Chapter

6
Detection Fundamentals
As was noted in Chap. 1, the primary functions to be carried out by a radar sig-
nal processor are detection, tracking, and imaging. In this chapter, the concern
is detection. In radar, this means deciding whether a given radar measure-
ment is the result of an echo from a target, or simply represents the effects
of interference. If it is decided that the measurement indicates the presence
of a target, further processing is usually undertaken. This additional process-
ing might, for instance, take the form of tracking via precise range, angle, or
Doppler measurements.
Detection decisions can be applied to signals present at various stages of the
radar signal processing, from raw echoes to heavily preprocessed data such as
Doppler spectra or even synthetic aperture radar images. In the simplest case,
each range bin (fast-time sample) for each pulse can be individually tested to
decide if a target is present at the range corresponding to the range bin, and the
spatial angles corresponding to the antenna pointing direction for that pulse.
Since the number of range bins can be in the hundreds or even thousands, and
pulse repetition frequencies can range from a few kilohertz to tens or hundreds
of kilohertz, the radar can be making many thousands to literally millions of
detection decisions per second.
It was seen in Chap. 2 that both the interference and the echoes from com-
plex targets are best described by statistical signal models. Consequently, the
process of deciding whether or not a measurement represents the inuence
of a target or only interference is a problem in statistical hypothesis testing.
In this chapter, it will be shown how this basic decision strategy leads to
the concept of threshold testing as the most common detection logic in radar.
Performance curves will be derived for the most basic signal and interference
models.
Clutter (echoes from the ground) is sometimes interference and sometimes
the target. If one is trying to detect a moving vehicle, ground clutter, along
with noise and possibly jamming, is the interference; but if one is trying to
295
296 Chapter Six
image a region of the earth, this same terrain becomes the desired target and
only noise and jamming are the interference.
An excellent concise reference for modern detection theory is given by John-
son and Dudgeon (1993, Chap. 5). When greater depth is needed, another ex-
cellent modern reference with a digital signal processing point of view is the
work by Kay (1998). An important classical textbook in detection theory is the
book by Van Trees (1968), while Meyer and Mayer (1973) provide a classical
in-depth analysis and many detection curves specically for radar applications.
6.1 Radar Detection as Hypothesis Testing
For any radar measurement that is to be tested for the presence of a target, one
of two hypotheses can be assumed to be true
1. The measurement is the result of interference only.
2. The measurement is the combined result of interference and echoes from a
target.
The rst hypothesis is denoted as the null hypothesis H
0
and the second as
H
1
. The detection logic therefore must examine each radar measurement to be
tested and select one of the hypotheses as best accounting for that measure-
ment. If H
0
best accounts for the data, the system declares that a target was
not present at the range, angle, or Doppler coordinates of that measurement; if
H
1
best accounts for the data, the system declares that a target was present.

Because the signals are described statistically, the decision between the two
hypotheses is an exercise in statistical decision theory. A general approach to
this problem is described in many texts (e.g., Kay, 1998). The analysis starts
with a statistical description of the probability density function (pdf) that de-
scribes the measurement to be tested under each of the two hypotheses. If the
sample to be tested is denoted as y, the following two pdfs are required
p
y
( y|H
0
) = pdf of y given that a target was not present
p
y
( y|H
1
) = pdf of y given that a target was present
Thus, part of the detection problem is to develop models for these two pdfs. In
fact, analysis of radar performance is dependent on estimating these pdfs for
the system and scenario at hand. Furthermore, a good deal of the radar system
design problem is aimed at manipulating these two pdfs in order to obtain the
most favorable detection performance.

In some detection problems, a third hypothesis is allowed: dont know. Most radar systems,
however, force a choice between target present and target absent on each detection test.
Detection Fundamentals 297
More generally, detection will be based on N samples of data y
n
forming a
column vector y
y [ y
0
. . . y
N1
]
t
(6.1)
The N-dimensional joint pdfs p
y
(y|H
0
) and p
y
(y|H
1
) are then used.
Assuming the two pdfs are successfully modeled, the following probabilities
of interest can be dened
Probability of Detection, P
D
: The probability that a target is declared
(i.e., H
1
is chosen) when a target is in
fact present.
Probability of False Alarm, P
FA
: The probability that a target is declared
(i.e., H
1
is chosen) when a target is in
fact not present.
Probability of Miss, P
M
: The probability that a target is not
declared (i.e., H
0
is chosen) when a
target is in fact present.
Note that P
M
= 1 P
D
. Thus, P
D
and P
FA
sufce to specify all of the proba-
bilities of interest. As the latter two denitions imply, it is important to realize
that, because the problem is statistical, there will be a nite probability that
the decisions will be wrong.

6.1.1 The Neyman-Pearson detection rule

The next step in making a decision is to decide what the rule will be for deciding
what constitutes an optimal choice between our two hypotheses. This is a rich
eld. The Bayes optimization criterion assigns a cost or risk to each of the
four possible combinations of actual state (target present or not) and decision
(select H
0
or H
1
) (Johnson and Dudgeon, 1993; Kay, 1998). In radar, it is more
common to use a special case of the Bayes criterion called the Neyman-Pearson
criterion. Under this criterion, the decision process is designed to maximize the
probability of detection P
D
under the constraint that the probability of false
alarm P
FA
does not exceed a set constant. The achievable combinations of P
D
and P
FA
are affected by the quality of the radar system and signal processor
design. However, it will be seen that for a xed system design, increasing P
D
implies increasing P
FA
as well. The radar systemdesigner will generally decide
what rate of false alarms can be tolerated based on the implications of acting
on a false alarm, which may include using radar resources to start a track on
a nonexistent target, or in extreme cases even ring a weapon! Recalling that

A fourth probability can be dened, that of choosing H

0
and thus declaring a target not present
when in fact the test sample is due to interference only. This probability, equal to 1 P
FA
, is not
normally of direct interest.
298 Chapter Six
the radar may make tens or hundreds of thousands, even millions of detection
decisions per second, values of P
FA
must generally be quite low. Values in the
range of 10
4
to 10
8
are common, and yet may still lead to false alarms every
few seconds or minutes. Higher-level logic implemented in downstream data
processing, beyond the scope of this book, is often used to reduce the number
or impact of false alarms.
Each vector of measured data values y can be considered to be a point in
N-dimensional space. To have a complete decision rule, each point in that space
(each combination of N measured data values) must be assigned to one of the
two allowed decisions, H
0
(target absent) or H
1
(target present). Then, when
the radar measures a particular set of data values (observation) y

, the system
declares either target absent or target present based on the preexisting
assignment of y

to either H
0
or H
1
. Denote the set of all observations y for
which H
1
will be chosen as the region
1
. Note that
1
is not necessarily a
connected region. General expressions can now be written for the probabilities
of detection and false alarm as integrals of the joint pdfs over the region
1
in
a N-dimensional space
P
D
=
_

1
p
y
(y|H
1
) dy
(6.2)
P
FA
=
_

1
p
y
(y|H
0
) dy
Because probability density functions are nonnegative, Eq. (6.2) proves a
claim made earlier, namely that P
D
and P
FA
must rise or fall together. As the
region
1
grows to include more of the possible observations y, either integral
encompasses more of the N-dimensional space and therefore integrates more
of the nonnegative pdf. The opposite is true if
1
shrinks. That is, as
1
grows
or shrinks, so must both P
D
and P
FA
.

In order to increase detection probability,

the false alarm probability must be allowed to increase as well. Loosely speak-
ing, to achieve a good balance of performance, the points that contribute more
probability mass to P
D
than to P
FA
are assigned to
1
. If the system can be
designed so that p
y
(y|H
0
) and p
y
(y|H
1
) are as disjoint as possible, this task
becomes easier and more effective. This point will be revisited later.
6.1.2 The likelihood ratio test
The Neyman-Pearson criterion is motivated by the goal of obtaining the best
possible detection performance while guaranteeing that the false alarm proba-
bility does not exceed some tolerable value. Thus, the Neyman-Pearson decision
rule is to
choose
1
such that P
D
is maximized, subject to P
FA
(6.3)

The exception occurs if points are added to or subtracted from

1
for which p
y
(y|H
0
), p
y
(y|H
1
),
or both are zero, in which case the corresponding probability is unchanged.
Detection Fundamentals 299
where is the maximum allowable false alarm probability.

This optimization
problemis solved by the method of Lagrange multipliers. Construct the function
F P
D
+( P
FA
) (6.4)
To nd the optimum solution, maximize F and then choose to satisfy the
constraint criterion P
FA
= . Substituting Eq. (6.2) into Eq. (6.4)
F =
_

1
p
y
(y|H
1
) dy +
__

1
p
y
(y|H
0
) dy
_
= +
_

1
{ p
y
(y|H
1
) +p
y
(y|H
0
)} dy (6.5)
Remember that the design variable here is the choice of the region
1
. The
rst term in the second line of Eq. (6.5) does not depend on
1
, so F is maxi-
mized by maximizing the value of the integral over
1
. Since could be nega-
tive, the integrand can be either positive or negative, depending on the values
of and the relative values of p
y
(y|H
0
) and p
y
(y|H
1
). The integral is there-
fore maximized by including in
1
all the points, and only the points, in the
N-dimensional space for which p
y
(y|H
1
)+p
y
(y|H
0
) > 0, that is,
1
is all points
y for which p
y
(y|H
1
) > p
y
(y|H
0
). This leads directly to the decision rule
p
y
(y|H
1
)
p
y
(y|H
0
)
H
1
>
<
H
0
(6.6)
Equation (6.6) is known as the likelihood ratio test (LRT). Although derived
from the point of view of determining what values of y should be assigned to
decision region
1
, it in fact allows one to skip over explicit determination of
1
and gives a rule for optimally guessing, under the Neyman-Pearson criterion,
whether a target is present or not based directly on the observed data y and
a threshold (which must still be computed). This equation states that the
ratio of the two pdfs, each evaluated for the particular observed data y, should
be compared to a threshold. If that likelihood ratio exceeds the threshold,
choose hypothesis H
1
, i.e., declare a target to be present. If it does not exceed
the threshold, choose H
0
and declare that a target is not present. Under the
Neyman-Pearson optimization criterion, the probability of a false alarmcannot
exceed the original design value P
FA
. Note again that models of p
y
(y|H
0
) and
p
y
(y|H
1
) are required in order to carry out the LRT. Finally, realize that in
computing the LRT, the data processing operations to be carried out on the
observed data y are being specied. What exactly the required operations are
depends on the particular pdfs.
The LRT test is as ubiquitous in detection theory and statistical hypothesis
testing as is the Fourier transform in signal ltering and analysis. It arises as
the solution to the hypothesis testing problem under several different decision

Some subtleties that can arise if the pdfs are noncontinuous are being ignored. See the book by
Johnson and Dudgeon (1993) for additional detail.
300 Chapter Six
criteria, such as the Bayes minimum cost criterion, or maximization of the
probability of a correct decision. Substantial additional detail is provided by
Johnson and Dudgeon (1993) and Kay (1998). As a convenient and common
shorthand, it is convenient to express the LRT in the following notation
(y)
H
1
>
<
H
0
(6.7)
From Eq. (6.6), (y) = p
y
(y|H
1
)/p
y
(y|H
0
) and = .
Because the decision depends only on whether the LRT exceeds the threshold
or not, any monotone increasing

operation can be performed on both sides of

Eq. (6.7) without affecting the values of observed data ythat cause the threshold
to be exceeded, and therefore without affecting the performance (P
D
and P
FA
). A
well-chosen transformation can sometimes greatly simplify the computations
required to actually carry out the LRT. Most common is to take the natural
logarithm of both sides of Eq. (6.7) to obtain the log likelihood ratio test
ln(y)
H
1
>
<
H
0
ln (6.8)
To make these procedures clearer, consider what is perhaps the simplest
example, detection of the presence or absence of a constant in zero-mean Gaus-
sian noise of variance
2
. Let wbe a vector of i.i.d. zero mean Gaussian random
variables. When the constant is absent (hypothesis H
0
) the data vector y = w
follows an N-dimensional normal distribution with a scaled identity covariance
matrix. When the constant is present (hypothesis H
1
), y = m+ w = m1
N
+ w
and the distribution is simply shifted to a nonzero, positive mean

H
0
: y N(0
N
,
2
I
N
)
(6.9)
H
1
: y N(m1
N
,
2
I
N
)
where m > 0 and 0
N
, 1
N
, and I
N
are respectively a vector of N zeros, a vector
of N ones, and the identity matrix of order N. The model of the required pdfs
is therefore
p(y|H
0
) =
N1

n=0
1
_
2
2
exp
_

1
2
_
y
n

_
2
_
(6.10)
p(y|H
1
) =
N1

n=0
1
_
2
2
exp
_

1
2
_
y
n
m

_
2
_

A monotone decreasing operation would simply invert the sense of the threshold test.

