Session 2 1

21. I

Constant False Alarm Rate

(CFAR)
Detection
 2 1 . 2
• When we considered the problem of optimal signal
detection, we assumed we knew the distribution of
the noise under the null hypothesis.

• We not only assumed we knew the form of the

distribution (e.g., Gaussian or Rayleigh), but we
assumed we had perfect knowledge of the
parameters of the distribution.

• What if we are wrong about the parameters (or

the distribution for that matter?)

• Incorrect knowledge of the noise distribution

results in errors in the likelihood ratio threshold.

• This can drastically effect actual detection.

2 1.3
Assume that we must decide between two simple hypotheses

H0 : X ∼ exp (µ0 ),
(assume i t , > Mo.)
H1 : X ∼ exp (µ1 ).
The resulting test will be a threshold test of the form
!
1, for X ≥ x0 ;
φ(X) =
0, for X < x0 .

The threshold x0 that yields a probability of false alarm α is

x0 = −µ0 ln α.

If we have an error in µ0 , we will have a significantly different false alarm

probability:

α = (αnom )µnom /µ0 .

 21.4

The effects of inaccurate noise estimates

α = (αnom )µnom /µ0

αnom = 10−6

µ0 /µnom (dB)

21.5
The exponential Detection Problem Revisited
Assume that we must decide between two simple hypotheses

H0 : Y exp (µ0 ),

H1 : Y exp (µ1 ).

Now if we think of
µ1 = µ0 + µs ,
where
µs = signal component of µ1 ,
then if we define the signal-to-noise ratio as

S = µs /µ0 ,

we can rewrite µ1 as
µ1 = µ0 (1 + S),
and our simple hypotheses can be rewritten as

H0 : Y exp (µ0 ) versus H1 : Y exp (µ0 (1 + S)).

 2 1.6

The most powerful test of size α is given by

!
1, for Y > Y0 ,
φ(Y ) =
0, for Y ≤ Y0 ,

where the threshold Y0 is given by

Y0 = −µ0 ln α.

The power of the test is given by

⇥ = P (Y > Y0 |H1 ) = · · · = 1/(1+S)

.

n.b. The threshold Y0 is a function of µ0 and the probability of false alarm .

2 1 . 7

If we don’t know the value of µ0 , we cannot set the threshold Y0 that will yield
our size test. How should we proceed?

In principle, µ0 could take on a broad range of positive values.

We could view H0 as a the composite hypothesis that µ101 ⇥ (0, ).

We could then use a generalized likelihood ratio test to solve the problem.

Under hypothesis H0 , this would correspond to finding the maximum likelihood

estimate µ̂0 and using it in place of µ0 . But for one sample measurement, this
does not yield a good estimate.

However, if we had N i.i.d. measurements X1 , . . . , XN of the noise, we could

use the maximum likelihood (and minimum variance unbiased) estimate
N
1
µ̂0 = Xi
N i=1

in place of µ0 .
 2 1.8

• In a “typical” radar scenario, targets are sparsely

located against a background of noise and clutter.

• There tends to be regions of local statistical

homogeneity in this noise/clutter background
because the physical environment giving rise to it
often has homogeneous statistics.

• However, there can be significant changes in the

local scattering characteristics as you move through
the scattering environment.

• There can be sharp boundaries between scattering

regions.

21.9

• This suggests that one approach to estimating the

background noise power for target detection in a
particular resolution cell is to average the measured
noise power in surrounding resolution cells.

• This is an example of a class of detection

techniques called Constant False Alarm Rate
(CFAR) techniques.

• We will see where the term Constant False Alarm

Rate comes from, but more important than the
constant false alarm rate is a robustness to changes
in the average noise power.
 21.10
12
A Generic CFAR Processor
Cell Under Test (CUT)

Square-Law
Input Signal XN ··· X N +1 Y XN ··· X1
Detector 2 2

··· ··· Comparator Detection Decision

Sort into Order Statistics X(1) , . . . , X(N )

fˆ(X1 , . . . , XN )
X(1) ··· X(N )

f (X(1) , . . . , X(N ) )

