Near Far Targets Using Antenna Arrays: Superresolution of or Field Coherent Aperiodic
Near Far Targets Using Antenna Arrays: Superresolution of or Field Coherent Aperiodic
Superresolution of Near or Far Field Coherent Targets Using Aperiodic Antenna Arrays
A
Abstruct-A technique for obtaining high resolution images of pointis developed. like coherent targets using aperiodic antenna arrays Whereas existing superresolution techniques can be used to image coherent targets located only in the far field of periodic receiving arrays, this technique can be used to obtain high resolution images of targets in the near field also. The key to this technique is the use of multiple transmitters. Completely correlated (coherent) targets are decorrelated by transmitting sequentially from different locations, thereby enabling the use of modern superresolutiontechniques to estimate the target locations. An expression for estimating the variance of the target locations is derived, and is verified by empirical methods. For the same amount of transmittedpower and thesamenumber of receivers, this variance is shown to be an order of magnitude smaller that than of existing superresolution techniques. The practicality of this technique is demonstrated with experimental microwave data.
Both CM and MCM require the antennas to be placed3t equidistant points. In order to avoid aliased images, the maximum interelement spacing should be M 2 , where X is the wavelengthofthe transmitted radiation. By removing the restriction of locating the antennas on a X/2 grid, the length of the array can be increased for a fixed number of receivers. It is intuitive to expect the increase in length to 1) decrease the mean squared error in the target location estimates, and 2) improve the resolving power of the system. Therefore, it is advantageous to develop a SR technique that does not impose any restrictions on array geometry. It will be shown later that by using N element aperiodic arrays, it is possible to spread the elements over a length much larger than (N- 1)M2 and still obtainunaliased images. Among the existing SR techI. LWRODUCIION niques, the MUSIC algorithm 171and the maximum likelihood N MOST LONG wavelength microwave imaging scenarios, method [8] are independent of array geometry; however both a transmitter illuminates the field of view, and the complex techniques fail to resolve coherent targets [9], [ 101. A field amplitude (CFA) of the echoes from the targets is technique that uses moving aperiodic arrays is described in recorded along a line array. These CFA values are then [ll]. The MUSIC algorithm wasmodified to accomodate processed to estimate the positions and amplitudes of the coherent targets in [12]; this modified version is computatargets. If the targets are stationary over a period of several tionally intensive as it requires an M-dimensional search for 4 transmitted pulses, their echoes will be unchanging relative to locating A targets. Therefore, existing SR techniques that are not computationallyintensive and can resolve coherent targets each other, and they may be modeled as coherent targets. Conventional (Fourier or Fresnel) microwave imaging require the receiving array to be periodic and/or moving. techniques estimate the amplitudes and locations of the targets These techniques could be used for detecting near field targets by focusing the array to a particular point in the near field and by focusing the array to various points in the field of view [20]. This procedure is adequate to resolve widely separated subsequently forming an all-pole estimator for the target targets; however, if the targets are within a beamwidth of each locations; however, there will be a residual error in the target other, one has to resort to superresolution (SR) techniques in location estimates even under noiseless conditions if the order to resolve them. Several SR techniques have been targets are away from the focal point in the near field [13]. In this paper, we develop an SR technique for resolving developed over the last two decades; most of them model the received CFA as a complex valued autoregressive (AR) coherent targets in the near or far fields of arbitrarily shaped process, andconsequently obtain a high resolution all-pole arrays. Multiple transmitters are employed to achieve this image by estimating the AR parameters. Among the AR goal. Both CM and MCM change the center of phase position the array by using identical periodic subarrays to superresolution techniques, the covariance method (CM) 111- along [4], and the modified covariance method (MCM) (also called perform spatial smoothing. This subarray smoothing procethe least squares method) [2], [5], [6] have proved effective dure decorrelates the coherent targets and estimates the AR for resolving coherent targets located in the far field (Fraunho- parameters to generate an all-pole estimator for the target that the subarray smoothing fer zone) of a periodic array. In many applications, the array locations 111-[4]. We note may be too close to the targets for the far field approximation procedures limit the size of the receiving array to (N - 1)M2, to be valid. Therefore, it is useful to develop a SR technique where N is the number of receivers. The SR technique developed in this paper (called RAM for random array that can be used to detect targets in the near field also. method) decorrelates the coherent targets by sequentially transmitting from several different locations. During each Manuscript received January 24, 1986; revised December 1, 1986. The authors are with the Valley Forge Research Center, The Moore School transmission, the distance from the transmitter to the target of Electrical Engineering,University of Pennsylvania,Philadelphia, PA corresponds to an equivalent phase delay. The sum of this 19104. phase delay and the phase of the reflection coefficient of the IEEE Log Number 8716772.
