This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY 1

Probabilistic Anticipation and Control in

Autonomous Car Following
Nianfeng Wan , Chen Zhang, and Ardalan Vahidi, Member, IEEE

Abstract— Human-driven and autonomously driven cars of The CACC allows participating vehicles to communicate
today act often reactively to the decisions of the cars they follow, with each other and improve their collective performance.
which could lead to uncomfortable, inefficient, and sometimes However, the CACC requires vehicle connectivity, some level
unsafe situations in stop and go traffic. This paper proposes
methods for probabilistic anticipation of the motion of the of autonomy, and a complex coordination scheme. In the near
preceding vehicle and for the control of motion of the ego vehicle. future, it is unlikely to replace all road vehicles with these
We construct: 1) a Markov chain predictor based on the observed capabilities, and the traffic will remain mixed.
behavior of preceding vehicle and 2) a maximum likelihood This paper proposes an autonomous car following system
motion predictor based on historical traffic speed at different that by anticipating the most likely motion of its preceding
locations and times. Heuristics are proposed for combining the
two predictions to determine a probability distribution on the vehicle performs better on average. For instance, such an
position of the preceding vehicle over a future planning horizon. autonomous car following function can switch to its “cautious”
A chance-constrained model predictive control framework is mode when it determines, from recent sensory data, that a pre-
employed to optimize the motion of the ego vehicle, given the ceding vehicle is being aggressively driven; or in anticipation
probabilistic prediction of motion of preceding vehicle. Effective- of a recurrent bottleneck (e.g., a congested intersection during
ness of the proposed approach is evaluated in multiple simulation
scenarios. morning rush hour), the car following function can smoothly
slow down instead of a reactive and sudden deceleration when
Index Terms— Autonomous vehicle, chance constrained model
predictive control (MPC), Markov chain, maximum likelihood faced with a queue [5]. Proliferation of connected vehicles and
estimation, Monte Carlo sampling, predictive cruise control. connected road infrastructure systems provides abundant data,
which could aid in prediction of imminent events down the
I. I NTRODUCTION road.
A UTONOMOUS vehicle functions or full vehicle auton-
omy could enable safer, more comfortable, and fuel-
efficient driving. For instance in automated car following,
There is a large body of the literature on predicting an
individual vehicle behavior. For instance, in [6], the authors
predict driving aggressiveness based on personality and age
precision sensing and tight control enable smoother perfor-
of the driver and power of the vehicle. An alternative way
mance in maintaining a safe distance to a preceding vehicle
is to learn the driving pattern from historical driving data.
and as a result could improve ride comfort and energy effi-
A widely used method is a Markov chain in which transitions
ciency of the vehicle. Due to these benefits, the adaptive cruise
of states such as velocity are assumed to be a random process
control (ACC) is now available on many vehicle models and
and in which the probability distribution of the next state
a large body of research on the ACC exists [1]. A vehicle
depends only on the current state and not on the sequence of
equipped with the ACC can detect the preceding vehicle’s
events that preceded it. A probability transition matrix between
position and velocity and maintain a safe desired distance or
consecutive states is built using historical data and then used
headway. However, to the best of our knowledge, current ACC
to predict the probability distribution of a state in near future.
systems rely on instantaneous sensor information and therefore
Liu and Salvucci [7] use a Markov model to predict driver’s
engage reactively rather than proactively. Although the reac-
intended actions. Guo et al. [8] use a Markov chain to predict
tion time of an onboard computer is shorter than a human
lane changing intention of a neighboring vehicle. Our group
driver, the motion based on instantaneous information can be
has proposed a Markov chain Monte Carlo method to sample
jerky, uncomfortable, and inefficient. One way to solve this
from the transition matrix to predict the preceding vehicle’s
problem is to resort to the Cooperative ACC (CACC) [2]–[4].
trajectory [9]. Similar ideas can also be seen in [10] in which
Manuscript received January 24, 2017; revised August 2, 2017; accepted the authors further take road grade into consideration in their
October 3, 2017. Manuscript received in final form October 8, 2017. This Markov model.
work was supported by the BMW Information Technology Research Center in
South Carolina. Recommended by Associate Editor K. Butts. (Corresponding A vehicle’s motion is also affected by surrounding traffic.
author: Nianfeng Wan.) The general traffic pattern is recurrent [11] thus predictable to
N. Wan and A. Vahidi are with the Department of Mechanical Engineering, some degree. On the other hand, traffic is not evenly distributed
Clemson University, Clemson, SC 29634 USA (e-mail: [email protected];
[email protected]). along the road; it is reasonable to expect high density and
C. Zhang is with Ford Motor Company, Dearborn, MI 48120 USA (e-mail: delay around intersections and faster traffic at midlinks [12].
[email protected]). Considering the recurrent pattern of upcoming traffic, one
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. can use historical traffic data to predict traffic conditions for
Digital Object Identifier 10.1109/TCST.2017.2762288 different sections of a road. For instance, an autonomous car
1063-6536 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

