Professional Documents
Culture Documents
Laroussi 2006
Laroussi 2006
Abstract: The problem of adaptive constant false alarm rate detection of a pulse-to-pulse partially
correlated target with 2K degrees of freedom in pulse-to-pulse partially Rayleigh correlated clutter
and multiple-target situations is addressed. Both the target and the clutter covariance matrices are
assumed to be known and are modelled as first-order Markov Gaussian processes. An exact
expression for the probability of false alarm (Pfa) for the mean level detector is derived. It is
shown that it depends on the degree of pulse-to-pulse correlation of the clutter samples. The
probability of detection (Pd) is shown to be sensitive to the degree of correlation of the target
returns and the degree of correlation of the clutter returns as well. Swerling’s well-known cases
I, II, III and IV are handled as extreme limits of the proposed model.
1 Introduction and overview of previous work four degrees of freedom. In this case, the limiting bounds
of complete correlation and complete decorrelation of the
Various models combining correlated target and/or corre- target returns yield the Swerling models III and IV. The
lated clutter returns have been widely used in the radar work done in [1, 2] used a fixed threshold detection. It is
systems literature [1 –10]. From the experimental data, known that radar detectors with fixed threshold cannot
clutter is modelled by either Rayleigh, log-normal, maintain a constant false alarm rate, and thus adaptive
Weibull or K-distribution [11 – 16]. It is known that although threshold detection is considered. Kim and Lee [3] and
the probability density function (pdf) of the target signals Han [4] exploited the results in [1, 2] and used the residues
from modern high-resolution radar systems does not to develop the generalised order statistics (GOS) CFAR
exactly agree with the concept of chi-square pdf, we can detector and the OS-CFAR detector, respectively.
always approximate it by the chi-square function, using Considering the case of partially correlated targets returns,
the central limit theorem, when integrating many pulses. Himonas and Barkat [5, 6] proposed different architectures
Dealing with either single pulse and uncorrelated data of time diversity and distributed CFAR detectors to mini-
sample models or multiple pulse or even distributed mise the effect of the correlation factor among the received
architectures with correlated data sample models and target pulses. In [5], they first considered the detection per-
multiple-target situations, aims for improved detection formance of the generalised censored mean level detectors
while maintaining a constant false alarm rate, have led to (GCMLD), which is known to assume no prior knowledge
the consideration of several types of adaptive constant of the number of interfering targets, for time diversity and
false alarm rate (CFAR) detectors. In [1 – 6], cases of non- multiple-target situations. Then, they proposed in [6] a
adaptive and adaptive detectors, where the target returns distributed architecture for the generalised two-level
are partially correlated, were investigated. In [1], Kanter (GTL)-CMLD CFAR when a clutter power transition is
has studied the detection performance of a non-coherent present in the range-resolution cells. They showed that
integration detector accumulating M-correlated pulses an appropriate choice of the fusion rule yields a detection
from a Rayleigh target with two degrees of freedom. performance to be less sensitive to the degree of correlation
Complete correlation and complete decorrelation of the among target returns.
target returns yielding Swerling models I and II, respect- The problem of Rayleigh targets and partially correlated
ively, were treated as extreme cases of the target correlation clutter has been treated in [7 –9]. In [7], Himonas and
coefficient. The noise was assumed to be uncorrelated both Barkat proposed the generalised CA-CFAR detector that
spatially and temporarily. Weiner [2] extended the work in adapts not only to changes in the clutter level but also to
[1] by deriving exact expressions for the probabilities of changes in the clutter covariance matrix, that is, they con-
detection for partially correlated chi-square targets with sidered the case where the correlated statistics of the
received clutter returns are unknown. In their analysis,
# IEE, 2006 they assumed a single pulse, processing a Swerling I type
IEE Proceedings online no. 20045057 target embedded in a spatially correlated clutter, whereas
doi:10.1049/ip-rsn:20045057
in the mean level (ML)-CFAR detection, we treat multiple
Paper first received 27th July 2004 and in revised form 10th November 2005
pulses, processing pulse-to-pulse correlated chi-square
T. Laroussi is with the Département d’Electronique, Université de Constantine,
target with 2K degrees of freedom in a pulse-to-pulse corre-
Constantine 25000, Algeria lated Rayleigh clutter. They also assumed Markov’s type
M. Barkat is with the Department of Electrical Engineering, American homogeneous clutter of much higher power than thermal
University of Sharjah, PO Box 26666, Sharjah, United Arab Emirates noise. Al Hussaini and El Mashade [8] studied in the detec-
E-mail: toufik_laroussi@yahoo.fr tion performance of the CA-CFAR and OS-CFAR detectors
44 IEE Proc.-Radar Sonar Navig., Vol. 153, No. 1, February 2006
considering a non-conventional time-diversity technique. number of interfering targets present in the range-resolution
Their schemes processed correlated clutter in the presence cells on the detection performance of the detector.
of interfering targets. El Mashade [9] introduced the case Swerling’s cases I, II, III and IV are handled as extreme
where the use of the moving target indication (MTI) filter limits of the proposed model. As a conclusion, Section 5
output introduces a correlated clutter, even though if its summarises the results of our contribution.