The resolution cell under test (CUT) with measurement Y is tested for the
Fig. 2.2. A Typical CFAR Processor.
presence of the target using a threshold computed using neighboring resolution
cell measurements X1 , . . . , XN .
the context of CFAR processors, then, these quantities are expectations with respect
The statistic Z = fˆ(X , . . . , XN ) is an estimate of the noise power.
to the the estimate 1Z.
Another very important note regarding CFAR processors is that in practice, the
The threshold scaling factor T sets the threshold level by scaling the statistic
assumption of homogeneous clutter samples in the reference window is often an in-
the statistic Z. This works because the threshold is the product of a constant
and the valid
average noise
one. The sizepower.
N of the reference window significantly affects the probability
of achieving a homogeneous clutter sample within the reference window. In the
following chapters, a review of common CFAR techniques will be given, as well as
2 1 . 1 1
an analysis of their performance when the homogeneity assumption is violated. Fi-
A More Specific Class of CFAR Processors 12
nally, several new techniques will be presented that are designed to improve the
performance of existing CFAR processorsCell
byUnder
giving the processor additional flexibil-
Test (CUT)

ity for dealing with inhomogeneity by adaptively switching between various CFAR
Square-Law
Input Signal
techniques, by varyingXthe · ·size
orDetector N· X Y
of the X ···
reference
N
2
+1X
window, depending on the homo-
N
2
1

··· Comparator Detection Decision

geneity of the sample in the reference window.· · ·
Sort into Order Statistics X(1) , . . . , X(N )

X(1) ··· X(N )

f (X(1) , . . . , X(N ) )

This processor can compute

Fig. 2.2. A Typical CFAR Processor.
• Mean
• Median
the context of CFAR processors, then, these quantities are expectations with respect
to the the estimate Z.
• Arbitrary Order Statistics
Another very important note regarding CFAR processors is that in practice, the
• Linear Combination of Order Statistics.
assumption of homogeneous clutter samples in the reference window is often an in-
valid one. The size N of the reference window significantly affects the probability
of achieving a homogeneous clutter sample within the reference window. In the
 2 1 . 1 2

Cell-Averaging CFAR (CA-CFAR)

14

Cell Under Test (CUT)

Square-Law
Input Signal XN ··· XN/2+1 Y XN/2 ··· X1
Detector

··· ··· Comparator Detection Decision

N
i=1 Xi

T
Adaptive Threshold

Fig. 3.1. Cell-Averaging CFAR Processor.

In CA-CFAR, we have that the statistic Z/N is just the sample mean.

3.2 CA-CFAR Performance Analysis

It can be shown that Z/N is the maximum-likelihood estimate of µ0 . (It is
also theAsminimum
in previousvariance unbiased
analyses, let estimate (MVUE)
the null hypothesis, of the
H0 , represent andwhen
µ0 case an efficient
no
estimate—satisfying the Cramer-Rao lower bound.)
target is present in the CUT (clutter only), and let H1 represent the case when a
target in addition to clutter is present. Let Y be the CUT. Under H0 , then, Y
is a single random observation drawn from the exponential clutter distribution with
mean µ. Under H1 , it is drawn from an exponential distribution with mean µ(1 + S),2 1 . 1 3
where S is the average target SNR. Additionally, it is assumed that the N cells in
If we assume that X1 , . . . , XN are i.i.d exponential with mean µ (drop subscript
the CA-CFAR
for simplicity) weprocessor’s
have reference window contain only i.i.d. exponential clutter
observations Xi , i = 1, . . . , N , with mean µ. Thus,
1
fXi (x) = 1e x/µ
x/µ
· 1[0,⇥) (x).
fXi (x) =µ e · 1[0,⇥) (x). (3.1)
µ
The moment generating function (MGF) of X is then
⇥
1
Xi (s) = . (3.2)
The moment generating function of each1 Xµs i is
⇥ the sum of N cell values in
The mean clutter power estimate, Z is made from
1
the reference window. Xi (s) = .
N 1
⇤ µs
Z= Xi . (3.3)
i=1

Note that this is equivalent to using the sample mean, and multiplying the threshold

The moment generating

scaling factor derived in function ofNZ. is
Eq. (3.7) by

⇥N
1
Z (s) = .
1 µs
 21.14

Because the test is a threshold test comparing the CUT Y to the threshold T Z,
the probability of false alarm is

= EZ [P [Y > T Z|H0 ]]
⇤⇧ ⇥ ⌅
1 y/µ
= EZ e dy
TZ µ
= EZ [exp( T Z/µ)]
⇧ ⇥
TZ
= e µ fZ (z)dz
inf ty
. ⇥
T
= Z ,
µ

where EZ [·] denotes expectation w.r.t. Z.

