Transportation Research A, 24A, pp. 345-370.
Payne, H. J. (1979). FREFLO: A Macroscopic Simulation Model of Freeway Traffic. Transportation Research
Record, Vol.722, pp. 68-77.
Prigogine, I. and R. Herman (1971). Kinetic Theory of Vehicular Traffic. Am. Elsevier Publ., New York.
Sanwal, K. K., Petty, K., Walrand, J., & Fawaz, Y. (1996). An extended macroscopic model for traffic flow.
Transportation Research Part B: Methodological, 30(1), 1-9.
Underwood, R.T. 1961. Speed, volume, and density relationships: Quality and theory of traffic flow. Yale
Bureau of Highway Traffic, pp. 141-188.

34
Chapter 2 Car-Following Models
The car-following model is the fundamental theoretical basis of most traffic flow models, including
those examined in Chapter 1: it arises from a simple and logical behavioral interpretation of the driver's
driving action within a traffic stream.
The car-following model, describing the driving behavior of the individual drivers, reproduces the
dynamics of the motion of the individual vehicles and thus can also represent the unstable stop-and-go
conditions that the macroscopic models, which reproduce the average characteristics of the traffic
stream, are cannot reproduce. Moreover, the car-following model can provide a global representation of
the traffic stream, which allows us to mathematically derive the different macroscopic models that were
obtained as experimental relationships between the average variables of traffic flow in stationary
conditions.
Although the model is relative to traffic streams with no overtaking, so it is not general, it reproduces
the most relevant aspects of traffic flow: longitudinal conditioning between vehicles in dense traffic
situations, which is at the origin of the congestion.
Car-following models are commonly classified into:
• Stimulus-response model: expresses the acceleration of the following vehicle as a function of the
speed relative to the previous one. The first model developed, attributed to Chandler et al. (1958),
was of the mono-regime type. The result of the first systematic experiments on the behavior of
drivers, promoted by General Motors in the early 1960s, is the GHR model which takes its name
from the initials of the authors Gazis, Herman and Rotery (1959) and is also known as General Car
Following Model. This model was then applied in the Transmodeler microsimulation software.
Countless evolutions have been proposed to the present day, differentiating the different values of
the parameters (Helly, 1959; Edie, 1960; Aron, 1988; Ozaki, 1993).
• Collision avoidance model: the vehicle that follows remains at a safe distance from the vehicle in
front of it. The first formulation is due to Kometani and Sasaki (1959), while the most famous
evolution of this model was introduced by Gipps (1981), which provides two different functions for
the free flow and the car-following regime.
• Perception thresholds models (or Psycho-physical 13or Action Points models): human behavior is
represented through different perception thresholds of differences in speed and distance, within
which the queuing conditions produce different reactions of the driver. The first discussion that led
to the development of these models is due to Michaels (1963); then Wiedemann and Reiter (1992)
presented a way to calculate the thresholds and then carry out a simulation. The model he proposed
is the basis of the VISSIM microsimulation software. Subsequently, Fritzsche (1994) presented a
model similar to the previous one, implemented in the PARAMICS microsimulation software.

2.1 Formulation of the Car-Following Models

2.1.1 Formulation of the Stimulus-Response Model

It is a simple linear formulation in which the response of the vehicle is directly proportional to the
relative speed between the vehicle and its leader (stimulus).
The car-following model identified as the stimulus-response model is the product of a series of
experiments on pairs and columns of vehicles, the most significant of which were carried out in the

13
The definition of the perception thresholds model as “psycho-physical” is not preferable because other models also represent the
human behavior in the task of driving a mechanical means and so they also have a psycho-physical nature.

35
 Car-Following Models

1960s by General Motors, which made it possible to calibrate the parameters and validate them the
results. The equation of the car-following model expresses the response of the driver of the generic
vehicle n+1 in the traffic flow following a stimulus from the previous vehicle, n, according to the
sensitivity (psychophysical and mechanical) of the vehicle-driver system (Figure 2–I).
The response is represented by acceleration (dvn+1/dt) that the driver controls by acting on the
accelerator or brake;
Sensitivity λ is the proportionality factor (or more generally a function) which equals the stimulus
function to the control function.
The stimulus is represented by the relative speed vn(t)–vn+1(t); i.e., the rate of spacing variation:
• with a positive relative speed, the vehicles move away, and the driver accelerates;
• in the case of a negative relative speed, the vehicles approach and the driver decelerates.

Remark. The car-following assumes that the behavior of a generic driver in the traffic flow consists
in performing the following tasks:
• follow the vehicle in front of it, which implies a relative speed u≈0;
• avoid a collision, which implies the largest possible collision time tc.
Since tc=s(t)/u (where s is the spacing between successive vehicles), both of these tasks require the
driver to check that in the short term it is tc≈0, i.e. to check that the relative speed is as small as possible.
Since the driver of the following vehicle cannot respond immediately to a stimulus, the driver’s
response is delayed by reaction time T after the stimulus from the leading vehicle. The following basic
equation of the GHR car-following model is obtained (Chandler et al., 1958):
dvn+1 (t + T )
= l [vn (t ) - vn+1 (t )] (2.1)
dt
This differential equation represents the adaptation of the generic driver to variations in spacing
from the vehicle in front. Therefore, it allows us to study the driving dynamics of a couple of vehicles,
given the law of motion of the first of these, and, considering the index n identifying the generic vehicle
to be variable, to study the dynamics of the traffic stream.
Note that the sensitivity has been introduced here as a constant but the GHR model considers also
the case in which it is a function of speed and spacing (Gazis et al. 1963).

al , m [vn +1 (t + T )]
l

l= (2.2)
[sn +1 ]m

Figure 2–I. Platoon of vehicles and relative notations used in the car-following model.

2.1.2 Collision Avoidance Models

Collision avoidance models are based on the assumption that the following vehicle n+1 always
keeps a safe distance from the vehicle n that precedes it. This distance ∆𝑥, coincident with the spacing,
is a function of the speeds 𝑣/ and 𝑣/01 of the two vehicles and the drivers’ reaction time 𝑇 (Kometani
and Sasaki, 1958).

36
 Traffic Engineering and Intelligent Transportation Systems

5 (𝑡)
∆𝑥(𝑡 − 𝑇) = 𝛼𝑣/5 (𝑡 − 𝑇) + 𝛽1 𝑣/01 + 𝛽𝑣/01 (𝑡) + 𝑏8 (2.3)
where α, β1, β, and b0 are calibration parameters.
This expression originates from the safety condition for vehicles on the road with no overtaking that
has already been dealt with in Chapter 1. The safe distance is given by the braking distance
;<
𝑠 = 5= + 2𝑣𝜏 (2.4)

Then, going back to the first expression we can be recognized in the terms 𝛼 5 𝑣/5 (𝑡 − 𝑇) and
5 (𝑡)
𝛽1 𝑣/01 the uniformly accelerated phases of motion, with rate α for the leading vehicle and rate β for
the follower vehicle, and in the term 𝛽𝑣/01 (𝑡) the constant speed section that the follower vehicle
travels during the reaction time τ.
The most well-known evolution of this model, due to Gipps (1981), introduces two different
functions that model two different driving regimes:
• Free-flow regime: The driver’s speed tends to the desired speed so that the acceleration rate
decreases and becomes nil once the desired speed has been reached.
• Car-following regime: the follower tends to change his/her speed to get a safe distance from the
vehicle in front; such a distance is considered safe if the follower can respond to any action of
the leader and avoid the collision.

2.1.3 Perception Threshold Models

The perception thresholds car-following models are based on the assumption that drivers can react
to changes in spacing or relative velocity only when some thresholds are reached. These thresholds
represent different perception levels, corresponding to different driving regimes, that characterized
different driver’s behavior. Multi-regime models are more realistic but have some drawbacks: introduce
discontinuities in the mathematical formulation and require cumbersome solution methods. In the last
years, great advances in computer science opened the opportunity to develop even complex models to
implement sophisticated microsimulation models.
Wiedemann and Reiter (1992) formalized a model that classifies such regimes and introduced
different mathematical relationships to describe the corresponding drivers’ behavior to apply them in a
simulation model. Specifically, Wiedemann’s model distinguishes four driving regimes, in which of
which a driver behaves differently:
• Uninfluenced driving (very far vehicles)
• Closing process (consciously influence by a slower front vehicle)
• Following process (unconscious influence)
• Emergency braking (avoid collision)
Wiedemann’s model represents human perceptions and reactions through a set of thresholds of the
desired distances that delimit these four driving regimes, illustrated in Fig.2.I. The types of driver’s
reaction are different depending on the two cases of either approaching or leaving vehicles, respectively
represented on the right and the left part of the plane. On each part, the different regimes are identified
by thresholds defined by specific mathematical functions that express the desired distance.
The feasible set of solutions as a whole is limited on both sides by a lower bound corresponding to
the distance between stopped vehicles and an upper bound corresponding to the visibility limit.
The desired distance between standing vehicles AX is the lower bound that is given by the sum of
the length L of the vehicle and the desired distance between the bumpers of the two vehicles, which is
defined by the combination the calibration parameters α and β, the latter of which multiplies a normally
distributed driver-dependent variable 𝜀A that considers the safety need for the driver:

37
 Car-Following Models

𝐴𝑋 = 𝐿 + 𝛼 + 𝛽𝜀A (2.5)
By definition, behind this horizontal line traffic states are unfeasible and would correspond to a
collision.

Dx
Visibility Limit

Free Flow SDV

Free Flow

SDX
Closing in
Car
Following
No Reaction Because
Leaving
OPDV ABX

Emergency braking
AX
Stopped Vehicles
–Dx Leaving Vehicles Approaching Vehicles +Dv
(Increasing distance) (Decreasing distance)

Figure 2–II. Different traffic regimes in the car-following model for perception thresholds (simplified figure
from Wiedemann and Reiter, 1992).

Let us now consider the case of approaching vehicles, represented on the right part of the plane.
From the bottom (low distance), the desired minimum distance for approaching vehicles is limited
by a lower bound (represented by the line ABX in the figure) that denotes the perceptual threshold for
the minimum distance that requires an emergency braking to avoid a possible collision:
𝐴𝐵𝑋(𝑡) = 𝐴𝑋 + 𝐵𝑋(𝑡) (2.6)

𝐵𝑋(𝑡) = (𝛾 + 𝛿𝜀A )H𝑣/ (𝑡) (2.7)

where AX is the distance between stopped vehicles defined above and BX is a speed-depending term;
𝑣 is the time-dependent speed of the slowest vehicle, e a normally distributed driver-dependent
parameter, and g and d are calibration parameters. The threshold ABX, because of the BX term, is an
increasing less-than linear function of the vehicle speed. In Figure 2–II, however, it is represented by a
straight line because Figure 2–II represents only relative speed, not the individual’s speed.
At the top, the desired distance for approaching vehicles is limited by an upper bound, (represented
by the curve SDV in the figure) that denotes the point where the driver notices that he is approaching a
slower vehicle and selects the desired spacing. This perceptual threshold of speed variations is not
constant but increases with the relative distance Δ𝑥:
MN(O)PQR
𝑆𝐷𝑉(𝑡) =
S0T·(VW 0VX )
(2.8)
where 𝜂 + 𝜁 are calibration parameters and 𝜀[ is a normally distributed random variable that
considers the drivers’ estimation ability. The left part of the plane describes the case of leaving vehicles,
whose relative distances are increasing.

38
 Traffic Engineering and Intelligent Transportation Systems

Apart from the lower bound that delimits the unfeasible region corresponding to standing vehicles,
in the semi-plane that describes the case of leaving vehicles, there are large portions in which the
following driver has no reaction because the distance from the leader or the relative speed are too high,
and even increasing. The existence of conscious reactions by a driver following a faster vehicle is limited
by two thresholds, corresponding to the perception capability of growing distances.
The upper threshold (represented by the line SDX in the figure) identifies the perceptual capability
of a driver to recognize spacing differences when he is leaving the following process because the leading
vehicle is becoming too far. The mathematical expression is very similar to the function ABX, with the
remarkable difference to consider the speed of the following vehicle instead of the leader, other than the
inclusion of an additional random term 𝜀 that is independent of the driver characteristics and the
corresponding calibration parameter 𝜆:
𝑆𝐷𝑋(𝑡) = 𝐴𝑋 + 𝜆 · (𝜀[ + 𝜀)𝐵𝑋(𝑡) (2.9)

𝐵𝑋(𝑡) = (𝛾 + 𝛿𝜀A )H𝑣/01 (𝑡) (2.10)

The lower threshold (represented by the line OPDV in the figure) identifies the perceptual capability
of a driver to recognize small speed differences at short but increasing distances.
𝑂𝑃𝐷𝑉(𝑡) = 𝜀_ · 𝜉 · (𝜀[ + 𝜀) · 𝑆𝐷𝑉(𝑡) (2.11)
being 𝜉 a calibration parameter.
After having defined the different thresholds we can now describe the drivers’ behavior in the traffic
regimes bounded by such thresholds.
Free-Flow regime. Traffic is in a free flow regime when no conditioning is perceived by the driver
who tends to reach or maintain his desired free speed by applying an acceleration of the vehicle.
𝑑𝑣𝑛+1 𝑣𝑚𝑎𝑥
= 𝜑 · 𝑣𝑚𝑎𝑥 − 𝑣 (2.12)
𝑑𝑡 𝑣𝑑 +𝜗·(𝑣𝑚𝑎𝑥 −𝑣𝑑 ) 𝑛+1

which is a function of the current speed 𝑣𝑛+1 , the desired speed 𝑣𝑑 , and the maximum speed 𝑣𝑚𝑎𝑥 , and
𝜑 and 𝜗 are calibration parameters.
In the case of approaching vehicles (decreasing distances), the free flow regime is above the SDV
curve; in the case of leaving vehicles (increasing distances), the free flow regime may occur even for
small distances, if the leading vehicle is significantly faster (that is, below the lower threshold OPDV),
other than for big distances, when the driver cannot perceive differences in spacing (that is, above the
upper threshold SDX).
Car-following regime. In the portion of the Δ𝑥 -Δ𝑣 plane comprised between the thresholds SDX
and OPDV, the combination of spacing and relative speed is small enough that they are perceived by
the following driver. The driver follows the leader at quite the same speed, and he does not consciously
react to the movements of the leader but tries to keep acceleration low. This is represented by the lowest
value of acceleration and deceleration 𝑎𝑚𝑖𝑛
𝑑𝑣𝑛+1
= 𝑎𝑚𝑖𝑛 (𝜀3 + 𝜀) (2.13)
𝑑𝑡

That multiplies the random component composed by a driver-dependent term 𝜀_ , which considers
the driver’s ability to control acceleration and a driver-independent normalized random term 𝜀 .
Wiedemann assumed that unconscious reactions occur for small differences of speed and distance
modeled in the region between the thresholds OPDV, ABX, or SDX. He modeled this unconscious car-
following behavior by assuming an oscillating process and keeping the acceleration constant until one
of the thresholds OPDV, ABX, or SDX that limit the unconscious car-following region is reached.
Closing-in regime. Traffic states between ABX and SBV correspond to a traffic regime consisting
in a closing-in process in which the following driver realizes that he is temporarily faster than the leader.
After a short delay due to the physical reaction, he begins to decelerate to reduce his speed up to that of

39
 Car-Following Models

the leader and keep a distance no lower than ABX. The deceleration of the follower exceeds that of the
leader by a quantity depending on the relative speed ∆𝑣 and the remaining distance (𝐴𝐵𝑋 − ∆𝑥) from
the minimum desired spacing ABX:
𝑑𝑣𝑛+1 1 ∆𝑣2 𝑑𝑣𝑛 𝜇·(𝜀2 +𝜀)
= · + + (2.14)
𝑑𝑡 2 𝐴𝐵𝑋−∆𝑥 𝑑𝑡 𝜌𝑛+1

The last random term considers the human error in estimating the distance from the preceding
vehicle and introduces a factor 𝜌𝑛+1 that models the learning process of the driver during the conscious
approach: the longer a driver is accelerating, the better he will estimate the motion of the leader vehicle.
Emergency braking. Under the ABX threshold of minimum desired distance, the traffic regime
consists in emergency braking, when vehicles are so close that the follower brakes at a faster rate than
normal; the expression is similar to the previous one, but since the spacing has exceeded the ABX
threshold, the intensity of the deceleration includes an additional term that reduces the maximum
deceleration 𝑏𝑚𝑎𝑥 depending on how much the desired distance has been exceeded:
𝑑𝑣𝑛+1 1 ∆𝑣2 𝑑𝑣𝑛 𝐴𝐵𝑋−∆𝑥
= · + + 𝑏𝑚𝑎𝑥 (2.15)
𝑑𝑡 2 𝐴𝐵𝑋−∆𝑥 𝑑𝑡 𝐵𝑋

Stopped Vehicles. This no-motion condition is the lower boundary of the emergency braking
regime. No traffic state is feasible under this threshold.

Figure 2–III exemplifies a hypothetical vehicle trajectory in the Δv-Δx plane that starts from very
far distance from the vehicle in front, then approaches it and finally follows it from behind. The diagram
can be analyzed proceeding along the trajectory from the top to the bottom. Assuming the relative
distance with the front vehicle is decreasing, the following driver moves along these traffic states:
• Free-flow traffic regime: front-to-rear distance is bigger than the sight distance (green region in
the figure) and there is no reaction by the vehicle;
• Closing-in regime: the vehicle perceives a vehicle ahead (orange region in the figure) and starts
decelerating as a reaction;
• Unconscious car-following regime: for small Δv and Δx, the driver tries to keep the stable state:
the relative speed Δv and the spacing Δx oscillate around the equilibrium in the unconscious
zone.

Figure 2–III. Hypothetical trajectory of a vehicle that approaches and then follows a leading vehicle in Δ𝑥 -Δ𝑣
plane.

40
 Traffic Engineering and Intelligent Transportation Systems

2.1.4 Lower Order Newell’s model

Newell’s model is not properly a new model but it is a basic relationship between vehicle
trajectories, which is demonstrated having a correspondence with Chandler’s and other models in the
literature; in such a way, Newell provides a coherent approach to the basics of traffic in a synthetic way
and introduces a different and wide perspective to look at the car-following behavior starting form very
simple assumptions that reduce the complexity of the previous elaborate models.
The very simple rule introduced in the model is that, if an n-th vehicle is following an (n–1) vehicle
in a homogeneous highway, the time-space trajectory of the nth vehicle is essentially the same as the
(n–1)-th vehicle except for a translation in space and in time14.
On a homogeneous highway, the trajectories of two consecutive vehicles traveling at a constant
speed are:

xn (t ) = x n (0) + vt (2.16)
In order to derive the relation between the two trajectories, it is necessary to specify the spacing
between the vehicles. It is well known that drivers adjust their spacing from the vehicle ahead to reach
a safe distance and that such€safe distance increases as the speed increases. To relate the trajectory of
the following vehicle to that of the vehicle ahead we analyze the adjustment of speed of the follower
when the leader changes his or her speed from v to v’.
We are not interested in the exact trajectory drawn by the vehicle, but we can admit an even rough
approximation with a piecewise trajectory. The actual trajectory may look like the path shown in Figure
2–IV by the broken line but suppose the two constant velocity segments are extrapolated until they
intersect, the solid lines, and that the actual trajectory stays close to the solid lines. The n-th vehicle will
do likewise with the junction of the solid lines for the nth vehicle displaced relative to the (n−1)th vehicle
by a space displacement dn and time lag tn.

