Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

SIViP (2012) 6:223–230

DOI 10.1007/s11760-010-0207-3

ORIGINAL PAPER

Optimum second threshold for the CFAR binary integrator


in Pearson-distributed clutter
Hilal A. Meziani · Faouzi Soltani

Received: 25 June 2009 / Revised: 24 December 2010 / Accepted: 24 December 2010 / Published online: 12 January 2011
© Springer-Verlag London Limited 2011

Abstract In this paper, we propose to analyze the binary by scaling the noise level estimate with a constant to achieve
integration of the cell-averaging constant false-alarm rate a desired probability of false alarm. This is the conventional
(CA-CFAR) and order statistics constant false-alarm rate CA-CFAR (cell-averaging constant false-alarm rate) detector
(OS-CFAR) detectors in the presence of non-Gaussian spiky proposed by Finn and Johnson [1]. This detector is optimal
clutter modeled as a Pearson distribution. We derive new when the clutter power in the reference window is Gaussian
closed form expressions for false alarm and detection prob- modeled (Rayleigh envelope distribution), independent, and
abilities for the CA-CFAR detector in the presence of identically distributed.
Pearson-distributed clutter backgrounds. We first show that In practical situation, multi-pulses transmission is
the use of binary integration improves the detection proba- employed in order to combat the fading in the power of
bilities of the detectors considered. Secondly, the maximum the received signal at the antenna for single-pulse case. The
of detection probability occurs for an optimum choice when above problem is overcome by using binary integration that
the second threshold is set to be equal to M = (3/4)L. For is also referred to as coincidence detector or M-out-of-L
this optimum M-out-of-L rule, the comparison analysis of detection [2,3].
the CA-CFAR and OS-CFAR binary integrators showed that The binary integrator employs a two-step thresholding
the latter has better performance in homogeneous Pearson- technique. In the first step, a preliminary binary decision is
distributed clutter. made about each pulse. The number of decisions where the
first threshold is exceeded is then counted. If this number
Keywords CFAR detection · Pearson distribution · exceeds a second threshold M, a target is declared present.
Optimum binary integration In the adaptive thresholding techniques used in the first deci-
sion step, the threshold is proportional to the mean power
of the local clutter. Hence, the knowledge of the statistics of
1 Introduction the clutter is important in order to properly design a CFAR
detector. In real situations, the clutter statistics deviate from
In radar automatic detection, the received signal is sampled in Gaussian assumption. In many previous studies, the clutter
range by the range resolution cells. The clutter background in statistics were modeled as Log-normal, Weibull, and K-dis-
the cell under test is estimated by averaging the outputs of the tribution [4–7]. Recent studies [8–10] showed that the clut-
nearby resolution cells. The detection threshold is obtained ter amplitude distribution is more impulsive, when the radar
operates in high resolution and/or at low grazing angle and
H. A. Meziani (B) · F. Soltani suggests that the clutter can be modeled as a Pearson distri-
Laboratoire Signaux et Systèmes de Communication, bution (also called Lévy-distribution) which is a heavy-tailed
Faculté des Sciences de l’Ingénieur, Département d’Electronique,
distribution.
Université de Constantine, Route d’Ain Bey,
Constantine 25000, Algeria This paper is organized as follows. In Sect. 2, we pres-
e-mail: hameziani@yahoo.fr ent a background on the positive alpha-stable distribution.
F. Soltani In Sect. 3, the problem is formulated. In Sects. 4 and 5,
e-mail: f.soltan@yahoo.fr we analyze the performance of the binary integrator for the

123
224 SIViP (2012) 6:223–230

Fig. 1 Block diagram of the


CA-CFAR binary integrator

CA-CFAR and OS-CFAR detectors in non-Gaussian spiky In the remaining of the paper, we will refer to this distribution
clutter modeled as a Pearson distribution. Section 6 presents simply as the Pearson distribution.
the results and gives the conclusion.

