Beruflich Dokumente
Kultur Dokumente
Eran Fishler, Alex Haimovich , Rick Blum , Len Ciminio , Dimtry Chizhik , Reinaldo Valenzuela
Model Classification Now, if the receiver uses a beamformer to steer towards di-
rection , then the output of the beamformer is
MIMO radar signal models can be classified into three gen-
eral groups: y = 1/ 2 a (0 ) r + v 0
Conventional radar array modeled with an array at = 1/ 2 a (0 ) a (0 ) bT (00 ) b (0 ) s + v 0 (10)
the receiver and a single antenna or an array at the
transmitter. The array elements are spaced at half- This model represents a bistatic radar where
wavelengths to enable beamforming and DF. bT (00 ) b (0 ) plays the role of the transmit antenna pat-
tern, whereas a (0 ) ha (0 )i is the receive antenna pattern.
MIMO radar for DF. Transmit antenna elements are 2 2
widely spaced to support spatial diversity aspects of Since, E [] = 0, E || = 2, |s| = 1/M, the signal
the target. Receiver array performs DF. power in (10), is given by
a (0 ) a (0 ) bT (00 ) b (0 ) s2 M N 2 . Note that the
MIMO radar for detection of multiple targets. In this
instantaneous signal power a (0 ) a (0 ) bT (00 ) b (0 ) s
scenario, there are multiple targets in the range cell
has a 22 distribution (chi-square with 2 degrees of freedom).
under test. Antennas at the transmitter are less sepa-
rated than in the previous scenario, such that scatter-
ers belonging to the same target are not individually MIMO Radar: Direction Finding
resolved. Yet, the separation is sufficient to resolve In MIMO radar for direction finding (DF), the transmit an-
multiple targets in the same range cell. tennas are sufficiently separated to meet the orthogonality
The first two signal models are detailed in the sequel. condition (5) for targets of interest. The columns of the
The third will be discussed in a subsequent publication. transmit matrix G meet the orthogonality condition in (1).
In contrast, elements of the receive array are closely sepa-
rated to enable DF measurements. Assume that the target is
Conventional Radar Array
at angle 0 with respect to the receive array normal. The re-
Conventional systems are systems in which the elements of ceive matrix is given by K = 1Q aT (0 ) . Since the goal
the transmitting and receiving arrays are closely spaced. At is to illuminate the target to achieve spatial diversity, phase
the transmitter, that means that the inter-element spacing shifts at the transmitter are set to zero, b (0 ) = 1M . From
does not meet (5) or, equivalently, that multiple elements (6), it follows that the channel matrix is given by
are contained within one target beamwidth. At the receiver,
1
the spacing is dr /2 to enable unambiguous estimation H = aT (0 ) , (11)
of the angle of arrival. 2
Let the target bearing with respect to the transmit and re-
ceive arrays be respectively, 00 and 0 . The transmit matrix Tm , of the M 1 vector , is de-
where the component
fined m = 1/ Q gm , and was defined previously.
is given by G = b (00 ) 1TQ . The receive matrix is given Due to the orthogonality among the transmit vectors gm , the
by K = 1Q aT (0 ) . It follows that the channel matrix is variates m are uncorrelated. Moreover, for Q , the
given by
random variables 1/ Q m are zero-mean, unit-variance
1
H = b (00 ) aT (0 ) , (8) (per dimension), i.i.d. complex normal.
2 The signal model is given by
where = 1/ Q 1T , and is the vector formed by the
T 1
T a (0 ) S + v
target gains q , = [0 , . . . , Q1 ] . By assumption, q are r =
zero-mean, unit-variance per dimension, i.i.d.; hence by the 2
M1
central limit theorem approaches a zero-mean, complex 1 X
2 = a (0 ) i si + v. (12)
normal distribution. Subsequently, the targets RCS || , 2
2 i=0
follows a 2 chi-square distribution with 2 degrees of free-
dom. Note that with this model, there is no diversity gain The signal model, with all normalization factors specified
in the target RCS. so far, ensures that the average transmitted power
With a conventional radar array, all antennas transmit P 2
M1
E 1/ 2 i=0 i si = 1. Conditioned on the target
the same waveform s. Beamforming at the transmitter is
vector , the received vector r is complex, multivariate nor- A common figure of merit for comparing the perfor-
1 2
mal with correlation matrix (2M ) kk a (0 ) a (0 ) + mance of different systems is the estimator mean square
1
SNR IN . error (MSE). The systems MSE depends on the exact es-
To gain better insight, consider a specific case with M = timation method, e.g., ML, MUSIC, beamforming, used. In
2, N = 1. The signal model is given by order to have a fair comparison between different systems,
for each system, we evaluate a lower bound on the perfor-
1 mance of all unbiased estimators.