1207
I -
B. Superresolution of Two Coherent Targets A technique to resolve two coherent targets at u = u1and u = u2using two transmit/receiver antennas at x = xI and x = I . SUPERRESOLUTION I UNDER NOISELESS CONDITIONS x2, and a receiving antenna at x = x3will now be developed. .A. Assumptions and Notation The objective is an estimator that tends to infinity at the target The SR technique willnowbe formulated for noiseless locations and is smalleverywhere else. This is the familiar allconditions. Consider an array of K transmitlreceive antennas pole estimator usedin various SR techniques [l], [6], [15]. locatedat x = xi, i = 1, , K , and (N - K ) receiving These estimators are the reciprocal of the squared magnitude
target can be considered to be the overall phase of the target for that particular transmission. Because this phase depends on the distance between the target and the transmitter, the phase difference between any two targets can be changed by changing the locationof the transmitter, thereby decorrelating the coherent targets. An all-pole estimator for the target locations can be formed after decorrelating the coherent targets. Wenotethatthis procedure for decorrelating the coherent targets is independent of the geometry of the receiving array. Therefore the size of the receiving array can be increased for a fixed number of receivers by placing the receivers aperiodically over the aperture. (If the array is periodic, the grating lobes destroy the potential for useful imaging.) This increase in the size of the aperture leads to a lower mean squared error; however, the lower sampling rate will introduce spurious peaks in the all-pole estimator for the target locations. The RAM overcomes this problem by using receiver location diversity. Several all-pole estimators for the target locations are formed by splitting the receiving array into subarrays (which are not identical to each other) and using the data measured at these subarrays to estimate the AR parameters. Each ofthese estimators peaks at the target locations, and has spurious peaks at locations that depend on the geometry of the individual subarrays. Therefore, a parallel resistor type average of the all-pole estimators obtained by using subarrays with different receiver geometries will contain not any significant spurious peaks. Further, because the size ofthe array can increased be to more than (N - l)h/2 if the receivers are aperiodically located over the aperture, the mean squared error (MSE) in the target location estimates will be smaller than the MSE for the MCM and CM. In SectionIV we showthat the RAM performs as well as the MCM and is superior to the CM when all methods use arrays of the same length. Further, we show that by increasing the length of the array used for the RAM, it is possible to decrease the mean squared error in the RAM target location estimates by more than an order of magnitude as compared to the MCM and CM. The paper is structured in the following format. The RAM is formulated for noiseless conditions in Section II and is then qualitativelycomparedwiththeMLM. The mean squared error in the target location estimates is derived in Section ILI. Further, a criterion to determine the number of targets is developed, and the statistics of the RAM estimator near its maxima are derived. Section IV presents somesimulation results and discusses the dependence of the mean squared error on the length of the array, the SNR, and separation the between thetargets, for the two target case. Simulation results are also presented to compare the RAM with Fourier processing, MCM, and CM. Section V confirms the theory using measured microwave data obtained from real targets at the Valley Forge Research Center, University of Pennsylvania.
M TARGETS
K
TRANSMIT/RECEIVE ANTENNAS
N-K RECEIVERS
I
I ,
/i
I1 : ' /
k-th
Transmzt/Receive
Antenna
7 11 1
YYY Y Y Y i Y Y Y Y Y
i
n-th
Receiver
YY Y "Y
y yy
antennas at x = xi, i = K + 1, * , N.(We note here that N receiving and K transmitting antennas could also be used; the present format is used for convenience of notation.) Let M targets be located in the same range bin at a distance R from the origin, at u = ui, i = 1, * * , M , where ui = sin Siand Bi is the angle from broadside. Each target may be parametrized by a complex amplitude Si, = 1, * , M , which depends on i its cross-section area and reflection coefficient. The geometry is depicted in Fig. 1. Let theCFA at the nth antenna, located atx = x,, while the mth antenna is transmitting, be denoted by em,. If the reciprocal of the bandwidth of the transmitted signal is much greater than the transit time of the RF pulse along the array, e,, may be written as
--
(1)
is the distance between the mth antenna and the ith target, k = 2n/h where h is the wavelength of the transmitted radiation, and the term inside the exponent in (1) is a phase delay corresponding to the round trip distance from the transmitter to the receiver [ 141.
--
1208
of the dot product ofaweight vector anda beamsteering vector, the weights being determined from the measured CFA so that they are orthogonal to the beamsteering vector at the target locations. With this in mind, consider the following system of equations:
where S is a K x M matrix whose elements are , = s, exp s ( -jkr,(u,)), T i s an M x K matrix with elements tmn= exp (-jkrn(um)), W;+1 = [ w K + I ( ~ ..*,~K+~(K)I, ), andXz+l = [exp(-jkrKcl(u1)), ..., exp ( - j k r ~ + ~ ( u ~ ) ) I . Itmaybe shown that if
rm(Uk)-rm(u/)+r,(ul)-rn(uk)#ih,
(3a) (3b)
i = + 1 , +2,
*.-
(9)
where w3( 1 ) and w3 (2)are unknown weightsto be determined by solving( ) and the subscript 3 indicates that the CFA at the 3, receiving element located at x = x3 is assigned a weightequal to unity. Using (l), we may rewrite (3) as
then the rank of E , given by (7), is M . Because the likelihood is small that the sum of the four distances in (9) is exactly equal to an integral number of wavelengths, (9) will be satisfied for most target scenarios. If (9) is satisfied, the rank of E and hence T equals M , and (8) may be written as
s 2
TWK+,+XK+,=O,
orDK+I(ui)=O, = l , * - - , M i
(10)
where
K+ 1 where Wis the weight vector equal to [w3(l), w3(2), B(u) 11 T, &+1(u) = w + l(i) exp (.hkfu)) (1 1) is a beamsteering [exp vector equal to i= I ( - j k r ~ ( u ) ) ,exp (-.J%(u)), exp ( - j k ( u ) ) l and the superscript T indicates transposition of a matrix. is the dot productof a beamsteering vector and a weight vector The constants attached to the WTB(u2) term in the two and wK+l(K + 1 ) = 1. Therefore, any all-pole intensity equations in (4) will be different if spectrum estimator I(u) of the form
-t 2,
-- -.