can learn in advance of a potential slow down near an on described by

ramp at rush hour. Our group has proposed a method, which ⎧
⎪
⎪ ẋ = v
only uses sparse probe vehicle data to estimate the most likely ⎨
traffic speed at different road segments [13]. v̇ = a (2)
⎪
⎪
The literature on prediction of driver behavior has often ⎩ȧ = − 1 a + 1 u
overlooked the dependence on traffic conditions. A contri- τ τ
bution of this paper is illustrating that individual driving where u is the acceleration command input and τ is a time
signatures as well as historic location- and time-specific traffic constant. The state vector can be written as z = [x, v, a]T .
patterns could be combined to predict the motion of a pre- After discretization, the state space model becomes
ceding vehicle. Combining the two predictors is a challenge
z d (k + 1) = Ad z d (k) + Bd u(k) (3)
and here we propose a heuristic solution. A number of papers
address the problem of combining forecasts with many predic- where Ad and Bd are discretized state space matrices.
tors [14]–[17].
With prediction results, various control schemes can be used C. Chance Constrained MPC Formulation
for car following. The model predictive control (MPC) can be
a very effective method [18] because of its anticipatory nature. We propose to use an MPC approach for controlling the
Such an MPC-based car following model uses predictions of speed of the ego vehicle. Given a cost function J , the
the motion of the preceding vehicle and considers safety gap MPC solves a finite moving horizon optimization problem at
constraints, to determine the optimal trajectory of ego vehicle. each time step and only applies the current time control input
Note that the predictions need not be deterministic. In fact, to the system. At the next step, the horizon is moved forward
a prediction represented by a probability distribution could be by one step and the process is repeated. Let z d (i +k|k) denote
used to handle partial and imperfect information of the motion the i th step prediction at time step k. With a horizon of Nc time
of preceding vehicle. This in turn would introduce probabilistic steps, a sequence U = [u(k|k), u(1+k|k) . . . , u(Nc −1+k|k)]T
constraints to be enforced. In this paper, we adopt a Chance- is to be found, which minimizes the cost function. But only
Constrained MPC approach [19] to handle probabilistic predic- the current time step control input u(k|k) is applied. As the
tions and constraints. A similar approach was first introduced system moves one time step forward, another finite horizon
by our group and was published in [9]. We build on this previ- optimization problem needs to be solved, with the current
ous work by introducing our combined predictor. We compare states as initial conditions. States and control inputs have
the performance of the proposed approach to methods that constraints.
use only instantaneous sensory information in a number of In this paper, the cost function is defined as
simulated scenarios. c −1
N
The rest of this paper is organized as follows: In Section II, J= a 2 (i + k|k)
the chance-constrained MPC problem is formulated. i=0
Sections III-A and III-B introduce the trajectory prediction + q(r (Nc + k|k) − x(Nc + k|k) − T ẋ(Nc + k|k) − Dmin )
approaches based on probe vehicle data and acc driving data, (4)
respectively. Section III-C discusses how predictions from
the two predictors are combined. A number of simulation where r and x denote the position of the preceding and ego
scenarios is presented in IV. Section V provides conclusions vehicles, respectively, and q is a penalizing coefficient. Penal-
of this paper. izing acceleration in the cost function is meant to improve
ride comfort. The last term penalizes the gap between the
II. C HANCE C ONSTRAINED MPC FOR C AR F OLLOWING two vehicles at the end of control horizon and encourages
A. Problem Statement maintaining a safe and reasonable distance between the two
vehicles. It has been shown that this cost function also reduces
The goal is for the ego vehicle to maintain a desired distance
fuel consumption [9], because the first term discourages unnec-
to its preceding vehicle. Determining conditions to enter or exit
essary braking (deceleration). Hard constraints on vehicles
this car-following mode is out of scope of this paper. Here,
states and on the following distance must be enforced at each
we define:
step in time. For example, at any time, the acceleration should
Ddesired = Dmin + T v (1) be between its maximum and minimum values, that is
amin ≤ a(i + k|k) ≤ amax . (5)
where Ddesired is the desired distance between the ego vehicle
and the preceding vehicle, Dmin is the minimum required safe The following distance between two vehicles must also be
distance, T is the reaction time (also known as headway), and constrained:
v is the speed of the ego vehicle.
Dmin ≤ r (i + k|k) − x(i + k|k) − T ẋ(i + k|k) ≤ Dmax (6)
B. Vehicle Kinematic Model where Dmin is the minimum allowable gap for safety reasons
The kinematics of a vehicle longitudinal motion is modeled, and Dmax prevents large gaps that reduce road capacity.
along with a first-order lag between the acceleration command If the preceding vehicle’s trajectory is known, this becomes
input and the vehicle’s acceleration. The state space model is a standard linear MPC problem and can be efficiently solved as
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WAN et al.: PROBABILISTIC ANTICIPATION AND CONTROL IN AUTONOMOUS CAR FOLLOWING 3