All of the following development is fairly easily generalized for the case when m is negative or
is of unknown sign. It will be seen later that radar detection generally involves working with the
magnitude of the signal, thus it is sufcient to work with a positive value of m.
Detection Fundamentals 301
The likelihood ratio (y) and the log-likelihood ratio can be directly computed
from Eq. (6.10)
(y) =

N1
n=0
exp
_

1
2
_
y
n
m

_
2
_

N1
n=0
exp
_

1
2
_
y
n

_
2
_ (6.11)
ln(y) =
N1

n=0
_

1
2
_
y
n
m

_
2
+
1
2
_
y
n

_
2
_
=
1

2
N1

n=0
my
n

1
2
2
N1

n=0
m
2
(6.12)
Because of its simpler form, the log likelihood ratio will be used. Substituting
Eq. (6.12) into Eq. (6.8) and rearranging gives the decision rule
N1

n0
y
n
H
1
>
<
H
0

2
m
ln() +
Nm
2
(6.13)
Note that the right hand side of the equation consists only of constants, though
not all are yet known. Equation (6.13) thus species that the available data
samples y
n
be integrated (summed) and the integrated data compared to a
threshold. This integration is an example of how the LRT species the data
processing to be performed on the measurements. Note also that Eq. (6.13)
does not require specically evaluating the pdfs, let alone determining what
exactly is the region in N-space comprising
1
or whether the observation y
is in it.
In many cases of interest, the specic form of the log-likelihood ratio can be
further rearranged to isolate on the left hand side of the equation only those
terms explicitly including the data samples y
n
, moving all other constants to
the right hand side. Equation (6.13) is such a rearrangement of Eq. (6.12). The
term

y
n
is called a sufcient statistic for this problem, and is denoted by
(y). The sufcient statistic, if it exists, is a function of the data y that has the
property that the likelihood ratio (or log-likelihood ratio) can be written as a
function of (y), i.e., the data appear in the likelihood ratio only through (y)
(Van Trees, 1968). This means that in making a decision that is optimal under
the Neyman-Pearson criterion (and the many others that lead to the LRT),
knowing the sufcient statistic (y) is as good as knowing the actual data y.
In particular, the decision criterion in Eq. (6.8) can be expressed as
(y)
H
1
>
<
H
0
T (6.14)
302 Chapter Six
Note that Eq. (6.13) is in the form of Eq. (6.14) with (y) =

y
n
and T =
(
2
/m) ln() +(Nm/2).
The idea of a sufcient statistic is quite rich. For example, it can be inter-
preted as a geometric coordinate transformation chosen to place all of the useful
information in the rst coordinate (Van Trees, 1968). Procedures for verifying
that a statistic (a function of the data y) is sufcient are given by Kay (1993),
as is the Neyman-Fisher Factorization theorem for identifying sufcient statis-
tics. Detailed consideration of the properties of sufcient statistics is beyond the
scope of this text; the reader is referred to the previous references for greater
depth.
The specic value of the threshold = that will ensure that P
FA
=
as desired has not yet been found. The original expression for P
FA
was given in
Eq. (6.2), but this is not very useful, since its evaluation requires the
N-dimensional joint pdf of y and an explicit denition of the region
1
, which
have been dened only implicitly as the points in N-space for which the LRT
exceeds the still-unknown threshold. Since they are functions of the random
data y, and are also random variables and thus have their own probability
density functions. Because of the similarity of the sufcient statistic and the log
likelihood ratio in this problem, only and need be considered. An alternate
approach to computing the LRT threshold is thus to express P
FA
in terms of
or , and then solve those expressions for , or equivalently for T. The required
expressions are
P
FA
=
_
+
=
p

(|H
0
) d = (6.15)
or P
FA
=
_
+
T
p

(|H
0
) d = (6.16)
As one would expect, the result depends only on the pdf of the likelihood ratio (if
using Eq. (6.15)) or the sufcient statistic (if using Eq. (6.16)) when a target is
not present. Given a specic model of that pdf, a specic value can be computed
for (equivalently, ) or T.
To illustrate these results, continue the constant in Gaussian noise example
by nding the threshold and then evaluating the performance, working with
the sufcient statistic (y). In this case, (y) is the sum of the individual data
samples y
n
. Under hypothesis H
0
(no target), the samples are independent and
identically distributed (i.i.d) N(0,
2
). It follows that N(0, N
2
). A false
alarm occurs anytime > T, so
= P
FA
=
_
+
T
p

(|H
0
) d
=
_
+
T
1
_
2N
2
e

2
2N
2
d (6.17)
Detection Fundamentals 303
Equation (6.17) is the integral of a Gaussian pdf, so the error function erf(x)
will appear in the solution. The standard denition is (Abramowitz and Stegun,
1972)

erf (x)
2

_
x
0
e
t
2
dt (6.18)
Also dene the complementary error function erfc(x) corresponding to erf(x)
erfc(x)
2

_
+
x
e
t
2
dt = 1 erf (x) (6.19)
One will generally be interested in nding the value of x that results in a
certain value of erf(x) or erfc(x); thus the inverse error and complementary
error functions, denoted by erf
1
(z) and erfc
1
(z), respectively, are of interest.
It follows from Eq. (6.19) that the two are related by erfc
1
(z) = erf
1
(1 z).

With the change of variables t = /

_
2N
2
, Eq. (6.17) can be written as
= P
FA
=
1

_
+
T /

2N
2
e
t
2
dt
=
1
2
_
1 erf
_
T
_
2N
2
__
(6.20)
Finally, Eq. (6.20) can be solved to obtain the threshold T in terms of the
tabulated inverse error function
T =
_
2N
2
erf
1
(1 2P
FA
) (6.21)
Equations (6.20) and (6.21) show how to compute P
FA
given T and vice-versa.
All of the information needed to carry out the LRT in its sufcient statistic
form of Eq. (6.14) is now available. (y) is just the sum of the data samples,
while the threshold T can be computed from the number N of samples, the
variance
2
of the noise, which is assumed known, and the desired false alarm
probability P
FA
.
The performance of this detector is evaluated by constructing a receiver op-
erating characteristic (ROC) curve. There are four interrelated variables of in-
terest: P
D
, P
FA
, the noise power
2
, and the constant m whose presence is to
be detected. The latter two are characteristics of the given signals, while P
FA
is
generally xed as part of the system specications at whatever level is deemed
tolerable. Thus it is necessary only to determine P
D
. The approach is identical
to that used for determining P
FA
: determine the probability density function of

The denitions of Eqs. (6.18) and (6.19) are the same as those used in MATLAB

The erf
1
() function will often be used here, even when erfc
1
() gives a slightly more compact
expression because of the wider availability of erf
1
() functions than erfc
1
() in MATLAB

and
similar computational software packages.
304 Chapter Six
the sufcient statistic under the hypothesis H
0
and integrate the area under
it from the threshold to +.
Continuing the example, note that the only change under hypothesis H
1
is
that the individual data samples y
n
now each have mean m, so their sum has
mean Nm. Thus, (y) N(Nm, N
2
) and
P
D
=
_
+
T
p

(|H
1
) d
=
_
+
T
1
_
2N
2
e
(Nm)
2
2N
2
d (6.22)
Again applying the denition of the error function in Eq. (6.18) leads to
P
D
=
1
2
_
1 erf
_
T Nm
_
2N
2
__
(6.23)
Since the primary concern is the relationship between the performance metrics
P
D
and P
FA
, Eq. (6.21) can be used in Eq. (6.23) to eliminate the threshold T
and arrive at
P
D
=
1
2
_
1 erf
_
erf
1
(1 2P
FA
)

Nm
_
2
2
__
=
1
2
erfc
_
erfc
1
(2P
FA
)

Nm
_
2
2
_
(6.24)
Nm is considered to be the signal component of interest (since the goal is to
detect its presence) in the sufcient statistic (y) (the sum of the individual
data samples y
n
). Since Nm is treated as a voltage, the corresponding power
is (Nm)
2
. The power of the noise component of (y) is N
2
. Thus, the term
m

N/ is the square root of the signal-to-noise ratio for this problem, and
Eq. (6.24) can be rewritten as
P
D
=
1
2
_
1 erf
_
erf
1
(1 2P
FA
)
_
/2
__
=
1
2
erfc
_
erfc
1
(2P
FA
)
_
/2
_
(6.25)
Figure 6.1 illustrates how the detection and false alarm probabilities follow
from the pdfs under the two hypotheses and the threshold, and how their rel-
ative values depend on the relation between the two pdfs. Two Gaussian pdfs
with variance equal to one are shown. The leftmost has a zero mean, while the
rightmost has a mean of 1. The left pdf is N(0, 1), while the right pdf is N(1, 1).
With N
2
= 1 and m = 1/N, these t the model of Example 1.3. P
D
and P
FA
are the areas under the right and left pdfs, respectively, from the threshold
Detection Fundamentals 305
Figure 6.1 Gaussian probability density functions of the suf-
cient statistic when N
2
= 1 under hypothesis H
0
(left) and
H
1
(right).
(shown as a vertical line, in this case at about = 1.5) to +. The receiver
design thus consists of adjusting the position of the threshold until the black
area equals the acceptable false alarm probability. The detection probability is
then the gray area (which includes the black area as well). This gure again
makes it clear that P
D
and P
FA
must increase or decrease together as the
threshold moves left or right. The achievable combinations of P
D
and P
FA
are
determined by the degree to which the two distributions overlap.
Figure 6.2 illustrates the ROC for this problem with the SNR as a param-
eter. Part a of the gure plots the ROC using linear scales for both P
FA
and
P
D
. Several features are worth noting. First, P
FA
= P
D
when = 0 (implying
m = 0). This is to be expected since in that case, the pdf of (y) is the same
under either hypothesis. For a given P
FA
and > 0, P
D
increases as the SNR
increases, a result which should also be intuitively satisfying. Finally, note how
abrupt the transition between near-zero and near-unity detection probabilities
becomes as the SNRincreases. This is a little misleading, since radars normally
operate with very low values of P
FA
; depending on the type of system, P
FA
is
typically no higher than 10
3
and very often is in the range of 10
6
to 10
8
.
Figure 6.2b plots the same data on a logarithmic scale for P
FA
, which better
reveals the characteristics of the ROC for false alarm probabilities of interest
in radar signal processing.
If the achievable combinations of P
D
and P
FA
do not meet the performance
specications, what can be done? Consideration of Fig. 6.1 suggests two an-
swers. First, for a given P
FA
, P
D
can be increased by causing the two pdfs to
move further apart when a target is present. That is, the presence of a target
must cause a larger shift in the mean m of the distribution of the sufcient
statistic. It was shown that mis proportional to the signal-to-noise ratio. Thus,
one way to improve the detection/false alarm tradeoff is to increase the SNR.
306 Chapter Six
Figure 6.2 Receiver operating characteristic for the Gaussian
example. (a) Displayed on a linear P
FA
scale. (b) Displayed on
a logarithmic P
FA
scale.
This conclusion is borne out in Fig. 6.2. Figure 6.3a illustrates the effect of
increased SNR on the pdfs and performance probabilities under the two hy-
potheses using the same threshold as Fig. 6.1.
The second way to improve the performance tradeoff is to reduce the over-
lap of the pdfs by reducing their variance. Reducing the noise power
2
will
reduce the variance of both pdfs, leading to the situation shown in Fig. 6.3b
Detection Fundamentals 307
Figure 6.3 Two ways to modify the pdfs of Fig. 6.1 to improve
the tradeoff between detection and false alarms. (a) Increasing
the system SNR. (b) Reducing the noise power.
(where the area corresponding to P
FA
is too small to be seen) and again im-
proving performance. As with the rst technique of increasing m, reducing
2
again constitutes increasing the signal-to-noise ratio. Thus, consistent with
Eq. (6.25), improving the tradeoff between P
D
and P
FA
requires increasing the
SNR . This is a fundamental result that will arise repeatedly.
Radar systems are designed to achieve specied values of P
D
and P
FA
sub-
ject to various conditions, such as specied ranges, target types, interference
environments, and so forth. The designer can work with antenna design, trans-
mitter power, waveform design, and signal processing techniques, all within
cost and form factor constraints. The job of the designer is therefore to develop
308 Chapter Six
a radar system design which ultimately results in a pair of target absent and
target present pdfs at the point of detection with a small enough overlap to
allow the desired P
D
and P
FA
to be achieved. If the design does not do this, the
designer must redesign one or more of these elements to reduce the variance
of the pdfs, shift them further apart, or both until the desired performance is
obtained. Thus, a signicant goal of radar system design is controlling the two
pdfs analogous to those in Fig. 6.1, or equivalently, maximizing the SNR.
6.2 Threshold Detection in Coherent Systems
The Gaussian problemconsidered so far is useful to introduce and explain all of
the major elements of Neyman-Pearson detection, such as the likelihood ratio
test, probabilities of detection and false alarm, receiver operating character-
istics, and the major design tradeoffs that follow. Furthermore, the problem
seems radar-like: under one hypothesis, only Gaussian noise is observed;
under the other, a constant was added to the noise, which could be interpreted
as the echoes from a steady target. Unfortunately, this example is not a good
model for any radar detection problem due to at least three major limitations.
First, coherent radar systems that produce complex-valued measurements
are of most interest. The approach, demonstrated so far only for real-valued
data, must therefore be extended to the complex case; and in doing so, the
complex signals and interference measurements must be modeled.
Second, there are unknown parameters. The analysis so far has implicitly as-
sumed that such signal parameters as the noise variance and target amplitude
are known, when in fact these are not known a priori but must be estimated
if needed. To complicate matters further, some parameters are linked. Speci-
cally in radar, the (unknown) echo amplitude varies with the (unknown) echo
arrival time according to the appropriate version of the radar range equation.
Thus, the LRT must be generalized to develop a technique that can work when
some signal parameters are unknown. This extension will introduce the idea
of threshold detection.
Finally, as seen in Chap. 2, there are a number of established models for radar
signal phenomenology that must be incorporated. In particular, it is necessary
to account for uctuating targets, i.e., statistical variations in the amplitude of
the target components of the measured data when a target is present. Further-
more, while the Gaussian pdf remains a good model for noise, in many problems
the dominant interference is clutter, which may have one of the distinctly non-
Gaussian pdfs discussed in Chap. 2.
The next subsections begin addressing these shortcomings by extending the
LRT to coherent systems.
6.2.1 The Gaussian case for coherent receivers
An appropriate model for noise at the output of a coherent receiver was de-
veloped in Chap. 2. It was shown there that if the noise in the system prior
Detection Fundamentals 309
to quadrature signal generation is a zero mean, white Gaussian process with
power
2
=kT,

