Documente Academic
Documente Profesional
Documente Cultură
Signal Processing
journal homepage: www.elsevier.com/locate/sigpro
a r t i c l e i n f o abstract
Article history: In this paper a multistatic radar system with n transmitters and one receiver is
Received 13 November 2009 considered and several constant false alarm rate (CFAR) algorithms for detection are
Received in revised form introduced. The decision statistics of the proposed detectors are the sum of the n largest
26 March 2010
returning signals in an array of N +n range cells. It is shown that the proposed decision
Accepted 1 June 2010
Available online 16 June 2010
statistic satisfies the CFAR condition and it is justified that the sum of the largest
returning signals is the optimal statistic. The proposed CFAR detectors are simulated
Keywords: both in homogenous and non-homogenous backgrounds and their performances are
Multistatic radar compared with the performance of a monostatic radar of higher power. It is shown that
CFAR detection
a multistatic radar outperforms a monostatic radar under equal transmit power
Homogenous background
condition.
Non-homogenous background
& 2010 Elsevier B.V. All rights reserved.
0165-1684/$ - see front matter & 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.sigpro.2010.06.003
V. Amanipour, A. Olfat / Signal Processing 91 (2011) 28–37 29
we assume that the transmitted waveforms are the same. .Finally, Section7 summarizes the main results of the
In comparison with a fixed monostatic radar, these n paper.
transmitters are assumed to be the same, except for their
power. They are assumed to have a fixed fraction of the 2. CFAR model
power of the monostatic transmitter.
One difficulty in dealing with multistatic radar was the We will assume that n transmitters and one receiver
lack of a CFAR detection algorithm in the literature. The form our multistatic radar. Note that using three trans-
earlier work only includes the case where the radar mitters the exact location of the target may be detected.
system includes n spatially distributed detectors, or n The transmitters and receiver are assumed to be synchro-
receivers, such that each detector sends its information nized and the square-law detector output for any range
about the presence or absence of the target to a data cell is exponentially distributed with probability density
fusion center. More specifically, among the earlier works function (pdf) [11]:
on this subject we should mention that Barkat and
f ðxÞ ¼ ð1=2lÞexpðx=2lÞ, xZ0 ð1Þ
Varshney [4,5], have developed the theory of decentra-
lized CFAR detection. Elias-Fuste et al. [6] have extended The null hypothesis oH0 of no target in a range cell
the data fusion rule [5]. The performance of distributed would mean that l is the total background clutter-plus-
CFAR detection in homogenous and non-homogenous thermal noise power, which will be denoted by m. Under
backgrounds are analyzed in [7–9]. Liu et al. [10–12] the alternative hypothesis H1 of presence of a target, l is
proposed a novel data fusion by using genetic algorithm. equal to m(1+ S), where S is the average signal-to total
For a monostatic radar CFAR algorithms are studied noise ratio (SNR) of a target. In other words, we are
extensively [13–15] and also adaptive robust detectors assuming a Swerling I model for radar returns from the
are introduced [16]. Sheikhi and Zamani [17] have target and Gaussian statistic for the background. Sum-
developed a CFAR detector for MIMO radars when the marizing the above assumptions we have
transmitters are assumed to transmit orthogonal wave- (
m under H0
forms. This assumption reduces the system of n trans- l¼ ð2Þ
mð1 þ SÞ under H1
mitters and m receivers to mn independent bistatic radars,
which may basically be studied independently. Their We will test windows of N + n cells in order to detect
CFAR algorithm is a generalization of cell averaging CFAR the target. Depending on the locations of the transmitters
algorithm which is simulated only in homogenous back- and the receiver (i.e. the geometry of the multistatic
ground. radar) the signals sent by the n transmitters will return
The main theoretical achievement of this paper is the from the target in a particular pattern which also depends
introduction of a family of CFAR detection algorithms on the coordinates of the target. In particular, if two
(called homogenous CFAR algorithms of order one) for a signals from two different transmitters are sent simulta-
multistatic radar. In fact, we show that a family of neously, some delay between their returns will be
statistics, called homogenous statistics of order one, may observed by the receiver. The number N +n is chosen so
be used in our algorithm so that the result has the CFAR that the maximum possible delay between any such two