Substituting this into the expression for Z (s), the false alarm probability is

= (1 + T ) N
.

Note that the false alarm rate is not a function of the mean noise power µ.
Hence the term constant false-alarm rate.

2 1 . 1 5

The threshold scaling factor yielding a size test is

T =( ) 1/N
1.

Similarly, the detection probability can be calculated under the alternative hy-
pothesis H1 and given by

= EZ [P [Y > T Z|H1 ]]
⇤⇧ ⇥ ⌅
1
= EZ e y/µ(1+S)
dy
T Z µ(1 + S)
= EZ [exp( T Z/µ)]
⇥
T
= Z
µ(1 + S)
⇤ ⌅ N
T
= 1+
(1 + S)
⇥N
1+S
= .
1+T +S
 21.16

Combining these results, we find that

⇥N
1+S
⇥= 1/N + S
.

In the limit, as N ⇥ ⌅, we have

lim = lim (1 + ⇤/N ) N

N ⇥⇤ N ⇥⇤
= exp{ ⇤}
lim ⇥ = lim (1 + ⇤/N (1 + S)) N
N ⇥⇤ N ⇥⇤
= exp{ ⇤/(1 + S)}
⇤ ⇥ ⇥ 1/(1+S)
, as N ⇥ ⌅

2 1.1 7

CA-CFAR Detection Performance

1

0.9

0.8

0.7
Detection probability

0.6 CFAR loss

0.5

0.4
Infinity
16
0.3 32
8

4
0.2

0.1 2

0
0 5 10 15 20 25 30
SNR (dB)
 30 0

Fig. 3.2. CA-CFAR Pd versus N and SN R (dB) for a desired Pf a =2 0.5.

1 . 1 8

Fig. 3.3. CA-CFAR Pd versus N and SN R (dB) for a desired Pf a = 1 10 6 .

19
CA-CFAR, Pfa = 1e-06

CA-CFAR Threshold Map, N = 8, Pfa = 0.5

Pd
1e+07
Clutter
Targets
1e+06 TZ
1
0.9
0.8
100000
0.7
0.6
0.5
0.4
10000
0.3
0.2
0.11000
0
100 30
25
20
0 10 15
5 SNR (dB)
10 10
15
1 20 5
N (samples) 25
30 0
0.1
Figure from: Michael F. Rimbert, Constant False Alarm Rate Detection Techniques Based on Empirical Distribution Function Statistics, Ph.D Thesis,
School of Electrical and Computer Engineering, Purdue University, August 2005.
0.01
0 20 40 60 80 100 120 140 160 180 200
Fig. 3.3. CA-CFAR Pd versus N and SN R (dB) for a desired Pf a = 1 10 6 .
21.19
19
Fig. 3.4. CA-CFAR Threshold Map, N = 8, PF A = 0.5.

CA-CFAR Threshold Map, N = 8, Pfa = 0.5

1e+07
Clutter
CA-CFAR Threshold Map, N = 8, Pfa = 1e-06 Targets
1e+06 TZ
1e+08
Clutter
Targets
1e+07
100000 TZ

1e+06
10000

100000
1000
10000
100
1000

100 10

10 1

1
0.1

0.1
0.01
0 20 40 60 80 100 120 140 160 180 200
0.01

0.001 F. Rimbert, Constant False Alarm Rate Detection Techniques Based on Empirical Distribution Function Statistics, Ph.D Thesis,
Figure from: Michael
Fig. 3.4. CA-CFAR Threshold Map, N = 8, P
School 0of Electrical
20 and Computer
40 60
Engineering, 80 University,
Purdue 100 August120 2005. 140 160 F A 180 200 = 0.5.

Fig. 3.5. CA-CFAR Threshold Map, N = 8, PF A = 1 10 6 .