the I and Qchannels will each contain independent, identically

distributed zero-mean white Gaussian processes with power kT/2 =
2
/2. That
is, the noise power splits evenly but independently between the two channels. A
complex noise process for which the real and imaginary parts are i.i.d. is called
a circular random process (Dudgeon and Johnson, 1993). The expression for
the joint pdf of N complex samples of the circular Gaussian random process is
p
y
(y) =
1
det{S
y
}
exp
_
(y m)
H
S
1
y
(y m)
_
(6.26)
where m is the N 1 vector mean of the N 1 vector signal y = m+ w, S
y
is
the N N covariance matrix of y
S
y
= E{yy
H
} mm
H
(6.27)
and H is the Hermitian (conjugate transpose) operator. In most cases the noise
samples are i.i.d. so that S
y
=
2
I
N
, which in turn means that det{S
y
} =

2N
. Treatment of the case where the noise samples are not equal-variance,
and the colored noise case where S
y
is not even diagonal, is beyond the scope of
this text. The reader is referred to the books by Dudgeon and Johnson (1993)
and Kay (1998) for these more complex situations.
Equation (6.26) now reduces to
p
y
(y) =
1

2N
exp
_

2
(y m)
H
(y m)
_
(6.28)
Further simplications occur when all of the means under H
1
are identical so
that m = m1
N
, where m can now be complex-valued. In this case Eq. (6.28)
reduces slightly further to
p
y
(y) =
1

2N
exp
_

2
(y m1
N
)
H
(y m1
N
)
_
(6.29)
The LRT test for the coherent version of the previous Gaussian example can
be obtained by repeating the steps in the example of Eqs. (6.22) through (6.25)
using the pdf of Eq. (6.28), with m = 0
N
under hypothesis H
0
and m = 0
N
under H
1
. The log likelihood ratio is
ln =
1

2
{2Re[m
H
y] m
H
m}
=
2

2
Re
_
N1

n=0
m

y
n
_

2
N|m|
2
(6.30)

Here T refers to receiver temperature, not the detection threshold. Which meaning of T is
intended should be clear from context throughout this chapter.
310 Chapter Six
where the second line of Eq. (6.30) applies only to the case where the means
are identical (m= m1
N
).
Some interpretation of Eq. (6.30) is in order. The term m
H
y is the dot prod-
uct of the complex vectors m and y. As seen in Chap. 1, this dot product rep-
resents an FIR ltering operation, evaluated at the particular instant when
the equivalent impulse response m
H
and the data vector y completely overlap.
Furthermore, since the impulse response of the lter is identical to the signal
to be detected under hypothesis H
1
, namely the presence of the mean m in
the data, it is a matched lter, a concept discussed in detail in previous chap-
ters. Restated, the impulse response is directly related (matched) to the signal
component whose presence one seeks to detect. In this example, the signal is
just the vector of means, but the same reasoning applies if the elements of m
are the samples of a modulated waveform or any other function of interest.
The second term in ln, which is the complex dot product of m with itself,
expands to m
H
m =

N1
n=0
|m
n
|
2
, which is just the energy in m. Denote this
quantity as E. In the equal means case, this is just E = N|m|
2
.
Finally, note the Re{} operator applied to the matched lter output m
H
y.
Because m and y are complex, one might be concerned that the dot product
could be purely imaginary or nearly so, such that Re{m
H
y } 0. The measured
data y would then have little or no effect on the threshold test. For this example,
under hypothesis H
0
m= 0
N
and the Re{} operator is inconsequential. Under
hypothesis H
1
, each element of m is a complex number of the form m
n
e
j
n
. If
the target really is present, i.e., H
1
is in fact true, the elements of the measured
data vector y = m+ w will be of the form m
n
e
j
n
+w
n
, where w
n
is a zero mean
complex Gaussian noise sample. It then follows that
m
H
y = m
H
m+m
H
w
=
N1

n=0
|m
n
|
2
+
N1

n=0
w
n
m
n
e
j
n
(6.31)
The rst term is again the energy E in the signal m; this is real-valued and
therefore unaffected by the Re{} operator. The second term is simply weighted
and integrated noise samples. While the phase of this noise component and
therefore the effect of the Re{} operator is random, the sum will tend to zero as
more samples are integrated.
It is evident by inspection of Eq. (6.30) that the sufcient statistic is now
Re{m
H
y}. Expressing the LRT in its sufcient statistic form for the complex
case
= Re{m
H
y}
H
1
>
<
H
0

2
2
ln() +
E
2
= T (6.32)
Note that if m = m1
N
, the term Re{m
H
y} = m

y
n
and Eq. (6.32) reduces
to Eq. (6.13) again. To complete consideration of the complex Gaussian case,
Detection Fundamentals 311
its performance, i.e., P
D
and P
FA
, must be determined. The sufcient statistic
= Re{m
H
y} = Re{

n
y
n
} is just a sum of Gaussian random variables, and
so will also be Gaussian. To determine the performance of the coherent detector,
the pdf of must be determined under each hypothesis. To do so it is useful to
rst consider the quantity z = m
H
y, which will be a complex Gaussian. First
suppose hypothesis H
0
is true. In this case the {y
n
} are zero mean, and therefore
so is z. Because the {y
n
} are independent, the variance of z is just the sum of
the variances of the individual weighted samples
var(z) =
N1

n=0
var(m

n
y
n
)
=
N1

n=0
|m
n
|
2

2
= E
2
(6.33)
Thus, under hypothesis H
0
, z N(0,E
2
). Similarly, under hypothesis
H
1
y = m + w and z N(E,E
2
). Note that the mean of z is real in both
cases. The power of the complex Gaussian noise splits evenly between the
real and imaginary parts of z (Kay, 1998). Since = Re{m
H
y}, it follows that
N(0,E
2
/2) under H
0
and N(E,E
2
/2) under H
1
. Following the pro-
cedure used in Example 7.2, it can be seen that
P
FA
=
1
2
_
1 erf
_
T
_

2
E
__
(6.34)
Repeating the development of Eqs. (6.22) to (6.24) gives the probability of de-
tection
P
D
=
1
2
_
1 erf
_
erf
1
(1 2P
FA
)

2
__
=
1
2
erfc
_
erfc
1
(2P
FA
)

2
_
(6.35)
Note again that the last term in Eq. (6.35) is the square root of the energy in
the signal m, divided this time by the noise power
2
, i.e., the signal-to-noise
ratio. Thus Eq. (6.35) can be written as
P
D
=
1
2
[1 erf {erf
1
(1 2P
FA
)

}]
=
1
2
erfc {erfc
1
(2P
FA
)

} (6.36)
Finally, in the equal means case when m= m1
N
, Eq. (6.35) is similar (but not
identical) to Eq. (6.24). The coherent case includes the term
_
Nm
2
/
2
instead
of
_
Nm
2
/2
2
because all of the signal energy competes with only half of the
noise power.
312 Chapter Six
Figure 6.4 shows the receiver operating characteristic for this example. It
is identical in general form to the real-valued case of Fig. 6.2, however, for a
given signal-to-noise ratio the performance is better because in the coherent
receiver, the signal competes with only half the noise power. For example, in
this coherent case, a signal-to-noise ratio of 13 dB produces P
D
= 0.94 at a
P
FA
= 10
6
. Figure 6.2 shows that in the real case the same and P
FA
produce
a P
D
of just under 0.39.
6.2.2 Unknown parameters and threshold detection
In general, perfect knowledge of each of the parameters of the pdfs p

(|H
0
)
and p

(|H
1
) is required to carry out the LRT, which usually means having
perfect knowledge of p
y
(y|H
0
) and p
y
(y|H
1
). In the Gaussian example, for
instance, it was assumed that the expected signal y is known under the various
hypotheses, as well as the noise sample variance
2
. This is not the case in the
real world, where the pdfs that form the likelihood ratio may depend on one or
more parameters that are unknown. Depending on our degree of knowledge,
three cases arise
1. is a random variable with a known probability density function.
2. is a random variable with an unknown probability density function.
3. is deterministic but unknown.
Different techniques are used to handle each of these cases. The rst is the most
important, because it has the greatest effect on the structure of the optimal
Neyman-Pearson detector.
To illustrate the approach for handling a random parameter with a known
pdf, consider yet again the complex Gaussian case. The optimal detector
Figure 6.4 Performance of the coherent receiver on the Gaussian
example.
Detection Fundamentals 313
implemented a matched lter operation m
H
y, followed by the Re{} operator.
The success of the matched lter structure depended on knowing exactly the
constant component of y = m + w under hypothesis H
1
, so that the lter
coefcients could be set equal to m and the lter output would be real-valued.
Recall that, when applied to radar, y under H
0
is considered to consist only of
samples wof receiver noise, and under H
1
to consist of noisy samples m+ wof
the echoes from a radar target over multiple pulses, or alternately successive
fast-time samples of the waveform of one pulse echo from a target.
Claiming perfect knowledge of mimplies knowing the range to the target very
precisely, since a variation in one-way range of only /4 causes the received echo
phase to completely reverse, i.e., to change by 180. At microwave frequencies,
this is typically only 15 to 30 cm(at L band to UHF) to a fraction of a centimeter
(at millimeter wave frequencies). Because this precisionis usually unrealistic, it
is more reasonable to assume mis known only to within a phase factor exp( j),
where the phase angle is considered to be a random variable distributed
uniformly over (0, 2 ] and independent of the random variables {m
n
}. In other
words, m = mexp( j), where m is known exactly but is a random phase.
(Note that the energy in m is the same as that in m, that is, m
H
m = m
H
m.)
This unknown phase assumption cannot usually be avoided in radar. What is
its effect on the optimal detector and its performance?
The goal remains to carry out the LRT, so it is necessary to return to its basic
denition of Eq. (6.6) and determine p
y
(y|H
0
) and p
y
(y|H
1
), both of which
now presumably depend on ,

and use the technique known as the Bayesian

approach for random parameters with known pdfs (Kay, 1998). Specically,
compute the pdf under H
i
by separately averaging the conditional pdfs p
y
(y|H
i
)
over
p
y
(y|H
i
) =
_
p
y
(y|H
i
, ) p

() d i = 0, 1 (6.37)
The unconditional pdfs p
y
(y|H
i
) are then used to dene the likelihood ratio in
the usual way.
As an example of the Bayesian approach for random parameters, consider
again the complex Gaussian case, but now with an unknown phase in the data,
m= mexp( j). The conditional pdf of the observations y becomes, under each
of the two hypotheses,
p
y
(y|H
0
, ) =
1