Figure 2–IV. Picewise linear approximation to vehicle trajectories according Newell (2002).

The figure shows the relation between sn, dn, v and tn as well as between sn’, dn, v’ and tn. We focus
our attention on the rectangle having sides dn and tn. From geometrical considerations on the picture it
is possible to express the spacing s and s’ before and after the speed change. In fact:

sn = d + vnτ
(2.17)
sn = d + vnτ

14
€
In the description of Newell’s model, unlike in the rest of this chapter, we denote the follower as the n–th and the leader as (n–1)th
to simplify mathematical manipulations applied to the trajectory of the follower.

41
 Car-Following Models

The following kinematic considerations correspond to the evidence of the geometric construction.
At time t, when the (n–1)-th vehicle changes its speed, the vehicle n is at a distance sn. After the time
lag tn it will have covered a distance v·tn and it will be at the distance d from the position that the vehicle
ahead had at time t.
At the same time lag tn, the (n-1)-th vehicle will have covered a distance v’·tn. The spacing with the
following will be the sum of this distance and the space displacement d where, by definition, the
following vehicle will be at time t+tn, when it will adjust its speed to the new value v’.
Now, the model predicts the trajectory of the n-th vehicle xn as a simple linear translation of the
trajectory of the vehicle ahead by the space displacement dn and a time lag tn:

xn (t + τ n ) = xn−1 (t ) − d n (2.18)
The equation is a simple model of inter-vehicle dynamics, which does not introduce any differential
equation, but includes the time displacement as an argument of the first term. If the lead vehicle should
increase (decrease) its speed,
€ the following vehicle does not need to respond immediately. It can wait
until the spacing has increased (decreased) to a value comparable with its value for the new speed.

Correspondence with macroscopic models. In the analysis of observations, one would not
typically observe the motion of every vehicle, but only some suitable “macroscopic” behavior. The τn
and dn are expected to vary considerably from one vehicle to the next because some drivers like to follow
closely at perhaps a minimum safe driving distance whereas others like to leave a comfortable cushion.
For a traffic stream it is possible to write:

( ) ()
xn t + τ n + τ n−1 + ... + τ 1 = x0 t − d n − d n−1 − ... − d1 (2.19)

If we let the arithmetic average of the τ’s and d’s

1 n 1 n
τ = ∑ τ
n i=1 i
, d = ∑d
n i=1 i
(2.20)

and we consider equations (2.17) in the form of the average values, then the average wave speed w is

d
s = d + vτ ⇒ w = (2.21)
τ
The wave, however, propagates like a “random walk” with independent increments vector (τi, di) in
the two-dimensional space of distance–time. It is to notice that the average wave speed does not depend
on the specific traffic state but only on features that describe the drivers’ behavior.
A “stationary flow” will be interpreted here as some region in the x, t plane where all vehicles are
traveling at nearly the same speed v, but possibly random spacings and headways.
To describe the macroscopic behavior in terms of flows q and density k rather than speeds and
spacings, as well-known, the density k is interpreted as the inverse of spacing:

1
k= (2.22)
s
and the speed v can be expressed through the state equation:

q
v= (2.23)
k
Thus, we get:

42
 Traffic Engineering and Intelligent Transportation Systems

q 1 q 1 d
s=d+ τ ⇒ =d + τ ⇒ q= + k (2.24)
k k k τ τ
This establishes the connection between this model and some “fluid models”. It should be noticed
this equation concerns only with the behavior of vehicles if the speed is less than the desired speed.
In Newell’s lower order model, the flow q is a linearly decreasing function of density k with
coefficients depending on the means of the space and time displacements. It is notice also that the ratio
of these displacements determines the average speed of the shock wave w.

Correspondence with traditional car-following models. To see the correspondence between

equation (2.18) and some of the more traditional car-following models, suppose that the n-th vehicle
could follow the equation (2.18)exactly and the (n−1)-th vehicle followed some smooth trajectory.
The equation (2.21) represents the relationship between the two extreme states before and after the
change of speed of the following vehicle. Thus, t is not the driver’s reaction time, which should be less
than t: if the speed change maneuver would be smoother, the driver would start accelerating before the
time t, at an intermediate instant t+Tn between t and t+tn. This can be formulated mathematically by
applying the mean value theorem of calculus between xn and xn(t+τn):

xn (t + τ n ) = xn (t ) + vn (t + Tn )τ n (2.25)
where vn is the derivative of xn in an intermediate point Tn within the segment [t, t+tn].
Substitution of the last equation in the model gets:
€
xn (t ) + v n (t + Tn )τ n = xn−1 (t ) − d n (2.26)
If we express the speed as a function of spacing

1 d
€ vn (t + Tn ) =
τn
[ ]
x n−1 (t ) − xn (t ) + n
τn
(2.27)

and derive with respect to time, we obtain the well-known GHR car following model:
1
€ an (t + Tn ) = [
v (t ) − vn (t )
τ n n−1
] (2.28)

as introduced by Chandler et al. (1958) assuming the sensitivity be constant.

Remarks. Newell€(2002) provides a different perspective to look at car-following behavior and a

new approach to derive both microscopic and macroscopic models.
From the microscopic point of view, provides the following correspondence with traditional car
following models:
• the reaction time Tn is different from the time lag tn that describes the time of the speed
change maneuver through an approximate finite discrete scheme
• the driver’s sensitivity l corresponds to the inverse of the time lag 1/tn
From the macroscopic point of view, the following correspondence holds with fluid models:
• the maximum density kJ corresponds to the inverse of the space displacement dn
• the average speed of the shock wave is given by the ratio dn/tn, that is ln/kJ
Note that Newell’s model only considers speeds below the desired speed which describe congested
traffic states. The macroscopic model may be completed by assuming that uncongested states are
described by the simple condition v=vf.
We get so a triangular fundamental diagram with the increasing branch that describe the
uncongested states and the decreasing branch that describe the congested states, as described in Figure

43
 Car-Following Models

2–V. Such a triangular model was used by Daganzo (1994) to develop his cell transmission model, that
provide a discrete linear formulation of the macroscopic traffic flow.
If we could put together microscopic and macroscopic features, we would have the following
macroscopic model:

q = v f k, 0 ≤ k < kc
⎛ k ⎞ (2.29)
q = λ ⎜ 1+ ⎟ , kc ≤ k < k J
⎝ kJ ⎠

q
1/t= l

qmax

vf w

kc kJ k

Figure 2–V. Triangular fundamental diagram derived from lower order Newell’s model and applied in the Cell
Transmission Model (Daganzo (1994).

2.2 Stability Analysis

The stability analysis concerns the GHR model. From a theoretical point of view, it is of interest to
determine, on the one hand, what are the conditions for system instability (hiccup or collision between
vehicles) and, on the other, which law regulates the system at steady state. The latter aspect is
fundamental for studying the system's performance in “normal” conditions, when the vehicles all travel
at the same speed, regardless of the behavioral differences between one driver and another, which are
considered negligible compared to the macroscopic behavior of the traffic stream. The conditions that
determine the establishment of instability are an equally important aspect since the system performance
is extremely degraded. From an engineering point of view, the determination of the stationary state is
mainly relevant for the planning and design of the system components; the instability conditions, on the
other hand, have relevance mainly for the design of the regulation and control systems.
The following paragraphs are dedicated to the study of the transient, which allows determining the
conditions for the stability of the system, to the study of the stationary state and to a brief mention of
the experiments performed for the calibration of the model.
In physics, an equilibrium system is said to be stable if, once perturbed, it returns to its initial
position at the end of the perturbation; unstable if, following the disturbance, it moves away indefinitely.
In mathematical terms, a system subjected to perturbations is described by a system of differential
equations. The solution of the system is said to be stable if an arbitrarily small variation of the initial
values implies, after an interval of anyone, a sufficiently small variation of the solution. The solution of
the system is said to be asymptotically stable if not only is it stable, but, after an infinite time, it tends

44
 Traffic Engineering and Intelligent Transportation Systems

to any other solution of the system15. In an asymptotically stable system, therefore, all solutions finally
converge to a single value.

2.2.1 Local Stability

To study the local stability of the motion of a vehicle n+1 with respect to a vehicle n, we must
integrate the differential equation (2.1). The integral of the homogeneous equation represents the
characteristics of the system regardless of any external perturbations and therefore reproduces the
behavior of a generic pair of vehicles (n, n+1).
The meaning of the solution is illustrated in Figure 2–VI. From top to bottom, it shows the
acceleration and the absolute speed of the leader and the follower (represented by the continuous and
the dashed curve, respectively), and therefore the relative speed and the distance between the two
vehicles.
The integration of (2.1) in a closed-form presents prohibitive difficulties if the sensitivity, as in
general expected, is a function of the speed. In fact, it would be a nonlinear differential equation in the
dependent variable. Instead, it is rather simple if we assume that the sensitivity is constant.

Figure 2–VI. From above: acceleration, speed, relative speed and relative distance of a pair of vehicles according
to the model of the car-following vehicle in the case of a stable pair of drivers (Rothery, 1998).

In the case in which the sensitivity λ is a constant, the solution can be obtained in an exact form by
using the Laplace transform method or, in an approximate form, by expanding the (1) in Taylor series
starting from instant t with increment T. The second method is undoubtedly more interesting for the
analysis of the solution since it allows to transform equation (1) into a second-order differential equation
with constant coefficients. This is a mathematical form that is common to many well-known stability
problems of the physical systems, like mechanics systems or electrical circuits.
Let be λ=c and J = t+T (to derive v instead of dv/dt in the Taylor series expansion), we can rewrite
equation (1) as follows:
dvn +1 (J )
= c[vn (J - T ) - vn +1 (J - T )] (2.30)
dJ

15
Cfr. for example: Elsgolts L.E. (1981). Differential equations and variation calculation. Reunited Publishers, MIR Editions, Rome.
pp.205-206.

45
 Car-Following Models

By developing the two terms of the second member with respect to the time interval –T:

dvn (J ) 1 2 d 2vn (J )
vn (J - T ) = vn (J ) - T + T
dJ 2 dJ 2 2 (2.31)
dv (J ) 1 2 d vn +1 (J )
vn +1 (J - T ) = vn +1 (J ) - T n +1 + T
dJ 2 dJ 2
we get a II-order differential equation with constant coefficients:

dvn+1 dv 1 d 2 vn dvn+1 1 2 d 2 vn+1

= cvn - cT n + cT 2 - cv n+1 + cT - cT (2.32)
dJ dJ 2 dJ 2 dJ 2 dJ 2

By moving to the first member the terms related to the vehicle n + 1, which constitute the unknown
function of the problem, and placing to the second member the members related to the II vehicle, which
represent the external disturbance, we get:

1 2 d 2 vn+1 dvn+1 1 2 d 2 vn dv
cT (
+ 1 - cT ) + cvn+1 = cT - cT n + cvn (2.33)
2 dJ 2
dJ 2 dJ 2
dJ
The general integral of the homogeneous equation:

1 2 d 2 vn+1 dv
cT + (1 - cT ) n+1 + cv n+1 = 0 (2.34)
2 dJ 2
dJ
is the sum of two exponential functions, whose coefficients are the roots of the associated
characteristic equation:

cT 2 2
m + (1 - cT )m + c = 0 (2.35)
2
The roots m1.2 of this second-degree algebraic equation are real if the discriminant of (6) is non-
negative (∆ ≥ 0); however, they are complex if it is ∆ < 0.

=-
(1 - cT ) æ (1 - cT ) ö
±
2

-
2
= b ± D = b ±g (2.36)
m1, 2 2 ç 2 ÷
cT è cT ø T2
The general integral of (4), as known from mathematical analysis, is a linear combination of
exponential functions with a real or complex argument, according to the value of the discriminant ∆ of
(6):

D > 0 Þ m1, 2 Î Â Þ vn+1 (J ) = c1e m1J + c2e m2J =c1e( b +g )J + c2e( b -g )J

(2.37)
D £ 0 Þ m1, 2 Ï Â Þ vn+1 (J ) = c1e m1J + c2e m2J =c1e bJ +igJ + c2e bJ -igJ

Case 1. ∆>0 (Real Roots)

The condition ∆>0 holds if the term cT has value comprised between 0 and 0.141. In fact, the
expression of the discriminant results in the following inequality:

æ (1 - cT ) ö
2
2
- 2 > 0 Þ (1 - cT ) > 2c 2T 2 Þ c 2T 2 + 2cT - 1 < 0
2
D= ç 2 ÷
(2.38)
è cT ø T
whose solution is:

cT < 2 - 1 = 0,141 (2.39)

46
 Traffic Engineering and Intelligent Transportation Systems

The general integral of (2.34) is an exponential function with a real argument, which is positive or
negative depending on the value of the model coefficients. In this case, the roots are both negative, as it
is always:

æ 2 ö
g = D = çb 2 - ÷<b (2.40)
è T2 ø
It follows that the solution

vn+1 (J ) =c1e ( b +g )J + c2 e ( b -g )J (2.41)

is a damped function and is therefore stable.

Case 2. ∆<0 (Complex Roots)

If the roots of the characteristic equation are complex (∆≤0, i.e. cT>0,141), the general integral of
equation (4) is a linear combination of two exponential functions with complex argument:

vn+1 (J ) = c1e bJ +igJ + c2 e bJ -igJ (2.42)

having an oscillatory shape. The coefficient of the imaginary component

g= D
determines the frequency of the oscillations while the coefficient of the real component determines its
amplitude; they are damped if β <0, amplified if β >0, periodic if β=0.

Summary of results
The previous16 results are summarized in Table 2–1 depending on the product of sensitivity c and
reaction time T. It can be observed that the solution is locally stable if cT ≤ 1.
Table 2–1. Results of the local stability study (approximation to the 2nd order).

Range of cT b ∆ Roots Trend Stable

2 2
0<cT<0,414 b<0 0<∆=√(b -2/T )<|b | Real negative Not damped oscillatory Yes

0,414≤cT<1 b<0 ∆≤0 Complex Damped Oscillatory Yes

cT=1 b=0 ∆≤0 Complex Periodic Oscillatory Yes

cT>1 b>0 ∆≤0 Complex Amplified Oscillatory No

It is worth noting that the stability of the model depends jointly on sensitivity and reaction time; in
particular, it requires that their product be less than a threshold value. A high value of the sensitivity
gives rise to an exaggerated reaction to the stimulus received by the leader. Thus, it gives rise to abrupt
accelerations and decelerations, which can cause, for certain values, the following vehicle to crash even
when the previous vehicle has made a temporary deceleration. On the other hand, a long reaction time
delays the deceleration maneuver and leads the following vehicle to get too close to the previous vehicle.
In such a condition, the following driver can do intense braking and, in some cases, cannot avoid a

16
In Tab.2-I the results of the approximate solution are shown. The corresponding results of the exact solution obtained with the
transformations of Laplace are: for 0<cT<e-1 (≈0,368): non-oscillatory damped trend; For e-1<cT<π/2 (≈1,571): oscillatory damped
trend; for cT=π/2: periodic oscillatory trend; for cT>π/2: amplified oscillatory trend. It is to notice that the approximation of the 2nd
order gives rise errors of up to 50%. It must be said, however, that the study of local stability has a more theoretical than a practical
value since the condition that acts in the reality is the most restrictive one; that is, that on the asymptotic stability, the solution of
which will be determined in the following paragraph without approximations, apart from that introduced with the hypothesis of
constant sensitivity

47
 Car-Following Models

collision. Figure 2–VII and Figure 2–VIII show the general integral of the car-following model for the
three possible different cases of the solution: non-oscillatory damped motion; dampened oscillatory
motion; periodic and amplified.

2 2

1.5
1.6
1

0.5
1.2
0

v
v

0.8 -0.5

-1
0.4
-1.5

0 -2
0 2 4 6 8 10 0 10 20 30 40 50
t t

Figure 2–VII. General solutions of the GHR car-following model: on the left, non-oscillatory dampened trend for
c=0,35s-1 and T=1s; on the right, oscillatory damped solution for c=0,70s-1 and T=1s.

2.5 8
2
6
1.5
4
1
0.5 2

0 0
v

-0.5
-2
-1
-4
-1.5
-2 -6

-2.5 -8
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10
t t

Figure 2–VIII. General solutions of the GHR car-following model: on the left, periodic oscillatory solution for
c=1s-1 and T=1s; on the right: Amplified oscillatory solution for c=0.80s-1 and T=1,5s.

Remark. It is interesting to note that the equation (2.33) is analogous to the equation of the
dynamics of a body of mass m subject to a force f(t), an elastic force kx, a force of friction cx! , and the
force of inertia m!x! :

m!x! - µx! + kx = f (t ) (2.43)

Their oscillations are determined by the integral of the associated homogeneous equation. The
coefficients of the solution are still the roots of the characteristic equation:

µ µ2 k
a1, 2 = b + D = - ± - (2.44)
2m 4m 2 m
The following correspondence holds between the coefficients of the car-following model and the
coefficients of the mechanic system:

48
 Traffic Engineering and Intelligent Transportation Systems

cT 2 =>m [mass]
(cT) => [friction]
c => k [elastic constant]
For the stability of the system, sensitivity plays a similar role to that of the elastic constant of a rigid
body recall spring. However, it also enters the other two terms: as a reducing element of friction and as
a proportional element to the mass, so its role is not of immediate and unequivocal interpretation.