 0.001 Fig. 3.4. CA-CFAR Threshold Map, N = 8, PF A = 0.5.
0 20 40 60 80 100 120 140 160 180 200

21.20
Fig. 3.5. CA-CFAR Threshold Map, N = 8, PF A = 1 10 6 .

CA-CFAR Threshold Map, N = 8, Pfa = 1e-06

1e+08
Clutter
Targets
1e+07 TZ

1e+06

100000

10000

1000

100

10

0.1

0.01

0.001
0 20 40 60 80 100 120 140 160 180 200

Figure from: Michael F. Rimbert, Constant False Alarm Rate Detection Techniques Based on Empirical Distribution Function Statistics, Ph.D Thesis,
6
Fig. 3.5. CA-CFAR Threshold Map, N = 8, PF A = 1
School of Electrical and Computer Engineering, Purdue University, August 2005. 10 .

2 1.2 1

CA-CFAR Limitations

• The performance of CA-CFAR suffers when

statistical homogeneity of the reference window
samples is violated. This commonly occurs when:

1. Reference window contains interfering targets

2. Reference window contains “clutter edges”—

boundaries between regions with differing
scattering characteristics.
 1e-04
0 20 40 60 80 100 120 140 160 180 200

2 1 . 2 2
Fig. 3.6. CA-CFAR – Interfering Targets 20
21

CA-CFAR Threshold Map, N = 24, Pfa = 1e-06

1e+10 CA-CFAR Threshold Map, N = 8, Pfa = 0.01
Clutter
1000 Targets
ountered at physical boundaries between differing types of vegetation, foliage, Clutter TZ
1e+08 Targets
TZ
d, or bodies of water, each having different radar reflectivity, and thus different
100 1e+06

an clutter power.
10000
N
The clutter power
10 estimate, Z = i=1 Xi , has a breakdown value of zero, which
O
100

he proportion of the sample that can be increased to infinity without making

1
1 O -

statistic infinite [12]. That is, it is extremely sensitive to statistical outliers in

0.01
sample. Thus, interfering targets in the reference window can severely bias the
0.1
1e-04
l clutter power estimate.
0 20This
40 leads
60 to80 “target
100 masking”,
120 140 160 where
180 a200target in the
Figure from: Michael F. Rimbert, Constant False Alarm Rate Detection Techniques Based on Empirical Distribution Function Statistics, Ph.D Thesis,
rence window
0.01
can
Schoolraise
of Electrical the
andFig.threshold
3.6.
Computer CA-CFAR
Engineering, enough – August
Purdue University, to prevent
Interfering
2005. Targets a target in the CUT
0 20 40 60 80 100 120 140 160 180 200

m being detected. Thus, closely-spaced targets reduce the detection performance

2 1 . 2 3
Fig. 3.7. CA-CFAR – Clutter Edges
he CA-CFAR processor.
encountered at physical boundaries between differing types of vegetation, foliage,
21
land, or bodies of water, each having different radar reflectivity, and thus different
The reference window will contain an inhomogeneous sample at edges between
mean clutter power. CA-CFAR Threshold Map, N = 8, Pfa = 0.01

onslower mean clutter

of different power.
mean
1000 These power.
clutter lower-valued
As samples will reduce
the reference
The clutter power estimate, Z =
the clutter
window
Clutter
power
sweeps acrossN
Xi , has a breakdownTargets
value of zero, which
TZ i=1
estimate, lowering the threshold, and consequently lead to excessive false-alarms at
transition, theproportion
is the CUT100 willof at
thesome
samplepoint be be
that can a sample
increasedfrom the region
to infinity without with larger
making
the clutter edges.
an clutter the statistic
power, infinite
while [12]. That
samples is, reference
in the it is extremely sensitive
window to statistical
come from theoutliers
region in with
Fig. the
3.6sample.
illustrates CA-CFAR’s
Thus, inability
interfering targets to reference
in the discriminate closely
window spaced bias
can severely targets
the
O
10

(i.e. it’slocal
sensitivity
clutter to interfering
power estimate.targets). Three
This leads closely
to “target spaced targets
masking”, where asignificantly
target in the
raise thereference window
mean clutter1 can raise
power the threshold
estimate, raising enough to prevent
the threshold a target
above two ofinthe
the three
CUT

