You are on page 1of 8

968 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 39, NO.

I, JULY 1991

Near-Field Multiple Source Localization by


Passive Sensor Array
Yung-DarHuang, Member, IEEE, and Mourad Barkat, Senior Member, IEEE

Abstract-A study is presented of the localization of multiple However, in [8] and [9], they did not consider the possible
near-field sources in a spatially white Gaussian noise environ- algorithms to localize near field sources nor did they study
ment. First, we use a modiiied two-dimensional (2-D) version of performance characteristics. In [101, Weiss and Friedlander
the multiple signal classification (MUSIC) algorithm to localize
the signal sources; range and bearing. Then, we consider a proposed a self-calibrating algorithm, for simultaneously esti-
global-optimum maximum likelihood searching approach to mating the DOAs and sensor locations. The technique is
localize these sources. It is shown that id the single source limited to the one-dimensional direction finding problem
situation, the covariances of both the 2-D MUSIC, estimator only.
and the maximum likelihood estimator (MLE) approach the Localization of far-field sources in a plane requires that
Cramer-Rao lower bound (CRLB) as the number of snapshots
increases to infinity. In the multiple source situation, we observe only the DOAs or bearings of the sources be estimated using
that for a high signal-to-noise ratio (SNR) and a large number a sensor array. Consequently, two subarrays sited apart in
of snapshots, the root mean square errors (RMSEs) of both sufficient distances are needed to obtain the intersections of
localization techniques are relatively small. However, for low the straight lines of DOAs. The use of subarrays requires
SNR and/or small number of snapshots, the performance of the some communication between them. Near-field source loca-
MLE is much superior than that of the modified 2-D MUSIC.
tion requires that both the range and the bearing of the
sources be estimated simultaneously. This can be accom-
I. INTRODUCTION plished by using one sensor array only, even though the

A N important problem in radar, sonar, radio-astronomy,


oceanography, geophysics and seismology is the local-
ization of radiating/reflecting sources by passive sensor ar-
number of parameters to be estimated is doubled. Hence,
there is no communication load.
In this paper, we present an analysis of multiple source
rays. The estimation of the directions of arrival (DOAs) of localization in the near field. Two techniques are considered
far-field narrow-band sources, with the same center fre- and studied. First we use the modified version of the two-di-
quency, by an array of sensors with arbitrary locations and mensional (2-D) multiple signal classification (MUSIC) algo-
arbitrary directional characteristics has received a great of rithm, then, the maximum likelihood estimator is employed
attention. Several techniques [13- [101 have been proposed in to determine range and bearing of sources. The root mean
the literature to solve this estimation problem. In the above square error (RMSE) evaluations of the two localization
literature, most of the work done has considered far-field techniques both in terms of signal-to-noise ratio (SNR) of the
sources. Very little has been done for near field in which sources at the sensor array and the number of snapshots are
both range and bearing of sources are unknown parameters. performed by computer simulations. In Section II, we formu-
Rockah and Schultheiss [8] presented an array shape calibra- late the problem. In Section III, the modified version of
tion technique using spectrally disjoint near-field calibrating MUSIC localization is presented. In Section IV, we present
sources whose locations are not known a priori. The joint the global-optimum searching technique to compute the exact
Cramer-Rao lower bound (CRLB) on the source locations maximum likelihood localization. In Section V, we present
and the array shape errors were calculated. It was shown that the simulation results. A summary and conclusions are given
the array shape errors can be made arbitrary small by using a in Section VI.
sufficient number of sufficiently strong calibrating sources. In
[9], Rockah et al. also presented the phase calibration tech- 11. PROBLEM FORMULATION
nique of estimation error due to random perturbations in the In this section, we formulate the problem for the near-field
phase shifts at the sensors. By using far-field or near-field localization of narrow-band sources. By near field, we mean
calibrating sources, and the calculations of the joint CRLBs that the wavefronts emitted from the signal sources are
of the estimations of both the source locations and the phase spherical and cannot be approximated to simply plane waves
shifts, the phase calibration can be achieved provided the as in the case of far-field sources. In addition, by narrow-
number of calibrating sources exceeds a certain minimum. band, we mean that the bandwidths of the signals and of the
system is much smaller than the reciprocal of the propagation
Manuscript received July 3, 1990; revised March 12, 1991. This work delay of the waves across the array.
was supported in part by National Science Foundation Research Award Consider D near-field sources at unknown locations and
ECS-8907176. with arbitrary directional characteristics. The sources are
The authors are with the Department of Electrical Engineering, State
University of New York at Stony Brook, Stony Brook, NY 11794-2350. emitting spherical waves impinging into a sensor array of L
IEEE Log Number 9100566. elements, L > D , of arbitrary locations and arbitrary direc-