3 Problem formulation

2 Positive alpha-stable distribution For the analysis, we assume that L adjacent pulses hit the
point target inside the same range cell. The first decision
Many real data show unsymmetrical and impulsive is based on a single-pulse basis, and the second decision is
characteristic that cannot be accommodated by the Gauss- based on the L pulse train, Figs. 1–2. To achieve that the
ian or symmetric alpha-stable distributions such as radar output of the square law detector of each channel is sampled
signal processing and telecommunications modeling [11]. by the range resolution cells. The received signal is sampled
A skewed α-stable distribution can be used for modeling such in range by the N + 1 range resolution cells resulting in
data. The positive alpha-stable distribution family is defined L × (N + 1) observations that are denoted by yi j . Under the
only by their characteristic function given by assumption that the observations are statistically independent
   π  and identically distributed, the probability density function
φ(ω) = Exp −γ |ω|α 1 + i sign(|ω|) tg α (1) (pdf) is
2
γ2
where sign(ω) is the sign function defined as √γ 1 − 2yi j
pYi j (yi j ) = e i = 1, . . . , L
⎧ 2π yi3/2
j
(4)
⎨ 1 ω>0 j = 1, . . . , N
sign(ω) = 0 ω=0 , (2)
⎩ For the i eme channel, the data available in the reference win-
−1 ω < 0
dow are processed to obtain the statistic Z i that is the estimate
0 < α < 1 is the characteristic exponent, and γ > 0 is the of the clutter level. To maintain the overall probability of false
dispersion parameter. The Fourier transform of (1) leads to a alarm (Qfa) at a desired constant value when the clutter is
compact form only in the special case, α = 0.5, namely the homogeneous, the detection threshold is obtained by scaling
Pearson distribution type V, given by the statistic Z i with a scale factor T which keeps the overall
probability of false alarm equal to the desired value. The test
γ 1 − γ2y2 cell Yi0 is compared with the adaptive threshold to make a
pY (y) = √ e (3)
2π y 3/2 decision Di according to

123
SIViP (2012) 6:223–230 225

Fig. 2 Block diagram of the


OS-CFAR binary integrator

H1 where Pfa and Pd are the probability of false alarm and


> the probability of detection for single-pulse transmission,
Yi0 T Z i i = 1, . . . , L (5)
< respectively.
H0
The preliminary decision Di i = 1, . . . , L is made about the
presence H1 or the absence H0 of a target in the cell under 4 Analysis of the CA-CFAR detector
test. That is Di is equal to one if hypothesis H1 is decided
to be true and equal zero if hypothesis H0 is decided to be In the CA-CFAR detector, the clutter power is estimated by
true, respectively, of the i eme channel. In the second step, the summing the contents of N cells surrounding the cell under
processing of preliminary decisions D1 , D2 ,D3 , . . . , D L is test yi0 ; for this detector, we have
considered. The number of decisions where the first thresh-
1 
N
old is exceeded is counted and compared to the second
threshold M. A final decision is made on the presence of Zi = yi j (9)
N
j=1
a target according to
H1 the clutter level estimate Z i is the sum of N Pearson-distrib-

L
> uted random variables with γ as the dispersion parameter.
Di M (6)
< Hence, the pdf is [10]
i=1
H0 √
γ N 1 − N2zγ 2
and for the final decision, the overall probability of false p Z i (z i ) = √ e i (10)
2π z i3/2
alarm and probability of target detection are given by
L 
 L 4.1 The probability of false alarm
Qfa = (Pfa )m (1 − Pfa ) L−m (7)
m
m=M
The probability of false alarm (Pfa) for the CA-CFAR detec-
and tor is given by (See Appendix A)
L 
 
L 2 1
Qd = (Pd )m (1 − Pd ) L−m (8) Pfa = arctg √
CA
(11)
m π
m=M NT

123
226 SIViP (2012) 6:223–230

This expression of the Pfa has a closed form compared to the estimate Z of the clutter level is there taken to be the kth
the one given in [10]. The overall Pfa can be obtained by largest samples Y(k)
substituting (11) into (7) which yields
Z i = Yi,(k) (19)
L 
 
 m
L 2 1
QfaCA = arctg √ The pdf of the estimate Z i is given by [10]
m π NT
j=m

  N −k
  L−m N γ
Z i (z i )
p OS =k 2 √ −1 2k−1
2 1
× 1 − arctg √ (12) k zi
π NT   k−1
γ γ 1 − 2zγ2
using the binomial development, we obtain × 1− √ √ e i (20)
zi 2π z i3/2
L  L−m 
L  L −m where N is the number of reference cells, k is the rank of the
QfaCA = (−1)t
m t representative cell, and (x) denotes the cumulative den-
m=M t=0
  t+m sity function of the normal Gaussian distribution, which is
2 1
× arctg √ (13) given by
π NT
x
1 t2
4.2 The probability of detection (x) = √ e− 2 dt (21)