a quadratic program. However, at time step k, r (i + k|k) is not

known. In Sections III-A and III-B, we propose two methods
for estimating the probability distribution of r (i + k|k) at
any given future step i . It is then reasonable to specify a
chance constraint on the minimum following distance where
the constraint is satisfied with a given probability α:
Pr (x(i + k|k) + T ẋ(i + k|k) ≤ r (i + k|k) − Dmin ) ≤ 1 − α
(7)
which means that the chance of violating the constraint should
Fig. 1. Schematic of segment travel times crowd-sourced from probe vehicle
be less than 1 − α. Note that at any instant time step i , x, ẋ updates [13].
and r are all deterministic rather than predicted, and the
MPC finds solutions that do not violate constraints. In other
words, MPC solutions will not violate any constraints in as shown in Fig. 1. After aggregating months of data, the bus
real time. The probabilistic constraint can be converted to a GPS position information is transformed to a distance to a ref-
deterministic constraint when the probability distribution of erence point, and the whole path is divided into 5-m segments.
j
r (i + k|k) is known. If we denote R1−α as the position where The travel time for each segment is denoted by ttxi for the
the cumulative distribution function of r (i + k|k) is equal j th bus at the x i th segment. For each bus, two consecutive
to 1 − α, then the equivalent deterministic constraint is updates provide a travel time observation, which is denoted
j
by tt[xa ,xa ] from the beginning segment x a1 to the end
x(i + k|k) + T ẋ(i + k|k) ≤ R1−α − Dmin . 1 n
(8) segment x an . The sum of each segment travel time should
Similarly, we can obtain the maximum distance constraint equal the total observed travel time, that is
at a given probability β, therefore j

an
j
tt[xa ,xa ] = ttxi . (10)
1 n
Rβ − Dmax ≤ x(i + k|k) + T ẋ(i + k|k) ≤ R1−α − Dmin . i=a1
(9) The approach is to allocate the total travel time into each
segment such as to maximize the likelihood of the observation.
With this transformation of the probabilistic constraints to
We proposed an expectation maximization (EM) approach to
deterministic ones, we end up with a standard MPC problem.
iteratively allocate the observed travel time into each segment
The challenge is obtaining a probability distribution for the
and subsequently calculate the statistics of travel time for
position of the preceding vehicle at future steps in time, which
each segment. We are assuming that the travel time for
is the focus of Section III.
each segment has a Gaussian probability density function.
Therefore, after each allocation, we can easily calculate its
III. P ROBABILISTIC P REDICTION OF P OSITION OF likelihood. The segment travel time is reallocated to maximize
P RECEDING V EHICLE this likelihood. After initialization, the EM algorithm has the
A. Vehicle Trajectory Prediction Based on following steps.
Historical Probe Data E step: Having allocated travel times from different buses,
In this section, we propose to use sparse data from probe for each segment x i , we calculate the mean μxi and vari-
vehicles to estimate travel time statistics across different ance σxi . Then, assuming a normal distribution, the probability
segments of the road and to use these probabilistic estimates density for tt kxi can be calculated as
 2
to predict the motion of a preceding vehicle in traffic. A probe tt kx −μx i
vehicle is designed to drive in the traffic, and collect and report   1 − i
2σx2
p tt kxi |μxi , σxi = √ ·e i . (11)
real time traffic information. This prediction method is built 2πσxi
on our previous work in [12], [13], and [20], in which we
M step: In this step, we go back to each travel time
proposed an approach to reconstruct the most likely trajectory
observation and reallocate the segment travel time such that
along an arterial road between two consecutive probe data
the likelihood function is maximized. Suppose an update pair
updates. The example data source we use is a sparse data
starts at segment x a1 and ends at x an ; its log likelihood function
feed from transit buses in the city of San Francisco that
can be represented by
was publicly available. But the method proposed here can
be applied to other vehicle data streams. Each bus in San   an
 
log p tt kX |μ X ,  X = log p tt kxi |μxi , σxi (12)
Francisco has a GPS device, which reports its location, speed,
i=a1
and time stamp every 200 m or 90 s whichever comes first.
The position and time of an update is at random. In [12], where X = [x a1 , x a2 , . . . , x an ]T , μ X = [μxa1 , . . . , μxan ]T , and
we proposed to divide a road into smaller segments to find ⎡ 2 ⎤
σxa 0 .. 0
each segment travel time statistics, as shown in Fig. 1. ⎢ 1
σx2a .. 0 ⎥
Different bus passes cover different sections of the path; X = ⎣ 0 ⎦.
2
however, they may cover several same small segments, 0 0 .. σxan
2
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 4. Prediction example including instantaneous velocity information.