(5)
(12) I(u)=C/IDK+I(u)l2
where C is a positive constant, tends to infinity as u tends to ui, i = 1 , * ,M. Because T and X,+ are independent of the Si, we observe from (10) that the weight vector is also independent of the target amplitudes. Until now no far field restrictions have been imposed; hence the SR technique developed above can be usedto detect targets I(u) = c/ I W%(u) 12 (6) located in the near and far field of the array. However, from this point onwards, we shall assume that all the targets are where C is a positive constant, tends to infinty as u present in the far field of the array. This assumption simplifies approaches uI and u2. In the presence of additive Gaussian the mathematical expressions to be derived and enables the noise in the measured CFA, E{I(ui)}is proportional to ISi I , comparison of this technique with the MCM and CM, both of as shown in Appendix I (eq.(40)). This desirable property of which yield erroneous estimates of the locations of near field I(u) is absent in the least squares estimator [ 161. targets even under noiseless conditions. If the targets are in the far field of the array, the all-pole C. Superresolution of M Coherent Targets estimator I(u), given by (12),simplifies considerably by virtue The SR technique willnow be extended to resolve M of the fact that the denominator reduces to coherent targets, using K transmitlreceive antennas (K 2 M ) K+ 1 and one receiving antenna. The weights wK+ (n), n = 1, , DK+l(u)= wK+l(i) exp (jkmi). (13) K are obtained by solving the following matrix equation: Therefore, if (5) is satisfied, the two equations in (4) demand that WTB(ul)= WTB(u2)= 0. Furthermore, if (5) holds, the system of equations (3) may be solved uniquely for w3(1) and w3(2) so that WTB(ul) = WTB(u2) = 0. Therefore, an all-pole intensity spectrum estimator of the form
--
i= 1
D. Artifacts of the All-Pole Ektimator If the poles of I(u) are located only at u = ul, , uM, then I(u) may be considered to be a good estimator for the target or locations. However, as shown below, I(u) will have poles for EWK+1 = - E K + 1 (7) u different from the target locations, especially if the array is where the e,, are given by (1). Further insight may be gained sparsely filled (L % (N - 1)N2 where L is the length of the array). The locations of the spurious poles will be determined. by decomposing E and EK+ in (7). After some manipulations, Armed with this knowledge, a technique will be developed to we may rewrite (7) as suppress these poles of I(u). Because the poles of I(u) occur at the zeros of DK+ (u), is 1 it STWK+ = - SXK+ I 1 (8)
---
1209
sufficient to enumerate the zeros of D K + 1 (u) to find the poles of Z(u). It will now be shown that DK+(u) is zero for those values of u where the quantity Aui 4 M - u;, i = 1, M, simultaneously equals the K quantities mk/(Xk - X I ) , k = 2, K + 1, where the m k are integers. In other words, we shall prove that the zeros of DK+l(u) enumerated by the are following:
e,
elements ((N - K ) > 1). Let the transmitters radiate sequentiallyand for eachtransmission let the N receivers record the CFA samples simultaneously. A total of KN CFA samples are stored. Only K(K + 1) < KNsamples are needed to solve ( 7 ) . Therefore (N- K ) weight vectors are obtained from (7). Each weight vector identifies .the proper target poles plus a different set of spurious poles, different because the receiver locations are different. Denote the (N- K ) weight Am2 Am3 hmK+ 1 vectors by W j , i = K + 1, * , N.The dot product of Aui 4 u-uj=---=. = (14) appropriate beamsteering vectors with each of these weight X2 -X1 X3 -X1 X K + 1 -X1 vectors, denoted by Di(u), will be zero at all the target wherei = 1, M , andmk = 0, kl, k 2 , *.., k = 2, locations; however, the spurious zeros of each of the Di(u) * * * , K 1. will, in general, occur at different values of M . Therefore, by There are two conditions under which the equalities in (14) using an estimator I @ ) , which contains the sum of the squared are satisfied. The condition m k = 0,all k,implies M = ui all i; modulii of Di(u),i = K 1, * N, its denominator, one in in other words the proper pole locations correspond to m k = can suppress a l the spurious poles of I(u) in the visible range. l 0. The second condition includes all sets of nonzero values of In other words, we obtain (N- K ) weight vectors Wi, i = K mk that satisfy (14). For each such' set all the mk./(Xk - x l ) + 1, * * ,N, each of length K , by solving the following (N take on a nonzero value, implying u # ui; in other words we K ) matrix equations: shall prove that a false target or spurious pole occurs for each EWi= - E j , i = K + 1, N (16) of those sets. If u # ui, we may write (14) as Xk = x1 h n k / Aui, m k = k l , k 2 , * - - , k 2, * - * , K 1. Usingthisin = + where E is a K X K matrix defied by (7), Wi and Et are (13) we obtain vectors'of length K given by
e ,
..