r = (1 s1 + 2 s2 ) + v. (13)
2
If both antennas transmit the same waveform, s1 = s2 = s, Cramer-Rao Bound
the received signal is given by In what follows, we analyze the performance of a MIMO
1 radar used as an active direction finder with M 1 trans-
r = (1 + 2 ) s + v. (14) mitting elements. The received signal is given by the model
2
(12). Conditioned on , M
P
m=1 m si is a zero mean, white,
Since the channel parameters 1,2 are unknown at the re- complex normal random variable with variance (1/M ) ||||2 .
ceiver, it is impossible to take advantage of the target spa- In our model there are three unknown parameters, the direc-
tial diversity. This system fails to achieve the target diversity tion parameter , the target parameters , and the SNR. Let
sought. denoting the unknown parameters, = [, SNR,
the vector
Conversely, for orthogonal transmitted waveforms such ||||2 .
2 2
that s1 s2 = 0, |s1 | = |s2 | = 1/2, the received signal can The Cramer-Rao bound is probably the best known lower
be processed to yield the test statistic bound on the MSE of unbiased estimators [12]. Denote by
p(r|) the family of distributions of the received signal pa-
2 2
= |s1 r| + |s2 r| rameterized by the vector of unknown parameters . The
1 2 2
Cramer-Rao lower bound for estimating is given by,
= |1 | + |2 | + v 00 . (15)
4 2
log p(r|)
1
1
CRB() = J = E . (16)
2
Since for Q , the random variables |i | , i = 1, 2, T
have a 2 distribution, and they are i.i.d. (due to the orthog- We are interested only in the direction , whereas the
onality between g1 and g2 ), the target component in (15) others are nuisance parameters. We denote the correspond-
has a 24 (chi-square with 4 degrees of freedom) distribu- ing bound CRB(|), where the notation indicates the con-
tion. This is a consequence of the different and uncorrelated ditioning of the bound on the unknown parameters . As
RCS presented by the target to the different elements of the already noted, conditioned on , r is a complex normal ran-
transmitting antenna. Thus MIMO radar results in a diver- 1
dom vector with correlation matrix (2M ) ||||2 a()a ()
sity gain that manifests itself through a more advantageous 1
+SNR IN . The CRB conditioned on can be computed,
distribution of the target component in the received signal. and it is given by [13, 12],
" #
N 2M 4M 2
III. MIMO DF Analysis CRB(|) =
2L N SNR||||2
+ 2
(N SNR) ||||4
!1
In a radar DF system, an omnidirectional antenna illumi-
4
2
nates the space, and based on the energy reflected from the
a()
a ()a()
/N
,
target, the receiver estimates the targets bearing. In this
section we examine the achievable performance of a MIMO where L is the number of snapshots used by the array for
radar when used as a DF system. For simplicity and mathe- estimating .
matical tractability we make the following assumptions: We can lower bound the MSE of any unbiased estimator
by averaging the CRB with respect to . We denote this
1. The transmitted signal vector s is random with a com- bound by ACRB() = E [CRB(|)]. By using (16) the
plex normal distribution, and a spatially white, sta- ACRB() is given by
tionary power spectral density with correlation matrix
(1/M ) IM . ACRB() = E [CRB(|)]
" #
N 2M 4M 2
2. The elements of both the transmitting and the receiv- = E +
ing arrays are omnidirectional. 2L N SNR||||2 (N SNR)2 ||||4
4 !1
2
3. Multiple, independent snapshots of the received sig-
a()
a ()a()
/N . (17)
nal are available for processing.
The following lemma is essential for deriving a closed form as
expression for ACRB().
M M
10 log10 FL(M ) 10 log10 p .