Fig. 2. Segment travel time and rms error evolution of our EM algorithm
in [12].
Mathematically, if time is fixed, the distance traveled is propor-
tional to velocity, and velocity distribution is the reciprocal of
travel time. Therefore, the position distribution is a reciprocal
Gaussian distribution, which in certain area is analogous to
a Gaussian distribution but usually skewed to right, and the
first-order and higher order moments (mean and variance) do
not exist [21]. In this problem, since the travel time is far
from zero and its variance is small, we can approximate the
position distribution (between A and C) by assuming that
it is Gaussian. Such assumption introduces some error and
downgrade the performance of the predictor. Another way of
solving this problem is to assume that travel time distribution
is log-normal. Then, the position distribution is reciprocal of
Fig. 3. Trajectory prediction schematic. log-normal, which is also log-normal. This is the future work
of [13] and is out of the scope of this paper.
The predictor requires a starting update (T (x a1 ), x a1 ) as the
This problem is a constrained quadratic programming input and use historical traffic data to predict the position
problem. The solution to this problem is the maximum like- probability distribution at the end of the prediction time
lihood segment travel time allocation for each observation. horizon. However, such prediction is merely based on histor-
In Fig. 2, we show the segment travel time evolution and root ical average travel times and does not take the instantaneous
mean square (rms) errors through multiple iterations. velocity information at start into consideration. Intuitively,
Note that the travel time statistics in a road is sensitive to the a faster-than-average vehicle at start is likely to maintain the
time of the day. Here, we generate the statistics for different fast behavior for a period of time, but the speed gradually
times of the day. Furthermore, queue patterns (e.g., queuing approaches the average speed. Based on this idea, we propose
and discharging characteristics) can also be learned and once a method to take the starting velocity information into account.
traffic signal timings become available, stop position and We first calculate the difference between the instantaneous
duration can be estimated as shown with more detail in [13]. velocity v(x a1 ) and the average velocity l/tt¯ (x a1 ), where l is
Having obtained the mean segment travel times along the segment length, that is
path of interest, and given a start time, we can predict the tra-
¯ a1 )|.
diff = |v(x a1 ) − l/tt(x (13)
jectory of a preceding vehicle by assuming that the vehicle is
following the mean travel time for each segment. If a lower and Then, we are assuming that such difference decays 20% every
upper probability thresholds are set, the lower and the upper prediction step and become zero after five steps, which is
trajectory boundaries can be obtained. An example is shown shown in Fig. 4. With this rule, we can generate the trajectory
in Fig. 3. As shown in Fig. 3, the travel time probability dis- based on both historical and instantaneous information and
tribution can be obtained at the end of the position prediction reduce the prediction error significantly (later shown in Fig. 7).
horizon (between A and B), which is also Gaussian, since it is
the sum of the segment travel times (all Gaussian distributions)
that the path covers. However, due to the nature of the MPC B. Vehicle Trajectory Prediction Based on
framework, a prediction of the position probability distribution Individual Driving Data
at a fixed time (between A and C) is required as the input rather In [10] and in our previous work [9], a Markov chain-based
than the time probability distribution at a fixed position. In the approach is used to predict the future position distribution
example, the travel time distribution between A and B needs to of a car. The approach uses historical driving data to train
be transformed into a position distribution between A and C. a Markov model. The individual driving data set includes
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WAN et al.: PROBABILISTIC ANTICIPATION AND CONTROL IN AUTONOMOUS CAR FOLLOWING 5

Fig. 5. Example of a position sample histogram using Monte Carlo sampling. Fig. 6. Prediction error samples of the two predictors versus velocity.

multiple trajectories’ GPS points with the 10-Hz updating C. Combination of Predictors
frequency. Each update point includes a time stamp, the Combining multiple predictors has the potential to improve
vehicle’s heading direction, its global coordinates, velocity and the prediction accuracy as shown in the literature [16], [17].
acceleration information. By counting historical occurrences, In Section III-A and Section III-B, we have proposed two
probabilities are assigned to the transition between two different predictors: predictor 1 is built on historical probe
states (i.e., preceding vehicle velocity) from consecutive time vehicles data, and predictor 2 is derived from individual
steps. The assumption is that the next step’s state only depends driving behavior data. The outputs of these two predictors
on the current step’s state and is independent from the previous are position distributions, both of which are assumed to be
steps. The element p(i, j ) in the transition matrix represents Gaussian. However, they are based on different data sources,
the probability of transitioning from the i th to the j th state. with different estimation conditions and assumptions.
Previous work in our group [9] only considered velocity as a One way to combine Gaussian distributions is to form a
state, but more precise predictions can be made if acceleration Gaussian mixture model (GMM) [23]. The probability density
is also included as a state. However, including acceleration will function of a GMM can be represented by
increase the dimension of the transition matrix by one, and