2.2.2 Asymptotic Stability

It has already been said that a system of n elements is asymptotically stable if all its elements tend
to assume the same value as the time tends to infinity. A traffic stream is therefore asymptotically stable
if, after acceleration or deceleration of a generic vehicle, all the following vehicles tend to assume, after
adequate time, the same speed. To study the asymptotic stability of the traffic stream, therefore, it is
sufficient to verify that a generic perturbation is not amplified by propagating along with the traffic
stream.
The response of a system to a given perturbation is determined by the particular integral of the
differential equation (2.33), complemented by some specific boundary conditions. In fact, the integral
of the homogeneous equation represents the characteristics of the system, regardless of the known term.
The integral of the non-homogeneous equation, on the other hand, describes the behavior of the system
in the presence of an external disturbance, described by the known term. To this end, it is necessary to
verify that, for a given oscillatory disturbance of the generic vehicle n, the amplitude of the oscillations
of the following vehicle n+1 is less than the amplitude of the previous one.
To determine the particular integral of the basic equation, we assume a variation in vehicle speed n
that first gradually decreases and then increases, according to an oscillatory type law with amplitude Vn:

vn (t ) = Vn e iwt (2.45)
The speed of the vehicle n+1 will also have a similar oscillatory pattern, of different size Vn+1 and
of equal frequency, as it will have to adapt its speed to the speed changes of the vehicle n:

vn+1 (t ) = Vn+1e iwt (2.46)

Then, we substitute the assumed solution, consisting of equations (2.45) and (2.46), in Taylor's
development of the basic equation (2.33) and, after the derivation operations, we obtain:

é 2 cT 2 ù iwt é 2 cT
2
ù
iwt
Vn+1e ê- w + iw (1 - cT ) + cú = Vn e ê- w - iwcT + cú (2.47)
ë 2 û ë 2 û
The oscillatory terms appear on both sides of the equation and can be simplified. For the system to
be stable, the amplitude of the oscillations must not grow with n. This corresponds to the condition that
the module Vn+1 of the speed of the follower n+1 is less than or equal to the module Vn+1 of the speed of
the leader n:

cT 2
-w 2 - iwcT + c
Vn+1 2
= £ 1Þ - cT £ 1 - cT (2.48)
Vn cT 2
-w 2
+ iw (1 - cT ) + c
2
Since the asymptotic stability always implies the local stability, it is cT≤1. Thus, we can write:
Þ cT ≤ |1−cT| = 1−cT Þ 2cT ≤ 1

This finally provides the condition for the asymptotic stability of a traffic stream of n vehicles:

49
 Car-Following Models

cT ≤ 1/2 (2.49)
It is to observe that the condition of asymptotic stability includes, of course, that of local stability.
Figure 2–IX illustrates the condition of asymptotic stability in the plane sensitivity-time reaction.

0.75 Flusso
instabile
c cT=0,5
0.5
Flusso
stabile
0.25

0
0 0.5 1 1.5 2
T
Figure 2–IX. Condition for asymptotic stability in the plane sensitivity (c) – reaction time (T).

Figure 2–X illustrates an example of asymptotically stable speeds for a platoon of 9 vehicles. In this
example, the perturbation consists of a maneuver of the first vehicle that at the time t = 9s decelerates
for 3s with an acceleration equal to –1ms-2; then, it accelerates again for 3s with an acceleration equal
to +1ms-2. The reaction time and the sensitivity are assumed to be 1s and 0.44s-1, respectively.

22

21

20
v (m/s)

19

18
n =9
n =7
n =5
17
n =3
n =1
16
5 10 15 20 25 30 35 40 45 50
t (s)

Figure 2–X. Speed profile of an asymptotically stable platoon (parameter values: c-0.44s-1; T-1s), as a result of a
disturbance represented by a uniform deceleration of the first vehicle as -1m/s-2 for 3s starting at t=9s and a similar
acceleration of 1m/s-2 again for 3s.

50
 Traffic Engineering and Intelligent Transportation Systems

In the stable case, the amplitude of the oscillations is reduced both for the same vehicle over time
(local stability) and from one vehicle to another along with the traffic stream (asymptotic stability).
In the asymptotically unstable case (at least for values cT <1, as in this example), the amplitude of
the oscillations of the same vehicle first increases, then decreases with time, but increases from one
vehicle to the next.
The amplitude of the oscillations is naturally limited by physical constraints. First of all, the speed
and the distance between vehicles cannot take on negative values. Moreover, we must consider that
while the speed of the platoon varies over time, at the same time the spacing between vehicles varies
accordingly; since the spacing is the integral of the speed it has an analogous oscillatory profile but out
of phase with respect to that of the speed.
It follows that, in the case of unstable traffic, both the speed and the spacing between vehicles
oscillate until either the speed or the spacing becomes zero:
• if the speed is reduced to zero by first and the distance is different from zero, vehicles will go
according to a stop-and-go profile;
• if the spacing is reduced to zero by first, while the speed is greater than zero, we will have a rear-
end collision.

Remark. In general, a rear-end collision does not occur between the second and the first vehicle
but between two successive vehicles along the traffic stream, because the local stability is a less
restrictive condition than the asymptotic one. This condition is more often observed in the reality,
especially on a highway when, after a relatively abrupt slowdown of the first vehicle, the second
manages to brake, the third also, while a subsequent vehicle along the traffic stream ends up hitting the
vehicle in front of it.
The previous considerations are exemplified in Figure 2–XI, which shows the trajectory of a platoon
of vehicles subject to the same perturbation assumed in the graph of Figure 2–X. In this case, however,
the values of the car-following parameters are c = 1s-1; T = 1s; therefore, the traffic stream is
asymptotically unstable and, after about 13s (at time t = 22s), the model predicts a collision between the
8th vehicle and the 7th.

400

350

300

250
x (m)

Collisione
200 n =1
n =3 n =5 n =7 n =9
150

100

50
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
t (s)

Figure 2–XI. Trajectories of an asymptotically unstable platoon of vehicles (parameter values: c=1s-1; T=1s),
showing a collision between the 8th vehicle and the 7th after about 22s, as a result of a disturbance represented by

51
 Car-Following Models

a uniform deceleration of the first vehicle as -1m/s-2 for 3s starting at t=9s and a similar acceleration of 1m/s-2
again for 3s.

Figure 2–XII highlights the oscillatory trend of relative spacing and speeds at the same perturbation
assumed in the previous figures, assuming a sensitivity value of 0.80s-1. This is an unstable run-off
condition, in which the seventh vehicle goes to buffer the sixth: after 29.4s the distance from the sixth
vehicle is canceled, while the speed is greater than zero, and equal to about 5.4m/s. Conversely, the
eighth and non-vehicle are able to stop (after 30.8 and 32.4s respectively), without buffering the vehicle
in front of them, as can be seen in the figure, observing how the spacing in those same moments is
greater than zero.

50

40

30
s (m)

20

n =1
10
n =3
n =5
0 n =7
5 10 15 20 25 30 35 40 45 50
s7 (t = 29,4 ) = 0 t (s)

40

30
v (m/s)

20

v8 (t = 30 ,8 ) = 0
n =1
10 n =3 v9 (t = 32,4) = 0
n =5
n =7
v7 (t = 29,4) = 5,4m/s
0
5 10 15 20 25 30 35 40 45 50
t (s)

Figure 2–XII. Spacing profile (top) and speed profile (bottom) of an asymptotically unstable platoon having
parameters c=0.80s-1; T=1s, with a collision of the 7th vehicle, as a result of a disturbance represented by a uniform
deceleration of the first vehicle as -1m/s-2 for 3s starting at t=9s and a similar acceleration of 1m/s-2 again for 3s.

52
 Traffic Engineering and Intelligent Transportation Systems

2.3 Analysis of the Stationary State

Although the car-following model is a dynamic model it can supply also a functional relationship
between acceleration and speed in the stationary state, i.e. in the absence of perturbations, if we simply
set T = 0 in equation (2.1):
dvn+1 (t )
= l[vn (t ) - vn+1 (t )] (2.50)
dt
and then we integrate it with respect to time.
Compared to (2.1), (2.50) is a much simpler equation since all the quantities are referred to the same
time instant. For simplicity, the corresponding subscript is omitted.
An important property of the car-following model is that, if we express the kinematic quantities of
a generical vehicle in the variables of the traffic stream, we can derive the different macroscopic models
described in Chapter 1 by assuming suitable values for the coefficients of the sensitivity in equation
(2.2).

2.3.1 Linear Model

By integrating equation (2.50) with respect to time we can get a functional relationship between the
speed and the positions of two consecutive vehicles. Under the same hypothesis of constant sensitivity
λ adopted so far, the model assumes a simple linear form17:
vn+1 = l ( xn - xn+1 ) + a (2.51)
being xn the abscissa of the generic vehicle n. Equation (2.52)can be more simply expressed by indicating
in term of with s:
s n+1 = xn - xn+1
(2.52)
vn+1 = ls n+1 + a
To determine the value of the integration constant we impose that at zero speed the spacing equals
L, which is the average length of the vehicle plus a minimum safe distance between two standing
vehicles.
vn+1 (L ) = lL + a = 0 Þ a = -lL (2.53)
So, equation (2.52) can be written as:
vn+1 = l (s n+1 - L ) (2.54)
Since the spacing is the inverse of the density, equation (19’) can be written as a macroscopic
relationship between flow quantities:

æ1 1 ö
v = l çç - ÷÷ (2.55)
è k kJ ø
where kJ is the maximum density; that is the inverse of the minimum spacing L.

17
This model is often cited in literature as Pipes (1953), which formulated it directly for the stationary state before Chandler et al.
(1958) presented the dynamic formulation (1).

53
 Car-Following Models

Equation (2.55) expresses a hyperbolic link between speed and density, which is consistent with a
decreasing trend, even if not fully responsive to the experimental observations (Figure 2–XIII).

80

60

v (veic/h)
40

20

0
0 5 10 15 20 25
k (veic/km)

Figure 2–XIII. Comparison between the hyperbolic speed-density function derived from the linear car-following
model and the experimental data collected by Edie et al. (1963) and reported in Rothery (1998).

The inconsistency of the linear model is predictable if we think that equation (2.54) implies a linear
relationship between speed and spacing; however, an inverse quadratic-type function would be
expected, provided that the function between braking distance and speed is quadratic.
Multiplying both members of (2.55) by k and applying the state equation, a linear relationship is
obtained between flow and density:

æ k ö
q = l çç1 - ÷÷ (2.56)
è kJ ø
This trend is clearly in contrast to the experimental data, except as an approximation of only one of
the two branches of the flow-density18curve, as can be seen from Figure 2–XIV.

18
However, some authors considered the model of Pipes, imposing that for k<kc=qmax/vl Is: v=vl And q=vlk. So there is a link q-k
triangular shape.

54
 Traffic Engineering and Intelligent Transportation Systems

1400

1200

1000

q (veic/h)
800

600

400

200

0
0 20 40 60 80 100
k (veic/km)

Figure 2–XIV. Comparison between the linear flow-density function resulting from the linear car-following
model and the experimental data (qualitative reproduction from Rothery, 1998).

Remark. Both logical considerations and experimental observations lead us to reject the model of
the linear car-following vehicle derived when assuming a constant sensitivity.
Since the stability study was possible because it was assumed that sensitivity was constant, at this
point we might think that the results obtained so far be completely useless. In the following paragraph,
we will see how an in-depth analysis of the possible functional form of sensitivity will allow us to derive
a likely field of applicability of the stability condition.

2.3.2 Non-Linear Models

The hypothesis that the sensitivity is constant, introduced to derive the linear model, demonstrated
to not responds to logic and experimental results. Indeed, this hypothesis corresponds to the assumption
that the acceleration response of a vehicle is a function of the relative speed of the vehicles only,
regardless of the speed at which they are located. However, at a very low speed, the vehicles are very
close to each other, at a distance of even the order of 4 or 5m; at a high speed, on the contrary, the
spacing can reach distances of the order of even 80 or 100m.
It is reasonable to think that the response to a variation of the relative speed is greater if the vehicles
are very close. Moreover, a speed variation of 10 km/h would be barely noticeable for vehicles that, for
example, travel at 130 km/h and are spaced of 80 m, while it is immediately noticeable by the driver of
a vehicle traveling at 30 km/h spaced, for example, by 15m.
Thus, it is reasonable to assume that the sensitivity decreases as the vehicular spacing increases. The
simplest model that complies with this assumption is obtained by assuming that the sensitivity is
inversely proportional to the spacing.

a
l= (2.57)
sn+1 (t )
where a is a proportionality parameter.
It is now convenient to express the basic equation of the car-following at the stationary state (2.50)
in terms of spacing rather than relative speed.

55
 Car-Following Models

dv n+1 (t ) a ds n+1 (t )
= (2.58)
dt sn+1 (t ) dt
By integrating the (2.59) with respect to t and omitting both the subscript n and the dependency on
the time, we get a logarithmic function:

ò ò
ds æ1ö
dv = a Þ v = a ln(s ) + b = a lnç ÷ + b (2.59)
s èkø
Now, we impose that the speed is zero at the maximum density kJ and we get the value of the
constant b:

æ 1 ö æ 1 ö
v(k J ) = 0 Þ a lnçç ÷÷ + b = 0 Þ b = -a lnçç ÷÷ = a ln(k J ) (2.60)
è kJ ø è kJ ø
Finally, by applying the properties of logarithms, we obtain the following expression of the velocity-
density function that corresponds to Greenberg's model (1959).

æk ö
v = a lnç J ÷ (2.61)
è k ø
It is possible to demonstrate that the constant a corresponds to the critical value of the speed vc that
corresponds to the maximum flow. To this end, we impose that the derivative of the flow is zero q(k) at
the critical point, when it is v = vc.

dq dk dv(k ) dv(k ) d é æk öù ,
= 0 Þv + k =v+k = v + êa lnç J ÷ú = 0
dk dk dk dk dk ë è k øû
(2.62)
d é æ kJ öù é k æ 1 öù æ aö
v+ a lnç ÷ú = v + k êa × k J ç - 2 ÷ú = v + k ç - ÷ = vc - a = 0 Þ a = vc
dk êë è k øû ë kJ è k øû è kø

Remark. The hypothesis that the sensitivity is inversely proportional to the spacing, other than
logically convincing, is also appropriate to reproduce the experimental data from an experimental point
of view.
As a matter of fact, Greenberg's model allows a better approximation of the experimental data than
the linear model, both in the velocity-density plane, where the inverse logarithmic curve approximates
the observed values very well (Figure 2–XV), and in the plane flow-density, where the convex curve
follows the first increasing and then decreasing trend of the flow (Figure 2–XVI).
Furthermore, Greenberg's model appears especially satisfactory from an engineering point of view,
since:
• provides a physical interpretation of sensitivity, defined as the relationship between critical speed
and distance;
• expresses the stationary flow model as a function of the most significant quantities for design
purposes: the critical speed, at which to measure the capacity value, and the maximum density, which
defines the density of vehicles in the queue.

56
 Traffic Engineering and Intelligent Transportation Systems

80

60

v (km/h)
40

20

0
0 20 40 60 80 100
k (veic/km)

Figure 2–XV. Comparison of the speed-density function of the Greenberg model derived from the hyperbolic
model of the car-following model and the experimental data of Edie et al. (1963), reported in Rothery (1998).

1400

1200

1000
q (veic/h)

800

600

400

200

0
0 20 40 60 80 100
k (veic/km)

Figure 2–XVI. Comparison of the flow-density function of the Greenberg model derived from the hyperbolic
model of the car-following model and the experimental data of Edie et al. (1963), reported in Rothery (1998).

It is possible to show that the general form of sensitivity (2) due to Gazis et al. (1963):

al , m [vn +1 (t + T )]
l

l= (2.63)
[sn +1 ]m
allows deriving many of the congestion models proposed in the literature by assigning appropriate
values to the parameters of the sensitivity function. The different formulations are shown in Table 2–2
and graphically described in Figure 2–XVII.
It should be noted, however, that the general form of sensitivity is theoretically unsatisfactory, since,
being v=v(s), it is not correct to express the sensitivity as a function of both speed and spacing, since it

57
 Car-Following Models

can be reduced to a function of only one of the two through a simple operation of variable substitution
(Papola, 1988).

Table 2–2. Several speed-density flow models corresponding to different values of the parameters l and m of the
sensitivity function.

m L Model Formula m L Model Formula

æ1 1 ö
0 0 Pipes model v = cçç - ÷÷ 1 2 Edie Model v = vl e - k kc

è k kJ ø
æk ö
2
Greenberg æ k ö
0 1 v = vc lnç J ÷ 1 3 Drake Model -0, 5çç
è kc
÷
÷
Model è k ø v = vl e ø

Greenshields æ k ö
0 2 v = vl çç1 - ÷÷
Model
è kJ ø

Figure 2–XVII. Several flow-density models corresponding to different values of the parameter l of the
sensitivity function (Rothery, 1998).

2.4 Experimental Results

The car-following model was developed between the late 1950s and early 1960s based on the results
of numerous experiments conducted on vehicle platoons, in which the spacing, speed, and acceleration
of each vehicle were simultaneously measured. Most of these experiments were performed on a track
with volunteer drivers. The most famous experiments were performed by General Motors, both on cars
(Chandler et al., 1958) and buses (Rothery et al., 1964). These experiments were divided into two parts,
relating to a pair of vehicles and the platoon. Recently, thanks to the availability of the GPS satellite
tracking system for measurements, the interest in experimental studies on car-following behavior
renewed. In this subsection, the earliest experiments at General Motors and a recent study conducted by
La Sapienza University are described.

2.4.1 General Motors Experiments

Experiments on pairs of vehicles were organized by recording on magnetic tape the driving
instructions necessary for the driver of the first vehicle to carry out a predetermined driving diagram,
including a given sequence of accelerations and decelerations. Drivers who were called to follow the
first vehicle from time to time received the only indication of driving as they considered normal and

58
 Traffic Engineering and Intelligent Transportation Systems

safe. The data obtained were used to verify and calibrate different functional forms of the queued vehicle
model. Sensitivity was estimated by the detection of the other terms of the basic equation (2.1). In most
experiments, the best result was achieved by adopting the assumption of sensitivity inversely
proportional to space. Figure 2–XVIII shows the calibration of the sensitivity function on the
experimental data collected by Chandler et al. (1958)19. This result further confirms the good
performances of Greenberg's model.

0,5

0,4
y = 12,9x
2
0,3 R = 0,4355
λ (s )
-1

0,2

0,1

0
0,000 0,008 0,016 0,024 0,032 0,040 0,048 0,056
-1
1/s (m )

Figure 2–XVIII. Calibration of the sensitivity as a function of the inverse of spacing using data from Chandler et
al., 1958).

Experimental results on pairs of vehicles, allowing a measure of sensitivity, also provided the
possibility to verify the stability conditions experimentally. The following figures reports experimental
data of sensitivity and reaction time, which are compared with local and asymptotic stability conditions,
respectively for the experiments conducted in the Holland and Queens Midtown Tunnels with cars
inserted into the real traffic stream (Figure 2–XIX) and for the experiment with a platoon of buses on a
specific track by Rothery et al. (Figure 2–XIX).