0 1991 IEEE
0018-926X~91~0700-0968$01.~
1

969
HUANG AND BARKAT: NEAR-FIELD MULTIPLE SOURCE LOCALIZATION

source to the ith sensor defined as


I' k t h source
z$ = ri -k d? - 2 r k di COS 8i, (4b)
where
ei, = e; - e, (44
and the radial coordinates of sensor (i) and source ( k ) are
( d i , ei) and (rk, e,), respectively. The coefficients of the
response, cik, are determined by the wavelength, X, the
0 0 X radiating patterns of the sources and the directivity patterns
of the sensors. If both sensors and sources are omnidirec-
tional, the coefficients C;k, are independent of the sources
and the sensors. In this case, we can set the coefficients
cik = 1 for i = 1, 2;**, L and k = 1, 2;.., D .
The observation interval between adjacent samples is larger
than the correlatio? time and thus the finite-sampled data
Fig. 1. Sensor-array geometry in near field. covariance matrix R,(r, 8) is
I M
tional characteristics as shown in Fig. 1. The observed data
sequence at the ith sensor can be expressed, by its complex
envelope, as In (5), ii,(r, e) is the estimate of the true data covariance
D matrix Rx(r, 8) given by
xi(m) = ui(rk,ek)sk(m) + ni(m) (l)
k= 1 R,(r,B) = A ( r , 8 ) R , A H ( r , B ) + u2I (6)
i = 1,2,.*.,L
m = 1,2;--,M where R, is the signal covariance of the sources. The
problem to be solved is to estimate the number of near-field
where M is the number of snapshots. Equation (1) can be sources and their location given the M-s-pled data vector
rewritten in vector form or matrix form as sequence, { x ( m ) } ,m = 1, 2; M, or R,(r, 8) obtained
e ,

in (5) using two different techniques.

HI. TWO-DIMENSIONAL
MUSIC ALGORITHM
In signal subspace techniques, the number of sources and
and their locqtions can be estimated by appropriately processing
x(r,O; m ) = A(r,B)s(m) + n(m) (34 the eigenvalues and the eigenvectors of the finite-sampled
data covariance matrix. Since the sample size is finite, the
where
eigenvalues are all different with probability one. Therefore,
the number of signal sources may be difficult to determine
due to the multiplicity of the smallest eigenvalue by simply
sk(m)is the scalar complex waveform of the kth incoming observing the structure of eigenvalues of R,. For a high
signal, a(r,, e,) is the (L x 1) steering vector characterized SNR, a subjective threshold may be set to determine the
by the sensor array and rk and 0, are the unknown range and number of sources since the eigenvalues corresponding to the
the bearing of the kth source with respect to a reference signals and the eigenvalues corresponding to the noise differ
origin; respectively. n(m) is the ( L x 1) additive complex greatly. When the SNR is moderate or low, and these two
noise vector. The noise components are stationary, ergodic groups of eigenvalues are no longer distinct, the subjective
Gaussian with zero mean and covariance matrix u21, and threshold setting can not be used. The model-selection tech-
uncorrelated from the impinging signals, where u 2 is un- niques [l 11, which do not require any subjective threshold
known and I is the identity matrix. The elements in the setting, such as rpinimum description length (MDL) and
steering vector, a(r,, e,), are fully dependent on the geomet- Akaike information criterion (AIC), can be employed to
rical structure of the array and the sources. Thus, the com- estimate the number of sources. Then, the MUSIC algorithm
plex response in the ith sensor element to the kth impinging can be used to determine the locations. In this section, we
signal can be expressed as modify the conventional 1-D MUSIC algorithm to its 2-D
version to estimate the near-field range and the bearing of the
sources simultaneously.
To determine the numbe! of the sources [ll], we rank
order the eigenvalues U, of R , in descending order to obtain
where j is the unitary complex number, X is the wavelength
of the impinging waves, and Z ; k is the dis@ce from the kth
970 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 39, NO. 7, JULY 1991