property. We then give some theoretic indications that returning signals (and for any geometric location of the
one particular statistic should have a relatively better target) is at most (N + n)T, where T is the delay propor-
performance. tional to the radar range resolution. The observations in
We simulate the performance of a multistatic radar the N +n cells are assumed to be statistically independent,
composed of n transmitters and one receiver using these including the cell under test. Let li for i=1,2,y,N +n be the
CFAR detection algorithms and compare it with the corresponding value of l for the ith cell. If in such a
performance of a monostatic radar as the power ratio window the target is present (i.e. under H1 hypothesis) for
varies. In particular, when there are three transmitters n of the N + n cells we will have li = m(1+ Si) and for the rest
and the power ratio is equal to 1/3, or 1/2 we compare the of them li = m. Let X1, X2,y,XN + n be the exponential
performances in both homogenous and non-homogenous random variables which model the values of the range
backgrounds for proposed algorithms. In reasonable cells. We will sort these variables in the increasing order
signal to noise ratio the performance of the multistatic to obtain Y1,Y2,...,YN + n. We will assume that the n largest
radar is better, even for power ratio equal to 1/3. values, i.e. YN + 1,YN + 2,...,YN + n, correspond to the returns
Moreover we show, for a fixed signal to noise ratio, how from the target. Fix a statistic Y= f (YN + 1,YN + 2,...,YN + n). The
the detection probability increases with the growth of rest of Yi in CFAR processor form the statistic Z. A target is
power ratio. declared to be present if Y exceeds a threshold TZ. Here
The remainder of the paper is organized as follows. In T is a constant scale factor used to achieve a desired
Section 2, the problem is formulated and we introduce our constant false alarm probability.
CFAR algorithms for detection, which includes a cell- The n synchronized transmitters send identical signals
averaging CFAR, an order statistic CFAR and a trimmed simultaneously. So the receiver does not need to distin-
mean CFAR. Then in Sections 3 and 4 the probability of guish which returning signal is associated with what
false alarm for these algorithms are computed. In Section 5, transmitter. Of course, we may send the signals using
we give an argument why the chosen statistic (the sum of orthogonal codes in order to be able to distinguish them
largest returning signals) in the algorithm is a logical in the receiver. Also note that for each returning signal
choice. Simulation results are presented in Section 6 corresponding to a particular transmitter, we could have
30 V. Amanipour, A. Olfat / Signal Processing 91 (2011) 28–37
chosen one cell from the window of length N and test all
N N
the ... ¼ Nn possible n-component vectors to
1 1
form the ‘‘cell under test’’. So, the complexity and
computational cost of the process would significantly
increase. However, the proposed algorithm does not force
such a complexity (and cost) for forming the ‘‘cell under
test’’. Yet, combining both techniques (i.e. using orthogo-
nal codes and the fast algorithm for forming the cell under
test) we may locate the precise position of the target if the
number of transmitters is more than two.
Among the CFAR processors we will examine a class of
CFAR processors called homogenous detectors of order
one. Examples of such CFAR processors include OS-CFAR,
CA-CFAR, TM-CFAR and Censor Mean Level detectors. OS-
CFAR has a relatively good performance, both in
homogenous and non-homogenous environments for
monostatic radar, while CA-CFAR is the optimal choice Fig. 2. Block diagram of proposed homogenous CFAR detector of order
in homogenous environments. We will compare the one.
fk ðzÞ ¼ fZ ðzÞ
Nk þ n þ 1 k1
Fig. 1. Block diagram of proposed CA-CFAR detector. The returning k Nþn z z
signals from the target are n of the N + n values X1, X2,yXN + n, which
¼ exp 1exp
2m k 2m 2m
would be the n largest values YN + 1, YN + 2,y,YN + n under the assumption
of the algorithm. ð5Þ
V. Amanipour, A. Olfat / Signal Processing 91 (2011) 28–37 31
And using (5) the probability of false alarm is parameter m such that
2 3
Z 1 Z1 fY ðaÞ ¼ ð1=mÞffix ða=mÞ ð11Þ
Nþn 6 7
Pfa ¼ k ½expðtÞNk þ n þ 1 ½1expðtÞk1 4 fY ðyÞ dy5 dt
k 0 From this formula we may compute
2mtT
Z1 Z1 Z1
ð6Þ 1 y x ¼ y=m
fY ðyÞ dy ¼ ffix dy ¼ ffix ðxÞ dx ð12Þ
where fY(y) is the probability density function for the m m
2mtT 2mtT 2tT
random variable Y= f (YN + 1,YN + 2,y,YN + n). In order for this
problem to be CFAR, the value of this integral should be If we replace (12) in (6), we obtain
2 3
independent of m.