−∞
Assuming that the target in the cell under test is modeled as
a Swerling case II, the probability of detection is given by 5.1 The probability of false alarm
(See Appendix B)
γ √ From [10], using the fact that Erfc(x) = 1 − Erf(x) and the
PdCA = e− σ NT
(14)
binomial development, the probability of false alarm for
As previously, the same remark can be make for the closed the OS-CFAR detector in homogeneous Pearson clutter is
form expression of the probability of detection compared to expressed as
one given in [10]. The overall probability of detection is  k−1 

2 N k−1
L  PfaOS = k (−1)t
  √ m  γ√
 L−m π k t
L − σγ
Qd CA = e NT
1 − e− σ N T t=0

m=M
m +∞    N −k+t
z z z2
(15) × Erf √ Erf √ e− 2 dz (22)
2T 2
0
using the binomial development, we obtain
Hence, the overall probability of false alarm is obtained by
L 
 
L−m √
L L −m − γσ (t+m) N T
substituting (22) into (7), and using the binomial develop-
Qd CA = (−1)t e ment, we get
m t
m=M t=0
(16) L 
 L−m 
L  L −m
QfaOS = (−1)h
m h
as in [10], we define the generalized signal-to-noise ratio m=M h=0

(GSNR) as  k−1 
2 N  k−1
 ×⎣ k (−1)t
σ π k t
GSNR = 20 log10 (17) t=0
γ +∞
⎤ L+h
   N −k+t 2
z z z
× Erf √ Erf √ e− 2 dz ⎦
2T 2
0
5 Analysis of the OS-CFAR detector
(23)
In the OS-CFAR detector proposed by Rohling [12], the con-
tents of the range cells are first ordered according to their 5.2 The probability of detection
magnitude.
The single-pulse probability of detection for the OS-CFAR
Yi,(1) ≤ Yi,(2) ≤ · · · ≤ Yi,(k) ≤ · · · ≤ Yi,(N ) (18) detector in homogeneous Pearson background is given

123
SIViP (2012) 6:223–230 227

by [10]
 +∞
Tγ2
2 N −
PdOS = k e 2z 2 σ 2
π k
0
  N −k  k−1
z z − z2
× Erf √ Erfc √ e 2 dz (24)
2 2
which can be expressed using the fact that Erfc(x) =
1−Erf(x) and the binomial development as
 k−1 
2 N  k−1
PdOS = k (−1)t
π k t
t=0

+∞ Tγ2 2   N −k+t
− + z2 z
2z 2 σ 2
× e Erf √ dz (25)
2
0
Fig. 3 Probability of detection of the binary integration with
the overall probability of detection of the OS-CFAR detec- CA-CFAR detector versus GSNR for different second threshold
tor is then obtain by substituting (25) into (8), and using the M, N = 16, L = 8, overall design Pfa = 10−4
binomial development, we get
 L  L−m 
L  L −m
Qd OS = (−1)h
m h
m=M h=0

 k−1 
⎣ 2 N  k−1
× k (−1)t
π k t
t=0

+∞ 
⎤ L+h
Tγ2 2   N −k+t
− + z2 z
× e 2z 2 σ 2
Erf √ dz ⎦
2
0
(26)