e . . ,
a ,
e - . ,
Wi= [ wi(l),
* * *
wi(K)]
(174 (17b)
l where m = 0. Therefore, the zeros of D K + 1 (u) occur for all u given by (14). The spurious zeros can be ignored if all of them occur outside the visible range, i.e., for JuI > 1. We can ensure this by using a periodic array withaninterelement spacing of M2. However, it is shown later that under noisy conditions, the mean squared error in the estimate of the target locations decreases as the length of the array increases. We would therefore like to usearrays whose interelement spacings are much larger than X/2 while at the same time keep all the spurious zeros of DK+I(u) outside the visible range. Overspacing the elements in a periodic manner leads to grating lobes, which replicate the target poles abouteach grating lobe. If the element locations are randomized it will be difficult to find a set of integers mk, k = 2, , K + 1 , which satisfies (14) for IuI < 1. However, there is a nonzero probability that some ofthe spurious zeros of DK+(u) occur inside the visible range, and this probability increases as the length of the array increases. Thus we would like to have a stronger solution towards eliminating the spurious zeros of DK+](u) the in visible range, rather than just depending upon randomizing the element locations. One way to accomplish this would be to let the denominator of I(u) equal the sum of several terms, all of which have zeros at the target locations, but whose spurious zeros occur at different values of M . To this end, we reformulate the SR technique in the following manner.
where the e,,,, in (1%) are given by (1). We present now the random array method (RAM) for estimating the intensity spectrum. From the preceding paragraph we conclude that if the xi, i = 1, * -,N , are random numbers, it is highly unlikely that the RAM estimate of the intensity spectrum given by
x
K n=l
wi(n)eiklrxn (18)
l2
where C is a positive constant, will exhibit any spurious poles in the visible range. Z(u) may exhibit local maxima at the locations of the spurious zeros of each of theDi(u);however, by increasing the number ofelements, it is possible to decrease the heights of these local maxima. A general procedure for positioning the transmitters and receivers so as to eliminate all the poles and zeros of Z(u) in the visible range will not be attempted in this paper. Computer simulations indicate that for thinning ratios (the thinning ratio is defined as the ratio of the length of the array L to the length of the filed array with the same number of receivers Nand is equal to 2L/((N - 1)h) less than 10, (14) may used be to obtainthe desired element positions if the ratio of the number of receivers to the number of transmitheceive antennas is greater than 1. From (18) we observe that the RAM estimator is a parallel resistor type summation of (N- K ) all-pole estimators, each of order K. It is shown in Appendix I that this summation makes it possible to directly estimate the target strengths from
1210
IEEE TRANSACTIONS ON ANTENNAS A D PROPAGATION, VOL. AP-35, NO. 11, NOVEMBER 1987 N
RAM. The MLM estimator can also be written as the parallel resistor summation ofseveral all-pole estimators, however the order of these estimators varies from 1 to K [17].Because M targets can be correctly represented by anall-polemodel whose order is atleast M , we would expect RAM to yield a the better estimate of the target locations than the MLM. For future use, we note that (16) may also be written as STWi= - S X ; , i = K + l ,
-.e,
(19)
where S is a K X M matrix, T i s an M x K matrix, and X ; is a vector of length M. These quantities are defined below:
smn = Sn
exp (jkunxrn)
process. The validity such of a linear approximation is supported by computer simulation results in Section IV. The variance of the peak position error is derived in Appendix I, andisgivenby (40). This expression does not lend itself toward making simplifying observations about the nonlinear RAM intensity spectrum estimator. It will now be simplified for the two target case. Assume that thetwo targets to be resolved are at u = u1 and u = u2 and have complex amplitudes S and S2, respectively. Let the imagingsystem I consist of two transmitheceive antennas at x = 0 and x = L , and ( N - 2) receivers at x = X k , k = 3, N, where the xk are independent uniformly distributed random variables between 0 and L . Equation (40) will now be simplified for two cases, viz: closely separated and widely separated targets. When Iul - u21 < 0.15X/L, i.e., the two targets are separated by less than 0.15 beamwidths, (40) can be approximated by [18]
a ,
0;
P E{(u-u;)~}=(~u~/IS~~~)(~/L)~(CI
i=l, 2
(23)
where C1is a dimensionless constant which dependson the x;, i = 3, N , and usually is between and When the targets are separated by more than a beamwidth, i.e., (ul - z.421 = (n + p)X/L, where n is a nonzero integer and 0 < p < 1 (40) can be approximated by [ 181
--
e,
0:
& E{(~-u,)~)=(~cT:/IS~~~)(X/L)~
gp, =
i= 1
where C2 is a dimensionless constant which depends on xk, k = 3, , N andisusuallybetween 0.002 and 0.20. The where the quadrature components of np, are assumed to be presence of the ( 1 - cos 2 ~ pterm in the denominator of (24) ) independent identically and distributed Gaussian random indicates that the RAM cannot resolve targets which are variables each with zero mean and variance 0:. The SR separated by integral numbers of beamwidths. This condition technique involves solving thefollowing (N - K) matrix was already derived in (5). equations to obtain (N- K ) weight vectors W ; B. Prediction Error Energy EWi= -Eil i = K + l , -..,N (22) When the number of targets M . exceeds the numberof where is a K X K matrix whose pqth element is equal to available transmitters K , it is notpossible to resolve the targets. gP,, J!?L = [&li, -,6 K ; I T , and R. = [R.