targets. from being

Thus, onlydetected. Thus,
one of the closely-spaced
three targets wastargets reduce
detected. Thetheother
detection performance
two are masked,
0.1
of the CA-CFAR processor.
and represent missed detections. Fig. 3.7 illustrates CA-CFAR’s poor detection
The reference window will contain an inhomogeneous sample at edges between
performance at clutter
0.01 edges. At the clutter edge to the left, a false-alarm occurs
0 20 40 60 80 100 120 140 160 180 200
regions of different mean clutter power. As the reference window sweeps across
where the Figure
threshold
from: Michael F.falls below
Rimbert, Constant Falsethe value
Alarm Rate Detectionof a clutter
Techniques sample.
Based on Empirical The Statistics,
Distribution Function samePh.Dsituation
Thesis,
the transition, the CUT
School of Electrical will
Fig.
and Computer at
3.7.some
Engineering, Purdue point
CA-CFAR be a2005.
sample
– Clutter
University, August from the region with larger
Edges
nearly occurs at thepower,
mean clutter right while
clutter edge,inwhere
samples the threshold
the reference fallsfrom
window come verythe
close to with
region the
 2 1 . 222
4
GO-CFAR and SO-CFAR Processors
Cell Under Test (CUT)

Square-Law
Input Signal XN ··· X N +1 Y XN ··· X1
Detector 2 2

··· ··· Comparator Detection Decision

Xi Xi

Z1 Z2

SO-CFAR: min{Z1 , Z2 }
GO-CFAR: max{Z1 , Z2 }

• Greatest
Fig. of
3.8.CFAR
GO-CFAR and SO-CFAR
(GO-FAR) Processors.
reduces excessive
false alarms near clutter edges, but poor detection
performance
N
in the presence of interfering targets.
timator Z = i=1 Xi , the threshold is extended far beyond both clutter edges and
targets. This •
Smallest
severely of CFAR
reduces the(SO-FAR) has
probability of improved
detection in surrounding regions,
performance
thus limiting the usefulnessin
of the presence
larger reference of interfering
windows targets, environ-
in inhomogeneous
ments.
but high false alarm rate near clutter edges.
In general, we do not have a priori knowledge of whether or not the region in a
21.25
reference window is homogeneous. Even though the CA-CFAR is optimal for a finite
size reference window, its assumption of homogeneity in the reference window lead
to a search for more practical solutions. Variants of CA-CFAR have been designed23
to address the shortcomings of CA-CFAR, including Greatest-of CFAR (GO-CFAR)
[19] and Smallest-of CFAR (SO-CFAR) Table[20].
3.1 Each of these processors have been
CFAR Processor Performance Comparison
extensively analyzed in [18]. Therefore, detailed analysis of these variants will not
Processorprocessors is provided in
be given here. A schematic of SO-CFAR and GO-CFAR
Problem
Fig. 3.8. CA-CFAR GO-CFAR SO-CFAR OS-CFAR
Both GO-CFAR
Clutter Edges and SO-CFAR
Poor processors divide the reference
Good Poor window into two
Good
sub-windows, namely the leading
Interfering Targets Poor and trailing portions of Good
Poor the reference window
Good on
either side of the CUT. The sample mean from each sub-window are computed
separately. GO-CFAR uses the greater value, while SO-CFAR uses the smaller value.
better able to resolve closely spaced targets, but suffers from higher false-alarm rates
GO-CFAR prevents excessive false alarms at clutter edges, but has poor detection
at clutter edges. Order-Statistic CFAR, on the other hand, has been shown to be
probability in the presence of interfering targets. On the other hand, SO-CFAR is
robust in both the presence of statistical outliers (interfering targets) and clutter
Figure from: Michael F. Rimbert, Constant False Alarm Rate Detection Techniques Based on Empirical Distribution Function Statistics, Ph.D Thesis,
School of Electrical and Computer Engineering, Purdue University, August 2005.
edges [1], and is the subject of chapter 4. This behavior is summarized in Table 3.1.
 21.26

Order-Statistic CFAR (OS-CFAR)

Herman Rohling, “Radar CFAR Thresholding in Clutter and Multiple Target Situations,”
IEEE Transactions on Aerospace and Electronic Systems, vol. 19, pp. 608–621, 1983.

• In OS-CFAR, the average noise power in a region is

estimated using an order statistic, or ranked sample
of the noise power samples in the reference window.

• For example, we might use the sample median

instead of the sample mean to estimate the average
noise power.

• While an order statistic estimate is not the

maximum likelihood estimate if the samples are
independent and statistically homogeneous (i.i.d.),
order statistics (e.g., the sample median) are much
more robust to deviations from this ideal.