2N
exp
_

2
y
H
y
_
(6.38)
p
y
(y|H
1
, ) =
1

2N
exp
_

2
(y me
j
)
H
(y me
j
)
_

An alternative approach called the generalized likelihood ratio test (GLRT), in which the un-
known parameter(s) are replaced by their maximum likelihood estimates, is discussed in many
detection theory texts (e.g., Kay, 1998).
314 Chapter Six
Expanding the exponent in Eq. (6.38) gives
p
y
(y|H
1
, ) =
1

2N
exp
_

2
(y
H
y 2Re { m
H
ye
j
} + E)
_
=
1

2N
exp
_

2
(y
H
y 2| m
H
y| cos + E)
_
(6.39)
Notice that p
y
(y|H
0
) does not depend on after all (not surprising since
there is no target present in this case to present an unknown phase), so it is
not necessary to apply Eq. (6.37). However, in p
y
(y|H
1
) the dependence on is
explicit. Assuming a uniform random phase and applying Eq. (6.37) under H
1
gives, after minor rearrangement
p
y
(y|H
1
) =
1

2N
e
(y
H
y+E)/
2 1
2
_
2
0
exp
_
2

2
| m
H
y| cos
_
d (6.40)
Equation (6.40) is a standard integral. Specically, integral 9.6.16 in the book
by Abramowitz and Stegun (1972) is
1

_

0
e
z cos
d = I
0
(z) (6.41)
where I
0
(z) is the modied Bessel function of the rst kind. Using this result
and properties of the cosine function, Eq. (6.40) becomes
p
y
(y|H
1
) =
1

2N
e
(y
H
y+E)/
2
I
0
_
2| m
H
y|

2
_
(6.42)
The log-LRT now becomes
ln = ln
_
I
0
_
2| m
H
y|

2
__

2
H
1
>
<
H
0
ln() (6.43)
or, in sufcient statistic form
= ln
_
I
0
_
2| m
H
y|

2
__
H
1
>
<
H
0
ln() +
E

2
= T (6.44)
Equation (6.44) denes the signal processing required for optimum detection
in the presence of an unknown phase. It calls for taking the magnitude of the
matched lter output m
H
y, passing it through the memoryless nonlinearity
ln[I
0
()], and comparing the result to a threshold. This result is appealing in
that the matched lter is still applied to utilize the internal phase structure
of the known signal and maximize the integration gain, but then a magnitude
operation is applied because the absolute phase of the result cannot be known.
Also, note that the argument of the Bessel function is the energy in the matched
lter output divided by half the noise power; again, a signal-to-noise ratio.
Detection Fundamentals 315
Only half of the noise power appears because the total noise power in the com-
plex case is split between the real and imaginary channels.
As a practical matter, it is desirable to avoid having to compute the natural
logarithm and Bessel function for every threshold test, since these might occur
millions of times per second in some systems. Because the function ln[I
0
()] is
monotonically increasing, the same detection results can be obtained by simply
comparing its argument 2| m
H
y|/
2
to a modied threshold. Equation (6.44)
then becomes simply
| m
H
y|
H
1
>
<
H
0
T

(6.45)
Figure 6.5 illustrates the optimal detector for the coherent detector with an
unknown phase.
The performance of this detector will now be established. Let z = | m
H
y|. The
detection test becomes simply z
>
<
T

; thus the distribution of z under each of

the two hypotheses is needed. As in the known phase case, under hypothesis H
0
(target absent) m
H
y N(0, E
2
); thus the real and imaginary parts of m
H
y
are independent of one another and each distributed as N(0, E
2
/2). It was
seen in Chap. 2 (or see Kay, 1998, Sec. 2.2.7) that z is Rayleigh distributed
p
z
(z|H
0
) =
_

_
2z
E
2
exp
_

z
2
E
2
_
z 0
0 z < 0
(6.46)
The probability of false alarm is
P
FA
=
_
+
T
P
z
(z|H
0
) dz = exp
_
T
2
E
2
_
(6.47)
It is convenient to invert this equation to obtain the threshold setting in terms
of P
FA
T

=
_
E
2
ln P
FA
(6.48)
Now consider hypothesis H
1
, i.e., target present. In this case
m
H
y N(E, E
2
). Since E is real-valued, the real part m
H
y is distributed
Figure 6.5 Structure of optimal detector when the absolute signal phase is unknown.
316 Chapter Six
as N(E, E
2
/2) while the imaginary part is distributed as N(0, E
2
/2). It again
follows from Chap. 2 (or see Kay, 1998, Sec. 2.2.6 or Papoulis, 1984) that the
pdf of z is
p
z
(z|H
1
) =
_

_
2z
E
2
exp
_

1
E
2
(z
2
+ E
2
)
_
I
0
_
2 z

2
_
z 0
0 z < 0
(6.49)
where I
0
(z) is againthe modiedBessel functionof the rst kind. Equation(6.49)
is the Rician pdf. The probability of detection is obtained by integrating it from
T

to +.
In normalized form, the required integral is
Q
M
(, ) =
_
+

t exp
_

1
2
(t
2
+
2
)
_
I
0
(t) dt (6.50)
The expression Q
M
(, ) is known as Marcums Q function. It arises frequently
in radar detection calculations. A closed form for this integral is not known.
Algorithms for evaluating Q
M
(, ) are compared by Cantrell and Ojha (1987).
The Communications Toolbox optional package of MATLAB

includes a
marcumq function; another MATLAB

algorithm is given by Kay (1998).

By dening a change of variables, the integral of Eq. (6.49) can be put into
the form of Eq. (6.50). Specically, choose t = z/
_
E
2
/2 and =

2E/.
Substituting into Eq. (6.49) and doing the integration gives
P
D
= Q
M
_
_

2E

2
,

2T
2
E
2
_
_
(6.51)
Finally, noting that E/
2
is the signal-to-noise ratio and expressing the
threshold in terms of the false alarm probability using Eq. (6.48) gives
P
D
= Q
M
_
_
2,
_
2ln P
FA
_
(6.52)
It is usually the case that the energy E in m or m is not known. Fortunately,
Eq. (6.52) does not depend on E (or the noise power
2
) explicitly, but only on
their ratio , so that it is possible to generate the ROCwithout this information.
However, actually implementing the detector requires a specic value of the
threshold T

as given in Eq. (6.48), and this does require knowledge of both E

and
2
. One way to avoid this problemis to replace the matched lter coefcients
m with a normalized coefcient vector m= m/E
m
, where E
m
is the energy in
m. This choice simply normalizes the gain of the matched lter to 1. The energy
in this modied sequence is

E = 1, leading to a modied threshold

T =
_

2
ln P
FA
(6.53)
The reduced matched lter gain, along with the reduced threshold, results in
no change to the ROC, so that Eq. (6.52) remains valid. Setting of the threshold
Detection Fundamentals 317

T still requires knowledge of the noise power

2
; removal of this restriction
is the subject of Chap. 7. The handling of unknown amplitude parameters is
discussed in somewhat more detail in Sec. 6.2.4.
The performance of the envelope detector in this example is given in Fig. 6.6.
The general behavior is very similar to the known phase coherent detector case
of Fig. 6.4. Closer inspection, however, shows that for a given P
FA
, the coherent
detector obtains a higher P
D
. To make this point clearer, Fig. 6.7 compares the
detection curves for the coherent and envelope detectors (known and unknown
phase, respectively), plotted two different ways. Part (a) of the gure simply
repeats the 10-dB curves from the two earlier gures. At P
FA
= 10
4
and =
10 dB, for example, P
D
is about 0.74. This gure drops to 0.6 when the envelope
detector is used.
Figure 6.7b plots the detection performance as a function of with P
FA
xed
(at 10
6
inthis example). This gure shows that, to achieve the same probability
of detection, the envelope detector requires about 0.6 dB higher SNR than the
coherent detector at P
D
= 0.9, and about 0.7 dB more at P
D
= 0.5. The extra
signal-to-noise required to maintain the detection performance of the envelope
detector compared to the coherent case is called an SNRloss. SNRlosses can re-
sult frommany factors; this particular one is oftencalled the detector loss. It rep-
resents extra SNR that must be obtained in some way if the performance of the
envelope detector is to match that of the ideal coherent detector. Increasing the
SNRin turn implies one or more of many radar systemchanges, such as greater
transmitter power, a larger antenna gain, reduced range coverage, and so forth.
The phenomenon of detector loss illustrates a very important point in detec-
tion theory: the less that is known about the signal to be detected, the higher
Figure 6.6 Performance of the linear envelope detector for the Gaus-
sian example with unknown phase.
318 Chapter Six
Figure 6.7 Performance difference between coherent and envelope de-
tectors for the Gaussian example. (a) Difference in P
D
for = 10 dB.
(b) Difference in P
D
for P
FA
= 10
6
.
must be the SNRto detect with a given combination of P
D
and P
FA
. In this case,
not knowing the absolute phase of the signal has cost about 0.6 dB. Inconvenient
though it may be, this result is intuitively satisfying: the worse the knowledge
of the signal details, the worse the performance of the detector will be.
Detection Fundamentals 319
6.2.3 Linear and square-law detectors
Equation (6.44) denes the optimal Neyman-Pearson detector for the Gaussian
example with an unknown phase in the data. It was shown that the ln[I
0
(x)]
function could be replaced by its argument x without altering the performance.
In Sec. 6.3.2 a simpler detector characteristic than ln[I
0
()] will again be desir-
able for noncoherent integration, but it will not be possible to simply substitute
any monotonic increasing function. It is therefore useful to see what approxi-
mations can be made to the ln[I
0
()] function.
A standard series expansion for the Bessel function holds that
I
0
(x) = 1 +
x
2
4
+
x
4
64
+ (6.54)
Thus for small x, I
0
(x) 1 + x
2
/4. Furthermore, one series expansion of the
natural logarithm has ln(1 + z) = z z
2
/2+z
3
/3 + . Combining these gives
ln[I
0
(x)]
x
2
4
x 1 (6.55)
Equation (6.55) shows that if x is small, the optimal detector is well approx-
imated by a matched lter followed by a so-called square law detector, i.e., a
magnitude squaring operation. The factor of four can be incorporated into the
threshold in Eq. (6.44).
For large values of x, I
0
(x) e
x
/

2x, x 1; then
ln[I
0
(x)] x
1
2
ln(2)
1
2
ln(x) (6.56)
The constant termon the right of Eq. (6.56) can be incorporated into the thresh-
old in Eq. (6.44), while the linear term in x quickly dominates the logarithmic
term for x 1. This leads to the linear detector approximation for large x
ln[I
0
(x)] x x 1 (6.57)
Figure 6.8 illustrates the t between the square law and linear approximations
and the exact ln[I
0
()] functions. The square law detector is an excellent t for
x < 5 dB, while the linear detector ts the ln[I
0
()] very well for x >10 dB.
Finally, note that it is easy enough to compute the squared magnitude of
a complex-valued test sample as simply the sum of the squares of the real
and imaginary parts. The linear magnitude requires a square root and is less
computationally convenient. Section 6.5.2 presents a family of computationally
simple approximations to the magnitude function.
6.2.4 Other unknown parameters
The preceding sections have shown the effect of unknown phase of the received
signal on the optimal detector. However, other parameters of the received signal
are also unknown in practice. The amplitude of the echo depends on all of the
factors in the radar range equation, including especially the unknown target
radar cross section and, at least until it is successfully detected, its range.
320 Chapter Six
Figure 6.8 Approximation of the ln[I
0
] detector characteristic by
the square lawdetector when its argument is small, and the linear
detector when its argument is large.
In addition, the target may be moving relative to the radar, so that the echo is
modied by a Doppler shift.
The derivation of the magnitude-based detectors of Secs. 6.2.2 and 6.2.3 in-
cluded an assumption that the received signal amplitude was known. Specif-
ically, it was assumed that the received signal sample vector m was known,
except for its absolute phase. However, as noted the absolute amplitude is also
unknown in general. To determine the effect of an unknown amplitude, as-
sume that the received signal is A m, where Ais an unknown but deterministic
scale factor.

The analysis of Sec. 6.2.2 can be repeated under this assumption.