K
requires a much larger and richer data set. To circumvent this P(x|π, μ, ) = πi N (x|μi , i ) (15)
challenge, we define the Markov state as Si = [v i , αi ]T , i=1
where v i is the discretized velocity and αi is a categori-
where N represents a Gaussian distribution component,
cal variable describing whether the vehicle is accelerating,
K is the number of components, and μi , i , and πi are
decelerating, or cruising. That is, v i ∈ {0, 1, . . . , 15}m/s and
the mean, the covariance matrix, and the weight of the
αi ∈ {acceler ati ng, deceler ati ng, cr uisi ng}. In the training
i th component. The weights are such that
procedure, we label the training data by the three acceleration
categories. With these newly defined states, the Markov chain 
K
requires three times as much data as in [9] to be trained with, πi = 1. (16)
but does much better in predicting future states. The transition i=1
matrix, denoted by PNv is an Nv × Nv matrix, where Nv is We propose to assign the weight to each predictor based
the number of discretized states. on its performance. When a predictor is performing better,
Given the current state and the transition matrix, the posi- a higher weight should be assigned to it in the GMM.
tion distribution over a next finite horizon can be obtained Intuitively, predictor 1 is expected to predict better the stop
using a Monte Carlo sampling technique. Through Monte positions and durations close to an intersection, and when
Carlo sampling, a sequence of state realizations can be the traffic is heavy. Predictor 2 is expected to perform bet-
generated. For each velocity sample sequence, the position ter at midlinks, since the vehicle is less affected by other
at the i th future step can be calculated vehicles and its behavior is more likely to be dominated by
individual driver style. To test this hypothesis, we evaluate
1
n−1
r (n + k|k) = r (k) + (v(i + k|k)+ v(i +1+ k|k))t (i ) the prediction errors against the ground truth trajectory for
2 each predictor separately. We have selected seven ground
i=0
(14) truth trajectories as training data. In the future, when more
trajectories are collected as training data, the training results
to which a probability can be assigned. With sufficient number could improve. Figs. 6 and 7 show the samples and the
of samples [22], a distribution can be estimated from the average position prediction errors versus velocity. It could be
generated histogram; here, we assume a Gaussian distribution. concluded from Figs. 6 and 7 that, on average, predictor 1
Fig. 5 shows an example of the histogram of the preceding performs better than predictor 2 when the velocity is close
vehicle position at the end of a ten step horizon based on to zero. Therefore, we assign higher weights to predictor 1
high-frequency individualistic driving data (10 Hz). at low speeds. At higher speeds, predictor 2 performs better
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

TABLE I
PARAMETERS OF MPC

Fig. 7. Average position prediction error of the two predictors versus velocity.

Fig. 8. Weights of the predictors over velocity.

Fig. 9. Simulation case study. The horizontal red and green lines represent
red and green phases of the traffic signal at the intersection. The signal timing
is assumed to be fixed and known to the vehicle.
than predictor 1 and receives a larger weight. Suppose at
position x i , the average error of predictor 1 and predictor 2
is denoted by e1 and e2 , respectively (note that the average with the assumption that the preceding vehicle’s trajectory is
errors are obtained by utilizing instantaneous information); predetermined; a standard MPC is solved to determine the best
we calculate their weights π1 and π2 by trajectory that can be theoretically achieved.
|e2 |
π1 = IV. C ASE S TUDY AND R ESULTS
|e1 | + |e2 | The scenarios are designed under the following simulated
|e1 |
π2 = . (17) situation: a vehicle in autonomous cruise control mode is fol-
|e1 | + |e2 | lowing a preceding bus on Van Ness Avenue in San Francisco
The assigned weights are shown in Fig. 8. The average for which we had training data. The path consists of multiple
prediction errors shown in Fig. 7 are consistent with our intersections, of which traffic signal timings are available.
hypothesis. Note that the errors of Predictor 1 reduce a lot The simulated ego autonomous vehicle predicts the preceding
with the consideration of instantaneous velocity information. bus’s trajectory by using the combination of two predictors
Another reason that Predictor 1 is generally not as good discussed in Section III. The training data for Predictor 1 are
as Predictor 2 may be due to the fact that currently we only aggregated from six months of low-frequency bus data, which
have ground truth data from 15 vehicle trajectories. The lack was publicly available, while training data for Predictor 2 are
of training data may cause the inaccuracy of the prediction, from multiple high-frequency GPS ground truth data, which
especially when the vehicle is accelerating and decelerating. we collected on the same route in San Francisco as reported
In the next simulation study, we will implement the pro- in [13].
posed combined predictor to obtain probabilistic predictions To evaluate the effectiveness of the approach, multiple
of the motion of a preceding vehicle and solve a chance- high-frequency trajectory ground truth data are selected as
constrained MPC to find the best trajectory of the ego preceding vehicle’s trajectories and simulations are conducted
vehicle. To demonstrate the effectiveness of the proposed in different scenarios. At any time, the ego vehicle predicts
scheme, we use a classic proportional derivative (PD) control the preceding vehicle’s trajectory within a finite horizon and
car following approach as a benchmark to compare with. solves an MPC problem with chance constraints. Table I lists
All coefficients of this model are adopted from [24] where the parameters used in our MPC design.
the PD coefficients are obtained from real data and proved A scenario is shown where the preceding vehicle expe-
to be reasonable. Also a best case scenario is simulated rienced a stop near an intersection as shown in Fig. 9.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WAN et al.: PROBABILISTIC ANTICIPATION AND CONTROL IN AUTONOMOUS CAR FOLLOWING 7