Nþn
Z 1 Z1
We claim that for any homogenous function f : Rn -R Pfa ¼ k ½expðtÞNk þ n þ 1 ½1expðtÞk1 4 ffix ðxÞ dx5 dt
k 0
of order 1, the probability of the false alarm given by 2tT
dA’ are the volume forms for the hyper-surfaces Since we are not able to distinguish the values
c(x1,y,xn) =z and c(s1,y,sn)= z/m, respectively. S1,S2,y,Sn we make a further assumption that
The functions hf(x) do not depend on the noise parameter S1 ¼ . . . ¼ Sn ¼ S. Under these assumptions, there will be
m and depend only on the coefficients bi,i=1,y,N and the a choice of a set I =(i1 oi2 o?oin) of n indices such that
numbers N,n (and of course, on the linear function f). Using {Xi1,Xi2,y,Xin} ={YN + 1,YN + 2,y,YN + n}. Let us denote the
these formulas for ff(z)we can obtain the probability of false -
alarm in homogenous CFAR algorithm of order one, corre- vector (Xi1,Xi2,...,Xin) by XI . The set A may be written as a
sponding to the function f : RN -R: disjoint union of the sets AI of events, where AI is the
- -
2 3 subset of A where Y corresponds to XI .
Z 1 Z1 Z 1 Z1
6 7 Under both hypothesis H0 and A, the random variables
Pfa ¼ hf ðtÞ4 fY ðyÞ dy5 dt ¼ hf ðtÞffix ðsÞ ds dt ð19Þ
0 0 Xij are independent exponential random variables. In case
2mtT 2tT
of the hypothesis H0 the parameters of all of them are
Here fY(y) is the probability density function for the m and in the second case (under the hypothesis A) the
random variable Y= f(YN + 1,YN + 2,y,YN + n), and ffix(s) is parameters of these exponential random variables are
defined as before. Clearly, this integral is independent of equal to mi = m(1+ S), i= 1,y,n. We may thus define the
the noise parameter. following expression in place of
These computations show that if in the algorithm we
use any linear combination of random variables f- ðr1 ,r2 ,. . .,rn Þ
ðX I 9AI Þ
KI ðr1 ,r2 ,. . .,rn Þ ¼ ð22Þ
Y1,Y2,y,YN to define the statistic Z, the result would be a f- ðr1 ,r2 ,. . .,rn Þ
ðX I 9H0 Þ
CFAR algorithm. This proves the CFAR condition for CA-
CFAR, TM-CFAR and CMLD, and more generally for any We have just changed the variables y1,y2,y,yn from
homogenous CFAR detector of order one. Eq. (21) to r1,y,rn since these variables are no longer
subject to the condition y1 oy2 o?oyn. We have
5. An optimal choice of the function f fðXi 9AÞ ðr1 Þ. . .fðXin 9AÞ ðrn Þ
KI ðr1 ,r2 ,. . .,rn Þ ¼ 1
as a set {Xi1,Xi2,y,Xin} ={YN + 1,YN + 2,y,YN + n} while the Note that KI Z t if and only if
random variables Xij do not have an order, and are
X
n
1þS
consequently independent. ri r 2mn logðð1 þ SÞtÞ ¼ gðtÞ ð24Þ
In order to use the Neyman–Pearson test [19] for i¼1
S
finding the best possible statistic Y = f (YN + 1,YN + 2,y,YN + n) Then we have
we need to compute the likelihood ratio:
f - ðy1 ,. . .,yn Þ r tf - ðy1 ,. . .,yn Þ ð25Þ
fððYN þ 1 ,YN þ 2 ,...,YN þ n Þ9H1 Þ ðy1 ,y2 ,. . .,yn Þ ðY 9AI Þ ðY 9H0 Þ
Lðy1 ,y2 ,:::,yn Þ ¼ ð20Þ
fððYN þ 1 ,YN þ 2 ,...,YN þ n Þ9H0 Þ ðy1 ,y2 ,. . .,yn Þ for some given y1 oy2 o? oyn if and only if for all the
subset I= (i1 oi2 o? oin) of indices (corresponding to the
Then, for any given threshold t, we should describe the
n largest values for the random variables) and some
inequality L (y1,y,yn)4 t as a relation of the form
re-ordering r1,y,rn of y1,y,yn the following relation is
f (y1,y,yn)4 f(t) for an explicit function f (y1,y,yn) and
satisfied:
some function f(t) of the threshold t. The statistic Y =f
(YN + 1,YN + 2,y,YN + n) will then be an optimal statistic in our f- ðr1 ,. . .,rn Þ r tf - ðr1 ,. . .,rn Þ ð26Þ
ðX I 9AI Þ ðX I 9H0 Þ
algorithm.