1
h
1
i M 1 (M 1) (M 2)
Lemma 1 E |||| 2 = 2(M1) and
h i (22)
1 1 Consider the case M = 1. In this case the fading loss
E |||| 4 = 22 (M1)(M2)
is infinite. Furthermore, the ACRB in (21) is infinite. How-
Proof 1 We first note that ||||2 is distributed as a chi- ever, when, the unknown angle parameter is estimated us-
square random variable with 2M degrees of freedoms. ing say, the maximum likelihood estimation, the resulting
h This
MSE approaches zero as the SNR approaches infinity. This
i
1
allows us to write the expectations of interest as E |||| 2 =
1 n o discrepancy deserves additional consideration. The CRB
1
= E22M X12 .
E22M X and E |||| 4 bound is a small error bound, that is, it predicts the MSE
Given the density function of the 22M random variable based on the behavior of the log-likelihood function in the
vicinity of the true parameter vector. Since it is a small
x2M/21 ex/2 error bound, this bound ignores the full structure of the pa-
pX (x) = , (18) rameter space [12], which may result in nonsensical values.
(2M/2) 22M/2
For example, in the problem at hand, if ||||2 is low, the
where denotes the Gamma function, evaluate CRB, (16), might be much higher than 2 . However, since
[, ] the MSE of any estimator is lower than 2 .
1 x2M/21 ex/2 Hence in this case, the CRB is useless.
1
Z
E2(2M ) = dx The CRB approaches infinity at the rate of |||| 1
X x (2M/2) 22M/2 2 , that
0
Z
x(2M2)/21 ex/2 is, CRB = O(||||2 ). Therefore in order for the ACRB
= dx to yield a finite result, the density function pX (x), where
0 (2M/2) 22M/2
X = ||||2 is a chi-square random variable with 2M de-
((2M 2) /2) 2(2M2)/2 1 grees of freedom, should approach zero faster than (||||2 )
= = (19)
(2M/2) 22M/2 2 (M 1) as |||| 0. By examining the probability density func-
tion of X in (18), it is easy to see that this happens only for
and M 3. But in reality, since the MSE of any estimator is
1
Z
1 x2M/21 ex/2 smaller than 2 , averaging the performance of any estimator
E22M 2
= dx with respect to yields a finite result for any M .
X 0 x2 (2M/2) 22M/2
Now, consider the case M . Here, the fading loss
x(2M4)/21 ex/2
Z
= dx approaches zero, that is, the variations in the targets SNR
0 (2M/2) 22M/2 do not affect the systems MSE. This phenomena can be
((2M 4) /2) 2(2M4)/2 explained using the following intuitive argument. Without
= fading, the received signal is r = a()s + v, where s is
(2M/2) 22M/2
zero-mean, (1/M ) variance, complex normal. With fad-
1
= 2 (20) ing, the received signal is r = 1/ 2 a() M
P
2 (M 1) (M 2) m=1 m s +
v. However, according PM to the central limit theorem, as M
Using the results of the lemma, the ACRB is given by approaches infinity, m=1 m s approaches a zero-mean,
(2/M ) variance, complex normal random variable. Hence,
!1
2
N as M approaches infinity the received signal is equal to the
2
ACRB()
=
a()
a ()a()
/N
received signal of a non-fading system.
2L
M M
1+ .(21) Uniform Linear Array
(M 1)N SNR (M 2) (N SNR)
In this section we specifically consider the case of a uni-
It is easy to verify that if the targets RCS is constant, form linear array (ULA) with omnidirectional antennas at
that is ifkk2 = 2M deterministically, the CRB is indepen- the receiver. Consider a ULA with N elements with half a
dent of M . Having this in mind, it is only natural to define wavelength spacing. The n-th element of the array steering
the systems fading loss as the additional SNR necessary vector equals
to achieve the same MSE as a system that is not subject to [a()]n = ejn sin . (23)
fading.