The estimated number of sources, b, is then determined as global-optimum searching algorithm can be formed from the
the value for which either the MDL or the AIC criterion is decision function by some matrix manipulation.
- -
minimized. Let U B + ~ U, B + ~ , , u L be the noise eigenvec- Let the hypothesis, (r,, e,), k = 1, 2;.-, L - 1, repre-
tors corresponding to the ( L - b)smaller eigenvalues, ub+ 1 , sent the locations of k sources as
U B + ~ ,* , U=. We construct the noise eigenvector matrix,
U,, from the noise eigenvectors as (rk,Ok) = [(r131) (r232). * * (rkd3k)I. (11)
The steering-vector matrix corresponding to this hypothesis,
( r k , d k ) , is
The near-field steering vector, a(r, e), comprises all possible
near-field steering vectors, that is, A ( r k J , ) = [a(rI,Ol) a ( r 2 , e 2 ). - * a ( r k , e k ) ] . (12)

[
a ( r , e ) = -exp
The conditional joint probability density function of the M
sampled-data vectors given hypothesis (r. e,) is [13]

where T denotes matrix transpose, and zi,i = 1, 2, * * e , L,


is
zi = r2 + d? - 2rdi COS ( e - ei). (9b)
To search for the locations of sources, we construct a 2-D
orthogonality-measure spectrum, u2 is the unknown noise variance. The product L a 2 is the
determinant of the covariance matrix of the array noise
vector. Taking the logarithm of (13a), ignoring the constant
term, we obtain the log-likelihood function to be

where H denotes the complex conjugate transpose. The g ( r k , e k ) = -&flog u 2


peaks of the spectrum, P ( r , e), are the locations of the 1 M

sources, (r,, O k ) , k = 1, 2, * * * , D . Note that the search


procedure is carried in the (r, e ) plane starting from the
origin. Also, the 2-D MUSIC algorithm, like the 1-D ver- To compute the maximum likelihood estimator, we have to
sion, is based on the orthogonality property between the maximize g with respect to all unknown parameters. Even
signal steering vectors, a(r,, ek), and the noise eigenvector though the number of sources is not known, we can use the
of the data covariance matrix, R,. The eigenvalues and the alternating maximization procedure introduced in [131- [151
eigenvectors of the data convariance matrix are not the actual to obtain the log-likelihood function, in terms of the number
ones, but estimates, since the number of snapshots is finite. of sources and their locations, as
Consequently, this spectrum-search technique may suffer an
M
estimation error. Furthermore, if the number of snapshots is
small and/or the SNR of sources is low, the algorithm may g(rk9dk) = C
m= 1
IP.qr,,e,)X(m) I (15a)
lose its resolution capability. Nevertheless, the proposed 2-D
k = 1 , 2 * - -L - 1
MUSIC algorithm has some statistical efficiency. This was
shown by studying the asymptotic relative efficiency. In [12] where is the projection operator onto the space
it was shown that the 1-D MUSIC algorithm attains the spanned by the columns of matrix A@,, e), and given by
CRLB as the number of measurements approaches infinity.
The real parameter vector 8 in [12] is analogous to the 'A(rkrek) = 9 7 9 7 '

location coordinate ( r k , e,) in the plane. Both the estimations (15b)


of range, r , and bearing, 8, satisfy the asymptotic relative
efficiency as the SNR of sources becomes sufficiently high. After some matrix manipulation, (15a) becomes

LKELIHWDESTIMATOR
IV. THEMAXIMUM g(rk 9 = tr [pA(r,,B,)RX] (16)
In this section, we present a global-optimum searching where tr [e] denotes the trace of the matrix in brackets. To
algorithm to compute the exact maximum likelihood estima- maximize the nonlinear log-likelihood function of (15) or
tor of the number of sources and their locations in the near (16), we must find a number of sources, D , at locations (ib,
field by searching the maximum'likelihood estimates of the ek), among a set D which consists of subsets, {(r,, e,)),
number of sources and their locations which can both be k = 1, 2 ; * , , L - 1, such that g(r,, e), is maximum. That
found simultaneously. Based on the available sampled data is,
and the assumed noise statistics, the maximum-likelihood
(17)
decision function is determined. A recursive equation for the
HUANG AND BARKAT: NEAR-FIELD MULTIPLE SOURCE LOCALIZATION 971

2500 -

2000 -
w
p:
22 1500-
r:

Fig. 2. MUSIC localization of single source at (15 m, 65). SNR = 20 dB, number of snapshots = 100.