Computing the above likelihood ratio L (y1,y,yn) is This is equivalent to the simple relation:
quite complicated for all possible events. However, on a r1 þ r2 þ þrn r gðtÞ ð27Þ
relatively large subset of all events the outcome is
relatively simple. Let A be the subset of H1 consisting of Since r1 + r2 +?+ rn =y1 + y2 + ?+ yn, this is satisfied if we
- have
the events where Y ¼ ðYN þ 1 ,YN þ 2 ,. . .,YN þ n Þ corresponds to
the signals returning from the target. As the values of y1 þ y2 þ þ yn rgðtÞ ð28Þ
V. Amanipour, A. Olfat / Signal Processing 91 (2011) 28–37 33
Thus over the subset A of the events the function and for either of multistatic OS-CFAR, CA-CFAR,
f(y1 + ?+yn) =y1 + ?+ yn gives the optimal choice for the and TM-CFAR algorithms, as well as for the monostatic
statistic Y =f (YN + 1,YN + 2,y,YN + n). radar.
Throughout the simulations, a Swerling I model for
6. Simulation results fluctuating point target is used and for a comparison
between multistatic radar and monostatic radar we
assume that radar cross section (RCS) of target and
In this section, we will give through simulations a
bistatic cross section (BCS) of target are the same.
comparison between the performance of multistatic
radar, processed with either of OS-CFAR, CA-CFAR or
TM-CFAR algorithms of this paper, and the performance of 6.1. Homogenous background
monostatic radar. We will consider both the homogenous
and non-homogenous environments. Figs. 4 and 5 show detection probability versus signal
In case of multistatic radar, we will assume n =3, i.e. to noise ratio of target in homogenous background when
there are three transmitters. The transmitters of multi- the power of multistatic transmitters is 1/3 of the power
static radar and monostatic radar are the same except for of monostatic transmitter. In Fig. 6 power ratio is set to
their power. The power of each transmitter in multistatic 1/2. The SNR quoted in the figures refers to the signal to
radar will be assumed to be a fraction of the power of the noise ratio assumed for the monostatic radar. The
monostatic radar. The positions of transmitters are corresponding SNR for the multistatic radar is typically
assumed to be at the vertices of a regular triangle, and much smaller. In fact according to the radar equations in
the receiver would be located at the center of this triangle bistatic and monostatic cases [18], the ratio of SNRmult(i)
(see Fig. 3). We will simulate the algorithm using two (the value of SNR for the ith transmitter in multistatic
different functions f : R3 -R. First we consider the radar) to SNRmono (the value of SNR for the monostatic
following statistic (multistatic 1): radar)can be expressed as
Y ¼ f ðYN þ 1 ,YN þ 2 ,YN þ 3 Þ ¼ YN þ 1 þYN þ 2 þYN þ 3 SNRmult ðiÞ R4 R2
¼ PRðiÞ 2 2 ¼ PRðiÞ 2 ð29Þ
and then we use the second statistic(multistatic 2) SNRmono RR RT ðiÞ RT ðiÞ
defined as
where PR(i) is the power ratio between the ith multistatic
Y ¼ f ðYN þ 1 ,YN þ 2 ,YN þ 3 Þ ¼ maxðW1 ,W2 ,W3 Þ transmitter and the monostatic transmitter, RR = R is the
¼ maxðYN þ 1 ,YN þ 2 YN þ 1 ,YN þ 3 YN þ 2 Þ distance between the target and the receiver and RT(i) is
the distance between the ith transmitter and the target.