Fig. 11. Gaps of each approach.

Fig. 10. Upper and lower bounds examples of predictors.

The vertical axis is the distance with respect to a reference

point. The intersection position is at 377 m. Different trajec-
tories using different predictors are plotted in Fig. 9. Note that
the horizontal red and green lines represent the red and green
phases of the traffic signal at the intersection.
At each time step, each predictor predicts the upper bound
and the lower bound of the position of the preceding vehicle at
the end of the predicting horizon. Two prediction boundaries Fig. 12. Average cost percentage of each approach.
for Predictors 1 and 2 are shown in Fig. 10. It can be shown
that the prediction ranges for Predictor 1 is generally narrower
than Predictor 2. The actual trajectory, which is shown in blue
line, is usually out of the prediction range, which indicates
that the prediction is not very accurate. However, thanks to the
characteristics of the MPC, the prediction is recalculated and
updated at each time step. Even though the prediction results
are not at high accuracy, the MPC still performs reasonably
well.
The results in Fig. 9 show that the ego vehicle trajectory
based on the combined predictors is relatively close to the
trajectory based on the MPC with perfect information and out
performs the baseline PD controller. Note that the strategy
is minimizing the sum of squares of accelerations while
maintaining a reasonable distance with respect to the preceding
Fig. 13. Jerk percentages of each approach.
vehicle. Although the trajectory based on MPC with perfect
information is further away from the preceding vehicle, it is
the one that minimizes our defined cost function. On the other probabilistic predictive approach outperforms the PD strategy.
hand, the PD is trying to maintain the minimum following dis- Furthermore, observe that the cost is reduced by combining the
tance, without considering acceleration and drivability. Fig. 11 two predictors, which demonstrates the benefits of assigning
shows the gaps of all approaches. Although the gap is smaller situational weights to these predictors.
for the PD approach, the driving comfort and fuel economy Note that the cost in (4) does not have a physical meaning.
perform worse than the proposed approaches. The zoomed- However, it reflects the improvement in driving comfort and
in figures in Fig. 9 show that before and after stop predictive fuel consumption.
approaches react faster and smoother than the PD control. Jerk, which is the rate of change in acceleration, is often
Fig. 12 compares the average costs, that are being minimized, used to indicate driving comfort. To better demonstrate the
of each predictor based on six simulation case studies. The aforementioned improvements, we show jerk percentages of
PD method is selected as the baseline, which is 100%. Note each approaches in Fig. 13, from where we can see that all
that the first column is the theoretical best case scenario. predictive approaches have lower average jerk values than the
The numerical values confirm our earlier observation that the baseline PD approach.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

TABLE II prediction of the preceding vehicle as an input and minimizes

VALUES OF V EHICLE M ODEL PARAMETERS a car following cost function with probabilistic intervehicle
constraints in a receding horizon manner. Two predictors are
introduced. Predictor 1 is based on historical probe vehicle
data and reflects the effect of recurrent traffic conditions, and
Predictor 2 is based on individual driver behavior data. The
probabilistic predictions of the two predictors are combined
in a Gaussian mixture model based on their prediction perfor-
mance in training data sets. Simulation results show that the
combined predictor achieves reduced car-following costs.

ACKNOWLEDGMENT
The authors would like to thank the support from
Dr. A. Luckow. They would also like to thank Mr. S. A. Fayazi,
who has collected, parsed, and preprocessed the probe bus data
and traffic signal timings for using in this paper.