where arg denotes the argument of, fi and (fb, ab) are found when saturation begins; that is, when
the maximum likelihood estimates of the number of sources
and their locations, respectively. D is determined by a search g(%+19b+,) - g(rg,h) = 0 (20)
procedure as described below. Since the number of sources is is satisfied. Having estimated the number of sources fi, we
not known a priori, the conventional estimation technique now determine the locations of these sources by solving for
[l] and the hill climbing appoach [13] are no longer their range and bearing. Hence (17) becomes
applicable. Consequently, we propose a recursive algorithm
to search for both the number of sources and their locations
and guarantee the global-optimum solution. Using the projec-
tion-matrix decomposition and the projection matrix update The recursive log-likelihood function in (18) is first com-
formula [13], (16) can be written in a recursive form as puted for k = 1 and k = 2_since we initialize g(r,, e,) to be
zero. If (20) is satisfied at D = 1, the algorithm stops and we
g(rk3 = g(rk-l) + tr [ P B ( r ~ - l , ~ ~ - l ) a ( r ~ . B ~ ) R ~say
l that the number of sources is one (D= 1). The estimated
(184 location of this single source, ( rl , 8 I)ML, is then determined
from (21). Note in (18a) that g(r2, 8,) is obtained directly
where from g(rl, 8,). This redyces the computation load greatly. If
B(r, 8,) = I - PA(rk,ek)
7 (18b) (20) is not satisfied for D = 1, we increment k by one and
compute (18) for k = 2 and k = 3. If (20) is satisfied, the
k = l , 2 , . * * , L - 1. algorithm SLOPS. For this case, we say that the number of
PB(rk-l,#k- I)s(rk,Bk) is the projection operator onto the one- sources is D = 2, obtain the locations of sources, (r2, 82)ML
dimensional space generated by the residual (that is, ( r l , flI)MLand ( r , , from (21). Otherwise,
,,
B(r,- 8,- I)a(r k , 0,). This residual is the column vector we increment k by one again aFd continue the procedure
obtained from the projection of the stzering vector a(r,, e,) until we find the searched value D .
onto the ( k - 1)-dimensional space spanned by the columns The procedure of the global-optimum search algorithm
of matrix A(r,- 8,- ,, It can easily be shown that the which estimates the number of the sources and determines
scalar function g(rk, 8,) is a monotonically increasing func- their locations simultaneously can be summarized as follows.
tion of k. Since there are D sources, (18a) saturates when 1. Start from k = 1 with B(ro, e,) = I, PA(ro,Bo) = 0 and
+
k = D 1. In other words, when SNR goes to infinity, g ( t , , e , ) = 0.
l e l compute g(rl, = g(rO, + ~r[PB(ro,f30)a(rI,BI)
g(rD+,* - g(rD, = (19) R,I;
1.2 If g(rl,e,) = g ( r , , e,), exit; If not, compute
for 1 = 1, 2; --, L - D - 1. Since the number of sources
PA(rl,Bl) = P*(ro,O0) + PB(ro,Bo)a(r,,B,);
is not known a priori, we need to determine D,the maxi-
mum likelihood estimate of the number of sources. This is B(rl,el) = I - PA(rl,Bl);
912 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 39, NO. I, JULY 1991

0.2

0 5 10 15 20 5 10 15 20

SNR ( d B ) SNR ( d B )
(a) (b)
Fig. 3. (a) Range-RMSE. (b) Bearing-RMSE versus SNR for single source at (15 m, 65'). Number of snapshots = 25, Number
of Monte Carlo runs = 100.

NUMBER OF SNAPSHOTS NUMBER OF SNAPSHOTS


(a) (b)
Fig. 4. (a) Range-RMSE. (b) Bearing-RMSE versus number of snapshots for single source at (15 m, 65"). SNR = 6 dB, Number
of Monte Carlo runs = 100.