We assume N=24, k=21 and false alarm probability is
Thus, according to (29), the average SNR of multistatic
Pfa =10 4. The parameters of TM-CFAR are set T1 =20 and
radar (compared to monostatic radar) is proportional to
T2 =1. Setting these parameters, we compute the scale factor 2
power ratio PR and the average of R RðiÞ2 . The average of
T for each of these two statistics (multistatics 1 and 2), R2 T
RT ðiÞ2
depends on the geometry of the transmitters and the
receiver. For the geometry shown in Fig. 3, that is used in
2
our simulations the average of R RðiÞ2 is approximately 0.6
T
which was computed by the Monte-Carlo method. So, in
the our simulations, for power ratios PR(i) equal to 1/3
radar is at least 25% of the power of the transmitter of modeled by exponential random variables with parameter
monostatic radar, detection probability becomes larger for m0 and in 24 m cells modeled by exponential random
multistatic radar. variables with parameter m1, with m1 4 m0. The target may
The high probability of detection in low SNR is what be either in or outside the clutter region, and the
should be regarded as a measure for a good and satisfying returning signals may have parameters m1(1+ Si) or
performance. As these figures suggest, starting from m0(1+ Si) accordingly. The expected noise parameter is
power ratio 1/2 the performance of the multistatic radar estimated by examining the values of m random variables
corresponding to multistatic 1 becomes better than the with parameter m0 and 24 m random variables with
performance of a monostatic radar. When the power ratio parameter m1, so the estimated parameter is expected to
is raised to 2/3 the probability of detection using a be some value m0 o mexp o m1 .
multistatic radar, even in low SNR, is almost 10% better If the cells under test correspond to the clutter region
than the probability of detection by a monostatic radar. (i.e. their noise parameter is m1) both the probability of
detection and the probability of false alarm will be
6.2. Non-Gaussian background increased, when compared with the homogenous back-
ground, since the actual noise parameter is larger than the
estimated noise parameter mexp . As the number of cells in
In the study of a non-Gaussian background, we have
the clutter region increase, the value of mexp becomes
also considered a clutter with heavy tailed distribution in
closer and closer to m1, the false alarm probability will be
Fig. 9. More precisely, the distribution of the background
decreased, and will converge to the false alarm probability
noise is assumed to be Weibull with shape parameter
in homogenous background. If the cell under test is
equal to 0.5, and the performance of multistatic and
outside the clutter region both the detection and false
monostatic radars are compared.
alarm probability will become less than the homogenous
As it may be observed from Fig. 9, the performance of
case, by a similar reasoning.
the multistatic radar in a Weibull clutter is similar to the
In order to compute the probability of false alarm in
performance of a monostatic radar, while the performance
the multistatic case, we will need to use the following
becomes better when the signal to noise ratio increases.
formula for conditional probability:
0.7
static structure, gets confused. Yet, our simulations using
0.6 the above formulas, which are illustrated in Fig. 10,
0.5 indicate that the performance of the multistatic radar
under the assumption of a clutter edge is almost as
0.4
multistatic1-sum (os)
good as the monostatic radar. In simulations shown
0.3 multistatic2-max (os) in Figs. 11 and 12 a clutter edge of 10 dB is assumed
0.2 monostatic (os) (typically larger than the signal to noise ratio assumed for
multistatic1-sum (ca)
the radar in homogenous background).
0.1 multistatic2-max (ca)
monostatic (ca) Fig. 13 illustrates the changes in detection probability
0 as the number of cells in the clutter region increases, for
10 12 14 16 18 20 22 24 26 28 30 different CFAR algorithms and fixed signal to noise ratio.
SNR (db) When the clutter region is small, OS-CFAR has a relatively
Fig. 9. Detection performance of monostatic radar versus multistatic
better performance for monostatic radar, while the
radars in Weibull clutter. performance of our multistatic CA-CFAR algorithm is
36 V. Amanipour, A. Olfat / Signal Processing 91 (2011) 28–37
-0.5
-1
-1.5
multistatic1-sum (os)
multistatic2-max (os)
monostatic (os)
-2
multistatic1-sum (ca)
multistatic2-max (ca)
monostatic (ca)
-2.5
1 2 3 4 5 6 7 8 9 10 11 12
No. of cells in clutter edge
Fig. 10. Total false alarm probability versus the number of cells in the
clutter region for 10 dB clutter edge. Fig. 12. False alarm probability versus the number of cells in the clutter
region for 10 dB clutter edge and CUT in the clutter region.
Fig. 11. False alarm probability versus the number of cells in the clutter
region for 10 dB clutter edge and CUT in the clutter region. Fig. 13. Probability of detection versus number of cells in the clutter
region for 10 dB clutter edge and CUT outside the clutter region.