19
It must be said that the calibration of the sensitivity function was performed by Chandler et al. (1958) using only 6 of the 8 available
data; excluding in addition to the first vehicle, another particularly anomalous data. Moreover, the estimation of reaction time, which
varies between 1s and 4.5s, would suggest some inaccuracies in the measures. Despite the shortcomings of the Chandler experiment
et al., the trend of the sensitivity function obtained by them was confirmed by all subsequent experiments, despite the due differences
in the value of the parameter: 45km/h according to Chandler et al.; from 29.2 to 32.7km/h secondo Herman and Potts (1959),
respectively, in the Holland Tunnel and Lincoln Tunnel experiments, conducted in a traffic stream; about 58km/h according to
Rothery et al. (1964) in an experiment on pairs of buses on the track.

59
 Car-Following Models

Figure 2–XIX. Correspondence between Figure 2–XX. Correspondence between

sensitivity and reaction time for stability sensitivity and reaction time for stability
experiments in the Holland and Queens Midtown verification in experiments on bus pairs (Rothery
Tunnel experiments (Rothery, 1968, in Rothery, et al. , 1964, in Rothery, 1998).
1998).
From the observation of the previous figures, one can see that about half of the users involved in the
experiment with cars (pictured left) have values of the parameters (pictured on the right) that ensure an
asymptotic stability while the other half is in the range between asymptotic and local stability; all the
drivers involved in the experiment with busses (pictured right) have values of the parameters that verify
the stability condition, implying that the professional drivers have a greater ability to calibrate their
reaction to stimuli from previous vehicles.
Experiments on vehicle platoons were also aimed at describing the stationary state. For this reason,
the macroscopic variables of the traffic stream –average speed, density, and flow– were measured,
respectively, by detecting the number of vehicles present in certain road segments of known length and
the number of vehicles traveling through some fixed sections. The resulting calibrated Greenberg’s
model is illustrated in Figure 2–XXI.

Figure 2–XXI. Calibration of the speed-density function according to the Greenberg model (Rothery et al.,
1964).

60
 Traffic Engineering and Intelligent Transportation Systems

Observation. The curve in Figure 2–XXI was calibrated by transforming the measures of the traffic
variables into a logarithmic plane relative to the speed, thus obtaining a linear model in density. In the
logarithmic plane, the line approximates the experimental points well. This is not, however, a guarantee
that the calibrated function is as satisfactory in the ordinary speed-density plan. In fact, the calibration
procedure was achieved by minimizing the sum of the deviations between observations and the natural
logarithm of the values predicted by the model, which in general does not coincide with the sum of the
deviations between observations and expected values, and tends to underestimate the differences
between the larger values. The calibration results could be improved by using another method than the
linear regression of the function logarithm, which is to directly minimize the sum of the quadratic
deviations between expected values and observed values using, for example, a gradient procedure.

2.4.2 Experiments with Differential GPS

The rising diffusion of satellite tracking techniques with centimeter precision20 allows making
observations on drivers' behavior that until a few years ago were possible only with very laborious
techniques and would be unenforceable on the road ordinary within a traffic stream. In addition to
satellite monitoring techniques, image processing techniques offer the ability to follow vehicle
trajectories with great accuracy and continuity of detection. It is thus possible to combine different
monitoring techniques for the combined detection of the different variables.
Thanks to these new technologies it is possible to detect microscopic traffic variables, the location
of vehicles on the network, and the speed of vehicles, whose knowledge allows us to calibrate car-
following models already known in the literature as well as to develop new models. In recent years,
many scholars conducted experiments on car-following by using either onboard GPS devices or image
processing software to tracking vehicles during their ordinary trips within the traffic stream.
The necessary data, collected in the field, for this research are:
• Vehicle trajectory
• Spacing
• Speed
In this way, both the longitudinal and lateral positions are precisely checked to determine the exact
behavior for both the lane changing and the acceleration and deceleration process.
In a first experiment, Ossen and Hoongendoorn (2004) aimed at calibrating Gazis, Herman and
Rotary's (1961) model by using high-resolution images of a motorway taken from a helicopter with
intervals of 0.1 s for a duration of 300 s.
The object of the calibration was the parameter of the sensitivity function and the reaction time, the
calibration method applied to the sensitivity parameter for different reaction time values was that of the
least-squares.
A second experiment was carried out by Punzo and Simonelli (2005) to calibrate and compare
several car-following vehicle models, including Newell's simplified car-following model (2002) and
Gipps’s collision avoidance model (1981). The methodology for data collection involved experiments
on suburban and urban roads with a platoon of 4 vehicles equipped with differential GPS receivers, with
post-processing correction of errors with differential data. To reduce the effects of measurement errors,
the data were filtered using a Kalman filter, while the Least Generalized Squares method was used for
calibration.
A further experiment was conducted at Sapienza University of Rome by Colombaroni and Fusco
(2006) by using high-precision GPS onboard devices (Figure 2–XXII).

20
The accuracy of GPS devices was assessed preliminarily by comparing the measured positions of two devices, mounted on a steel
bar at 1 m distance (see the second picture from the left in the figure). Average errors resulted of 2 mm in stillness and ranged from
0.2 to 17 cm in motion.

61
 Car-Following Models

Figure 2–XXII. Equipment used in the experiments.

The average absolute error of the measurements in motion was 3 cm, therefore completely adequate
for the purpose of the experimentation. The diagrams of Figure 2–XXIII show the speed variation of
two vehicles (left figure) and the corresponding variation of the acceleration (right figure) during the
time interval considered. It is appreciable in these figures how the reaction of the follower is always
delayed compared to that of the leader of a time T, which is precisely the reaction time, a fundamental
variable for the models that are being studied. It is also interesting to note that, in reality, this is not
constant. It is possible to note also the high data noise present in the diagram representing the
accelerations of the two vehicles.
Variazione di velocità dei 2 veicoli Variazione di accelerazione dei 2 veicoli
60 3
]
Acceleration (m/s2)

50 2 2
] ^
Speed (m/s)

/s s
/
m 40
[ [m 1
à
ti e
30 n
c io 0
lo z
e
V 20 ra 0 10 20 30 40 50 60 70
e
l -1
e
10 c
c
A -2
0
0 10 20 30 40 50 60 70 -3
Te mpo [s]
Time (s)
Tempo[s] Time (s)
LeaderLeader
Velocità Follower
Velocità Follower Leader
Accelerazione Leader Follower Follower
Accelerazione

Figure 2–XXIII. Variations of speed and accelerations for a pair of vehicles in the experiment.

The following diagrams illustrate the results obtained after the calibration of Chandler's linear model
and the spacing inverse GHR model. Figure 2–XXIV shows the variation in the relative speed between
two successive vehicles within a platoon and the corresponding acceleration of the follower. In this case,
for a reaction time T=1s, by applying the least-squares method, a value R2=0.631 was obtained. The
value of the known term, close to zero, confirms the goodness of the calibration.
3
)
2
^
s/ 2
ay ==0.662 ∆v ++0.0019
0,6621x 0,0019
R²R=
2 =0,6312
Follower Acceleration (m/s2)

m 0.631
(
r
e 1
w
o
ll
o
f
l 0
e
d
e
n
o
iz -1
a
r
e
l
e
cc -2
A
-3
-3 -2 -1 0 1 2 3

Relative
Velocità Speed
relativa (m/s)
[m/s]

Figure 2–XXIV. Experimental data and corresponding Chandler’s model for a pair of consecutive vehicles with
a reaction time T = 1s.

62
 Traffic Engineering and Intelligent Transportation Systems

Chandler’s model was also applied to a leader-follower pair composed by two non-consecutive
vehicles (specifically, the 1st and the 3rd of the platoon) assuming a reaction time equal to 2s. The value
of the determination coefficient that was obtained in this case is 0.567 (Figure 2–XXV).
2

)
2 1,5 ay ==0.243 ∆v++0,0013
0.0013

Follower Acceleration (m/s2)

^
s/ 0,243x
R² R=2 0,5675
= 0.567
(m 1
r
e
w
lo
l
0,5
fo
l
e 0
d
e
n
io
z
-0,5
a
r
e
l -1
e
cc
A
-1,5

-2
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6
Velocità relativa
Relative Speed[m/s]
(m/s)

Figure 2–XXV. Experimental data and corresponding Chandler’s model for a pair of non-consecutive vehicles
with a reaction time T = 2s.

Figure 2–XXVI illustrates the results obtained after the calibration of the spacing-inverse GHR
model. In the diagram, the ratio between the relative speed and the spatial spacing between the two
vehicles is represented in the abscissa axis. In this case, again, a rather satisfactory R2 (0.655) is obtained.
The angular coefficient of this model corresponds to the critical speed, which is equal to 8.9 m/s (32
km/h). The known term is close to zero also in this case.
3

)
2
^
s/ 2
ya =
= 8.91 ∆v/s- –0,013
8,9124x 0.013
Follower Acceleration (m/s2)

m
( R²R2==0,655
0.655
r
e 1
w
o
ll
fo
l 0
e
d
e
n
io
z
-1
a
r
le
e
cc -2
A

-3
-0,3 -0,2 -0,1 0 0,1 0,2 0,3

Relative Speed/Spacing
Dv/s (1/s)

Figure 2–XXVI. Experimental data and corresponding GHR model with spacing inverse sensitivity.

Table 2–3 reassumes the results obtained in two different experiments, which involved 4 drivers
that assumed different positions in each experiment. For each couple of consecutive drivers (denoted
as GP, PA, AM in platoon 1 and PA, AG, GM in platoon 2) as well as non-consecutive drivers of
each platoon (GA, PM in platoon 1 and PG, AM in platoon 2), the table indicates the sensitivity
coefficient c of the linear model, the sensitivity coefficient c’ of the spacing inverse model, and the
determination index R2 of each model.

63
 Car-Following Models

Table 2–3. Summary of results for different pairs of drivers

LINEAR MODEL SPACING INVERSE MODEL

TR=1 TR=2 TR=3 TR=1 TR=2 TR=3

PLATOON 1 R² c R² c R² c R² c' R² c' R² c'

GP 0,45 0,43 0,45 0,44 0,19 0,28 0,43 8,19 0,38 7,63 0,14 4,63
PA 0,43 0,29 0,53 0,33 0,36 0,27 0,45 6,28 0,54 6,84 0,33 5,32
AM 0,36 0,28 0,31 0,28 0,12 0,18 0,49 5,08 0,35 4,42 0,12 2,69
GA 0,33 0,19 0,55 0,24 0,60 0,25 0,03 7,81 0,52 10,22 0,54 10,47
PM 0,39 0,15 0,54 0,18 0,47 0,17 0,52 6,32 0,63 7,07 0,47 6,20
PLATOON 2
PA 0,63 0,66 0,64 0,67 0,27 0,43 0,66 8,91 0,61 8,61 0,23 5,28
AG 0,35 0,12 0,43 0,13 0,36 0,12 0,29 3,37 0,36 3,76 0,21 2,76
GM 0,60 0,61 0,34 0,46 0,06 0,20 0,71 7,37 0,40 5,53 0,08 2,51
PG 0,44 0,21 0,57 0,24 0,54 0,24 0,45 10,19 0,58 11,51 0,53 10,93
AM 0,28 0,10 0,41 0,13 0,15 0,26 0,25 5,13 0,38 6,27 0,30 5,64
Drivers order in platoon 1: GPAM; Drivers order in platoon 2: PAGM

In order to capture the effect of the mutual interaction of the vehicles, the following considerations
can be done:
• as expected, the value of R2 usually increases for couples of non-consecutive vehicles when the
reaction time corresponds (values in a dotted circle) to 2s for the pairs composed by 1st and 3rd
vehicles and 2nd and 4th vehicles.
• Sensitivity usually decreases when pairs of non-consecutive vehicles are considered; that is, a
driver is usually more sensitive to actions of the vehicle immediately in front, rather than to
those in a further position in the platoon (values in the circle).
• Exceptions to the expectations provide interesting information on drivers' behavior: in order to
individuate the actual leader for each driver, it is searched for which pair of even non-
consecutive vehicles provide the highest sensitivity. By observing values highlighted by
rectangles, a higher determination coefficient of linear model for driver G in platoon 2 is
observed by assuming a reaction time of 2s (R2=0,57) instead of 1s (R2=0,43), when considering
the couple AG (2nd and 3rd vehicles); interestingly, a better model is obtained when considering
the couple PG (1st-3rd vehicles). So, the 3rd driver G follows the actions of the first vehicle in
the platoon rather than that immediately preceding.
• The sensitivity of driver G with respect to driver A (c=0,13) is much lower than that related to
driver P, which leads the platoon (c=0,24).
• Similar considerations can be done for the spacing inverse model (c’=3,37 and R2=0,29 for AG
with TR=1s; c’=11,51 and R2=0,58 for PG with TR=2s).
• A detailed analysis of vehicle trajectories confirms that driver G has a quite aggressive behavior
seeking to follow the vehicle leading the platoon, P. This explains also the observed reaction
time of 2s.
• Other considerations can be made by looking at the T reaction time values. To determine the
type of guide, you must also know the spacing. Having a long reaction time causes a delay in
starting the deceleration maneuver, if the distance is not high the follower gets too close to the
vehicle that is following performing a sharp braking.

64
 Traffic Engineering and Intelligent Transportation Systems

2.5 Bibliography
Aron, M., 1988. Car following in an urban network: simulation and experiments. In Proceedings of Seminar D,
16th PTRC Meeting, pp. 27-39.
Chandler R.E., Herman R., Montroll E.W. (1958). Traffic Dynamics: Studies in Car-Following. Operations
Research, Vol.6, pp.165-184.
Colombaroni, C. and Fusco, G. (2014). Artificial Neural Network Models for Car Following: Experimental
Analysis and Calibration Issues. Journal of Intelligent Transportation Systems, 18(1), 5-16.
Daganzo, C. F. (1994). The cell transmission model: A dynamic representation of highway traffic consistent
with the hydrodynamic theory. Transportation Research Part B: Methodological, 28(4), 269-287.
Edie, L. (1961). Car-Following and Steady-State Theory for Noncongested Traffic. Operations Research, 9(1),
pp.66-76.
Gazis D.C., Herman R., Rothery R.W. (1961). Non-Linear Follow the Leader Models of Traffic Flow.
Operations Research. Vol.9, No.4, pp.545-567.
Gazis, D. C., R. Herman, and R. W. Rothery (1963). Analytical Methods in Transportation: Mathematical Car-
Following Theory of Traffic Flow. Journal of the Engineering Mechanics Division, ASCE Proc. Paper 3724
89 (Paper 372), pp. 29-46.
Gipps, P.G. (1981). A Behavioral Car Following Model for Computer Simulation. Transpn. Res. B, 15, pp.105-
111.
Greenberg, H. (1959). An Analysis of Traffic Flow. Operations Research, Vol 7, pp. 78-85.
Gurusinghe G.S., Nakatsuji T., Azuta Y., Ranjitkar P., Tanaboriboon Y. (2002). Multiple Car-Following Data
with Real-Time Kinematic Global Positioning System. Transportation Research Record, No.1802, pp.166-
180.
Helly, W. 1959. Simulation of bottlenecks in single lane traffic flow. In Proceedings of the Symposium on
Theory of Traffic Flow, 207–238. Amsterdam, Netherlands: Elsevier. Research Laboratories, General
Motors.
Herman, R. and R. B. Potts (1959). Single Lane Traffic Theory and Experiment. Proceedings Symposium on
Theory of Traffic Flow. Ed. R. Herman, Elsevier Publications Co., pp. 120-146.
Kerner B. (2002). Empirical features of Congested Patterns at Highway Bottlenecks. Transportation Research
Record, No.1802, pp.145-154.
Kometani, E. and Sasaki, T. (1959). Dynamic Behaviour of Traffic with a Non- Linear Spacing–Speed
Relationship. in Herman R. (ed.) Proc. of Symp. Theory of Traffic Flow, New York, pp. 105–119.
Newell, G. F. (2002). A Simplified Car-Following Theory: a Lower Order Model. Transportation Research Part
B: Methodological, 36(3), 195-205.
Ozaki, H. Reaction and anticipation in the car following behaviour. Proc. of the 13th International Symposium
on Traffic and Transportation Theory, pp.349–366.
Ossen S., Hoogendoorn S.P. (2004). Car Following Behavior Analysis from Microscopic Trajectory Data,
Transport & Planning Department Faculty of Civil Engineering and Geosciences Delft University of
Technology. Journal of the Transportation Research Board, Vol. 1934, pp. 13-21.
Papola N. Theory of vehicular runoff and its use in design and regulation _ Part I. University of Rome La
Sapienza, Department of Hydraulics, Transport and Roads, Acts of Transportation. Hesagraphic, Rome.
Pipes L.A. (1953). An Operational Analysis of Traffic Dynamics. Journal of Applied Physics, Vol.24, No.3.
Punzo V., Simonelli F. (2007). Analysis and Comparison of Microscopic Traffic Flow Models with Real Traffic
Microscopic Data. Department of Transportation Engineering, University of Napoli Federico II.
Transportation Research Record: Journal of the Transportation Research Board, Vol. 1934, pp. 53-63.
Rothery R.W. (1998). Car Following Models, in Traffic Flow Theory, Nathan H. Gartner, Carroll Messer, Ajay
K. Rathi (Eds.). Transportation Research Board. http://www.tfhrc.gov/its/tft/tft.htm.
Rothery, R. W., R. Silver, R. Herman and C. Torner (1964). Analysis of Experiments on Single Lane Bus Flow.
Operations Research. Vol.12, pp. 913-933.

65
 Car-Following Models

Wiedemann, R. and Reiter U. (1992). Microscopic traffic simulation: the simulation system MISSION,
background and actual state, CEC Project ICARUS (V1052), Final Report, vol. 2, Appendix A. Brussels.
CEC.

66
Chapter 3 Traffic Models at Intersections
3.1 General
Drivers’ behavior at intersections differs significantly from the driving modes on road links that
were object of the previous chapters. At intersections, the safety conditions must be imposed against
vehicles that move in different directions than that of travel. In particular, when a vehicle coming from
a major road is approaching the intersection, every driver approaching the intersection from a minor
stream without right-of-way21 must slow down and eventually stop and wait for the intersection area to
be free for a sufficiently long time to cross the intersection safely. If a second vehicle arrives at the
minor approach while the first is waiting, it will stop in a queue and wait that the first will have started
crossing. Only when the first vehicle in the queue will have begun the maneuver to enter the intersection,
the following driver will be able to decide whether to enter or not. In any case, his decision will depend
on the comparison between the time required to carry out the crossing maneuver and the time intervals
between the successive vehicles on the major road22.
Thus, the representation of traffic flow at intersections requires structurally different models than
those conceived for the link flow. In particular, the intersection models have to capture these essential
aspects of the phenomenon: the decision-making mechanism of crossing or waiting; the duration of the
time gaps between the vehicles of the major traffic stream; the waiting time in a queue; the length of a
queue. The alternate use of the available road space in the intersection area makes the stationary flow
occasional: speed and density vary rapidly and considerably over time so that knowing the average value
of the density is useless because three significant values correspond to the respective traffic states:
arrival flow, queued flow, departure flow. The problem is instead to establish if, at a given instant, there
is a queue or not; or, during a time interval, what are the duration and the average length of the queue.
From what has been said, the traffic model at intersections consists of the following components,
which refer to the different elements that characterize the traffic at the intersection, as illustrated in
Figure 3–I:
• The gap acceptance model, which reproduces the choice behavior of drivers engaged in the
decision to enter the intersection; that is to say, the model that represents the behavior of the first
vehicle in the queue;
• The queue model, which represents the queuing process of the vehicles that follow the first and
wait to have the possibility to choose whether to enter the intersection, according to:
o the arrivals model, which provides the probability distribution of vehicle arrivals;
o the departures model, which provides the traffic flow deperting from the approach;
o the circulation rules, which defines the priority ranking and the operations at the
intersection.