(l), W i ( K ) ] by usingtheRAM estimator. Therefore, it is important to * . ' The Wj obtained by solving (22) are substituted into (19) to formulate a criterion to determine the number oftargets. Such obtain the RAM intensity spectrum I(u). Clearly, the presence a criterion has been developed for various SR techniques and of noise in the measured CFA will makethe weight vectors W ; is often referred to as the prediction error energy [ 101. We deviate from their noiseless values W;.Therefore the peaks in start by defining the prediction error energy at the pth update, I(u) will be shifted from the actual target locations, and the denoted by PE(p), as amount of the shift will depend on the noise samples np,. We P shall now evaluate the variance of this shift in the peak PE@) g p + ~ , p ++ 1 gp+ l , m G p + , ( m ) (25) locations of T(u). m=I Assume that W; = W; + 6 W;, i.e., the noisy weights are l(m),m = 1 , * , the sum of the noiseless weights plus an error weight vector. where the gP, are given by (21)and the Gp+ Assume further that the peak position error is a smooth p , are obtained by solving (22).It can be shown thatPE(p) is function ofthe noise samples. Then it may be represented by a of the order of the signalpowerwhen p < M (M is the Taylor's series within a finite region of convergence. For number of targets) and is of the order of the noise power when sufficiently highS N R the peak positionerror may be estimated p 2 M [ 1 8 ] .ThevaluesofPE(m), m = 1 , " - , K - 1 , can be evaluated by solving (22) for the weights and substituting by truncating this series after the linear term. Once this approximation is made, the variance of the peak positionerror these weights into(25).If all the PE(m) are of the order of the can be calculated giventhe second-order statistics of the noise noise power, we may conclude that only one target is present.
(21)
( c 2 / ( ( N - 2 ) ( l-cos ~ T P ) ) ] ,
i = 1 , 2 (24)
--
e ,
l2
121 1
/
INCOHEREST
-5.0-
-15.0-
-25.0-
-35.0-
Fig. 2.
' -81.5 ' -81.4 ' -8.3 ' -d.2 ' - d l ' d.0 ' a ' l ' 012 ' 0.'3 ' 014 ' n.'5 ' B.'Eu Comparison of RAM with Fourier beamforming. Intensity spectra obtained by using two transmitkeceive antennas and four receivers. Len,@ of array = 45 wavelengths. Two unit amplitude targets at u = - 0.05 and 0.05.
-45.0 -0.6
If PE(k) Q PE ( k - l), k > 1, wemayconcludethat k targets are present. If all the PE(m) are of the order of the signal power, we may conclude that more than K - 1 targets are present.
C. Statistics of I@) Near Its Maxima
In Appendix I it is shown that the expected value of I(&), where Eii is the estimate of the location of the ith target, is proportional to lSi12. Further, it is shownthatif the noise powercanbeestimated(by estimating the varianceof the received signal from empty rangebins), it is possible to obtain a value for the constant in the numeratorof I(u) so that E { I ( & ) } = I Si 1 *. This value of C is by given (42). Substituting (42) into ( l e ) , we obtain the final form of the RAM intensity spectrum estimate I(u), for N - K > 1 , as
I(u)= ( 4 U 3 N - K - l ) / K }
[
i=K+ I
lDi(u)l2] - 2 ,
E { [ I ( a i ) - ISil']'}ISi12/(A'-K-2),
N-K>2. (28)
From (27) we observe that the expected value of the peaks in the RAM intensity spectrum is proportional to the target intensity; this advantage is not found in the MCM and CM.
IV.
REsULTs
observed CM that
CoMPrUUSoN
MCM
In this section, the validity of the varianceof the target location estimation error, given by (40) is first confirmed by
computer simulations. This variance is also compared to the variance of the target location estimation error for the MCM and CM. Theoretical expressions for the variance for the MCMandCM are available for very closely separated and widely separated targets [19];however, we shall be dealing with target separations where the approximations leading to the expressions for the variancein [19] maynotbe valid. Therefore, empirical values for the variance for the MCM and CM are obtained from Monte Carlo simulations. In of most the simulations, two targets, each of unit magnitude,were located at u = -0.05 and 0.05: respectively. The imaging system consisted of two transmitheceive elementsat x = 0 and x = L , and four receivers at x = 0.93L, 0.69L, 0.54L, and 0.18L. Therefore, the beamwidth of the array is X / L , the separation between the targets in beamwidth units is O.lL/X, and the thinning ratio 2L/X(N 1 ) is 0.4L/h. The SNR is defined as 1/2a$ where ut,is the variance of each of the quadrature components of the noise. Because two transmitters are needed to estimate the intensity spectrum using the RAM and only one transmitter is required for the MCM and CM, the SNR for the MCM and CM was always doubled, thus keeping the total transmitted power constant. Six equispaced receiving elements used were for the MCM and CM. Theseparation between adjacent elements was M2. Therefore the length the array was of fixed at 2.5 X for both the MCM and CM, and was a parameter of the length L for the RAM. Fig. 2 is a plot of the RAM intensity spectrum for various SNR, for a thinning ratio of 18. The separation between the targets is 4.5 beamwidths. Fig. 2 demonstrates that thinned arrays can be used for imaging purposes without any significant spurious peaks in the intensity spectrum. be It may the sidelobe characteristics of all the images in as follows. The Fig. 2 are similar. Thismaybeexplained sidelobe artifacts occur at the spurious zerosof all the Di(u) in
IEEE TRANSACTIONS ON AhTENNAS AND PROPAGATION, VOL. Ap-35, NO. 11, NOVEMBER 1987
: TBEORY
10
20
10
$0
so
60
. I !O (SSR dB)
Fig. 3. RMS angular error in target location estimates versus SNR for various array lengths. Two unit amplitude targets at u = -0.05 and 0.05, Solid lines indicatetheoreticalvalues for RAM, CM, and MCM. The various symbols represent empirical values from Monte Carlo runs.