6 Numerical results

In this section, we present the numerical results of the Fig. 4 Probability of detection of the binary integration with
above analysis and compare the performances of the binary OS-CFAR detector versus GSNR for different second threshold
CA-CFAR integrator and binary OS-CFAR integrator in the M, N = 16, L = 8, overall design Pfa = 10−4
presence of homogeneous Pearson-distributed clutter.
To get the overall probability of detection, we first obtain that to achieve the minimum GSNR occurs when the M equal
the threshold multiplier T of the two detectors considered by 6 (6 out-of 8) over a wide range of overall probabilities of
solving numerically (13) and (23), respectively, for a desired detection for both CFAR binary integrator considered.
overall probability of false alarm and different M-out-of-L In Fig. 7, we compare the detection performances of the
rules. Then, the value of T is used to compute the proba- CA-CFAR and the OS-CFAR binary integrators for the opti-
bilities of detection of (17) and (26). Figures 3 and 4 show mum decision rule obtained in the previous analysis. We
the detection performances for an overall probability of false observe that the OS-CFAR binary integrator yields better
alarm equal to 10−4 , a number of references cells N = 16, a detection performance than the CA-CFAR binary detector in
number of pulses L = 8, (k = 3N /4 for OS-CFAR only), homogeneous Pearson-distributed clutter.
and different decision rules namely the 2 out-of 8, 6 out-of
8, and 8 out-of 8. The results show that for both detectors,
the 6 out-of 8 decision rule exhibits better performances than 7 Conclusions
the other decision rules for the Pearson distribution clutter.
To confirm this results, Figs. 5 and 6 show the minimum In this paper, we have presented a theoretical analysis of the
required GSNR versus the second threshold M, we observe performances of the CA-CFAR and OS-CFAR binary inte-

123
228 SIViP (2012) 6:223–230

Fig. 7 Comparison of detection performance of the binary integration


Fig. 5 Minimum required GSNR versus the second threshold M for with CA-CFAR and OS-CFAR detectors for the optimum decision rule,
the binary integration with CA-CFAR detector, N = 16, L = 8, overall N = 16, L = 8, overall design Pfa = 10−4
design Pfa = 10−4

Appendix A

In this Appendix, we present the derivation the Pfa of


the binary integrator with CA-CFAR detector in homoge-
neous Pearson-distributed clutter. The test statistic (5) can be
rewriten as

H1
Yi0 >
R= T i = 1, . . . , L (A.1)
Zi <
H0

the probability of false alarm is then given as

P f a C A = Pr ob(R > T |H0 )


+∞
Fig. 6 Minimum required GSNR versus the second threshold M for = p R (r ) dr (A.2)
the binary integration with OS-CFAR detector, N = 16, k = 12, L = 8,
overall design Pfa = 10−4 T

where p R (r ) is the pdf of the random variable R which is the


ratio of two random variables given by [13]

+∞
gration detectors in the presence of spiky clutter modeled as
a Pearson distribution. The analysis showed that the binary p R (r ) = pYi0 (r.z i ) p Z i (z i ) |z i | dz i (A.3)
integration improves the detection performances compared 0
to the single-pulse (L = 1) detectors, and the improvement
is obtained for optimum M out-of L. In our case, this opti- where pYi0 (yi0 ) and p Z i (z i ) are the pdf of the CUT and Z i
mum occurs for M = (3/4)L. Finally, we showed that the the estimate of the clutter, respectively. We assume under
OS-CFAR binary integrator has better performances than the hypothesis H0 that the clutter is homogeneous and the
CA-CFAR binary integrator for the optimum decision rule N + 1 reference range samples Yi0 , Yi1 , . . . , Yi N are inde-
in homogeneous Pearson-distributed clutter. pendent and identically distributed and follow the Pearson

123
SIViP (2012) 6:223–230 229

distribution given in (4). Substituting (4) into (A.3), we obtain given by

√ +∞ Pd CA = Pr ob(Yi0 > T.Z i |H1 )


γ2 N 1 − γ2
(N + r1 ) +∞
p R (r ) = e 2z i
dz i (A.4)
2π r 3/2 r2 = Pr ob[Yi0 > T.Z i |Z i , H1 ]. Pr ob[Z i |H0 ]dz i
0
0
⎛ ⎞
γ2 +∞ +∞
setting x = 2 z i (N + r ), we get
1
⎜ ⎟
= ⎝ pYi0 (yi0 |H1 )dyi0 ⎠ p Z i (z i )dz i (B.1)
√ +∞ 0 T.z i
N −x
p R (r ) = √ e dx (A.5) As in [10], we assume that the target in the cell under test
π (1 + r N ) r
0 is modeled as a Swerling case II, with σ as the parameter of
 +∞ the fluctuating target, and that the target power dominate the
or the integral 0 e− x dx = 1, hence the PDF of the ran- clutter background in the cell under test, the probability of
dom variable R is given by detection is given by
√ 1 − yi02
N pYi0 (yi0 |H1 ) = e 2σ (B.2)
p R (r ) = √ (A.6) 2σ 2
π (1 + r N ) r we substitute (B.2) and (10) into (B.1), we get