21
The priority rule holds not only for unsignalized intersections but also permitted turning movements at signalized intersections that
conflict with higher rank movements.
22
The time gap is often defined as the time difference between the rear of a vehicle and the front of its follower while the time headway
is the time interval between the fronts of two successive vehicles. In the gap acceptance theory, however, the time gap is coincident
to the headway because the probabilty distribution functions of time intervals between vehicles are always measured thorugh
headways (Troutbeck and Brilon, 1996).

67
 Traffic Models at Intersections

Conflict
Area area:
di conflitto:
REGOLE DI CIRCOLAZIONE
Circulation rules

STOP
Main
Stradaroad:
principale: Firstveicolo
Primo vehicle:
:
LEGGE DEGLI
Probabilistic ARRIVI
arrival function MODELLO DI A CCETTAZIONE
Gap acceptance model
DEGLI INTERVA LLI

Vehicles
Veico in line:
li in coda:
MODELLO DI CODA
Queue model

Strada secondaria:
Secondary road:
LEGGE DEGLI ARRIVI
Probabilistic arrival function

Figure 3–I. Schematic representation of the traffic model at road intersections.

A fundamental aspect of the problem is the relevance of the random component of the phenomenon.
The possibility of a vehicle to enter the intersection depends on the duration of the time gaps between
vehicles of the main traffic stream, not only on their average value23.
Another important difference between link and node flows concerns the capacity. While the capacity
of road links depends exclusively on the geometric and functional characteristics of the road, in the
intersections the capacity of a minor approach depends on the flows on the main approaches. Capacity
is, therefore, one of the characteristic quantities to be determined in the analysis of an intersection. The
maximum possible value of the flow at an intersection approach, achievable only if there was no flow
on the other approaches, is instead called saturation flow and coincides conceptually and numerically
with the capacity of the approach arc at the intersection.
As for the flow on links, also for that at the intersections, microscopic models were formulated that
aim to reproduce the behavior of the individual drivers, and macroscopic models that assimilate vehicle
traffic to a stream of fluid and reproduce the states of motion and queue in aggregate form, without
modeling drivers’ decision logic explicitly. Unlike the flow on the unidirectional road links without the
possibility of overtaking, which describe a single action of the driver, the longitudinal motion, in the
case of the intersections, the actions in question are substantially different for the first driver in the
queue, who must decide whether to accept or no the available gaps in the main traffic stream, and for
the following drivers, who can follow up the first one or remain in the queue. In this case, therefore, the
different behavioral and macroscopic models are not alternative, but complementary. In fact,
macroscopic queue models apply at an aggregate way the results of probability models that apply such
behavioral drivers’ logic.
The present chapter deals with the models that describe the traffic operations and estimate the
performances of a road intersection: the behavioral model of gap acceptance, which represents the
behavior of the first driver in the queue; the probability model of vehicular arrivals, which determines

23
To get an idea of the importance of the probabilistic aspect, just think of what paradoxical results a deterministic
model would produce if we assumed a critical gap of 7s and applied the gap acceptance theory to two slightly
different cases: a major stream of 500 vehicles/hour, spaced in time by 7.2s, would allow the passage of as many
vehicles from the secondary approach (1 for each interval because all interval would usable); a major stream of
550 vehicles/h, spaced in time by 6.5s, would instead saturate the intersection because it would not allow the
passage of even one vehicle!

68
 Traffic Engineering and Intelligent Transportation Systems

how traffic feeds the different approaches at the intersection; the queue model, which determines the
queue length and the waiting time of the vehicles in the queue, also known as delay models.
This chapter aims to describe the theoretical foundations necessary for modeling traffic at road
intersections and to provide the specification of these models to the different types of road intersections,
signalized or unsignalized. The description of the methods for the operational analysis of traffic
performances at the different types of intersections and the criteria for signal settings are instead the
subject of the subsequent chapters.

3.2 Gap Acceptance Model

The gap acceptance model24 is the theoretical foundation that represents the traffic flow of
conflicting vehicular movements. This theory assumes that the decision logic of the driver at the
intersection consists of comparing the time interval required to safely complete a crossing or turning
maneuver (this time gap is called the critical gap) with the duration of the vehicle gaps in the main
traffic stream. It is based on three basic elements: the random vehicle arrival pattern (i.e., the probability
distribution function of time gaps between successive vehicles) in the priority road; the usability of the
gaps in the major stream for the drivers of the minor stream (an interval is usable if it is greater than the
critical gap) and the determination of the levels of priority of the streams (priority raking).
The behavioral logic of drivers is defined by the following hypotheses:
• the behavior of the drivers is rational (they follow a logical rule), consistent (with the same
external conditions they do not change their behavior) and homogeneous (all drivers have the
same behavior when performing the same maneuver);
• no first driver in the queue will enter the intersection area if the interval between vehicles of the
hierarchically higher streams is not greater than or equal to the critical gap tc;
• the drivers in a next position in the queue will be able to exploit also shorter gaps than the critical
one if they can enter the intersection area with a speed v > 0 (following the first vehicle just
behind it); the minimum interval accepted by each subsequent driver is usually referred as the
follow-up time tf;
• the drivers of the hierarchically higher streams are not influenced by approaching vehicles that
have to yield the priority to them; thus, the rank 1 stream is not delayed.
The logic underlying the gap acceptance model is sketched in Figure 3–II. The upper part of the
figure shows the trajectories of the vehicles arriving on the main road, assuming for simplicity that all
they travel at a uniform speed v. The driver of the first vehicle in the queue (in red in the figure) estimates
whether the time gap h between the vehicles in the major stream is greater than the critical gap tc. If so,
the waiting vehicle can enter the intersection. After the interval tc, the first vehicle safely has passed the
intersection area. The subsequent vehicles in the queue start when their position is reached by the
kinematic wave w0: if a vehicular gap of the priority stream is greater than tc + tf the second vehicle can
enter the intersection in its turn and cross it safely. On the right side of the figure, the possible trajectories
of the queued vehicles if no conditioning were are shown.

24
The gap acceptance model has a wider field of application other than that, even very important, concerning the road intersections.
In fact, it reproduces the behavior of drivers who wish to exploit a gap in a different traffic stream for any lane changing maneuver
as overtaking, direction changing, merging or exiting from a freeway.

69
 Traffic Models at Intersections

v h h>t c?
s=vh

tc tf tf

t
STOP w0

Figure 3–II. Representative scheme of the gap acceptance model.

The gap acceptance model allows to represent the traffic flow at an usignalized intersection (yield-
controlled or a stop-controlled) and to determine the main operating characteristics, such as capacity,
delay, queue length, number of vehicles stopped, depending on the geometric characteristics of the
intersection and the distribution function of the vehicle gaps of the major traffic streams.
Without new arrivals in the major stream, i.e. assuming that all intervals are usable, the gap
acceptance model provides the possible departures in compliance with the safety rules at the intersection,
shown in Figure 3–III. The presence of vehicles arriving on the major road limits the number of vehicles
leaving the secondary road, as it makes unusable, for each arriving vehicle, a time interval of duration
equal or less than tc.
The capacity of the minor approach, therefore, must be assessed in probabilistic terms as the
expected number of usable intervals, according to the law of arrivals on the major road. This topic will
be discussed in Paragraph 3.5.2.

70
 Traffic Engineering and Intelligent Transportation Systems

N
2

0
0 5 10 15 20 25 30
t (s )

Figure 3–III. Maximum number N of vehicles departing at an approach of an unsignalized intersection as a
function of the duration of the vehicular interval on the main road.

Observation. The gap acceptance model, in its classical definition, assumes that the critical interval
for a given maneuver in a given intersection is constant and equal for all drivers. In fact, experimental
observations have shown that the process of accepting the intervals varies among drivers even
considerably and is significantly influenced by numerous variables, such as the gender of the driver, the
type of vehicle, the speed of the major stream traffic, the time spent in the queue, the presence of other
vehicles waiting in the queue. These aspects can be adequately considered by reformulating the model
in probabilistic terms.

3.3 Probabilistic Arrival Models

Vehicle headways of a traffic stream are characterized by a random nature, which is due both to the
different starting instants of the users and the resulting “history” of their movements up to that point
and, at a local scale, to the differences in driving behavior from one driver to another, which leads to
keep different distances even when traveling at the same speed, as it occurs at a stationary traffic state.
Figure 3–IV illustrates the distribution of the relative frequencies of the vehicle headways measured on
the different lanes of an American urban road. Various hypotheses have been formulated in the literature
on the probability of arrivals on major streams, giving rise to different flow patterns at intersections.

71
 Traffic Models at Intersections

Figure 3–IV. Distribution of the relative frequencies of the vehicle intervals on the 3 lanes of the Long Island
Expressway (source: HCM, 2000).

3.3.1 Negative Exponential Distribution

The most common hypothesis on the probability distribution of vehicle arrivals is that the time
headways between successive vehicles follow a negative exponential function. This distribution is based
on the assumption that the vehicles arrive randomly without any dependence on the instants of the arrival
of the previous vehicles and that in a small interval of time (at the order of 1s) the probability of arrival
of 1 vehicle is a constant and is impossible than more than 1 vehicle arrives. This assumption
corresponds to assume that the number of arrivals in a give time interval is distributed according to the
Poisson probability function.
As the Poisson distribution, also the negative exponential function has only one parameter,
corresponding to the traffic intensity q. The expected value and the standard deviation are both equal to
1/q.
The cumulative probability function FT , which indicates the probability of observing a vehicular
interval less than a given value t, can be determined by imposing that at least one vehicle arrives in the
time interval t. This is the complementary of the probability, according to a Poisson distribution25, that
no vehicle arrives during the time interval t, that is:
FT ( t ) = 1− e –qt (3.1)
The probability density function can be obtained immediately by derivation of (3.1):
FT ( t ) = qe –qt (3.2)
The negative exponential distribution has a very simple mathematical expression, which allows
deriving in a closed form the main variables that descrive the performances of the intersection. However,
this distribution assigns a non-zero probability even at unrealistic intervals (Fig. 3-V), shorter than the
minimum possible temporal spacing, tm , corresponding to the inverse of the saturation flow26.

25
The Poisson probability distribution is PX(x=0)=(qt)xe-qt/x! The probability that no vehicle arrives (X=0) in the time interval t, is:
PX(x=0)=(qt)0e–qt/0!=e–qt.
26
This unrealistic property of the negative exponential distribution is, in effect, of little relevance for the theory of acceptance of
intervals, since very short intervals are however discarded. On the other hand, it affects the form of distribution in the field of
acceptable ranges, which is defined by the only parameter constituted by the flow of traffic.

72
 Traffic Engineering and Intelligent Transportation Systems

0.35

0.3

0.25
q=0,3
0.2

f(h)
0.15

0.1
q=0,2
0.05

0
0 3 6 9 12 15
h
Figure 3–V. Density function of two negative exponential distributions of parameters q=0.2 and q=0.3.

3.3.2 Translated Negative Exponential Distribution

A better representation of the probabilistic distribution of the vehicular intervals of the traffic stream
can be obtained by translating the negative exponential distribution by an amount equal to the minimum
vehicular interval tm, to explicitly exclude the possibility of intervals shorter than tm. The hypothesis
behind this probability distribution is that, if a road is crossed by a vehicular flow q, in one hour, while
q vehicles are passing, there are also qtm seconds lost. The remaining time, which is actually available
for vehicle arrivals, is distributed randomly after every vehicle so that the random component of the
headway is on average:
1 - qt m
h= (3.3)
q
The cumulative distribution function then holds:

FT (t ) = 1 - e - l (t -tm ) ; t > t m (3.4)

where the parameter λ is the inverse of the average rate of the random component of the intervals h :
1 q
l= = (3.5)
h 1 - tm q
The probability density function, shown in Fig.3-VI, is easily obtained from (3.5) by derivation:

f T (t ) = le - l (t -tm ) ; t > t m (3.6)

73
 Traffic Models at Intersections

0.35

0.3

0.25

0.2

f(h) 0.15

0.1

0.05

0
0 3 6 9 12 15
h
Figure 3–VI. Translated negative exponential density function (continuous curve) of parameters λ = 0.3 and t =
m

2 and negative exponential density function of parameter λ = 0.3 (dashed curve).

The average value of the vehicle intervals distributed according to the translated exponential
distribution is:
1 1
E [H ] = = tm + (3.7)
q l
while the variance, being a second-order moment, is independent of the translation and is therefore equal
to that of the original negative exponential distribution:
1
Var [H ] = (3.8)
l2
The translated negative exponential function provides a good approximation of the distribution of
the vehicle headways in case of low traffic, while it becomes unrealistic for dense traffic streams, for
which the conditioning between vehicles becomes relevant and the hypothesis that the available time
(1−tm)/q can be distributed randomly and independently for each vehicle is no longer acceptable.

3.3.3 Dichotomized Distributions

More realistic probability distributions assume that in traffic flow at a given moment there are two
types of vehicles:
• free vehicles, sufficiently far from previous vehicles that they do not interact with the vehicles
ahead;
• grouped vehicles, which form a platoon following a previous vehicle.
Different formulations of dichotomized distributions have been proposed, for example by Schuhl
(1955) and Cowan (1975).
Cowan’s model, in particular, does not care to set a distribution function for bunched vehicles, since
short intervals are always rejected, and applies a translated negative exponential distribution only to the
fraction a of free vehicles.

FT (t ) = 1 - ae - l (t -tm ) ; t > t m (3.9)

fT (t ) = αλe ( m ) ; t > t m
− λ t−t
(3.10)

€
74
 Traffic Engineering and Intelligent Transportation Systems

To apply the model it is necessary to estimate the fraction of free vehicles depending on the average
flow. Brilon (1988) proposed the following equation:
a=e−Aq (3.11)
where A is a parameter whose values range between 6 and 9. This formulation can be explained
considering that the fraction of free vehicles is given by the probability that vehicle headway, distributed
according to a negative exponential, are greater than a given value, A, which represents the beginning
of the interaction between vehicles. Brilon’s expression is illustrated graphically on the left side of
Figure 3–VII. It had an experimental confirmation in the researches of Sullivan and Troutbeck (1993),
who discovered that the value A depends on the width and type of lane, with values ranging from 3.7s
to 7.5s.
The corresponding diagram of the dichotomized probability density function, for one lane of
standard size, is shown in the right part of Figure 3–VII. Note that, unlike the translated exponential
function, which expresses the probability density of the intervals of all vehicles, the dichotomized
function defines the probability density of the intervals only of the fraction α of free vehicles. In the
example in the figure (having assumed q = 0.3veh/s and A = 5.25s), it is approximately 0.20.
1 0.35

0.3
0.8 Wide kerb
kerb lanes
Wide
Corsia lanes
esterna larga
(A=3.7)
(A=3,7) 0.25
Esponenziale
Translated negativa
Negative traslata (q=0.3)
Exponential (q=0,3)
0.6 0.2
Average
Corsia kerbnormale
esterna lanes
f(h)
a

(A=5.25)
(A=5,25) 0.15
0.4 Dicotomizzata (q=0,3; A=5,25)
Dichotomized (q=0.3; A=5.25)
0.1
0.2
Corsia centrale 0.05
Median lanes
(A=7,5)
(A=7.5)
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 4 8 12 16 20
q (veic/s) h
Figure 3–VII. Dichotomized probability density function of free vehicle headways. Left: α fraction of free
vehicles not depending on the flow q for different types of lanes. Right: comparison between the negative
exponential function and the dichotomized function that models the headways of free vehicles for a flow q = 0.3.

A function that provides the probability distribution of the intervals of all vehicles, taking into
account both the fraction of free vehicles and that of conditioned vehicles, is the generalized Erlang
function, which is therefore appropriate to describe dense traffic streams (Dawson, 1969).

+ t − t .x
% t ( mp
% t (
− λ' t−
mp
−t mp * k −1
k- 0
− λ'' t− ml −t ml ** ' h * -, h p − t mp 0/ (3.12)
FT (t ) = 1 − αe & hl )
+ (1 − α )e & p )
∑ x!
x=0

The parameters of Erlang distribution tml, hl and tmp, h p represent the minimum interval and the
average value of the headways for the two components, respectively.
Due€to its complexity, the generalized Erlang function is not used in intersection queue models, for
which it is not necessary to know the distribution of all the intervals, but only those greater than the
critical value. Instead, it is used in simulation models.

75
 Traffic Models at Intersections

3.4 Queueing Model

The evolution of a queue of vehicles waiting at an intersection can be described in the more general
context of the queueing theory, the branch of operational research born in the early twentieth century
and developed especially in the 1960s that describes the relationships between the demand for a service
and the delays suffered by system users.
In the more general case, the provision of the service can be variously located in the space; in the
simplest case, it is located in correspondence with a given system, consisting of one or more service
units (called “server”) and one or more queues of waiting users. Users arrive in the system, wait for their
turn in the queue, then access the system where they are served and finally leave the system (Figure 3–
VIII).

•Circulation rules
Regole d i circo lazione
•Departures model
Modello di partenza

Server 1
2
3
Arrivo utenti
Userin Arrivals
coda
Queue
Coda

m-1

Server m

Impianto Punto di uscita

Punto di arrivo
Entrance in the Exit from the
nel sistema Queueing system
di servizio dal sistema
Queue system Queue system

Figure 3–VIII. Scheme of a generic queue system.

Kendall notation. The queueing system is characterized by a service time ts necessary to serve a
generic user, by the number m of service units and by the time interval between two consecutive user
arrivals, that is the headway h. These three pieces of information are synthetically expressed according
to the Kendall notation in the form A|B|m, which denotes the probability distribution of arrivals A, the
probability distribution of service time in place of letter B and with m, as mentioned, the number of
service units is indicated. The codes used to distinguish the various probability distributions are:
M: negative exponential distribution27,
D: deterministic function;
Ek: Erlang distribution of order k ;
G: “general” distribution, i.e., any unspecified distribution.