R EFERENCES
[1] A. Vahidi and A. Eskandarian, “Research advances in intelligent colli-
sion avoidance and adaptive cruise control,” IEEE Trans. Intell. Transp.
Syst., vol. 4, no. 3, pp. 143–153, Sep. 2003.
[2] V. Milanés, S. E. Shladover, J. Spring, C. Nowakowski, H. Kawazoe,
and M. Nakamura, “Cooperative adaptive cruise control in real traffic
situations,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 1, pp. 296–305,
Feb. 2014.
[3] B. van Arem, C. J. G. van Driel, and R. Visser, “The impact of
Fig. 14. Fuel consumption percentages of each approach. cooperative adaptive cruise control on traffic-flow characteristics,” IEEE
Trans. Intell. Transp. Syst., vol. 7, no. 4, pp. 429–436, Dec. 2006.
[4] J. Ploeg, B. T. M. Scheepers, E. van Nunen, N. van de Wouw,
To demonstrate the fuel economy improvement, a fuel con- and H. Nijmeijer, “Design and experimental evaluation of cooperative
sumption model is needed. Here, we adopt a fuel consumption adaptive cruise control,” in Proc. 14th Int. IEEE Conf. Intell. Transp.
model from [25], where the authors sampled sufficient data Syst. (ITSC), Oct. 2011, pp. 260–265.
[5] N. Wan, A. Vahidi, and A. Luckow, “Optimal speed advisory for
from a passenger vehicle and fit into third-order polynomial connected vehicles in arterial roads and the impact on mixed traffic,”
curves that describe the relation between fuel consumption rate Transp. Res. C, Emerg. Technol., vol. 69, pp. 548–563, Aug. 2016.
and speed and acceleration. The model can be described as [6] B. Krahé and I. Fenske, “Predicting aggressive driving behavior: The
⎧ role of macho personality, age, and power of car,” Aggressive Behavior,
⎪
⎨α0 + α1 v + α2 v + α3 v
2 3 vol. 28, no. 1, pp. 21–29, 2002.
[7] A. Liu and D. Salvucci, “Modeling and prediction of human driver
ṁ f = + (β0 + β1 v + β2 v )a, a  0
2 (18) behavior,” in Proc. Int. Conf. HCI, 2001, pp. 1–5.
⎪
⎩ [8] C. Guo, N. Wan, S. Mita, and M. Yang, “Self-defensive coordinated
α0 , a<0 maneuvering of an intelligent vehicle platoon in mixed traffic,” in
Proc. 15th Int. IEEE Conf. Intell. Transp. Syst. (ITSC), Sep. 2012,
where v is the vehicle velocity, a is the vehicle acceleration, pp. 1726–1733.
and αi and βi are model parameters. In this paper, the ego [9] C. Zhang and A. Vahidi, “Predictive cruise control with probabilistic
vehicle is assumed to have the parameters from those in [25], constraints for eco driving,” in Proc. ASME Dyn. Syst. Control Conf.
Bath/ASME Symp. Fluid Power Motion Control, 2011, pp. 233–238.
the fuel consumption rate may not be accurate, but it is able to [10] K. McDonough, I. Kolmanovsky, D. Filev, S. Szwabowski, D. Yanakiev,
show the improvement introduced by predictors. The vehicle and J. Michelini, “Stochastic fuel efficient optimal control of vehicle
parameter values of the ego vehicle and the fuel consumption speed,” in Optimization and Optimal Control in Automotive Systems.
New York, NY, USA: Springer, 2014, pp. 147–162.
model can be found in Table II.
[11] A. Skabardonis, P. Varaiya, and K. Petty, “Measuring recurrent and
Fig. 14 shows the average fuel consumption percentages, nonrecurrent traffic congestion,” Transp. Res. Rec., J. Transp. Res.
where the PD control is again the baseline approach. Note that Board, vol. 1856, pp. 118–124, 2003.
the combination method has higher fuel consumption than the [12] N. Wan and A. Vahidi, “Probabilistic estimation of travel times in arterial
streets using sparse transit bus data,” in Proc. IEEE 17th Int. Conf. Intell.
two predictors. One explanation is that the MPC is not directly Transp. Syst. (ITSC), Oct. 2014, pp. 1292–1297.
minimizing the fuel use; therefore, in some cases, the fuel use [13] N. Wan, A. Vahidi, and A. Luckow, “Reconstructing maximum like-
is sacrificed to maintain desired distance. Another reason is lihood trajectory of probe vehicles between sparse updates,” Transp.
Res. C, Emerg. Technol., vol. 65, pp. 16–30, Apr. 2016.
that the sample size is too small in this paper. With more [14] J. M. Bates and C. W. Granger, “The combination of forecasts,” OR,
samples obtained, it is expected that the fuel use is at the vol. 20, no. 4, pp. 451–468, 1969.
same level when combining the predictors. [15] A. Timmermann, “Forecast combinations,” in Handbook of Economic
Forecasting, vol. 1, Kidlington, U.K.: Elsevier, 2006, pp. 135–196.
[16] T.-H. Lee, “Combining forecasts with many predictors,” Adv. Econ.
V. C ONCLUSION Forecast., pp. 149–172, 2011.
[17] Y. Hong and T.-H. Lee, “Inference on predictability of foreign exchange
In this paper, a predictive cruise control approach is rates via generalized spectrum and nonlinear time series models,” Rev.
proposed. The approach takes the position distribution Econ. Statist., vol. 85, no. 4, pp. 1048–1062, 2003.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WAN et al.: PROBABILISTIC ANTICIPATION AND CONTROL IN AUTONOMOUS CAR FOLLOWING 9