(18), which depend on the element and target locations via (14). Since the array and target geometry were the same for all the three images in Fig. 2, their sidelobe characteristics are similar. Also shown in Fig. 2 is a noiseless intensity spectrum obtained by using conventional (Fourier) beamforming. The intensity spectra obtained from thetwo transmissions were averaged to form this image. Although the SNR is infinite, the peak artifact of the Fourier intensity spectrum, due to the small number of receivers, is - 1.3 dB. The RAM clearly outperforms conventional Fourier imaging. Fig. 3 confirms the validity ofthe theoretical variance of the target location estimation error u i = 1, 2, given by (40). : , The root mean squared angular error (RMSE) in estimating the location of the target at 1c = 0.05 is plotted against SNR for various array lengths. The solid lines represent the theoretical RMSE obtained from (40). The various symbolsrepresent empirical valuesobtained by usingMonte Carlo methods. Fifty trials were conducted for each SNR and , and the target L location estimates iij, i = 1, 2, were obtained by looking for the two highest peaks of I(u) in the visible range. Empirical values for the variance were obtained by averaging the 50 values of (ui It can be seen that the theoretical and empirical values of RMSE agree closely. The symbols + and x represent empirical values of the RMSE in estimating the location of the target at u = 0.05 by using the MCM and CM respectively. The RMSE for the RAM with a 165 wavelength array is about two orders of magnitude lower than the RMSE for the MCM and CM. When the RMSE is approximately 0.05, the widths of the two peaks in I(u) near the target locations are of the same order as the separation between them. When thishappens, the two peaks merge to form a single peak about midway between
the two targets, i.e., theRMSE in estimating each target location is half the distance between them. This phenomenon is evident in the data for the three 2.5 wavelength arrays, all of which taper off at low SNR to about 0.05. The empirical variance also deviates from theory for low SNR. This behavior may be explained in the following manner: while evaluating the theoretical variance, the peak position error was approximated by truncating its Taylors series after the linear term. This approximation is valid for large SNR, however, when the SNR is small, (40) will yield an inaccurate value for variance since the higher order terms of the Taylors series will also contribute to the peak position error. As mentionedearlier, the variance of the error in estimating the target locations is much higher for the MCM andCM. This indicates thatthe accuracy in estimating the target location can be vastly improved by using RAM as opposed to the MCM the and CM. Further, because the widthsof the peaks in the estimatedintensity spectrum are proportional to the RMS angular error, the RAM can be used to resolve targets which may be not resolvable by MCM the and CM. This is demonstrated in the following figure. Fig. 4 is a plot of the estimated intensity spectrum obtained by using the RAM, MCM, and CM. In order to keep the transmitted power constant, the noise power in the measured CFA for the MCM and CM was half the noise power in the CFA for the RAM. The narrow widths of the peaks in the intensity spectrum for large L clearly indicate that the RAM outperforms the MCM andCM. The spurious peaks thatoccur in the intensity spectrum when the length of the array is large have been suppressed to a level of - 19 d B , i.e., the noise power level. This confirms the observations made previously on suppressing the spurious peaks while using thinned arrays to estimate the target locations. V. EXPERIMENTAL RESULTS The theory developed previously will now be verified with real data. The following experiment was conductedat the field site of the Valley Forge Research Center, University of Pennsylvania. A 16 element array with adjacent elements separated by 0.4 m (lengthof array = 16 m) wasused for imaging purposes. The wavelength of the transmitted radiation was 3.125 cm. Three one foot comer reflectors were placed in thenear field ofthe array as shown in Fig. 5. The range resolution of the radar system was 75 cm. The distances in Fig. 5 and the element locations were measured by tape, and hence are only approximate. The SNR per element in the received signals was estimated to be 20 dB. For each range bin, two sets of echoes were recorded at each of the 16 elements by transmitting sequentially from two different locations. The comer reflector at the 62 m range bin was used as a beamformer to phase cohere the array. Fig. 6 is the intensity spectrum of the targets in the 65 m range bin obtained by 1) adding the Fresnel intensity estimates obtained from the two sequential transmissions, labeled Fresnel beamforming plus incoherent addition, 2) the RAM developed in Section IV-B, and 3) CM. The two targets were separated by eight beamwidths (length of array = 6 m) for the first image;
1213
5.0-
-5.0-
2
z,
-25.R
-35.