∞ Tγ2 N z i2
to obtain the Pfa, we substitute (A.6) into (A.2), we get 2N − +
2z i2 σ 2 2
Pd CA
= e dz i (B.3)
+∞ √ π
0
1 N
PfaCA = √ dr (A.7) to evaluate the integral, we use (eq. 1.3.3.20) [14], hence the
π (1 + r N ) r
T probability of detection is

√ γ
setting u = N r , we obtain PdC A = e− σ NT
(B.4)

+∞
2 1 References
Pfa CA
= du (A.8)
π√ (1 + u 2 )
NT 1. Finn, H.M., Johnson, R.S.: Adaptive detection mode with thresh-
old control as function of spatially sampled clutter level esti-
the evaluation of the integral gives mates. RCA Rev. 29, 414–463 (1968)
2. Weiner, M.A.: Binary integration of fluctuating targets. IEEE
2 π √  Trans. Aerosp. Electron. Syst. 27(1), 11–17 (1991)
PfaCA = − Arctg NT (A.9) 3. Himonas, S.D.: CFAR integration processors in randomly
π 2 arriving impulse interference. IEEE Trans. Aerosp. Electron.
Syst. 30(3), 809–817 (1994)
using the fact that 4. Schleher, D.C.: Radar detection in Weibull clutter. IEEE Trans.
Aerosp. Electron. Syst. 12(6), 736–743 (1976)

1 π 5. Goldstein, G.B.: False alarm regulation in log-normal and weibull
Arctg + Arctg(x) = (A.10) clutter. IEEE Trans. Aerosp. Electron. Syst. 9(1), 84–92 (1973)
x 2 6. Ward, K.D.: Compound representation of high resolution sea clut-
ter. Electron. Lett. 17, 561–563 (1981)
Hence, the probability of false alarm for the CA-CFAR in 7. Watts, S.: Radar detection prediction in sea clutter using the com-
distributed Pearson clutter is pound K-distribution model. IEE Proc., Part F 132, 613–620 (1985)
8. Pierce, R.D.: RCS characterisation using the alpha-stable distri-
 bution. In: Proceedings of the IEEE national Radar Conference,
2 1
PfaCA = Arctg √ (A.11) pp. 154–159 (1996)
π NT 9. Pierce, R. D.: Application of the positive alpha-stable distribution.
In: IEEE Signal Processing Workshop on Higher-Order Statistics,
pp. 420–424. Banff, Alberta, Canada (1997)
10. Tsakalides, P., Trinci, F., Nikias, C.L.: Performance assessment
Appendix B of CFAR processors in Pearson-distributed clutter. IEEE Trans.
Aerosp. Electron. Syst. 36(4), 1377–1386 (2000)
11. Kuruoglu, E.E.: Analytical representation for positive alpha stable
In this Appendix, we derive the probability of detection of the
densities. In: Proceedings of the IEEE International Conference
binary integrator with CA-CFAR detector in homogeneous in Acoustics, Speech and Signal Processing, vol. 6, pp. 729–773
Pearson-distributed clutter. The probability of detection is (2003)

123
230 SIViP (2012) 6:223–230

12. Rohling, H.: Radar CFAR thresholding in clutter and multiple 14. Prudnikov, A.P., Brychkov, Y.A., Marichev, O.I.: Integrals and
target situations. IEEE Trans. Aerosp. Electron. Syst. 19(4), 608– Series, Volume 1: Elementary Functions. Gordon and Breach sci-
621 (1983) ence publishers, New York (1986)
13. Papoulis, A.: Probability, Random Variables and Stochastic Pro-
cesses. McGraw-Hill, New York (1984)

123

You might also like