input signal is uncorrelated. He analysed the CA-CFAR
family for the multiple-target environments scenario. 2 Statistical model and assumptions
The problem of both partially correlated target echoes and
partially correlated clutter echoes was first introduced by The received signal r(t) is processed by the in-phase and
Farina and Russo [10]. They considered the problem of quadrature phase channels as shown in Fig. 1. Assuming a
detecting pulse-to-pulse partially correlated target returns correlated chi-square target with 2K degrees of freedom
in pulse-to-pulse partially correlated clutter returns. The embedded in correlated noise, the in-phase and quadrature
target model was assumed to be as a coherent Gaussian dis- phase samples fuijg and fvijg at pulse i and range cell j,
tributed process having a known covariance matrix. They respectively, i ¼ 1, 2, 3, . . . , M and j ¼ 1, 2, 3, . . . , N,
suggested a batch detector and a recursive detector to esti- are observations from Gaussian random variables. M and
mate the clutter covariance matrix. They showed, by Monte N are the number of radar processed pulses and the
Carlo simulations, that the threshold multiplier is indepen- number of reference range cells, respectively. Assuming
dent of the degree of pulse-to-pulse correlation among the that the total clutter plus noise power is normalised to
clutter samples for a number of range cells around ten. unity, the integrated output of the square law detector, is
The presence of interfering targets in the reference cells M n
1X o
degrades considerably the performance of the ML detector.
qj ¼ u2ij þ v2ij j ¼ 1; 2; . . . ; N ð1Þ
To alleviate this problem, a lot of detectors have been 2 i¼1
proposed in the literature. The aim of these algorithms
is to show how robust a CFAR detector can be in where
multiple-target situations by taking into account the non-
homogeneous background such as clutter edge and/or uij ¼ aij þ cij i ¼ 1; 2; . . . ; M and j ¼ 1; 2; . . . ; N
interferences, either analytically or by simulation. ð2aÞ
In summary, all of the detectors listed above did not show
the effect of the pulse-to-pulse partially Rayleigh correlated vij ¼ bij þ dij i ¼ 1; 2; . . . ; M and j ¼ 1; 2; . . . ; N
clutter on the analytical expression of the probability of ð2bÞ
false alarm Pfa . Furthermore, they did not consider the
general case of a pulse-to-pulse partially correlated chi- Or, by accommodating M 1 vectors
square target with 2K degrees of freedom. All of this has 1
led us to introduce, in this paper, a mathematical model to qj ¼ jU j j2 þ jV j j2 j ¼ 1; 2; . . . ; N ð3Þ
represent the general case of detecting a pulse-to-pulse 2
chi-square partially correlated target with 2K degrees of where
freedom embedded in a pulse-to-pulse partially Rayleigh
correlated clutter and Rayleigh but uncorrelated thermal U j ¼ Aj þ C j j ¼ 1; 2; . . . ; N ð4aÞ
noise. The paper is organised as follows. In Section 2, we V j ¼ Bj þ Dj j ¼ 1; 2; . . . ; N ð4bÞ
formulate the statistical model and set the assumptions.
Then, in Section 3, we investigate the effect of the temporal j.j is the Euclidean norm and T is the transpose operator.
correlation of the clutter on the probability of false alarm The aij , cij , bij and dij represent the in-phase and quadra-
and particularly on the CFAR parameter T for the ML detec- ture phase samples of the correlated clutter and the thermal
tor employing M-pulse non-coherent integration. Next, in noise, respectively. The clutter samples are assumed to be
Section 4, we first show, by means of computer simulation, first-order Markov processes with zero mean and variance
the effect of the degree of correlation of the clutter returns sc2 and are identically distributed but correlated (IDC)
and the number of processed pulses M, on the threshold from pulse-to-pulse and uncorrelated from cell-to-cell.
multiplier T and on the probability of false alarm. Then, The thermal noise samples are assumed to be independent
by deriving sets of detection curves, we show the effect of and identically distributed (IID) random variables with
the degree of correlation of the target returns, the degree zero mean and variance sn2 from pulse-to-pulse and from
of correlation of the clutter returns, the number of degrees cell-to-cell. The covariance matrices of the clutter and
of freedom, the number of processed pulses and the noise processes are denoted Lc and Ln, respectively.
target and ICR ¼ SCR. Such a problem can be overcome The curves shown in Figs. 9 – 11 have been evaluated at
just by increasing the number of processed pulses integer values of K, M and NI, respectively, and interpolated
M. Besides, we may note that as these situations happen to non-integer values. First, note that when the number of
for values of Pd , 0.4, they are of no great interest to the pulses increases, the undesirable situations occur at higher
performance of the ML-CFAR detector. numbers of interferers than for the case of M ¼ 2 pulses.
With the same operating conditions as for Fig. 7, we plot In Fig. 9, we plot the probability of detection Pd in terms
Fig. 8, where rt is substituted by rc . Here also, an increase of K for rt ¼ rc ¼ 0.5 and M ¼ 2, 4 and 8. Note that, in
of the number of interferers seems to help the detection of the presence of interferers, two pulses are not appropriate
targets with lower number of degrees of freedom. In other for a good detection even for primary targets with a large
words, whenever M ¼ 2 pulses, the ML-CFAR detector number of degrees of freedom. Very large values of K cor-
becomes vulnerable to cases in which the number of inter- respond to the non-fluctuating limit. In Fig. 10, we plot the
ferers is over 1. Consequently, here also, one has to increase probability of detection Pd in terms of M for rt ¼ rc ¼ 0.5
the number of pulses so as to avoid situations in which and K ¼ 1, 2 and 4. Note that in the multiple-target cases,
targets with lower degrees of freedom become more detect- more pulses are needed to achieve the same desired Pd
able than those with higher degrees. than in the no interferers’ case. For instance, when M 4
pulses, the ML-CFAR detector behaves normally in the
Fig. 9 Simulated probability of detection against number of Fig. 11 Simulated probability of detection against number of
degrees of freedom at SCR ¼ 15 dB, for N ¼ 16, rt ¼ rc ¼ 0.5, interferers at SCR ¼ 15 dB, for N ¼ 16, rt ¼ rc ¼ 0.5,
ICR ¼ SCR and Pfa ¼ 1024 ICR ¼ SCR and Pfa ¼ 1024