Go to Step 2. We note that, instead of computing PA(rk,ek)directly as


2. MAINLOOP: k + k + l given in (15b), we compute P B ( r k - , , e k - I ) a ( T y , e t )where B@,- 1,
2.1 Compute g(r,,b,) by (18a);
' 8,- l)a(rk,e,) is of dimension ( L x 1). This way, the com-
2.2 If (20) is satisfied, go to step 3; If not, compute putational complexity is greatly reduced. The global optimum
- of the maximum likelihood estimator is obtained for both the
,ek) - P A ( r k _ ,,ek-,) + P B ( r k _ , . e k - , ) a ( r l r . e k;) number of sources and their locations. The outlined proce-
dure requires more computation than MUSIC. It should be
Compute B(r,, e,) by (18b); mentioned that this optimum searching procedure is equally
Go back to MAIN LOOP. applicable to both the case of incoherent sources and the case
3. Obtain the maximum likehood estimate (rb, from of coherent sources since the structure of the signal-covari-
(21). ance matrix is not relevant.
HUANG AND BARKAT: NEAR-FIELD MULTIPLE SOURCE LOCALIZATION

Fig. 5. MUSIC localization of two uncorrelated sources at (15 m, 45") and


(30 m, 60").SNR = 20 dB, number of snapshots = 100.

12 - *

10 -
h 8- 'k,,,, y
n
t ,60 ): MUSIC

5 B-
z ?,
W
w - .
v
z W t
6- z
z 4-
4-

2-
2-

I
0 0 5 10 15 20 25
0 5 10 15 20 25
SNR (dB) SNR (dB)
(a) (b)
Fig. 6. (a) Range-RMSE. (b) Bearing-RMSE versus SNR for two uncorrelated sources at (15 m, 45") and (30 m, 60").Number
of snapshots = 100, number of Monte Carlo runs = 100.

V . SIMULATION
RESULTS

The performances of the near-field MUSIC and the maxi- mately 116 m, the sources at 15 and 30 m are clearly in the
mum likelihood localization are studied by various examples. near field. The signals from the sources are deterministic but
Consider 100 Hz-centered narrow-band acoustic sources at unknown. In each experiment, we perform 100 Monte-Carlo
15 and 30 m, emitting spherical waves into a linear array runs and compute the RMSE's of the estimates.
with five sonar elements uniformly spaced half a wavelength From Figs. 2-4 we observe that for a single source, the
apart. Since the far-field distance ( L - 1)*(X/2) is approxi- 2-D MUSIC localization is satisfactory for a large number of
914 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 39, NO. 7, JULY 1991

12 bearing of the near-field sources. A modified-version of the


2-D MUSIC algorithm to localize sources was considered
and investigated. A global-optimum searching algorithm for
10
simultaneously computing the exact maximum likelihood esti-
mators of the number of near-field sources and their locations
8
was also proposed. We have shown that both algorithms
h attain a sufficiently good performance in localization when
w
E
0
the SNR is high and/or the number of snapshots is large.
?!-S However, when the SNR is low or moderate and/or the
I2 number of snapshots is small, the global-optimum searching
ET technique is much superior than the 2-D MUSIC. While
4 MUSIC fails when the sources are coherent, the global-opti-
mum searching algorithm still maintains its good perform-
ance as in the case when the sources are uncorrelated. The
2 drawback, however, is the heavy computational load.
ACKNOWLEDGMENT
0
IO 102 I 03 The authors are grateful to the reviewers for their con-
NUMBER OF SNAPSHOTS structive comments that led to the enhancement of the quality
(a)
of the paper.