-0ls
i
-al5
I
-013
-0:2
-;.I
0 , ;
011
0.12 '
013
014
015
0;s
Fig. 4. Comparison of RAM w t MCM and CM. RAM intensity spectrum obtained by using ih two transmit/receive antennas and four receivers, CM and MCM w t six receivers. SNR = 20 dB for RAM. 23 dB for CM and MCM. Two unit amplitude targets at ih u = -0.05 and 0.05.
x j y2 . 5 ~ 1x4
65m
62m
__--RECEIVING ARRAY
-4Le
-2.5
/.
-$.B
-1.5
. '
-!.a
e.a
, . '
E l
x5
1.8
/-.----..-.
1.5
ZB 215
( 4 SECEIYERS)
-a5
CROSSRABGE (RETERS)
I-
6m
------I
for microwave experiment.
Fig. 5 . Arrayandtargetgeometry
Fig. 6 . Experimentalverification of RAM and comparisonof RAM w t ih Fourier beamforming and CM. Intensity spectra obtained by using RAM with two transmitters and three receivers, CM with one transmitter andfour receivers, and incoherent addition of hvo Fourier images obtained by using two transmitters and 16 receivers.
hence would we expect this image to estimate the target locations accurately. The weights for the RAM and CM were obtained by using three and four adjacent elements, respectively; therefore the two targets were separated by about one beamwidth for both these methods. RAM yields an accurate estimate of the target locations and intensities, and does not contain the sidelobe artifacts present in the diversity image. CM yields a distorted image as the quadratic curvature of the wavefront leads to errors in the subarray averaging performed during its impIementation.
VI. SUMMARY
An SR technique, called the random array method, for resoIving coherent targets located in the near or far fields of
periodic andfor aperiodic arrays has been developed in this paper. The key to its success is the use of sequential transmissions from different locations for decorrelating the coherent targets. Existing SR techniques that can be used to resolve coherent targets such as CM [ 11, [4], or the MCM [ 5 ] , [6] require that 1) the array be periodic, and 2 ) the targets be located in the far field of the array. Other techniques require a moving array [ 111 or a multidimensional search for the target locations which could be computationally cumbersome [121. In addition to removing these restrictions, it is shown that for the same amount of transmitted power andthe same numberof receivers, the MSE in the RAM target location estimates is more than an order of magnitude lower the MSE in the MCM and CM target location estimates. A theoretical expression for
1214
IEEE TRANSACTIONS ON AhTENNAS AND PROPAGATION, VOL. AP-35, NO. 11, NOVEMBER 1987
the contribution of noise to the MSE is derived, and confirmed by setting the derivative of the truncated series to zero, i.e., by empirical methods. The practicability of this technique is iii- ui= - A ' ( u ; ) / A "(u;). (34) demonstrated with measured microwave data. APPENDIX I The RAM intensity spectrum estimate in the presence of noise is Using AOk(u;) = 0 for k = K + 1, N,substituting (32) into (36), andneglecting terms containing A;,(ui) in favor of A (ui) obtain we
a ,
Ok
wheretheweight vectors W ; are obtained by solving (22). Substituting mi = Wi + 6 W i? V i is the noiseless weight ( vector given by (16)), I? = E + N (E is the noiseless CFA matrix defined by (7) and N is a K x K noise matrix whose pqth element is np4),and I?; = Ei + N; where E,? = [eli,* , eKj] ' = [nl;, , nKi] (29), and using E = STand and N into E; = SXi, we obtain
--
From (29) and (31) wecanshow that E{Alk(ui)) = 0, therefore the mean value of the right hand side of (35) is zero ) (note that A o ~ ( uis deterministic). Therefore E { & } = u;, i.e., I(u) is an unbiased estimator of the target locations. The mean squared error in the target location estimates is
where S , T, and X; are defined by (20). Using STW; = - SX; and ignoring the productof two small quantities viz.: the noise matrix and the error weight vector we obtain
~&(u;)]]/ 2
[5
lAik(Ui)l2]
2-
(36)
k=K+1
(29)
We now shall evaluate the terms on the numerator of the right hand side of (36). Using (20) for the matrices S and T , and (29) we can show that
i=K+ 1
where Aoj(u)=exp (jkux,)+ and (31b) where the superscript indicates transposition and complex conjugation of a matrix, and the subscript ii stands for the ith we wish to determine the maxima of I ( ~near the target diagonal element of a matrix. In a similar fashion we obtain ) locations, i.e., the minima of A (u) near u = u,, rn = 1, , E { ( A ~ ~ ( u = ) ){~ )~ ~ ( u ; ) A I I ( u ; ) ) = O(38) ;E A M y where and
n=l
x
K
n=I
x
K
A(u)=-=
( '1
j=K+1
x
N
IAoi(u)+Al;(u)I2*
(32)
E{A~~(~;)A~,(U;))=(~'TZ~/IS;~~)
Expanding A (u) in a Taylor's series around u = u;, i = 1 , .,M , and neglectingterms higher thanthe second order, we Substituting obtain A (u) A (u;) (U - u;)A'(u;) 0.5 = + +
(U - u
~ ) "(u;) (33) ~ A
whereA'(u;) = dA(u)/duevaluated at u = ui, etc. Letthevar minimum of A (u) in the neighborhood ofu; occur at ti;.If tii is sufficiently close to ui so that the truncated Taylor's series reasonably approximatesA (u) at ti;, then tij can determined be
k=K+I
.1215
where
REFERENCES
Equation (40) is the mean squared error in the estimate of the target located at u = ui, and is one of the main results of this appendix. As T , W,, and A i k ( u i )do not depend on the target strengths Si. we observe that the MSE isinversely proportional to the SNR of the target. We now evaluate the mean and variance of the target intensity estimate I(&) for large M and N - K , with K 2 M . Substituting (33) and (34) into (32),we obtain
,
I(ai)/(A(~i)-[(A(u;))/2A(ui)]).