Indendence Hypothesis. Independence is assumed for two successive intervals as well as two
successive service times.
Service Rules. There are also several service rules: according to the most frequent FIFO rule (first-
in, first-out), users are served in the order in which they present themselves to the system; according to
the LIFO rule (last in, first out) they are served in the reverse order of arrival; the abbreviation SIRO
(service in random order) indicates that the service is carried out randomly.

27
The letter M indicates the so-called memoryless property of a Poisson process, whose probability is independent of the previous
events. The same property holds for the exponential distribution function as it provides the probability distribution of two successive
arrivals of a Poisson.

76
 Traffic Engineering and Intelligent Transportation Systems

3.4.1 Fundamental Relations and Variables

Service times and the intervals between consecutive arrivals are linked respectively to the quantity
of supply and demand by two simple but fundamental relationships. The inverse of the service time
(time interval between two successive users served) is the capacity of the single server element (number
of users served):
1
µ= (3.13)
ts
just as the inverse of the interval between two arrivals is, as is known, the average arrival rate:
1
l= (3.14)
h
The relationship between the rate of arrivals and the overall capacity of the m servers of the system
is said degree of saturation :
l
r= (3.15)
mµ
The comparison between the functions of arrivals and departures completely describes the
performances of the system. If the system is initially empty, the number of users28 Q(t) present inside
it (queued or in the service unit) at time t, is given by the difference between the arrivals and the parties,
that is to say by the vertical segment that joins the two curves A and P in Figure 3–IX:
t

Q(t ) = A(t ) - P(t ) =

ò [l(t ) - µ (t )]dt
0
(3.16)

N N

A(t)
A(t)

Q(t)
d(t) d(t)
Arrivi
Arrivals
Departures
Partenze Arrivi
Arrivals
P(t) Q(t)
Departures
Partenze
P(t)

t t
Figure 3–IX. Curves of arrivals and departures in a generic queue system in the cases of a discrete process (on
the left) or in the case of uniform arrivals and departures (on the right).

The waiting time or delay of a user who enters the system at instant t is given by the difference
between the time instants of exit and arrival. Since these are random variables, so is the delay, D(t).

28
Traditionally, uppercase letters are used in traffic engineering to indicate absolute quantities, like the total number of vehicles, while
lowercase letters are used to denote unitary variables like flows and density. On the other hand, in probability theory capital letters
are used to indicate random variables while lowercase letters are used to indicate the generic value assumed by a random variable
in a single observation. In this chapter, which deals with the probabilistic models of vehicular traffic, we reserve capital letters to
indicate the random nature of a variable (like Q for the queue length) and we adopt different symbols for the unitary variables, even
if they describe the same size (for example, l for average arrival flow, which is often denoted as arrival rate in the queueing theory).

77
 Traffic Models at Intersections

The total time spent in the system (queued or in the service unit) by all users (called total delay and
indicated herewith the letter W) is represented by the area between the two curves of arrivals and
departures in Figure 3–IX:
t t t

W (t ) = ∫ Q(t )dτ = ∫ [ A(t ) − P(t )]dτ = ∫ [λ(τ)τ − µ(τ)τ]dτ (3.17)

0 0 0

that is, it is given to the area of the triangle between the two slope lines equal to λ and μ respectively.
The average unit delay D(t) of the users arriving in the time interval [0, t] is given by the ratio:
€ t

W (t )
∫ [λ(τ ) − µ(τ )]τdτ
0
D (t ) = = t
(3.18)
A(t )
∫ λ(τ )dτ
0

Little’s law. The average number of users waiting in the system in the interval [0, t] is similarly
obtained by dividing the total time of all users in the system by the duration of the interval. This
relationship can be €rewritten in a different form by dividing and multiplying by the number of arrivals
A(t). The result is an important relationship, known as Little's law 29:

W (t ) W (t ) A(t )
Q (t ) = = = D(t ) λ (t ) (3.19)
t A(t ) t
which links together the average number of users in the system, the average delay and the average arrival
rate. Little’s law can also be written in the form:
€ Q (t )
D (t ) = (3.20)
λ (t )
which indicates that the average delay of users queued in a range [0, t] is equal to the average number
of users queued in the same interval divided by the average rate of arrivals in that interval.

€
Special case: uniform arrivals. In the particularly simple case of a deterministic system D|D|1 with
uniform arrivals, illustrated in the graph on the right of Fig.3-IX, the previous relations are considerably
simplified. In fact, we have the following expressions, respectively for the number of users in the queue
Q(t), the total delay W(t) and the average unit delay d(t):
t

Q (t ) = ∫ (λ − µ)dτ = (λ − µ)t = (ρ −1)µt (3.21)

0

t
t2 t2
€
W (t ) = ∫ (λ − µ)τdτ = (λ − µ) 2 = (ρ −1)µ 2 (3.22)
0

W (t ) æ l - µ ö t æ 1öt
d (t ) = =ç ÷ = çç1 - ÷÷ (3.23)
€ lt è l ø2 è rø2

29
Little's law is valid regardless of the assumptions about both the probabilistic distributions of arrivals and the service time as well
as the service rule, FIFO or LIFO that is.

78
 Traffic Engineering and Intelligent Transportation Systems

3.4.2 Hypotheses on the Queue Model at an Intersection

The queue model described in the previous paragraph is absolutely general and is therefore
applicable to describe waiting phenomena that are also very different from each other, such as,
remaining in the field of transport, a queue at a road toll, at a bus or subway station, approaching an
airport runway or, finally, as is in our case, at a road intersection.
A queue at an intersection approach can be described by different probability distributions of service
time and vehicle arrivals. Generally, arrivals on a secondary road can be modeled according to a negative
exponential probability distribution function or by an Erlang distribution in the case of more congested
traffic conditions. At unsignalized intersection, the service time is given by the distribution of the
vehicular intervals of the major traffic stream. In a signalized intersection, the service time depends on
the duration of the green and red times and is a deterministic quantity (apart from the traffic actuated
signals).
The queue function differs significantly in the two possible cases of under-saturation or over-
saturation. Indeed:
• in conditions of equilibrium between demand and supply (that is, the arrival rates are less than
the capacity of the minor traffic stream), the queue is stationary (i.e., it has a possibly variable
but in any case limited length);
• if the demand exceeds the supply the queue is time-dependent and has an indefinitely increasing
length, until the demand becomes lower than the capacity.
The following paragraphs introduce the general expressions of the average delay and the average
queue length at a road intersection. Two main hypotheses are introduced, corresponding to two different
applications: the hypothesis of Poisson arrivals on both the minor and the major traffic streams (queue
M|M|1), which describes traffic operations at an unsignalized intersection; the hypothesis of Poisson
arrivals on only the minor traffic streams and no specific hypothesis on the probability distribution of
the service time (queue M|G|1), corresponding to a general expression, which is also adopted by the US
Highway Capacity Manual and is valid both for unsignalized and signalized intersections.

3.4.3 Stationary Queue Model M|M|1

The analysis of a stationary queueing system requires the assumption that the demand does not
exceed the capacity (λ<μ), a necessary condition for the queue to reach an equilibrium state. To derive
the stationary condition we should write the equations that describe the transition from one state to
another and then impose the equilibrium condition.
A stochastic process that has a transition probability from one state to another depends only on the
state of the system and not on its history, that is, on the order in which the states have followed one
another, is called a Markov process. It is usually represented as a chain whose states are represented as
ovals and the transitions from one state to another by arrows with associated the respective constant
probability. In the case in question, the state of the system is represented by the number of users within
the system, while the transitions are determined by the arrival or a departure of a user.
A queue whose arrivals and departures are both Poisson distributed is a particular example of a
Markov process, in which the transition probabilities between successive states assume only two
possible values, equal to the probability of an arrival in the unit of time d (assumed to be small enough
that it cannot occur more than one arrival) and the probability of a departure in the same unit of time d.

79
 Traffic Models at Intersections

1 λδ λδ λδ λδ λδ

0 1 2 … k … n-1 n

µδ µδ µδ µδ
Figure 3–X. Graphical representation of the transition between states in an M|M|1 queue.

The probabilities of transition between states are in fact :

P( j, j +1) = λδ P( j +1, j) = µδ P( j, j) = 1− λδ − µδ P(i,i + j) = 0, | j |≠ 1 (3.24)
At the equilibrium, the probability of passing from state n to state n +1 is equal to the probability of
the inverse passage from n +1 to n:
€ € € €
P(n,n + 1) = P(n + 1,n) (3.25)
The probability of transition from n to (n +1) is given by the probability that one arrival occurs,
conditioned to the probability of the system being in state n; analogously, the inverse probability is given
by the probability that a departure occurs conditioned to the probability of the system to be in the state
€
n +1. In formulas:
P(n,n + 1) = λP(n); P(n + 1,n) = µP(n + 1) (3.26)
Substituting the two expressions (3.26) in (3.25) we have, for r<1, the probability of state transition
in a system in equilibrium:
€ λP(n) = µP(n + 1) ⇒ P(n + 1) = ρP(n) (3.27)
Since the transition between states of a Poisson process is constant, we want to get an expression of
the probability of the system to be in a state n that does not depend on n. By iteratively applying the
state transition probabilities starting from the initial condition of empty queue, we obtain the probability
€
of the state n as a function of the probability of the initial state:
P(1) = ρP(0)
P(2) = ρP(1) = ρ 2 P(0)
P(3) = ρP(2) = ρ 3 P(0) (3.28)
...
P(n) = ρ n P(0)
By imposing the condition that the sum of the probabilities of all the states of the system is equal to
one, we obtain an expression of the state P(0) as a function of the degree of saturation of the system:
€ ∞ ∞ ∞

∑ P(n) = ∑ ρ P(0) = P(0)∑ ρ

n=0 n=0
n

n=0
n
=1 (3.29)

The geometric series in (3.29) has argument |r| <1 ; therefore its sum is:
∞
1
€
∑ρ n
=
1− ρ
(3.30)
n=0

It follows that the initial state of the empty queue has probability:

€
80
 Traffic Engineering and Intelligent Transportation Systems

P(0) = 1 − ρ (3.31)
Substituting the expression of P(0) in the last equation of (3.28), we finally obtain the desired
probability of the state P(n) as a function of the only degree of saturation of the system:
€
P(n) = ρ n (1 − ρ ) (3.32)
Once the system transition function is specified, we can easily derive the performance indicators of
the system, namely the average number of queued users and the average delay.
For this, we replace the expression
€ (3.32), which provides the probability of a state n of the M/M/1
process, in the definition of the average number of users in the system:
∞ ∞ ∞

Q = ∑ nP(n) = ∑ nρ (1 − ρ ) = (1 − ρ )∑ nρ
n n
(3.33)
n=0 n=0 n=0

Thus, we obtained a power series, which can be solved by exploiting the derivation properties of
the power functions, and we can derive the expression of the average queue:
€ ∞ ∞ ∞

(1 − ρ ) ∑ nρ n
∑ nρ
= (1 − ρ ) ρ n−1
∑ D( ρ ) =
= (1 − ρ ) ρ n

n=0 n=1 n=0

& ∞ ) (3.34)
& 1 ) ρ(1 − ρ ) ρ
= (1 − ρ ) ρD(
( ∑n+
ρ = (1 − ρ ) ρD(
+
+ = =
'1 − ρ * (1 − ρ )2 (1 − ρ )
' n=0 *
Where the symbol D(.) indicates the derivative of a generic function.
The (3.34) gives an expression of the average number of users in the system as a function of the
degree
€ of saturation. Multiplying the (3.34) for the rate of arrivals we obtain the following formulation,
which expresses that the average number of users in the system in an M/M/1 process as the ratio between
the arrival rate and the residual capacity:

λ
Q = (3.35)
µ−λ
The average delay of users before exiting the system is obtained by applying Little's law:
Q 1 1
D€= = ⇒ D = (3.36)
λ µ−λ µ(1 − ρ )
Note how (3.36) expresses that the average delay is the inverse of the residual capacity.
The expressions (3.37) and (3.38) provide the average number and the average delay of users in the
system, respectively;
€ therefore they include the time needed to serve the first user in the queue and the
waiting time of the other users. To obtain the time average of waiting in the queue, before being
served, it must subtract the average delay in the time system the mean service:

1 1 1 λ
DQ = D − = − = (3.37)
µ µ−λ µ µ( µ − λ )
By applying the Little’s Law, the average number of users in the queue NQ is obtained

λ2 ρ2
€ N Q = DQ λ = = (3.38)
µ( µ − λ) (1 − ρ)

€
81
 Traffic Models at Intersections

3.4.4 Stationary Queue Model M|G|1

To study the case of a stationary queueing system of type M|G|1 we follow an approach similar to
that of the case M|M|1: we analyze the transition from one state to another and we impose the
equilibrium condition, that is, that the average value is stable. Unlike the queue system M|M|1 previously
discussed, in this case we do not need to specify the probability function for the service time but only
the moments of the function, that is: expected value and variance.
The study of the queue dynamics is based on the specification of the states of the queue in two
successive time instants just after a vehicle leaves the queue30. Consider one of these instants. A vehicle
moves away from the queue and enters the intersection. At that moment the queue includes Q vehicles.
Let be H the service time of the next vehicle and let be R the number of vehicles arrived at the queue
during the time interval H. Now, consider the time instant when the next vehicle enters the intersection;
that is, after the service time H. At this time instant, as illustrated in Figure 3–XI, the queue will consist
of a number Q' of vehicles, given by the previous queue length plus the vehicles arrived in the meantime
minus the vehicle departed; that is, if Q ≠ 0:
Q" = Q + R −1 (3.39)
or, if it is Q = 0:
Q" = R (3.40)
€

Arrivals
€
Departures

R
R+Q

Figure 3–XI. Trend of a stationary queue in two successive departures.

The two previous expressions can be summarized in a single form by introducing a function δ such
that:

ì1 se Q = 0
d =í (3.41)
î0 se Q > 0
so that (3.39) and (3.40) can be synthetically written:
Q" = Q + R −1+ δ (3.42)
To describe this stochastic process mathematically we have to define the moments of the queue,
which depend on the probability of arrivals, which are assumed to be Poisson distributed, and on the
€

30
The study of the dynamics of the system, because of its time-dependent nature, focuses on an elementary change of state of the
system, analgously to the derivative of a deterministic function, which focuses on the variation of the function depending on an
elementary variation of the independent variable.

82
 Traffic Engineering and Intelligent Transportation Systems

moments of the service time, whose density function fH(h) is not specified but whose moments are
known.
The average number of vehicles E[R] arrived in the time interval H are determined by integrating
the probability distribution of the arrivals pR(r) with respect to h and summing up for the possible values
of the number of arrivals. Since the arrivals are distributed according to Poisson, the following
expression is obtained:
∞ ∞ t r
(λh) e−λh
E [ R] = ∑rp(r ) = ∑ ∫ r
r!
f H ( h ) dh (3.43)
r=1 r=1 0

Since it is:
∞ r−1

∑ ((r −1) )! = eλh

€ λh
(3.44)
r=1

by applying the definition of the average of the random variable H, we obtain that the average
number of arrivals E[R] is equal to the product between the average arrival rate l and the average
service time E[H]; since the€latter is the inverse of the capacity μ, the average number of arrivals is
finally equal to the degree of saturation of the queue ρ :
t
λ
E [ R] = λ ∫ hf H (h)dh = λE [ H ] = µ = ρ (3.45)
0

To derive a general expression of the average number of vehicles in the queue E[Q] and the average
delay E[D], you must describe the probability distribution of the time of service H by means of its mean
and its variance.€To exploit the expression that relates the variance to the square of the average, we raise
to the square the equation (3.42):
2
Q"2 = Q2 + ( R −1) + δ 2 + 2Q( R −1) + 2( R −1)δ + 2Qδ (3.46)

For the definition of δ it results:

d 2 =d
€ Qd = 0
so that:
2
Q"2 = Q2 + ( R −1) + δ + 2Q( R −1) + 2( R −1)δ =
2
(3.47)
2
= Q + ( R −1) + 2Q( R −1) + (2R −1)δ
Let's now pass to the average values of the quantities in question. From (3.42) we get:

E [Q"] = E [Q] + E [ R] −1+ E [δ ] (3.48)

€
We now impose the equilibrium conditions; that is, we impose that that the average queue is
independent from the particular state from which it is computed:
€ E [Q"] = E [Q] (3.49)
From the equation (3.42) we have:

E [δ ] = 1 − E [ R] (3.50)
€
The application of the average to (3.42) provides the following expression:

€
83
 Traffic Models at Intersections

[ ] [ ] [ ]
E Q"2 = E Q2 + E R2 − 2E [ R] +1+ 2E [Q] E [ R] −1 + 2E [ R] −1 E [δ ] ( ) ( ) (3.51)

and, applying the (3.49) and (3.50), we write:

€ [ ] (
E R2 − 2E [ R] +1 − 2E [Q] 1 − E [ R] + 2E [ R] −1 1 − E [ R] ) ( )( ) (3.52)

that is :

( ) [ ]
2
{ }
€ [Q] 1 − E [ R] = E R − E [ R] + 1 − E [ R] + 2E [ R] 1 − E [ R] − 1 − E [ R] = 0
2E ( ) { } (3.53)

After some algebraic passages, we obtain the expression of the average value of the queue:

€ E [Q ] = E [ R ] +
[ ]
E R2 − E [ R]
(3.54)
(
2 1 − E [ R] )
as a function of the average number E[R] of vehicles arrived during the service time, whose
expression is given by (3.45) and by the expected value of the square of R, whose expression is instead
to be determined. €
The latter quantity, E[R2], can be calculated by applying the definition of average value, similarly
to what was done in equation (3.43):
∞ ∞ t r
(λh) e−λh
[ ] = ∑r p(r ) = ∑ ∫ r
E R 2 2 2
r!
f H ( h ) dh (3.55)
r=1 r=1 0

dividing the series that appears in (3.42) into the following two series:
∞ r−1 ∞ ∞ r−1 ∞ ∞ r−1 r−2 r−1
(λh) =
∑(r −1) (r −1)! + ∑ (r −1)! = λh∑ (r − 2)! + ∑ ((r −1) )!
( λh ) ( λh ) ( λh ) λh
∑ €
r
(r −1)!
(3.56)
r=1 r=1 r=1 r=1 r=1

and still applying the expression (3.44), in this case, up to the term r –2:
∞ r−1 r− 2
(λh) = ∞ (λh) = e λh
€
∑ (r −1)! ∑ (r − 2)! (3.57)
r=1 r=1

the series (3.56) can be written:

∞ r−1
∞ ∞ r−2 r−1

∑€r (r −1)! = λh∑ (r − 2)! + ∑ ((r −1) )! = (λh +1)eλh

( ) ( )
λh λh λh
(3.58)
r=1 r=1 r=1

Substituting the equation (3.56) in (3.45) we obtain the following expression:

t
€ [ ] ∫ (λ h +1) f (h)dh = λ E[ H ] + λE[ H ]
E R = 2 2 2
H
2 2
(3.59)
0

By applying the relation between the mean value of the square and the variance of a random variable:
2
€ [ ]
Var [ H ] = E H 2 − E [ H ] (3.60)

we can finally express the expected value of the square of the number of vehicles arriving in the
queue in the service time as a function of the variance and the average value of the service time itself:
€

84
 Traffic Engineering and Intelligent Transportation Systems

# 2&
[ ]
E R2 = λ2 %Var [ H ] + E [ H ] ( + λE [ H ]
$ '
(3.61)

Finally, by replacing (3.61) in (3.54) we arrive at the sought expression of the expected value of the
length of the queue:
€ # 2&
λ2 %Var [ H ] + E [ H ] ( + λE [ H ] − E [ R]
$ ' (3.62)
E [Q ] = E [ R ] +
( )
2 1 − E [ R]
The (3.62) can be simplified by knowing that the average value of the number of arrivals during the
time of service, based on the (3.15), is equal to the degree of the queue saturation, that the average value
of the service
€ time is the inverse of the capacity:
1
E[ H ] =
µ
and denoting the variance with the symbol:

€ Var [ H ] = σ2H
the following expression of the average value of the queue is obtained:
% 1( 1
λ2 'σ H2 + 2 * + λ − ρ
€
& µ ) µ ρ 2 + λ2σ H2 (3.63)
E [Q ] = ρ + =ρ+
2(1 − ρ) 2(1 − ρ )
Finally, by applying Little's law (3.20):

E [Q] = λE [ D]
€
we obtain the expression of the average waiting time in queue, known as the Pollaczek-Khintchine
formula, which is at the base of the queue model used in the USA Highway Capacity Manual:
€ $ 1'
λ&σ H2 + 2 )
1 % µ ( (3.64)
E [ D] = +
µ 2(1 − ρ)
which is valid in the hypothesis λ < μ.
Note that the expression of the delay consists of a first term, which expresses the delay of the first
user in the queue, whose€ average waiting time is equal to the inverse of the average service time, and a
second term, which expresses the delay experienced by a user in a subsequent positions in the queue.
Such a second term is directly proportional to the arrival rate that feeds the queue, to the variance of the
service time, and to the square of the average service time, and is hyperbolically increasing with the
degree of saturation of the queue.
Figure 3–XII shows the shapes of the average queue and the average delay in function of the degree
of saturation and the capacity of the approach for a queueing system M|G|1.