The detector output under hypothesis H
0
is unchanged as would be expected,
since the target echo with its unknown amplitude is not present in this case.
Under hypothesis H
1
, the detector output is now Re( m
H
y) N( A
2
E, E
2
/2).
Note that the detector still considers the quantity m
H
y rather than A m
H
y
because the m arises from the matched lter applied to the data and thus does
not include the unknown amplitude factor Aof the signal echo. Also, the quan-
tity E = m
H
m is now the energy of the matched lter reference signal, while
the actual signal energy becomes A
2
E.
The equivalent of Eq. (6.51) is now
P
D
= Q
M
_
_

2A
2
E

2
,

2T
2
E
2
_
_
(6.58)

For instance, the earlier discussionof unknownsignal energy corresponds to choosing A= 1/E m
.
Detection Fundamentals 321
As before, the second argument of Eq. (6.58) can be written in terms of the
probability of false alarm. Furthermore, because the actual signal energy is
now A
2
E, the rst argument is still
_
2. Thus, the detection performance is
still given by Eq. (6.52). The unknown echo amplitude neither requires any
change in the detector structure, nor changes its performance.
Despite the unknown amplitude, the sufcient statistic was not changed.
Furthermore, the probability of false alarm could be computed without knowl-
edge of the amplitude. When both these conditions hold, the detection test is
called a uniformly most powerful (UMP) test (Dudgeon and Johnson, 1993).
A UMP does not exist for the case where the signal delay (range) is unknown,
which again is the only realistic assumption that can be made in radar. It is
therefore necessary to resort to a generalized likelihood ration test (GLRT),
in which the likelihood ratio is written as a function of the unknown signal
delay , and then the value of that maximizes the likelihood ratio is found.
Details are given by Dudgeon and Johnson (1993). The result simply requires
evaluation of the matched lter over a range of delays. In practice, the matched
lter is applied to the entire fast time signal; the lter output samples are
simply the matched lter response for eachcorresponding possible target range.
The maximumoutput is selectedandcomparedto the threshold. If the threshold
is crossed, a detection is declared. Furthermore, the value of at which the
maximum occurs is taken as an estimate of the target delay.
If the target is moving, an unknown Doppler shift will be imposed on the
incident signal. The received echo will then be proportional not to m, but to
a modied signal m

where the samples of the reference signal m have been

multiplied by the complex exponential sequence exp(j
D
n), where
D
is the
normalized Doppler shift. The required matched lter impulse response is now
m

; if mis replaced by m

in the derivations of Sec. 6.2.2, the same performance

results as before will be obtained. Because
D
is unknown, however, it is nec-
essary to test for different possible Doppler shifts by conducting the detection
test for multiple possible values of
D
, similar to the procedure used to test for
unknown range. If a set of K potential Doppler frequencies uniformly spaced
fromPRF/2 to +PRF/2 is to be tested, the matched lter can be implemented
for all K frequencies at once using the pulse Doppler processing techniques de-
scribed in Chap. 5.
6.3 Threshold Detection of Radar Signals
The results of the preceding sections can now be applied to some reasonably
realistic scenarios for detecting radar targets in noise. These scenarios will
almost always include unknown parameters of the signal to be detected (the
target), specically, its amplitude, absolute phase, time of arrival, and Doppler
shift. Detection on both a single sample of the target signal, and when multiple
samples are available, is of interest. In the latter case, as discussed in Chap. 2,
the target signal is often modeled as a random process, rather than a simple
constant; the discussion in this chapter will be limited to the four Swerling
322 Chapter Six
models to illustrate both the approach and the classical, and still very useful,
results obtained in these cases. Furthermore, it will be seen that the idea of
pulse integration is needed in the case of multiple samples. Finally, a square-
lawdetector will be assumed, though one important approximation that applies
to linear detectors will also be introduced. Figure 6.9 represents one possible
taxonomy of the most common variations on the radar detection problem. Each
of these will be discussed in turn in this section, with the exception of the two
diagonally hatched boxes, and of constant-false-alarm-rate (CFAR) techniques,
which are the subject of the next chapter.
6.3.1 Coherent, noncoherent, and binary integration
The ability to detect targets is inhibited by the presence of noise and clutter.
Both are modeled as random processes; the noise as uncorrelated from sample
to sample, the clutter as partially correlated (including possibly uncorrelated)
from sample to sample. The target is modeled as either nonuctuating (i.e., a
constant) or a random process that can be completely correlated, partially
correlated, or uncorrelated from sample to sample (the Swerling models).
The signal-to-interference ratio and thus the detection performance are often
Figure 6.9 Diagram of the taxonomy of detection problems considered. (Adapted from Levanon,
2002.)
Detection Fundamentals 323
improved by integrating (adding) multiple samples of the target and interfer-
ence, motivated by the idea that the interference can be averaged out by
adding multiple samples. This idea was rst discussed in Chap. 1. Thus, in
general detection will be based on N samples of the target+interference. Note
that care must be taken to integrate samples that represent the same range
and Doppler resolution cells.
Integration may be applied to the data at three different stages in the pro-
cessing chain
1. After coherent demodulation, to the baseband complex-valued (I and Q, or
magnitude and phase) data. Combining complex data samples is referred to
as coherent integration.
2. After envelope detection, to the magnitude (or squared or log magnitude)
data. Combining magnitude samples after the phase information is dis-
carded is referred to as noncoherent integration.
3. After threshold detection, to the target present/target absent decisions. This
technique is called binary integration.
A system could elect to use none, one, or any combination of these techniques.
Many systems use at least one integration technique, and a combination of
either coherent or noncoherent with postdetection binary integration is also
common. The major cost of integration is the time and energy required to ob-
tain multiple samples of the same range, Doppler, and/or angle cell (or multiple
threshold detection decisions for that cell); this is time that cannot be spent
searching for targets elsewhere, or tracking already-known targets, or imag-
ing other regions of interest. Integration also increases the signal processing
computational load. Modern systems vary as to whether this is an issue: the
required operations are simple, but must be performed at a very high rate in
many systems.
In coherent integration, complex (magnitude and phase) data samples y
n
are
combined to form a new complex variable y
y =
N1

n=0
y
n
(6.59)
As shown in Chap. 1, if the SNR of a single sample y
n
is
1
, the integrated
data sample y has an SNR that is N times that of the single sample y
n
, i.e.,

N
= N
1
. That is, coherent integration attains an integration gain of N.
Detection calculations are then based on the result for a single sample of
target+noise having the improved SNRequal to
N
. Thus, no special results are
needed to analyze the case of coherent integration; one simply uses the single-
sample detection results for the target and interference models of interest with
the coherently integrated SNR
N
.
In noncoherent integration, phase information is discarded. Instead, the
magnitude or squared magnitude of the data samples is integrated. (Some-
times another function of the magnitude, such as the log-magnitude, is used.)
324 Chapter Six
Most classical detection results have been developed for the square lawdetector,
which bases detection on the quantity
z =
N1

n=0
|y
n
|
2
(6.60)
Consideration will be largely restricted to the square law detector in this
section.
When coherent integration is used, detection results are obtained by using
single-sample (N = 1) results with
1
replaced by the integrated
N
. The situ-
ation for noncoherent integration is more complicated, and it will be necessary
to determine the actual probability density function of the integrated variate z
to compute detection results; this is done in the next subsection.
Binary integration takes place after an initial detection decision has taken
place. That initial decision may be based on a single sample, or on data that
have already been coherently and/or noncoherently integrated. Whatever the
processing before the threshold detection, after it the result is a choice between
hypothesis H
0
, target absent, and H
1
, target present. Because there are
only two possible outputs of the detector each time a threshold test is made,
the output is said to be binary. Multiple binary decisions can be combined in an
M out of N decision logic in an attempt to further improve the performance.
This type of integration is discussed in Sec. 6.4.
6.3.2 Nonuctuating targets
Now consider detection based on noncoherent integration of N samples of a
nonuctuating target (sometimes called the Swerling 0 or Swerling 5 case)
in white Gaussian noise. The amplitude and absolute phase of the target compo-
nent are unknown. Thus, an individual data sample y
n
is the sum of a complex
constant m = mexp( j) for some real amplitude m and phase , and a white
Gaussian noise sample w
n
of power
2
/2 in each of the I and Q channels (total
noise power
2
)
y
n
= m+w
n
(6.61)
Under hypothesis H
0
, the target is absent and y
n
= w
n
. The pdf of z
n
= |y
n
| is
Rayleigh
p
z
n
(z
n
|H
0
) =
_

_
2z
n

2
e
z
2
n
/
2
z
n
0
0 z
n
< 0
(6.62)
Under hypothesis H
1
, z
n
is a Rician voltage density
p
z
n
(z
n
|H
1
) =
_

_
2z
n

2
e
(z
2
n
+ m
2
)/
2
I
0
_
2 mz
n

2
_
z
n
0
0 z
n
< 0
(6.63)
Detection Fundamentals 325
Thus for a vector z of N such samples (not to be confused with the scalar sum
z of samples), the joint pdfs are, for each z
n
0
p
z
(z|H
0
) =
N1

n=0
2z
n

2
e
z
2
n
/
2
(6.64)
p
z
(z|H
1
) =
N1

n=0
2z
n

2
e
(z
2
n
+ m
2
)/
2
I
0
_
2 mz
n

2
_
(6.65)
The LRT and log-LRT become
=
N1

n=0
e
m
2
/
2
I
0
_
2 mz
n

2
_
= e
m
2
/
2
N1

n=0
I
0
_
2 mz
n

2
_
H
1
>
<
H
0
(6.66)
ln =
m
2

2
+
N1

n=0
ln
_
I
0
_
2 mz
n

2
__
H
1
>
<
H
0
ln() (6.67)
Incorporating the term involving the ratio of signal power and noise power on
the left hand side into the threshold gives
N1

n=0
ln
_
I
0
_
2 mz
n

2
__
H
1
>
<
H
0
ln() +
m
2

2
= T (6.68)
Equation (6.68) shows that, given N noncoherent samples of a nonuctuating
target in white noise, the optimal Neyman-Pearson detection test scales each
sample by the quantity 2 m/
2
, passes it through the monotonic nonlinearity
ln[I
0
()], and then integrates the processed samples and performs a threshold
test. There are two practical problems with this equation. First, as noted earlier,
it is desirable to avoid computing the function ln[I
0
()] possibly millions of times
per second. Second, both the target amplitude m and the noise power
2
must
be known to perform the required scaling. The test can be simplied by using
the results of Sec. 6.2.3. Applying the square law detector approximation of
Eq. (6.55) to Eq. (6.68) gives the test
N1

n=0
m
2
z
2
n

4
H
1
>
<
H
0
T (6.69)
Combining all constants into the threshold gives us the nal noncoherent inte-
gration detection rule
z =
N1

n=0
z
2
n
H
1
>
<
H
0

4
T
m
2
= T

(6.70)
326 Chapter Six
Equation (6.70) states that the squared magnitudes of the data samples are
simply integrated and the integrated sum compared to a threshold to decide
whether a target is present or not. Note that the integrated variate z is the
sufcient statistic for this problem.
The performance of the detector given in Eq. (6.70) must now be determined.
It is convenient to scale the z
n
, replacing them with the new variables z

n
= z
n
/
and thus replacing z with z

(z

n
)
2
= z
2
/
2
; such a scaling does not change
the performance, but merely alters the threshold value that corresponds to a
particular P
D
or P
FA
. The pdf of z

n
is still either Rayleigh or Rician voltage as
in Eqs. (6.62) and (6.63), but now with unit noise variance
p
z

n
(z

n
|H
0
) =
_
2z

n
e
z
2
n
z

n
0
0 z

n
< 0
(6.71)
p
z

n
(z

n
|H
1
) =
_
2z

n
e

_
z
2
n
+
_
I
0
(2z

) z

n
0
0 z

n
< 0
(6.72)
where = m
2
/
2
is the signal-to-noise ratio. Since a square law detector is
being used, the pdf of r
n
= (z

n
)
2
is needed (thus z

r
n
); this is exponential
under H
0
and a Rician power density under H
1
p
r
n
(r
n
|H
0
) =
_
e
r
n
r
n
0
0 r
n
< 0
(6.73)
p
r
n
(r
n
|H
1
) =
_
e
(r
n
+)
I
0
(2

r
n
) r
n
0
0 r
n
< 0
(6.74)
Since z

is the sum of N scaled random variables r

n
= (z

n
)
2
, the pdf of z

is the
N-fold convolution of the pdf given in Eq. (6.73) or (6.74) (Papoulis, 1984). This
is most easily found using characteristic functions (CFs). The characteristic
function C(q) corresponding to a pdf p(z) is given by (Papoulis, 1984)
C
z
(q) =
_
+

p
z
(z) e
jqz
dz (6.75)
Note that C
z
(q) is just the Fourier transformof p
z
(z), though with the sign of the
complex exponential kernel chosen opposite from the denition commonly used
in electrical engineering texts. It still follows, however, that the characteristic
function of the N-fold convolution of the pdfs is the product of their individual
characteristic functions.
Under hypothesis H
0
the CF of r
n
can be readily shown from Eqs. (6.73) and
(6.75) to be
C
r
n
(q) =
1
1 j q
(6.76)
Detection Fundamentals 327
The characteristic function of z

is therefore
C
z
(q) = [C
r
(q)]
N
=
_
1
1 jq
_
N
(6.77)
The pdf of z

is obtained by inverting its characteristic function using the inverse

Fourier transform
p
z
(z

|H
0
) =
1
2
_

C
z
(q)e
jqz

dq (6.78)
Using Eq. (6.77) in Eq. (6.78) and referring to any good Fourier transform table
(with allowance for the reversed sign of the Fourier kernel in the denition of
the characteristic function), the Erlang density (a special case of the gamma
density) is obtained (Papoulis, 1984)
p

z
(z

|H
0
) =
_

_
(z

)
N1
( N1)!
e
z

0
0 z

< 0
(6.79)
Note that this reduces to the exponential pdf when N = 1, as would be expected
since in that case z

is the magnitude squared of a single sample of complex

Gaussian noise.
The probability of false alarm is obtained by integrating Eq. (6.79) from the
threshold value to +. The result is (Abramowitz and Stegun, 1972)
P
FA
=
_