[18] L. Del Re, F. Allgöwer, L. Glielmo, C. Guardiola, and I. Kolmanovsky, Chen Zhang received the B.S. degree in mechanical
Eds., Automotive Model Predictive Control: Models, Methods and Appli- engineering from Chongqing University, Chongqing,
cations, vol. 402. New York, NY, USA: Springer, 2010. China, the M.S. degree from Shanghai Jiao Tong
[19] M. Cannon, B. Kouvaritakis, and X. Wu, “Probabilistic constrained University, Shanghai, China, in 2003 and 2006,
MPC for multiplicative and additive stochastic uncertainty,” IEEE Trans. respectively, and the Ph.D. degree in mechanical
Autom. Control, vol. 54, no. 7, pp. 1626–1632, Jul. 2009. engineering from Clemson University, Clemson, SC,
[20] N. Wan, “Estimation and control of traffic relying on vehicular connec- USA, in 2010.
tivity,” Ph.D. dissertation, Clemson Univ., San Diego, CA, USA, 2016. From 2011 to 2012, he was a Lead Control
[21] N. L. Johnson, S. Kotz, and N. Balakrishnan, Continuous Multivariate Systems Engineer with Navistar, Inc., Lisle, IL,
Distributions: Models and Applications, vols. 1–59. New York, NY, USA, for developing electrified school buses and
USA: Wiley, 2002. trucks. Since 2012, he has been a Research Engineer
[22] J. C. Spall, “Estimation via Markov chain Monte Carlo,” IEEE Control with the Research and Advanced Engineering Organization of Ford Motor
Syst., vol. 23, no. 2, pp. 34–45, Apr. 2003. Company, Dearborn, MI, USA. He has co-authored nine peer reviewed papers
[23] N. Wan, G. Gomes, A. Vahidi, and R. Horowitz, “Prediction on travel- and 22 U.S. patents and invention disclosures. His current research interests
time distribution for freeways using online expectation maximization include systems and controls of electrified vehicles and automated driving
algorithm,” in Proc. 93rd Annu. Meet. Transp. Res. Board, 2013, p. 14. vehicles.
[24] X. Fang, H. A. Pham, and M. Kobayashi, “PD controller for car-
following models based on real data,” in Proc. 1st Hum.-Centered
Transp. Simulation Conf., 2001, pp. 1–6.
[25] M. A. S. Kamal, M. Mukai, J. Murata, and T. Kawabe, “Ecological
vehicle control on roads with up-down slopes,” IEEE Trans. Intell.
Transp. Syst., vol. 12, no. 3, pp. 783–794, Sep. 2011.

Ardalan Vahidi (M’05) received the B.S. and

Nianfeng Wan received the B.S. degree in automa- M.Sc. degrees in civil engineering from Sharif Uni-
tion from Tianjin University, Tianjin, China, and the versity, Tehran, Iran, in 1996 and 1998, respec-
M.S. degree in control theory and engineering from tively, the M.Sc. degree in transportation safety from
Shanghai Jiao Tong University, Shanghai, China. George Washington University, Washington, DC,
He is currently pursuing the Ph.D. degree in mechan- USA, in 2002, and the Ph.D. degree in mechanical
ical engineering with Clemson University, Clemson, engineering from the University of Michigan, Ann
SC, USA. Arbor, MI, USA, in 2005.
He has been a Visiting Student Researcher with He is currently an Associate Professor with the
the University of California at Berkeley, Berkeley, Department of Mechanical Engineering, Clemson
CA, USA, and a Research Intern with the Toyota University, Clemson, SC, USA. He has been a
Institute of Technology, Nagoya, Japan. His current Visiting Scholar with the University of California at Berkeley, Berkeley, CA,
research interests include exploring machine learning and control technologies USA, and a Visiting Researcher with the BMW Group Technology Office,
to estimate and improve mobility, fuel efficiency, and safety of connected USA, from 2012 to 2013. His current research interests include the control
vehicles and the traffic. of vehicular and energy systems and connected vehicle technologies.