85
 Traffic Models at Intersections

100 100

80 80

s H2 = 1
60 60

E [D ]
s H2 = 5
E [Q ]

s H2 = 1
s H2 = 10
40 40
s H2 = 5
s H2 = 10
20 20

0 0
0,0 0,2 0,4 0,6 0,8 1,0 0,0 0,2 0,4 0,6 0,8 1,0
r r

Figure 3–XII. Average queue (left) and average delay (right) of an M|G|1 queue system in function of the degree
of saturation of the queue, for different values of the service time variance, assuming that the average capacity of
the service is μ = 0.20.

Special case: Exponential service times (queue M|M|1)

We can now demonstrate that the expressions of a queueing process obtained for any unspecified
service time provide the same results (3.34) and (3.36) derived by assuming that both the service time
and the arrivals are distributed according to a Poisson process. In this case, the variance of the service
time assumes the well-knonw expression:
1
s H2 = (3.65)
µ2
By substituting the (3.65) in the general equations (3.63) and (3.64), we get again the equations (3.34)
and (3.36) for average queue and the average delay of an M|M|1 queueing process:

2 λ2
ρ + 2
µ ρ2 ρ − ρ2 + ρ2 ρ (3.66)
E [Q ] = ρ + =ρ+ = =
2(1 − ρ) (1 − ρ) (1 − ρ) (1 − ρ)
ρ λ 1
E [ D] = = = (3.67)
€ λ (1 − ρ ) λµ(1 − ρ ) µ(1 − ρ)

3.4.5 Time-dependent queue model

€
If for a certain time interval the demand exceeds the capacity, the queue does not reach a stationary
condition but grows over time. If the demand is much greater than the capacity ( λ >> μ ), the contribution
due to the random components of the number of vehicles arrived and the service time is negligible
compared to the continuous growth of the queue, which, therefore, can be more simply described by a
deterministic law. If we indicate with Q(0) the number of vehicles queued at the initial time, the
evolution of the queue according to the deterministic model is written:

Q(t ) = Q(0) + ( λ − µ)t = Q(0) + ( ρ −1) µt (3.68)

In the case of the oversaturated queues, the queue grows indefinitely and the end time of the queue is
not known. Thus, when we compute the delay by applying as the ratio between the average number of
€

86
 Traffic Engineering and Intelligent Transportation Systems

users in queue and the average number of users served in the unit of time, we have to use the capacity
instead of the arrival rate. Thus:
Q(0)
d (t ) = + ( ρ −1)t (3.69)
µ
A queue model valid in both undersaturated and oversaturated conditions should integrate the stable
queue model, described by equations (3.66) and (3.67) and the growing queue model, described by
equations (3.68) and (3.69). €Joining two structurally different models would lead to a discontinuity of
the function and therefore poor mathematical tractability. An empirical method, called the coordinate
transformation method, was then introduced to get a single monotone function by transforming a
stationary queue function into a time-dependent function.
The coordinate transformation method is based on the principle that the function sought, which
represents the dynamic model, has the same deviation from the deterministic model as the static model
has from the vertical asymptote. This condition is graphically illustrated in Figure 3–XIII, which
provides a qualitative diagram of the queue length in function of the saturation degree. In the figure, ρt
denotes the degree of saturation of the sought time-dependent queue, ρs denotes the degree of saturation
of the stationary queue, ρd denotes the degree of saturation of the deterministic oversaturated queue, and
a denotes the distance between the stationary queue and its vertical asymptote passing through the
capacity as well as the distance between the time-dependent queue and the deterministic oversaturated
queue.

a a

(ρ−1)μt

ρs 1 ρt ρd ρ

Figure 3–XIII. Coordinate transformation method: from the stationary queue model with vertical asymptote
(left) to the time-dependent model having the deterministic model asymptote (right).

In algebraic terms, the method applies the equality condition of the distances, domputed for the
same value of the queue and always indicated by the letter a in the figure, between the stationary queue
and its vertical asymptote and between the time-dependent queue and the deterministic oversaturated
queue. The equality condition provides the following relation.
1 - r s = r d - rt (3.70)
Explicitating the deterministic model for the degree of saturation (3.68):

87
 Traffic Models at Intersections

Q(t ) - Q(0 )
Q(t ) = Q(0 ) + ( r d - 1)µt Þ r d = +1 (3.71)
µt
and the stationary model (3.34):
rs Q
Q= Þ Q - Qr s - r s = 0 Þ r s = (3.72)
(1 - r s ) 1+ Q
and substituting in (3.70) we have:

Q Q − Q(0) 1
1−
1+ Q
=
µt
+1 − ρt ⇒Q = ρ t + ρtQ −
µt
[
Q − Q(0) + Q2 − Q(0)Q ] (3.73)

We thus obtain a second-degree equation in Q of the type:

Q2 + bQ − c = 0 (3.74)
€
where:
b = 1+ µt − µtρt − Q(0)
€ (3.75)
c = µtρ t + Q(0)
whose solution is in the form:
1# &
€ Q = %−b ± b2 + 4c ( (3.76)
2$ '
the only positive sign being possible since it must necessarily be Q > 0.
Proceeding analogously with the delay model, replacing the stationary delay model of Pollaczek-
€
Khintchine (3.64) and the deterministic delay model (3.69) in the coordinate transformation equation
(37), we obtain the delay equation dependent on time:
1$ '
d= − b# ± b#2 + 4c# ) (3.77)
2 &% (
where:

€ t 1
b" =
2
[
(1 − ρ) − µ Q(0) +1 ]
(3.78)
4 %t 1 (
c" = ' (1 − ρ) + ρt *
µ &2 2 )
By developing (3.77) we obtain the following expression, adopted by the US Highway Capacity
Manual (HCM, 2000):
€
1 T$ 2 8ρ '
d= + &( ρ −1) + (ρ −1) + ) (3.79)
µ 4 &% µT )(
Fig. 3-XII graphically illustrates the delay model as a function of the degree of saturation for a
general queue process with capacity µ over an observation period T.
€

88
 Traffic Engineering and Intelligent Transportation Systems

250

200

150
d
100

50

0
0 0.25 0.5 0.75 1 1.25 1.5
r
Figure 3–XIV. Time-dependent delay model according to the HCM formula.

Remark. Equation (3.79) was obtained under the hypothesis of a M|M|1 queue process, which is
valid for unsignalized intersections, where the capacity μ of a minor approach is a random variable
distributed according to the exponential law. Interestingly, it can be extended to signalized intersections,
whose over-saturation component has a similar trend for the deterministic queue. However, the first
term must be modified because it represents the average delay of a stationary queue in conditions of
under-saturation and depends on the particular structure of the traffic law. Of course, the capacity of the
signalized intersection cannot be assimilated to an exponential distribution but must be determined by
the succession of times of green and red.

3.5 Traffic Model at Unsignalized Intersections

The general expressions of the queue model derived in the previous paragraph can be applied to
both the cases of signalized and unsignalized intersections, if the the model parameters are appropriately
specified. In the case of unsignalized intersections, knowing the parameters of the probability
distribution of the vehicular arrivals on the major road (at least, the average and the variance) we can
determine the parameters of the service time. In the case of signalized intersections, the service model
must be suitably modified with respect to a function that takes into account the periodicity of the green
and red light times implemented by a traffic signal regulation system.
Once the condition are met that allows the first vehicle in the queue and to some subsequent vehicles
to enter the intersection for a crossing or turning maneuver, if the average duration of the service time
is very short and thus allows just one or two vehicles entering the intersection, as generally occurs for
unsignalized intersections, the departures can be determined according to the dynamics of an isolated
vehicle; however, there is no need to know precisely the law of motion of the individual vehicle, as this
does not affect the way the queue is cleared, which is the relevant to determine the performance of the
intersection. Viceversa, if the service time is long enough, the departures follow the departure model of
a traffic stream starting from a standstill. By applying the departure model, the capacity of a signalized
approach will be determined.

3.5.1 Hierarchy between maneuvers

The circulation rules of an unsignalized intersection are based on a precise hierarchy between the
traffic stream (see also Fig.3-XII):

89
 Traffic Models at Intersections

• Rank 1 (crossing and turning movements to the right from the main road): have priority over all
other streams;
• Rank 2 (left turn movements from the main road and right turn from the side streets): they must
yield priority to the streams of rank 1;
• Rank 3 (crossing movements from the side streets): they must yield priority to the streams of
rank 1 and rank 2;
• Rank 4 (turning movements to the left from the side streets): they must yield priority to the
streams of rank 1, 2, and 3.

12 11 10

STOP

6
5
4
1
2
3

Rank Movement
Rango Manovre STOP

1 2, 3, 5, 6
2 1, 4, 9, 12
3 8, 11 7 8 9
4 7, 10

Figure 3–XV. Hierarchy between the different maneuvers at an usignalized intersection.

Since vehicles waiting at an intersection approach must yield to all vehicles of higher-ranking
approaches, the queue model of any minor stream must be specified by assuming as usable intervals all
those not used by vehicles of the higher-ranking streams.
In the case of rank 2 movements, the average headways of the main stream is given by the sum of
the conflicting flows represented by the movements of rank 1. For example, the headways of the major
stream for the movement 1 depend on the sum of the conflicting flows consisting of the movements 5
and 6.
In the case of the rank 3 movements, we must consider that the usable headways of the rank 1
streams are usable by the rank 3 movements only if no vehicles are waiting in the rank 2 approaches.
Therefore, the model is formulated using the joined probability of having usable headways in the major
stream and no vehicles on the queue of any rank 2 approach.
In the case of the rank 4 movements, the usable headways of rank 1 movememnts can be used by
the rank 3 streams only if no vehicles are waiting in the rank 2 and 3 approaches. Compared to the
previous case, however, the model is complicated further, since the different events are not independent
of each other: the probability of not having vehicles in the queue of the approaches of the streams of
rank 3 is correlated to the probability of having usable headways in the streams of rank 1 and 2. All
these factors are considered from an operational point of view by using empirical models.
In the queue models at unsignalized intersections, it is generally assumed that the major stream is
not affected by the minor ones. This assumption is acceptable in under-saturation conditions, while it
becomes unrealistic when the overall flow is near the capacity of the road. In this case, the major stream
loses the conditions at steady-state motion and ends up being conditioned by the minor, so much so that
an actual reversal of capacity is often observed.

90
 Traffic Engineering and Intelligent Transportation Systems

The following paragraphs provide the formulations of the main descriptive variables of the traffic
at the unsignalized intersections: the capacity, the delay, and the number of vehicles in the queue.

3.5.2 Capacity of a Rank 2 Traffic Stream

The mathematical derivation of the capacity qm of a stream of rank 2 can be obtained in a rather
general form starting from the following quantities:
• N(t): the number of vehicles of the minor stream that can use a headway of duration t of the
major stream;
• fT(t): the density function of the vehicle headways of the major stream;
• qp: the average flow of the major stream.
The average number of headways t in the major stream is qp fT(t). The permitted flow of the minor
traffic streams at this value t of the major stream ranges is, according to the previous definitions: qp
fT(t)N(t).
The expected value of the capacity QM is then given by the integral with respect to the duration of
the gaps on the major stream:
¥

qm = q p
ò f (t) N (t)dt
0
T (3.80)

According to the gap acceptance theory, the capacity of an intersection between two streams of
traffic in stationary conditions (i.e., in the absence of a monotonically growing queue) can be determined
by applying simple probabilistic methods, assuming:
• known and constant values for the critical gap tc and the follow-up tf ;
• a negative exponential distribution for the headways on the major traffic stream;
• an average steady flow (ie constant over time) for the major traffic stream.
Two possible functional forms have been assumed for the function N(t): a discrete step-wise
function (see Figure 3–III):
∞

N (t ) = ∑ np n (3.81)
n=0

where it is:
$1 se : t c + (n −1)t f ≤ t ≤ t c + nt f
pn (t€) = % (3.82)
&0 altrimenti
otherwise

and a continuous function of linear form:

$ 0 for t < t0
€ &
N (t ) = % t − t0 (3.83)
& t for t ≥ t0
' f
in which
tf
€ t0 = tc - (3.84)
2
The (3.84) is a linear approximation of the step function, which represents the cumulative number
of vehicles that accept a headway interval of duration t (Figure 3–XVI): t0 is the intercept of the axis of

91
 Traffic Models at Intersections

the durations of the intervals; 1/tf is the slope of the continuous line, being tf the average time required
by a driver to cross the intersection. The factor tf /2 is obtained by constraining the line that continues to
pass through the midpoint of each “step” of the function and applying the definition of critical gap tc as
the interval required by the first vehicle in the queue.

3
N

0
0 5 10 15 20 25 30
t (s )

Figure 3–XVI. Number of vehicles in the minor stream capable of entering the crossing area as a
function of the headway between the vehicles of the major stream.

Capacity of a minor approach with negative exponentially distributed arrivals

Assuming that the arrivals of both major and minor traffic streams have a negative exponential
probability distribution and replacing (3.80) in (3.81), we obtain the expression of the capacity of a
minor approach provided by the Harders model (1968), which is also used in the US Highway Capacity
Manual (HCM, 2000) for non-platooned traffic streams:
- q p tc
e
qm = q p -q pt f
(3.85)
1- e
The equation (3.85) can be obtained by thinking of the capacity qm as the average number of vehicle
headways N observed before having a smaller interval of tf , conditioned to the probability that the
vehicular interval is greater than tc (that is, the headways are usable) in a time interval in which qp
arrivals occur in the major stream :

[
qm = q p P [t > t c ] N | t1 ,t2 ,...,t N > t f ,t N +1 ≤ t f ] (3.86)

In fact, the average number of headways N before observing the occurrence of an event p is a random
variable with a geometric probability distribution, whose expected value is equal to the inverse of the
probability p€of occurrence of a single event:
1
[
E N ( p) = ] p
(3.87)

In the present case, the probability p of that event equals the probability that T <tf in a major stream
with average flow qp
€
−q t
[
p = P T ≤ t f =1− e p f ] (3.88)

€ 92
 Traffic Engineering and Intelligent Transportation Systems

The probability of the conditioning event, corresponding to the occurrence that the vehicular interval
is greater than tc is:
−q p t c
p = P [T < t c ] = e (3.89)

By replacing the relationships (3.87), (3.88), and (3.89) in (3.86) we obtain the (3.85).

Figure 3–XVII shows€an example of a diagram, taken from HCM (2000), which provides the
capacity of the minor approach depending on the average value of the conflicting flow on the major
stream31. It should be observed that of the priority rule, which imposes the drivers to stop and verify if
the gaps in the major stream are usable, implies, even if the flow on the main road were nil, a potential
capacity of about 1,000 veh/h; that is, it produces a reduction in the capacity of the road as about the
50% compared to the undisturbed flow conditions on a road link. As the conflicting flow of the major
streams increases, the capacity of the minor decreases rapidly and gets a value of about 500veh/h when
the major stream assumes the same value of 500veh/h.

Figure 3–XVII. Capacity of a minor stream as a function of the average flow of the overall conflicting
movements of the major stream (HCM, 2000).

3.5.3 Delay Model for Unsignalized Intersections

The delay expressions obtained in the paragrah 3.4 can be applied directly to a minor stream of an
unsignalized intersection, simply by replacing the generic symbol of the capacity μ with the value qm
obtained in par.3.5.2.
Thus, in the case of stationary queue ( ρ <1), we have the Pollaczek-Khintchine formula:

æ 1 ö
l çç s H2 + 2 ÷÷
1 µ ø (3.90)
E [D ] = + è
qm 2(1 - r )

31
More general expressions can be obtained by adopting a dichotomized distribution instead of the simple negative exponential
(Troutbeck and Brilon, 1996).

93
 Traffic Models at Intersections

being, as usual, E[D] the expected value of the average delay of the vehicles in the queue, ρ the degree
of saturation, and σ2 the variance of the vehicle intervals in the minor stream.
In the case of a time-dependent queue, the expression adopted by the USA Highway Capacity
Manual since 1990 (Akcelik, 1988) is instead:

1 Té 8r ù
E [D ] = + ê( r - 1) + (r - 1)2 + ú+5 (3.91)
qm 4 ë q mT û
that expresses the average unit delay in the queue in a time interval of length T. It should be noted that
compared to the formula (3.79) obtained from the queue theory, which expresses the waiting time in the
queue, an additional delay period of 5s has been added to the HCM operating formula, which includes
the times lost in deceleration and acceleration during the approach and departure from the intersection.
Fig. 3-16 shows a diagram taken from the HCM (2000), which shows the unit delay of the vehicles
of a minor stream as the flow varies, and for different capacity values of the minor stream.