By using the results of Lemma 1, it is easy to verify that
The n-th element of a() is
the fading loss (in dB) as a function of the number of ele- h i
a() = jn cos [a()]n . (24)
ments in the transmitting array is lower and upper bounded n
2
From these relations it follows that the squared norm of the 10
M= CRB
steering vector is given by a ()a() = N, and the squared M=16 Emp
M=16 CRB
norm of the derivative of the steering vector is given by M=4 Emp
M=4 CRB
N
X 1
a ()a()
= 2 cos2 n2
n=0
(N 1) N (2N 1) 2
RMS
cos2 . (25)
3
= 10
6
Finally, we have
N 1
!2
a ()a()
2
X
jn cos a ()a()
=
n=0
2
(N 1) N 2 2 4
10
= cos2 (26) 0 2 4 6 8 10
SNR [dB]
12 14 16 18 20
4
Substituting these results in (21), the ACRB for the case of Fig. 2. Average Cramer-Rao bound and empirical MSE for
a ULA at the receiver is given by, direction finding of a Gaussian source. Large transmit array.
6 M
ACRB() = +
(N 2 1)L 2 cos2 (M 1) N SNR a transmitting array with M = 4 elements is used, the fad-
ing loss is about 1.3 dB. This value is consistent with our
#
M2
(27) analysis based on (22) predicts that the fading loss will be
(M 1) (M 2) (N SNR)2 between 1.25 dB and 2.1 dB. When the array with M = 16
is used, the fading loss is negligible, as also predicted by
Let us investigate some special cases.
our analysis.
/2: Here, ACRB() , confirms that the
In Fig. 3, the ACRB and the RMS error of the MLE are
direction cannot be estimated at endfire, since the array has
a zero effective aperture (zero resolution). shown as functions of SNR for small transmit arrays with
= 0: This is the best case for estimating the direction M = 1 and M = 2 elements, respectively. In addition,
also shown is the CRB for the case of a Gaussian source
parameter. Indeed, at broadside the array has the largest
effective aperture (best resolution). without fading. For the case of M = 2 transmit antennas,
N = 1: The bound is infinite. Indeed, a single omnidi- we also plot a Modified ACRB (MACRB). The MACRB is
defined as the lower bound of the CRB at high SNR, such
rectional antenna cannot measure the angle of arrival.
that SNR1 SNR2 , and terms containing the latter are
neglected. From (16) we obtain
Numerical Results
N 2M
In this section, numerical results are provided on the ACRB MCRB(|) =
2L N SNR||||2
for a ULA with N = 6 elements at the receiver. Perfor- !1
2
mance is parameterized by the number of transmitting an-
2
tennas M . The transmit antennas are spaced sufficiently to
a()
a ()a()
/N (28)
.
achieve diversity.
Fig. 2 depicts the ACRB for various SNRs for large The MACRB is obtained by averaging (28) with respect to
transmitting arrays with M = 4 and M = 16 transmitting . It is easy to see that the MACRB for the ULA at the
elements, respectively. The CRB for the case of a Gaussian receiver equals
source without fading is depicted as well. Empirical results
are represented through root mean square (RMS) errors of 6 M
MACRB() =
the MLE. It is well known that (for spatially and tempo- (N 2 2 2
1)L cos (M 1) N SNR
rally white noise) the MLE for bearing estimation of a single (29)
source is given by a conventional beamformer. The MLE is We can observe from the figure that if only one trans-
the value of the angle of arrival that maximizes the output mitting element is used, the fading loss is very large and it
of the beamformer. That number of independent snapshots is about 15 dB. However, for large SNR, by adding an ad-
used in the estimation was L = 80. ditional transmitting element, the fading loss decreases to
It is evident from the figure that the ACRB values match about 2 dB. Moreover, when the SNR is large and (28) is
the empirical results for both M = 4 and M = 16. When tight, the MACRB fits the empirical results quite well.
0
10
M= CRB
Processing, Springer - Verlag, New York, 1st edition,
M=2 Emp
M=2 MACRB
1993.
M=1 Emp
2
10
processing for airborne surveillance radar syste,s,
IEEE Trans. on Aerosp. and Electron. Systems, vol.
30, no. 3, pp. 660669, July 1994.
3
10
[7] J. Ward, Cramer-Rao bounds for target angle and
doppler estimation with space-time adaptive process-
ing radar, in Proc. 29th Asilomar Conf. Signals, Syst.
4
10
0 2 4 6 8 10 12 14 16 18 20
Comput., Nov. 1995, pp. 11981202.
SNR [dB]