REFERENCES
F. C. Schweppe, Sensor array data processing for multiple signal
sources, IEEE Trans. Inform. Theory, vol. IT-14, pp. 294-305,
1968.
*\ C. H. Knapp and G. C. Carter, The generalized correlation method
10 - for estimating of time delay, IEEE Trans. Acoust., Speech, Signal
Processing, vol. ASSP-24, pp. 320-327, 1976.
M. Wax and T. Kailath, Optimum localization of multiple sources
-
w
8 - by passive Arrays, IEEE Trans. Acoust., Speech, and Signal
Processing, vol. ASSP-31, pp. 1210-1217, 1983.
K R. 0. Schmidt, Multiple emitter location and signal parameter
w estimation, IEEE Trans. Antennas Propagat., vol. AP-34, pp.
9 6- 276-280, 1986.
w
v1 G. Bienvenu and L. Kopp, Adaptivity to background noise spatial
H coherence for high resolution passive method, in Proc. ICASSP-80,
1980, pp. 307-310.
4- S. S. Reddi, Multiple source location-A digital approach, IEEE
Trans. Aerosp. Electron. Syst., vol. AES-15, pp. 95-105, 1979.
R. Kumaresan and D. W. Tufts, Estimating the angles of arrival of
2- multiple plane waves, IEEE Trans. Aerosp. Electron. Syst., vol.
AES-19, pp. 134-139, 1983.
...+.. .r;
Y. Rockah and P. M. Schultheiss, Array shape calibration using
-++ sources in unknown locations- Part 11: Near-field sources and estima-
0 tor implementation, IEEE Trans. Acoustics, Speech, Signal Proc-
101 102 I 03 essing, vol. ASSP-35, pp. 724-735, 1987.
NUMBER OF SNAPSHOTS Y. Rockah, H.Messer, and P. M. Schultheiss, Localization perform-
0) ance of arrays subject to phase errors, IEEE Trans. Aerosp.
Electron. Syst., vol. 24, pp. 402-410, 1988.
Fig. 7. (a) Range-RMSE. (b) Bearing-RMSE versus number of snapshots
A. I . Weiss and B. Friedlander, Array shape calibration using
for two uncorrelated sources at (15 m, 45) and (30 m, 60). SNR = 10 dB, sources in unknown locations-A maximum likelihood approach,
number of Monte Carlo runs = 100. IEEE Trans. Acoust., Speech, Signal Processing. vol. 37, pp.
1958-1966, 1989.
snapshots and a high SNR. The maximum likelihood estima- M. Wax and T. Kailath, Detection of signals by information theo-
retic criteria, IEEE Trans. Acoust., Speech, Signal Processing,
tor gives a much better performance than MUSIC for a small vol. ASSP-33, pp. 387-392, 1985.
number of snapshots and/or a low SNR. Also, the range and B. Porat and B. Friedlander, Analysis of the asymptotic relative
the bearing estimates of these two algorithms approach the efficiency of the MUSIC algorithm, IEEE Trans. Acoust., Speech,
Signal Processing, vol. ASSP-36, pp. 532-544, 1988.
CRLBs as the SNR goes to infinity even when the number of I. Ziskind and M. Wax, Maximum likelihood localization of multi-
snapshots is small. In Figs. 5-7, we show the performance in ple sources by alternating projection, IEEE Trans. Acoust.,
terms of bearing and range of the considered algorithms for Speech, Signal Processing, vol. 36, pp. 1553-1560, 1988.
J. Boheme, Estimating the source parameters by maximum likeli-
two uncorrelated sources. hood estimator and nonlinear regression, in Proc. ICASSP-84,
1984, pp. 7.3.1-7.3.4.
VI. SUMMARYAND CONCLUSION M. Wax, Detection and estimation of superimposed signals, Ph.D.
dissertation, Stanford Univ., Stanford, CA, 1985.
In this paper, we presented a study of the near-field Y. D. Huang, Ph.D. dissertation in process, Dept. Elec. Eng., State
localization where we simultaneously estimated the range and Univ. New York at Stony Brook, Stony Brook, NY.
HUANG AND BARKAT: NEAR-FIELD MULTIPLE SOURCE LOCALIZATION 915

Yung-Dsr Huang (S87-M90) was born in Mound Barkat (S82-M82-SM91) was born in
Hualien, Taiwan, on February 5, 1945. He re- Constantine, Algeria, on January 6, 1958. He re-
ceived the B.S. and M.S. degrees in electrical ceived the B.S. degree with honors, magna cum
engineering from National Taiwan University, laude, in 1981, the M.S.degree in 1983, and the
Taipei, Taiwan, in 1967 and 1970, respectively, Ph.D. degree in 1987, all in electrical engineering
and the M.S. and Ph.D. degrees in electrical engi- from Syracuse University, Syracuse, NY.
neering from the State University of New York, Currently he is an Assistant Professor with the
Stony Brook, NY, in 1981 and 1991, respectively. Department of Electrical Engineering at the State
He worked with the Telecommunication Labora- University of New York, Stony Brook. His current
tories, Taiwan, as a Research Engineer from 1971 research includes detection-estimation, detection
to 1976. He was an Associate Professor in the with distributed Sensors and data fusion. Dhased
Department of Electronic Engineering, National Taiwan Institute of Technol- arrays, adaptive antennas, and queueing for distributed networks.
ogy from 1976 tG 1987. He was a Research Assistant and Teaching Assistant Dr. Barkat is a member of Tau Beta pi and Eta Kappa Nu Honor
at the Department of Electrical Engineering, State University of New York, Societies, and Sigma Xi the Scientific Research Society. He is the author of
Stony Brook, from 1988 to 1991. His research interests include detection the book, Signal Detection and Estimation (Artech House), May 1991.
and estimation, array signal processing, and digital communication tech-
niques.

You might also like