T
(z

)
N1
( N1)!
e
z

dz

= 1 I
_
T

N
, N 1
_
(6.80)
where I (u, M) =
_
u

M+1
0
e

M
M!
d (6.81)
is Pearsons form of the incomplete gamma function. For a single sample
(N = 1), Eq. (6.80) reduces to the especially simple result
P
FA
= e
T
(6.82)
so that T = ln P
FA
.
Equation (6.80) can be used to determine the probability of false alarm P
FA
for a given threshold T or, more likely, to determine the required value of T
for a desired P
FA
. Now the probability of detection P
D
corresponding to the
same threshold must be determined. Start by nding the pdf of the normal-
ized, integrated, and square-law-detected samples under hypothesis H
1
. Each
individual data sample r
n
is Rician(Eq. (6.74)); the corresponding characteristic
function is
C
r
n
(q) =
1
q +1
e
[q/(q+1)]
(6.83)
328 Chapter Six
Figure 6.10 Effect of noncoherent integration on detection perfor-
mance for a nonuctuating target in complex Gaussian noise as a
function of single-sample SNR.
The CF of the sum of N such samples, z

, is
C
z
(q) =
_
1
q +1
_
N
e
N[q/(q+1)]
(6.84)
and the pdf of z

is

p
z
(z

|H
1
) =
_
z

N
_
( N1)/2
e
zN
I
N1
_
2
_
Nz

_
(6.85)
P
D
is found by integrating Eq. (6.85). One version of the result, given by Meyer
and Mayer (1973), is
P
D
=
_

T
_
z

_
( N1)/2
e
zN
I
N1
_
2
_
Nz

_
dz

= Q
M
_
_
2N,

2T
_
+e
(T +N)
N

r =2
_
T
N
_
(r1)/2
I
r1
_
2
_
NT
_
(6.86)
Note that the summation term in the second line of Eq. (6.86) only contributes
when N 2. Equations (6.80) and (6.86) dene the performance achievable
with a noncoherent integration using a square law detector (Meyer and Mayer,
1973; DiFranco and Rubin, 1980).
Figure 6.10 shows the effect of the number of samples noncoherently inte-
grated, N, onthe receiver operating characteristic when P
FA
= 10
8
. This gure

Note that I
N1
(x) is the modied Bessel function of the rst kind and order N 1, not to be
confused with the incomplete gamma function I (u, M).
Detection Fundamentals 329
shows that noncoherent integration reduces the required single-sample SNR
required to achieve a given P
D
and P
FA
, but not by the factor N achieved with
coherent integration. For example, consider the single-sample SNR required to
achieve P
D
= 0.9. For N = 1, this is 14.2 dB; for N = 10, it drops to 6.1 dB,
a reduction of 8.1 dB, but less than the 10 dB that corresponds to the factor
of 10 increase in the number of pulses integrated. In Sec. 6.3.3 an estimate
will be developed of this reduction in required single sample SNR, called the
noncoherent integration gain.
6.3.3 Albersheims equation
The performance results for the case of a nonuctuating target in complex
Gaussian noise are given by Eqs. (6.80) and (6.86). While relatively easy to
implement in a modern software analysis system such as MATLAB

, these
equations do not lend themselves to manual calculation. Fortunately, there
does exist a simple closed-form expression relating P
D
, P
FA
, and SNR that
can be computed by hand or with simple scientic calculators. This expression
is known as Albersheims equation (Albersheim, 1981; Tufts and Cann, 1983;
Levanon, 1988).
Albersheims equation is an empirical approximation to the results by
Robertson (1967) for computing the single-sample SNR
1
, required to achieve
a given P
D
and P
FA
. It applies under the following conditions
I
Nonuctuating target in Gaussian (i.i.d. in I and Q) noise
I
Linear (not square-law) detector
I
Noncoherent integration of N samples
The estimate is given by the series of calculations
A = ln
_
0.62
P
FA
_
B = ln
_
P
D
1 P
D
_
(6.87)

1
= 5log
10
N +
_
6.2 +
_
4.54

N +0.44
__
log
10
( A+0.12AB +1.7B) dB
Note that
1
is in decibels, not linear power units. The error in the estimate of
1
is less than 0.2 dB for 10
7
P
FA
10
3
, 0.1 P
D
0.9, and 1 N 8096,
a very useful range of parameters. For the special case of N = 1, Eq. (6.87)
reduces to
A = ln
_
0.62
P
FA
_
B = ln
_
P
D
1 P
D
_

1
= 10log
10
( A+0.12AB +1.7B) dB (6.88)
330 Chapter Six
Note that on a linear (not decibel) scale, the last line of Eq. (6.88) is just
1
=
A+0.12AB +1.7B.
To illustrate, suppose P
D
= 0.9 and P
FA
= 10
6
are required for a nonuc-
tuating target in a system using a linear detector. If detection is to be based
on a single sample, what is the required SNR of that sample? This is a direct
application of Albersheims equation. Compute A= ln(0.62 10
6
) = 13.34 and
B = ln(9) = 2.197. Equation (6.88) then gives
1
= 13.14 dB; on a linear scale,
this is 20.59.
If N = 100 samples are noncoherently integrated, it should be possible to
obtain the same P
D
and P
FA
with a lower single-sample SNR. To conrm this,
use Eq. (6.87). The intermediate parameters Aand B are unchanged.
1
is now
reduced to 1.26 dB, a reduction of 14.4 dB. Note that in this case, the noncoher-
ent integration gain of 14.4 dB, a factor of 27.54 on a linear scale, is much better
than the

N rule of thumb sometimes given for noncoherent integration, which

would give a gain factor of only 10 for N =100 samples integrated. Rather, the
gain is approximately N
0.7
. Albersheims equation will be used shortly to de-
velop an expression for estimating the noncoherent integration gain.
Albersheims equation is useful because it requires no function more exotic
than the natural logarithm and square root for its evaluation. It can thus be
evaluated on virtually any scientic calculator, and is convenient to program
into any programmable scientic calculator. If a somewhat larger error can be
tolerated, it can also be used for square-law detector results for the nonuc-
tuating target, Gaussian noise case. Specically, square law detector results
are within 0.2 dB of linear detector results (Robertson, 1967; Tufts and Cann,
1983). Thus, the same equation can be used for rough calculations over the
range of parameters given previously with errors not exceeding 0.4 dB.
Equations (6.87) and (6.88) provide for calculation of
1
given P
D
, P
FA
, and
N. It is possible, however, to solve Eq. (6.87) for either P
D
or P
FA
in terms
of the other and
1
and N, extending further the usefulness of Albersheims
equation. For instance, the following calculations show how to estimate P
D
given the other factors (
1
is in dB)
A = ln
_
0.62
P
FA
_
Z =

1
+5log
10
N
6.2 +
4.54

N+0.44
(6.89)
B =
10
Z
A
1.7 +0.12A
P
D
=
1
1 +e
B
In Eq. (6.89), A and B are the same values as in Eq. (6.87), though B cannot
be computed in terms of P
D
, since P
D
is now the unknown. A result similar to
Eq. (6.89) can be derived for computing P
FA
.
Detection Fundamentals 331
Albersheims equation can also be used to estimate the signal-to-noise ratio
gain for noncoherent integration of N samples. The noncoherent integration
gain G
nc
is the reduction in single-sample SNR required to achieve a specied
P
D
and P
FA
when N samples are combined; in dB, this is given by
G
nc
( N)(dB) =
1
|
1pulse

1
|
Npulses
= 5log
10
N
_
6.2 +
_
4.54

N +0.44
__
log
10
( A+0.12AB +1.7B)
(6.90)
+10log
10
( A+0.12AB +1.7B) dB
= 5log
10
N
__
4.54

N +0.44
_
3.8
_
log
10
( A+0.12AB +1.7B) dB
On a linear scale this becomes
G
nc
( N) =

N
k
f ( N)
(6.91)
where k = A+0.12AB +1.7B
(6.92)
f ( N) =
_
0.454

N +0.44
_
0.38
The constant k absorbs terms that are not a function of N, and the term f ( N)
is a slowly declining function of N.
Figure 6.11 plots this estimate of G
nc
in decibels for Albersheims nonuc-
tuating, linear detector case as a function of N. Also shown are curves corre-
sponding to N
0.7
and N
0.8
. The noncoherent gain is slightly better than N
0.8
for very few samples integrated (N = 2 or 3), with the effective exponent on N
declining slowly as N increases. G
nc
is bracketed by N
0.7
and N
0.8
to in excess
of N = 100 samples integrated; the gain eventually slows asymptotically to
become proportional to

N for very large N. Thus, noncoherent integration is

more efcient than the

N often attributed to it for a wide range of N, but
remains less efcient than coherent integration, which achieves an integra-
tion gain of N. Nonetheless, its much simpler implementation, not requiring
knowledge of the phase, means it is widely used to improve the SNR before the
threshold detector.
6.3.4 Fluctuating targets
The analysis in the preceding section considered only nonuctuating targets,
often called the Swerling 0 or Swerling 5 case. A more realistic model allows
for target uctuations, in which the target RCS is drawn from either the expo-
nential or chi-square pdf, and the RCS of a group of N samples follows either
332 Chapter Six
Figure 6.11 Noncoherent integrationgain G
nc
for a nonuctuating
target, as estimated using Albersheims equation.
the pulse-to-pulse or scan-to-scan correlation model, as described in Chap. 2.
Note that representing the target by one of the Swerling models 1 through 4
has no effect on the probability of false alarm, since that it is determined only
by the pdf when no target is present; thus Eq. (6.80) still applies.
The strategy for determining the probability of detection depends on the
Swerling model used. Figure 6.12 illustrates the approach. In all cases, the pdf
of the magnitude of a single sample is Rician, so that the CF of a single square-
law detected sample is still given by Eq. (6.83). However, the SNR is now a
random variable because the target RCS is a random variable.
Figure 6.12 Strategy for computing P
D
for Swerling target uctu-
ation models.
Detection Fundamentals 333
In the scan-to-scan correlation cases (Swerling models 1 and 3), the target
RCS is a xed value for all N pulses integrated to form z

. Thus, the CF of z

is
the N-fold product of Eq. (6.83) with itself
C
z
(q; , N) =
_
1
q +1
_
N
e
N[q/(q+1)]
(6.93)
This is the same expression as Eq. (6.84) except that nowC
z
is written explicitly
as a function of all of q, , and N. Next take the expected value of the CF over
the target RCS (and thus the SNR)

C
z
(q; , N) =
_

0
p

()C
z
(q; , N) d (6.94)
where p

() is the pdf of the SNR, and will be either exponential (Swerling 1)

or chi-square (Swerling 3).
For the Swerling 1 case, the pdf of the SNR is exponential
p

() =
1

e
/
(6.95)
where is the average SNR over the N samples. Using Eqs. (6.93) and (6.95)
in Eq. (6.94) gives the characteristic function, averaged over the signal uctu-
ations

C
z
(q; , N) =
1
(1 +q)
N1
[1 +q(1 + N )]
(6.96)
Fourier transform tables and properties can again be applied to compute the
inverse CF of Eq. (6.96), which is the pdf of z

under hypothesis H
1
for Swerling
case 1
p
z
(z

|H
1
) =
1
N
_
1 +
1
N
_
N2
I
_
z

(1 +1/N )