Remark. The delay model derived from the queue theory expresses the average waiting time in a
queue characterized by a completely general law of arrivals on the secondary road and by a negative
exponential distribution of the service time (that is, of the vehicle headways of the major stream). This
second hypothesis overlooks the fact that the service time differs for the first vehicle in the queue, for
which tc is valid and for the following ones, for which it is tf . A queue model M|G2|1 was thus developed,
characterized by two types of service time, one valid when the arriving vehicle finds an empty queue,
the other when it finds other vehicles. The expressions of delay obtained are an extension of (3.79).
However, they are excessively complicated to be accepted by the manuals and are instead more suitable
for use within a simulation model.

Figure 3–XVIII. Unitary delay of the minor stream as a function of the average minor flow for different capacity
values and an analysis interval of 15 minutes (HCM, 2000).

3.5.4 Queue length

The expressions of the length of the queue are obtained in a completely analogous way to that of
the delay. The average number of vehicles in the queue is then given by (3.63) in the stationary case and
by (3.76) in the case of an increasing queue over time.

94
 Traffic Engineering and Intelligent Transportation Systems

The probability that there are no vehicles waiting for a generic approach, necessary to calculate the
capacity of the streams of rank 3, in the rather general queue hypothesis M|G|1, holds:
P[Q = 0] = 1 - r (3.92)
being ρ the degree of saturation.
Another value of interest for traffic engineering is the queue length with a 95% probability of not
being exceeded; that is, the 95th percentile:

Té 8r ù
Q95 = ê( r - 1) + (r - 1)2 + ú qm (3.93)
4ë q mT û

3.6 Traffic Model at Signalized Intersections

3.6.1 Signal Settings at Road Intersections

Signal settings allows to reduce the points of conflict between the various traffic streams of the
intersection, thus obtaining a reduction in the waiting times of the non-priority streams which, for very
high opposite flows, could grow to become unacceptable.
The regulation of maneuvers with a traffic signal is based on the following rules:
a) the movements of the different traffic streams are organized according to successive phases: in each
signal phase the occupation of the intersection is allowed to a group of compatible maneuvers (two
different examples of organization in phases are shown in Fig.3-XVII);
b) the movements of each phase are allowed during the green light time (and in a fraction of the yellow
time, for the time necessary to clear the intersection by vehicles that when the yellow appears are
unable to stop ) and prevented during the red light time: this rule of circulation corresponds in the
theory of queues to a deterministic service time which is cyclically variable over time according to
the law illustrated in Fig.3-XVIII;
c) the green, yellow, and red signals for each stream i follow each other cyclically (the duration of the
overall period consisting of green Gi, yellow Yi, and red Ri is called cycle and is denoted as C).

Phase11
FASE Phase
FASE 22
12 11 10
A
IPOTESI A
Example

6
5
4
1 Phase11
FASE Phase 22
FASE Phase
FASE 33
2
Example BB

3
IPOTESI

7 8 9

Figure 3–XIX. Conflict points at a signalized intersection (on the left) with two different hypotheses for
organizing the maneuvers in signal phases (on the right).

95
 Traffic Models at Intersections

t
R G Y R G Y R
C C
Figure 3–XX. Service time of a queue regulated with a traffic signal.

Some constraints among the traffic signal parameters prevent the simultaneous execution of the
conflicting maneuvers of two different phases. For a two-phase intersection, we have:
R1 = G2 + Y2 ; R2 = G1 + Y1 (3.94)
As a result, the cycle duration is:
C = G1 + Y1 + G2 + Y2 (3.95)
The Figure 3–XXI shows the relationship of incompatibility which determines the alternation of
green and red times of a traffic signal in two stages in two successive cycles.

R1 G1 Y1 R1 G1 Y1
C C

G2 Y2 R2 G2 Y2 R2
C C

Figure 3–XXI. Sequence of the green and red times of a two-phase traffic signal.

For pretimed traffic signals, the sequence and duration of the various phases is predetermined on
the basis of statistical observations of the traffic flows of the various streams. Usually, a different
timing is provided for the different time slots (morning rush hour, morning off-peak hour, afternoon
peak hour, night hours), in order to adapt the regulation to the statistically foreseeable traffic conditions.
For actuated traffic signals, the flow of the various traffic streams is detected by automatic traffic
detectors which send the information to a control unit, which adjusts the duration of every phase from
time to time, in compliance with some minimum and maximum values.

3.6.2 Departure Model at a Traffic Signal

While the arrivals of the vehicles at an intersection are characterized by an inherent uncertainty,
regardless of whether the intersection is signalized or not, so that we can use the probability distributions
discussed in subsection 3.3, the departures of vehicles at green from a stopline can be considered
deterministic with a good approximation, at least in the case of pretimed traffic signals. The vehicles
start from a standstill and gradually increase their speed; at the same time, by assimilating the traffic
stream to a fluid, an increase in the flow at the stopline of the approach is observed, which tends to the
maximum flow (denoted as saturation flow). This means that a traffic stream that starts from zero speed
cannot fully exploit the available capacity during the green light time, as shown in Fig. 3-XVIII.

96
 Traffic Engineering and Intelligent Transportation Systems

This phenomenon is described with a good accuracy by the departure model developed in the
context of the macroscopic flow theory and described in par.1.3.2, whose final equation (1.45) is
reported here for the reader’s convenience:
at
" 2 % −v
#
(
q = $ at + vc ± ) (at + v ) c ( )
− 2a vc t + u ' k J e c
&
(3.96)

Representing the outflow with this equation would significantly complicate the queue model. On
the contrary, it would be very convenient to use the simple scheme described in Figure 3–XX.

t
R G Y R G Y R
C C
Figure 3–XXII. Service time of a queue regulated with a traffic signal.

On this regard, it is sufficient to substitute the real light time of right-of-way G + Y, with an effective
green that supplies, for a service rate equal to the saturation flow, the same number of vehicles actually
departed during the right-of-way time G + Y .
The effective green g is therefore equal to the ratio between the number of vehicles started, given
by the integral of (54) in that interval, and the saturation flow s .
G +Y

sg =
ò q(t )dt
0
(3.97)

The outflow at the start of the green of the signal is thus schematized assuming that for a certain
time l1 (called “start-up lost time”) the flow is zero, and therefore the flow occurs at a constant rate equal
to the saturation flow s (which numerically coincide with the capacity of lane with uninterrupted
traffic), until the queue is cleared of or until a vehicle is required to stop at yellow. That is, a fraction of
the yellow time is used for the outflow by those vehicles which, at the start of the yellow, were too close
to the stop line to stop, while the remainder is used to allow the last transited vehicle to clear the area
crossing. This second fraction of yellow is not used for the outflow and constitutes a second term of lost
time, l2, said “clearance lost time”.
The total lost time of any generic movement i is therefore:
Li = li1 + li2 (3.98)
The clearance lost time can be calculated assuming that the last vehicle crosses the clearing section
(that is, the hypothetical line where the intersection area ends) at the critical speed. This hypothesis
involves a completely negligible approximation, considering that in most intersections the clearing
section is not too far from the critical section and, moreover, that the speed tends asymptotically to the
critical value. The clearing lost time l2 is therefore simply given by the ratio between the distance xs of
the critical section (see Figure 3–XXIII) from the stopline and the critical speed v c :

97
 Traffic Models at Intersections

xs
l2 = (3.99)
vc
The start-up lost time is then obtained by simply subtracting the effective green and the clearance
lost time from the total time right-of-way G + Y :
l1 = G + Y - g - l 2 (3.100)

xs

vc

t
l2

Figure 3–XXIII. Clearance of the intersection area from a signal: the clearance lost time l2 is the time necessary
for the last vehicle in the queue to completely cross the whole intersection area.

Remark. It is not necessary to solve the integral (3.97) every time you have to study a vehicle queue
at a traffic signal. In fact, by replacing (3.97) and (3.98) in (3.99), we can express the calculation of the
start-up lost time as the only variable of the problem.
G +Y

ò q(t )dt
x 1
l1 = G + Y - s - (3.101)
vc s
0

Consider now that, for successive instants of time to a value τ next to 15s, the flow q is practically
coincident with the saturation flow s , so that, beyond this value τ, the contribution provided by the last
term of the second member of (3.101) equals the corresponding contribution of the free time, G + Y - τ .
Fig. 3-XXI provides a graphical interpretation of the start-up lost time, as an interval to be subtracted
from the green time to equal the two areas identified in the figure by the letters OABC and OF. For this
property, the areas between the two curves are equal to each other. It can be observed that the upper area
reduces itself out as time increases. It is clear then that by repeating the same procedure for an interval
greater than 20s, the same value would be obtained for the start-up lost time, since two almost equal
areas would be added and subtracted in (3.101), with a height approximately equal to the saturation flow
s.

98
 Traffic Engineering and Intelligent Transportation Systems

0.6
B C
0.55

0.5
0.45
l1 F
0.4
0.35

0.3
q

0.25

0.2

0.15

0.1
0.05
O A
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
t
Figure 3–XXIV. Departure model at a traffic signal: the start-up lost time l2 = OA is defined in such a way that
the broken line OABC subtends the same area of the OF curve.

The introduction of lost time allows us to model the operations of the traffic signal by referring to
the only two conditions that are effective from a functional point of view: the effective green, g , equal
to the available right-of-way time, minus the lost time:

g i = Gi + Yi - Li (3.102)
and the effective red , r , equal to the total unused time for the flow:

ri = C - g i (3.103)
Figure 3–XXV illustrates the correspondence from the real signal timing (with green, red, and
yellow lights) to the effective green and red scheme and the role of the start-up lost time l1 and the
clearance lost time l2.

C C

R1 G1 Y1 R1 G1 Y1

l1 l2 l1 l2
g1 r1 g1

g2 r2 g2

l1 l2 l1 l2

G2 Y2 R2 G2 Y2 R2
C C

Figure 3–XXV. Schematic diagram of the signal timing using effective green g and effective red r, defined by
subtracting for each phase i the start-up lost time l1 and clearing lost time l2, respectively to the light green
time Gi and the light yellow time Yi.

99
 Traffic Models at Intersections

As a result of previous assumptions and simplifications, the departure model is represented, as in

Figure 3–XXVI, assuming that:
• During the effective red the flow is zero;
• During the effective green the flow is equal to the saturation flow until the queue of waiting
vehicles is discharged;
• From the clearance time of the queue, the arriving vehicles cross the intersection without being
stopped and therefore arrivals coincide with the departing ones, so that the outflow is the same
as the inflow.
In formal terms, the number Pi(t) of vehicles departing from the approach I at time t is written:

0 (0 £ t < r ) i

Pi ( t ) = s (t - r ) ( r £ t < t )
i i i
*
i (3.104)
qt i( t £ t < C)
*
i

having indicated with t1* the clearance time of the queue.

N
A(t)

P(t)
s
t* t**
t

r1 g1 r1 g1
C C
Figure 3–XXVI. Schematization of a generic process of arrivals A(t) and departures P(t) at a traffic signal, where
t* and t** denote the queue clearance times at the first and second cycles respectively.

3.6.3 Delay model

The delay model at signalized intersections derives from the queuing theory and the related general
concepts described in par. 3.4, taking account of the traffic rules defined by the departure model
introduced in subpar.3.6.2.
It is clear that the considerations already done in par. 3.4 and reported in Figure 3–IX can be applied
to the flow scheme at a signalized approach, as illustrated in Figure 3–XXVII:
• The total delay of the vehicles arrived in every cycle C is represented by the area between the
two curves of arrivals A(t) and departures P(t);
• A horizontal segment d(t) joining the two curves represents the delay of one vehicle that has
reached the queue at the instant corresponding to the first extreme of the segment;
• A vertical segment NQ(t) that joins the two curves identifies the number of vehicles in the queue
at time instant t corresponding to the abscissa of the segment.
Thus, the basic concepts of the queueing theory that was applied to unsignalized intersections can
be applied also to signalized ones. In this the latter case, however, the alternation of the green over the
different signal phases determines a periodicity of the departures that allows us to model them as a

100
 Traffic Engineering and Intelligent Transportation Systems

deterministic process. If the arrivals also had a limited or perhaps a negligible random variability, the
delay model could be formulated in deterministic terms. This condition is not as rare as one might think
and is indeed an acceptable approximation of the real phenomenon for very high traffic flows, when the
vehicles are interact very closely, so that the statistic dispersion of the headways is very small.
The importance of the deterministic model also goes beyond the approximation of the flow for heavy
traffic: in fact, because of the periodic structure of the departure function, a deterministic component is
inherent in the delay model and corresponds to the periodic process of departures and to the average
value of the arrival flow.
To this deterministic component, a random component is superimposed due to the variability of the
arriving flow from one period to another. It follows that, in general, the delay of a traffic stream at an
intersection approach can be represented by the sum of two terms: a deterministic and a random one.

n
qt* = s(t* - r)

n A (t)=q t
d(t)
n P (t)=s(t-r)

nQ(t)
t*
r g t
C

nQ
n Q (t) = n A (t) - n P (t) dw=n Q (t)dt
== Queued
veico li invehicles
coda
× iltime
per in queue
tempo in coda
n Q (t)

t t
t*
dt

Figure 3–XXVII. Delay model for uniform arrivals: diagram of arrivals and departures (above); digaram of the
queue (below).

Deterministic component. The deterministic component of the delay can be determined once the
arrival distribution function and the method for calculating the average value of the flow have been
specified.
In general, the time profile of the arrival flow can be even highly variable within a cycle. This is the
case, for example, of an urban road artery with close intersections whose cycles have the same length.
From a generic intersection, a platoon of vehicles starts at the beginning of the green, ending at the latest
at the beginning of the red and traveling relatively compact along the artery. At the next intersection,
the profile of the vehicle arrivals is characterized by a high flow for a time interval having a duration
close to the green of the previous intersection and by a lower flow (with a greater uncertainty) during
the remaining part of the cycle. In this case, the arrivals function must be described with a time-
dependent function q (t), defined in the interval [0, C ], determined by averaging the flow arrived in the
various cycles at instants q(t + iC), being t the generic instant of time and the generic cycle considered.

101
 Traffic Models at Intersections

In the case of isolated intersections, on the other hand, incoming traffic is not characterized by any
regularity, so that for high flows the arrival process can be considered uniform. The deterministic
component of the delay is called uniform delay. This particularly simple case is actually very important,
both because it allows us to easily calculate and represent the delay graphically in these simple cases,
and because it is adopted as the first term of the function of the overall delay in most Manuals. Figure
3–XXVIII illustrates the delay model for a signalized approach with uniform arrivals.

Since for uniform arrivals and departures the queue has a triangular pattern, the total delay of the
traffic stream at the approach is:
1 *
w = qrt (3.105)
2
The instant of queue clearance t* can be determined by imposing the condition that at that instant
the numbers of vehicles arrived and deperted are equal:

( )
qt * = s t * - r Þ t * =
r
q
=
r
1- y (3.106)
1-
s
having indicated with y the ratio between arrival flow and saturation flow (degree of saturation) of the
traffic stream.
The formula of the total delay for uniform arrivals at the approach can then be written, after
replacing (3.106) in (3.105):

qr 2
w= (3.107)
2(1 - y )
The average unit delay for the approach is given by the ratio between the total delay and the total
number of vehicles arrived at the approach (whether they have been stopped or not):

w qr 2 r2
d= = = (3.108)
qC 2qC (1 - y ) 2C (1 - y )
The average delay of the intersection is of course the sum of the total delays at all the approach,
divided by the total number of vehicles arriving at the whole intersection in a cycle:
m m

å å wi
qi ri 2
2(1 - y i )
d= = i =1
m
i =1
m
(3.109)

åq åqi =1
i
i =1
i

being m the number of phases.

102
 Traffic Engineering and Intelligent Transportation Systems

100
90
80
r =35s
70
60
50
d

r =25s
40
30
20
10
0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
y

Figure 3–XXVIII. Deterministic term of the unit delay for uniform arrivals as a function of the degree of
saturation y, for C = 60s and r = 35s (dark blue curve) or r = 25s (light purple curve).

Random component
The random component of the delay at signalized intersections can be derived in closed form starting
from the queueing theory and applying the same procedure described for the unsignalized intersections,
the hypotheses of Poissonian arrivals still being valid (vehicular intervals distributed according to a
negative exponential) and assuming that the departures follow a generic law (queue M|G|1). Under these
assumptions, the average unit delay, as known, is given by the Pollaczek-Kintchine formula. By
applying the coordinate transformation method (Kimber and Hollis, 1979) we obtain the following
equation for the incremental delay that sums up the random component of the delay and the
oervsaturation component:

Té ù
E [D2 ] = ê( x - 1) + (x - 1)2 + 8 x ú (3.110)
4ë cT û
Where a different symbology has been introduced for the capacity ( c ) and the flow/capacity ratio ( x ),
as these are modified on the basis of the green share for the traffic stream under examination.
c = gs
q
y=
C (3.111)
qC
x=
gs
The only difference with the unsignalized intersections concerns the different definition of capacity
c in the case of the traffic signal.
The equation (3.111) is also adopted by the HCM (2000), apart from an empirical adjustment factor
that multiplies the last term, introduced to take into account the type of traffic signal, pretimed or
actuated.

103
 Traffic Models at Intersections

3.7 Bibliography
Akcelik, R. (1988). The highway capacity manual delay formula for signalized intersections. ITE journal, 58(3),
23-27.
Cowan, R. J. (1975). Useful Headway Models. Transportation Research, 9(6), pp. 371-375.
Harders, J. (1968). Die Leistungsfähigkeit Nicht Signalgeregelter Städtischer Verkehrsknoten (The Capacity of
Unsignalized Urban Intersections). Schriftenreihe Strassenbau und Strassenverkehrstechnik, Vol. 76. (from:
Troutbeck & Brilon, 1996).
HCM (2000). Highway Capacity Manual. Transportation Research Board, National Research Council.
Washington, DC, USA.
Schuhl, A. (1955). The Probability Theory Applied to the Distribution of Vehicles on Two-Lane Highways.
Poisson and Traffic. The Eno Foundation for Highway Traffic Centre Control. Sangatuck, CT.
Troutbeck, R. J., & Brilon, W. (1996). Unsignalized Intersection Theory. In Traffic Flow Theory. Washington,
DC: US Federal Highway Administration, 8-1.

104