N 1
, N 2
_
e
z

/(1+N )
(6.97)
Integrating the pdf of Eq. (6.97) from the threshold T to +, it is possible
to arrive at the following expression for the probability of detection in the
Swerling 1 case (Meyer and Mayer, 1973)
P
D
= 1 I
_
T

N 1
, N 2
_
+
_
1 +
1
N
_
N1
I
_
_
T
_
1 +
1
N
_

N 1
, N 2
_
_
e
T /(1+N )
N > 1 (6.98)
This expression can be simplied when P
FA
1 and the integrated average
SNR N > 1; both conditions are almost always true in any scenario where
target detection is to be successful. The result is
P
D

_
1 +
1
N
_
N1
e
T /(1+N )
( P
FA
1, N > 1) (6.99)
334 Chapter Six
Furthermore, Eq. (6.99) is exact when N = 1; in this case it reduces to
P
D
= e
T /(1+)
(6.100)
(Note that when there is only one pulse, the average SNR is just the single-
pulse SNR .) For the N = 1 case, Eq. (6.82) can then be used in Eq. (6.100) to
write a direct relationship between P
D
and P
FA
P
D
= ( P
FA
)
1/(1+)
(6.101)
In the Swerling 2 or 4 cases, the samples exhibit pulse-to-pulse correlation,
meaning they are in fact uncorrelated with one another. In this case, each of
the N samples integrated has a different value of SNR. Consequently, it is
appropriate to average over the SNR in the single-sample CF

C
r
(q; ) =
_
1
q +1
__

0
p

()e

_
q
q+1
_
d (6.102)
and then perform the N-fold multiplication of the averaged single-sample CF
to get the CF of the integrated data

C
z
(q; , N) = [

C(q; )]
N
(6.103)
For the Swerling 2 case specically, the exponential pdf can again be used for
the SNR (Eq. (6.95)), applying it this time in Eq. (6.102) to arrive at

C
r
(q; ) =
1
[1 +q(1 + )]
(6.104)
and thus

C
z
(q; , N) =
1
[1 +q(1 + )]
N
(6.105)
Inverse transforming Eq. (6.105) gives the pdf of z

under hypothesis H
1
for
Swerling case 2
p
z
(z

|H
1
) =
z
N1
e
z

/(1+ )
(1 + )
N
( N 1)!
(6.106)
Integrating Eq. (6.106) gives the probability of detection, which can be shown
to be (Meyer and Mayer, 1973)
P
D
= 1 I
_
T
(1 + )

N
, N 1
_
= e
T /(1+ )
N1

l=0
1
l!
_
T
1 +
_
l
(6.107)
The series approximation in the second line of Eq. (6.107) provides a convenient
computational implementation.
Detection Fundamentals 335
Results for Swerling 3 and 4 targets canbe obtained by repeating the previous
analyses above for the Swerling 1 and 2 cases, but with a chi-square instead of
exponential density function for the SNR
p

() =
4

2
e
2/
(6.108)
Derivations of the resulting expressions for P
D
can be found in the books by
Meyer and Mayer (1973), DiFranco and Rubin (1980), and many other radar
detection texts. Table 6.1 summarizes one form of the resulting expressions.
Figure 6.13 compares the detection performance of the four Swerling model
uctuating targets and the nonuctuating target for N = 10 samples as a func-
tion of the SNR for a xed P
FA
= 10
8
. Assuming that the primary interest
is in relatively high (>0.5) values of P
D
, the upper half of the gure is of
greatest interest. In this case, the nonuctuating target is the most favorable,
in the sense that it achieves a given probability of detection at the lowest SNR.
The worst case (highest required SNR for a given P
D
) is the Swerling case 1,
which corresponds to scan-to-scan correlation and an exponential pdf of the
target RCS. For instance, P
D
= 0.95 requires 6 dB for the nonuctuating
case, but 17 dB for the Swerling 5 case, a difference of about 11 dB.
TABLE 6.1 Probability of Detection for Swerling Model Fluctuating Targets with a Square-Law Detector
Case P
D
Comments
0 or 5 Q
M
__
2N ,

2T
_
+e
(T+N )
N

r =2
_
T
N
_r1
2
I
r1
_
2
_
NT
_
1
_
1 +
1
N
_
N1
e
T /(1+N )
Approximate for P
FA
1 and
N > 1; exact for N = 1
2 1 I
_
T
(1 + )

N
, N 1
_
3
_
1 +
2
N
_
N2
_
1 +
T
1 +( N /2)

2( N2)
N
_
e
T /(1+( N /2))
Approximate for P
FA
1 and
N /2 > 1; exact for N = 1 or 2
4
c
N
N

k=0
N!
k!( Nk)!
_
1 c
c
_
Nk
_
2N1k

l=0
e
cT
(cT)
l
l!
_
T > N(2 c)
1 c
N
N

k=0
N!
k!( N k)!
_
1 c
c
_
Nk

l=2Nk
e
cT
(cT)
l
l!
T < N(2 c)
c
1
1 +( /2)
P
FA
= 1 I
_
T

N
, N 1
_
in all cases
I (, ) is Pearsons form of the incomplete Gamma function; I
k
() is the modied Bessel function of the rst kind
and order k
336 Chapter Six
Figure 6.13 Comparison of detection performance for uc-
tuating (Swerling) and nonuctuating target models
using noncoherent integrationof 10 pulses (N = 10) and a
xed probability of false alarm P
FA
= 10
8
with a square
law detector.
At least two general conclusions can be drawn from Fig. 6.13. First, for
P
D
> 0.5, nonuctuating targets are easier to detect than any of the Swerling
cases; target uctuations make detection more difcult, i.e., require a higher
SNR for a given P
D
. Second, pulse-to-pulse uctuations (Swerling 2 and 4)
aid target detectability compared to scan-to-scan detectability. For instance, a
Swerling 2 target is easier to detect than the Swerling 1 that shares the same
pdf for target uctuations, and a Swerling 4 target is easier than a Swerling 3
target. Finally, note that the converse of these statements is true for detection
probabilities less than 0.5.
6.3.5 Shnidmans equation
The analytic results in Table 6.1 are too complex for back-of-the-envelope
calculations, or even for calculation on programmable calculators. Albersheims
equation provided a simple approximation for the nonuctuating target case,
but it is not applicable to uctuating targets ingeneral, andthe Swerling models
in particular. This is a serious limitation since, as seen for example in Fig. 6.13,
the nonuctuating case provides optimistic results for most parameter ranges
of interest.
Fortunately, empirical approximations have also been developed for the
Swerling cases. Probably the most useful result is Shnidmans equation
Detection Fundamentals 337
(Shnidman, 2002). Similar to Albersheims equation, this series of equations
gives the single-pulse SNR required to achieve a specied P
D
and P
FA
with
noncoherent integration of N samples. Unlike Albersheims equation, the re-
sults are for a square lawdetector; however, as noted previously, the differences
are small and in any event, the square law detector is used as much or more
than a linear detector.
Shnidmans equation is given by the following series of calculations
K =
_

_
, nonuctuating target (Swerling 0 / 5)
1, Swerling 1
N, Swerling 2
2, Swerling 3
2N Swerling 4
=
_
0 N < 40
1
4
N 40
(6.109)
=
_
0.8ln(4P
FA
(1 P
FA
)) +sign( P
D
0.5)
_
0.8ln(4P
D
(1 P
D
))
X

=
_
+2

N
2
+
_

1
4
_
_
(6.110)
C
1
= {[(17.7006P
D
18.4496) P
D
+14.5339]P
D
3.525}/K
C
2
=
1
K
_
exp(27.31P
D
25.14) +(P
D
0.8)
_
0.7ln
_
10
5
P
FA
_
+
(2N 20)
80
__
C
dB
=
_
C
1
0.1 P
D
0.872
C
1
+C
2
0.872 < P
D
0.99
C = 10
C
dB/10
(6.111)

1
=
C X

1
(dB) = 10log
10
(
1
) (6.112)
Note that several of the equations simplify in the nonuctuating case, K = .
Specically, in this case C
1
= C
2
= 0, so that in turn C
dB
= 0 and C = 1. The
function sign(x) is +1 if x > 0 and 1 if x < 0.
The accuracy bounds onShnidmans equationare somewhat looser thanthose
specied for Albersheims equation. Specically, except at the extreme values
of P
D
for the Swerling 1 case, the error in the estimate of
1
is less than 0.5 dB
for 0.1 P
D
0.99, 10
9
P
FA
10
3
, and 1 N 100. This is a muchbetter
range for P
D
than used in Albersheims equation. The range of N is much
smaller, but still large enough for almost all problems of interest. Figure 6.14
illustrates the error in the estimate of the single-pulse SNR
1
using Shnid-
mans equation for the case P
FA
= 10
6
and N = 5, and for P
D
over the specied
range of 0.1 to 0.99.
338 Chapter Six
Figure 6.14 Example of error in estimating
1
via Shnidmans
equation for P
FA
= 10
6
and N = 5.
A still more accurate approximation for the nonuctuating and Swerling 1
cases has recently beendeveloped(Hmam, 2003). However, it is not applicable to
all of the Swerling cases, and the computations, while easy, are more extensive.
Consequently, Shnidmans equation appears to be the most useful general tool
currently available.
6.4 Binary Integration
Any coherent or noncoherent integration is followed nally by comparing the
integrated data to a threshold. The result is a choice between two hypotheses,
target present or target absent, so the output is binary in the sense that
it takes one of only two possible outcomes. If the entire detection process is
repeated N times for a given range, Doppler, or angle cell, N binary decisions
will be available. Each decision of target present will have some probability
P
D
of being correct, and a probability P
FA
of being incorrect. To improve the
reliability of the detection decision, the decision rule can require that a target
be detected on some number M of the N decisions before it is nally accepted
as a valid target detection. This process is called binary integration, M of N
detection, or coincidence detection (Levanon, 1988; Skolnik, 2001).
To analyze binary integration, begin by assuming a nonuctuating target so
that the probability of detection P
D
is the same for each of N threshold tests.
Then the probability of not detecting an actual target (i.e., the probability of a
miss) on one trial is 1P
D
. If there are N independent trials, the probability of
missing the target on all N trials is (1 P
D
)
N
. Thus, the probability of detect-
ing the target on at least one of N trials, denoted as the cumulative probability
P
CD
, is
P
CD
= 1 (1 P
D
)
N
(6.113)
Detection Fundamentals 339
Table 6.2 shows the single-trial probability of detection required to achieve
P
CD
= 0.99 as a function of N. Clearly, a 1 of N decision rule achieves a high
cumulative probability of detection with relatively lowsingle-trial probabilities
of detection. In other words, the 1 of N rule increases the effective probability
of detection.
The trouble with the 1 of N rule is that it works for the probability of
false alarm also. The probability of at least one false alarm in N trials is the
cumulative probability of false alarm, P
CFA
P
CFA
= 1 (1 P
FA
)
N
(6.114)
Assuming that P
FA
1, Eq. (6.114) can be approximated as
P
CFA
= 1 (1 P
FA
)
N
= 1
_
1 N P
FA

N( N 1)
2
P
2
FA

_
1 (1 N P
FA
) = N P
FA
(6.115)
where the binomial series expansion was used to obtain the second line of
Eq. (6.115). Equation (6.115) shows that the 1 of N rule increases P
FA
by
a factor of approximately N.
A binary integration rule that reduces the required single-trial P
D
but in-
creases P
FA
does not accomplish the goal of decreasing the required single-
sample SNR to achieve a specied P
D
and P
FA
. What is needed is a binary
integration rule that increases P
CD
compared to P
D
, while leaving P
CFA
equal
to, or less than P
FA
. An M of N strategy provides better results.
Consider the cumulative probability P
C
of H successes in N trials, when the
probability of success on a single trial is p; it is
P
C
=
N

r =M
_
N
r
_
p
r
(1 p)
Nr
(6.116)
where
_
N
r
_

N!
( N r )! r !
(6.117)
Equation (6.116) can be applied to the cumulative probability of false alarm by
letting p = P
FA
, and to the probability of detection by letting p = P
D
. In the
former case, a success is a false alarm, i.e., the event that has a probability
of p; in the latter case, a success is a correct detection. Consider the specic
TABLE 6.2 Single-Trial P
D
Needed
to Achieve P
CD
= 0.99
N 1 2 4 10 20 100
P
D
0.99 0.90 0.68 0.37 0.2 0.045
340 Chapter Six
example of a 2 of 4 rule, that is, N = 4 and H = 2. Using these parameters
in Eq. (6.116) gives
P
C
=
4

r=2
24
(4 r )! r !
p
r
(1 p)
4r
= 6p
2
(1 p)
2
+4p
3
(1 p) + p
4
(6.118)
To determine the effect of this rule on the probability of false alarm, let
p = P
FA
. Assuming that P
FA
1, Eq. (6.118) can be approximated by its rst
term and simplied to obtain
P
CFA

_
N
M
_
P
M
FA
= 6P
2
FA
(6.119)
Thus, the 2 of 4 rule will result in a cumulative false alarm probability that is
less than the single-trial P
FA
as desired. Because single-trial values of P
D
are
not necessarily very close to 1, Eq. (6.118) cannot easily be approximated in a
simple form similar to Eq. (6.119). Table 6.3 shows the cumulative probability
obtained using a 2 or 4 rule for various values of the single-trial probability p.
The three cases above the dotted line are appropriate for considering the effect
on likely single-trial probabilities of detection, while the two cases below the
line are examples of the effect on likely single-trial probabilities of false alarm.
This table shows that the 2 of 4 rule not only reduces the probability of false
alarms, it also increase the probability of detection so long as the single-trial
P
D
is reasonably high.
This example illustrates the characteristics required of an M of N rule.
For small values of p, P
C
should be less than or equal to p so that the rule
reduces false alarm probabilities. For larger values of p, P
C
should be larger
than p so that detection probabilities are increased by binary integration. To
showthe effect of the M of N rule on large and small single-trial probabilities,
Fig. 6.15 plots the ratio of P
C
to p for N = 4 and all four possible choices of M.
If the ratio is greater than 1, P
C
is greater than p; this should be the case for
values of p appropriate to single-trial detection probabilities. Conversely, for
small values of p, appropriate to false alarm probabilities, the ratio should be
TABLE 6.3 Cumulative
Probability Using a 2 of 4
Rule
p P
C
0.5 0.688
0.8 0.973
0.9 0.996
..........................................................
10
3
5.992 10
6
10
6
6.0 10
10
Detection Fundamentals 341
Figure 6.15 Ratio of the cumulative probability to the
single-trial probability for an M of 4 binary integration
rule.
less than 1. Figure 6.15 shows that the ratio is greater than 1 for all values of p
for the 1 of 4 rule, consistent with the earlier discussion. Similarly, the 4 of 4
rule results in a ratio that is always less than 1, good for false alarm reduction
but bad for improving detection. The 2 of 4 and 3 of 4 rules both provide
good false alarm reduction for small values of p and detection improvement
for large values of p. However, the 3 of 4 rule increases probabilities only for
p equal to approximately 0.75 or higher, a relatively narrow range; and the
increase is very slight. The 2 of 4 rule improves detection for values of p down
to approximately 0.23, still well above any likely single-trial false P
FA
. Thus
the 2 of 4 rule appears to be the best choice when N = 4.
Note that a nonuctuating target has been implicitly assumed, since the
single-trial P
D
was assumed to be the same on each trial. The results can be
extended to uctuating targets (Weiner, 1991; Shnidman, 1998). One result of
these analyses is that for a given Swerling model, P
FA
specication, SNR, and
number of samples N, there is a value of M, M
opt
, that maximizes P
D
. M
opt
can
be estimated as
M
opt
= 10
b
N
a
(6.120)
where the parameters a and b are given in Table 6.4 for the various Swerling
models for P
D
= 0.9 and 10
8
P
FA
10
4
(Swerling 0 is the nonuctuating
case) (Shnidman, 1998).
Other binary integration strategies exist having somewhat different proper-
ties. For instance, the decision logic can require that the M hits be contiguous,
rather than just any M hits out of N. See the book by Skolnik (1988) for refer-
ences to the literature for these and other variations on binary integration.
342 Chapter Six
TABLE 6.4 Parameters for Estimating M
opt
Swerling model a b Range of N
0 0.8 0.02 5700
1 0.8 0.02 6500
2 0.91 0.38 9700
3 0.8 0.02 6700
4 0.873 0.27 10700
6.5 Useful Numerical Approximations
As has been seen, calculations of detection and false alarm probabilities, or of
signal-to-noise ratios required to achieve a given P
D
and P
FA
, are plagued by
difcult-to-evaluate functions; even the simplest calculations involve nontrivial
functions such as error functions or Marcums Q function, or their inverses or
integrals. Thus, one is drivento tabulated results or to computer calculations for
accurate estimation of radar performance probabilities. However, there do exist
some numerical approximations and computational techniques that allow esti-
mates of radar performance in some cases with nothing more complicated than
a scientic calculator. Two such particularly useful techniques, Albersheims
and Shnidmans equations, were discussed in Sec. 6.3.3 and 6.3.5, respectively.
In this section, a few more useful techniques are collected.
6.5.1 Approximations to the error function
For Gaussian problems, it has been seen that the error function is of criti-
cal importance in calculating probabilities. A number of approximations and
bounds to erf(x) are available in the book by Abramowitz and Stegun (1972). A
relatively simple approximation is given by
t =
1
1 +0.4707x
erf(x) 1 e
x
2
{0.3480242t 0.0958798t
2
+0.7478556t
3
} (6.121)
One pair of upper and lower bounds on erf(x) is
1
2

e
x
2
x +

x
2
+2
< erf(x) < 1
2

e
x
2
x +
_
x
2
+
4

(6.122)
Figure 6.16 plots erf(x), the approximation of Eq. (6.121), and the bounds of
Eq. (6.122). Equation (6.121) is an excellent t over the entire range. The lower
bound is fairly tight over the entire range, while the upper bound is best used
only for x > 0.8.
Detection Fundamentals 343
Figure 6.16 Error function erf(x) and the upper and lower
bounds of Eq. (6.122).
Another calculator-friendly approximation to the error function is given by
the equations (Vedder, 1988)
a =
_
8

b =
(4 )
3
_
2

(6.123)
erf(x)
2
1 +e

2x(a+2bx
2
)
1
This approximation to erf(x) is not plotted because the difference is visually
indistinguishable. Examination of the error in Vedders approximation shows
that it peaks at a value of approximately 6.27 10
4
when x = 1.64, and
6.2710
4
when x = +1.64. This is accurate enough to allow the approxima-
tion to be used for all values of x in most cases.
Vedder also gave a formula for the inverse error function
A =
a
3b
B =
ln
_
_
X+1
2
_
1
1
_
2b
(6.124)
erf
1
( X)

2Asinh
_
1
3
sinh
1
(BA
3/2
)
_
Again, the visual match between erf
1
( X) and Eq. (6.124) is excellent. The
error is greatest at the extremes, that is, X 1; it is less than 1 percent for
|X| < 0.98 and less than 2 percent for |X| < 0.995.
344 Chapter Six
6.5.2 Approximations to the magnitude function
In Secs. 6.2.2 and 6.2.3 it was shown that the problem of unknown phase leads
to a detector structure based on a function of the magnitude of the matched
lter output. Depending on the SNR regime, the ideal ln[I
0
(x)] detector char-
acteristic is approximated by either a linear or square law detector. Computing
the squared magnitude from the complex I and Q data is simple, requiring
only two multiplies and one add per sample. Efciently computing the linear
magnitude is more troublesome. Filip (1976) in his paper described a family of
13 simple approximations to the magnitude function and calculated the SNR
loss of each as compared to an exactly linear detector. If a complex data sample
is of the form z = I + jQ = Aexp( j ), the approximations to R = |z| are all of
the form

R =
_
a
1
max(|I |, |Q|) +b
1
min(|I |, |Q|) 0
B
a
2
max(|I |, |Q|) +b
2
min(|I |, |Q|)
B
/4
(6.125)
Clearly this is a two-region approximation, with
B
determining the breakpoint
between the regions. The max() and min() functions serve to reect an arbi-
trary complex number into the rst octant. Different members of the family are
dened by different choices of the parameters {a
i
} and {b
i
}.
Filip gave 13 variations corresponding to 13 different parameter choices.
Table 6.5 lists four of the 13 approximations, identifying them by the same
numbers used by Filip (1976). Approximation number 3 is an example of a
very simple estimate, using only one region and binary coefcients. The loss
at a relatively stressing P
D
= 0.999 and P
FA
= 10
6
is 0.0939 dB. Number 8
is an example of a two-region approximation with binary coefcients; the loss
is reduced to 0.0245 dB. Approximation number 10 is designed to minimize
the error variance while maintaining a zero mean error. It achieves a loss of
0.0629 dB. Finally, number 13 is an example of an equiripple approximation.
At 0.0021 dB, it has the lowest loss of any of Filips approximations, but also
requires the most computation, possibly defeating its purpose.
TABLE 6.5 Coefcients of Approximations to the Magnitude Function
SNR loss, dB at P
D
= 0.999,
Filips approx. No. a
1
b
1
tan(
B
) a
2
b
2
P
FA
= 10
6
3 1
1
/2 1 0.0939
8 1
1
/4
1
/2
3
/4
3
/4 0.0245
10 0.948 0.393 1 0.0629
13 0.9903 0.19698 0.414214 0.83954 0.56094 0.0021
Detection Fundamentals 345
References
Abramowitz, M., and I. A. Stegun, Handbook of Mathematical Functions with Formulas,
Graphs, and Mathematical Tables. National Bureau of Standards, U.S. Dept. of Commerce,
1972.
Albersheim, W. J., Closed-FormApproximation to Robertsons Detection Characteristics, Proceed-
ings of IEEE, vol. 69, no. 7, p. 839, July 1981.
Brennan, L. E., and I. S. Reed, A Recursive Method of Computing the Q Function, IEEE Trans-
actions on Information Theory, vol. IT-11, no. 2, pp. 312313, April 1965.
Cantrell, P. E., and A. K. Ojha, Comparison of Generalized Q-function Algorithms, IEEE Trans-
actions on Information Theory, vol. IT-33, no. 4, pp. 591596, July 1987.
DiFranco, J. V., and W. L. Rubin, Radar Detection. Artech House, Dedham, MA, 1980.
Filip, A. E., A Bakers Dozen Magnitude Approximations and Their Detection Statistics,
IEEE Transactions on Aerospace & Electronic Systems, vol. AES-12, no. 1, pp. 8689,
Jan. 1976.
Hmam, H., Approximating the SNRValue inDetectionProblems, IEEETransactions onAerospace
and Electronic Systems, vol. AES-39, no. 4, pp. 14461452, Oct. 2003.
Johnson, D. H., and D. E. Dudgeon, Array Signal Processing. Prentice Hall, Englewood Cliffs, NJ,
1993.
Kay, S. M., Fundamentals of Statistical Signal Processing, Vol. I: Estimation Theory. Prentice Hall,
Upper Saddle River, NJ, 1993.
Kay, S. M., Fundamentals of Statistical Signal Processing, Vol. II: Detection Theory. Prentice Hall,
Upper Saddle River, NJ, 1998.
Levanon, N., Radar Principles. Wiley, New York, 1988.
Meyer, D. P., and H. A. Mayer, Radar Target Detection. Academic Press, New York, 1973.
Papoulis, A., Probability, Random Variables, and Stochastic Processes, 2d ed. McGraw-Hill, New
York, 1984.
Robertson, G. H., Operating Characteristic for a Linear Detector of CW Signals in Narrow
Band Gaussian Noise, Bell System Technical Journal, vol. 46, no. 4, pp. 755774, April
1967.
Shnidman, D. A., Radar Detection Probabilities and Their Calculation, IEEE Transactions on
Aerospace and Electronic Systems, vol. AES-31, no. 3, pp. 928950, July 1995.
Shnidman, D. A., Binary Integration for Swerling Target Fluctuations, IEEE Transactions on
Aerospace and Electronic Systems, vol. AES-34, no. 3, pp. 10431053, July 1998.
Shnidman, D. A., Determination of Required SNR Values, IEEE Transactions on Aerospace and
Electronic Systems, vol. AES-38, no. 3, pp. 10591064, July 2002.
Skolnik, M. I., Introduction to Radar Systems, 3d ed. McGraw-Hill, New York, 2001.
Tufts, D. W., and A. J. Cann, On Albersheims Detection Equation, IEEE Transactions on
Aerospace and Electronic Systems, vol. AES-19, no. 4, pp. 643646, July 1983.
Van Trees, H. L., Detection, Estimation, and Modulation Theory, Part I: Detection, Estimation, and
Linear Modulation Theory. Wiley, New York, 1968.
Vedder, J. D., Approximation to the Normal Probability Distribution, NASA Tech Briefs, p. 96,
Feb. 1988.
Weiner, M. A., Binary Integration of Fluctuating Targets, IEEE Transactions on Aerospace and
Electronic Systems, vol. AES-27, no. 1, pp. 1117, Jan. 1991.
346