For large M and N - K , the second term denominator can be neglected [ 181, and
in the
where akiand bkiare the real and imaginary parts of A Ik(ui). The aki and bki are the weightedsumofGaussianrandom variables, therefore they are also Gaussian. It can be shown random that the a,; and bki are iid zero mean Gaussian variables whosevariance is half of (A-9). Further, for large K it can be shown that (T*T7)-l tends to I / K (lis the K X K 1) identity matrix), and 2) CIwk(k(n)l (the summationistaken over n from 1to K ) tends to unity. Using these in the preceding equation we obtain E{aii! = E { b i ; } = 2ui/ KlS, I 2. Finally, the denominator of I(ui) is the sum of 2(N K ) iid zero mean Gaussian random variables, and has a chisquared pdf with 2(N - K ) degrees of freedom [21].Using this density we obtain
E { l ( & ) } 03 for N - K = 1 , E{I(ti,)} = =CKISi12/4u2,(N-K-1), forN-K>l.
A. K . Luthra, Maximum entropy method in the space-angle domain and a new technique with superior performance, Ph.D. dissertation, Syst.Eng. Dept., Univ.Pennsylvania,Philadelphia,1981. J. E. Evans, J. R. Johnson, and D. F. Sun, Application of advanced signalprocessingtechniques ,to angle of arrival estimationin ATC navigation and surveillance systems, Lincoln Labs., Lexington, MA, Tech. Rep. 582, June 1982. J. Makhoul, Linear prediction: A tutorial review, Proc. IEEE, vol. 63, no. 4, Apr.1975. T. J.Shan, M.Wax, and T. Kailath, spatial On smoothing for direction of arrival estimation of coherentsignals, IEEE Trans. Acoust., Speech and Signal Processing, vol. ASSP-33, no. 4, Aug. 1985. A. H. Nutall, Spectral analysis of a univariate process with bad data points, Naval Underwater Syst. Center, NewLondon, CT,Tech. Document 5419, May 1976. T. J. Ulrych R. and W. Clayton, Times series modeling and maximum entropy. Phys. Earth Planetary Interiors, vol. 12, Aug. 1976. location and signal parameter R. 0. Schmidt, Multiple emitter estimation, in Proc. RADC Spectrum Estimation Workshop, Oct. 3-5, 1979, pp. 243-258. J. Capon, High resolution frequency-wavenumber spectrum estimation, Proc. IEEE, vol. 57, Aug. 1969. F.Haber,R. S. Berkowitz,and 3. S. Meagher,Radio location in multipath and interference environments, Valley Forge Research Center, Univ. Pennsylvania, Philadelphia, UP-VFRC-18-81, Rep. Mar.1981. S. M. Kay and S. L. Marple Jr., Spectrum analysis-A modem perspective, Proc. IEEE. vol. 69, Apr. 1981. F. Haber and M. Zoltowski, Spatial spectrum estimation in a coherent signal environment using an array in motion, IEEE Trans. Antennas Propagat., vol. AP-34, no. 3, Mar. 1986. M. Zoltowskiand F. Haber, A vector spaceapproach to direction finding in a coherent multipath environment, IEEE Trans. Antennus Propagat., vol.AP-34,no. 11, Sept. 1986. C.Jin, Adaptive superresolutioninthenearfieldusingperiodic arrays, Masters thesis, Elec. Eng. Dept., Univ. Pennsylvania, 1986. B. D. Steinberg: Principles o Aperture and Array System Design. f NewYork:Wiley,1976. J. P. Burg, Maximum entropy spectral analysis, in Proc. 37th Meet. Soc. Exploration Geophys., 1967. S . L. Marple Jr., A new autoregressive spectrum analysis algorithm, IEEE Trans. Acoust., Speech, andSignalProcessing, V O ~ .ASSP-28, Aug. 1980. J. P. B u g : The relationship between maximum entropy spectra and maximum likelihood spectra. Ceophys., vol. 37, Apr. 1982. H. M. Subbaram, Superresolution of coherent targets using aperiodic antenna arrays, Ph.D. dissertation. Elec. Eng. Dept., Univ. Pennsylvania,Philadelphia, 1984. S. W . Lang and J. H. McClellan, Frequency estimation with maximum entropy spectral estimators, IEEE Trans. Acousr., Speech, and Signal Processing, vol. ASSP-28, Apr. 1980. B. D. Steinberg. Microwave Imaging with Large Antenna Arrays. New York: Wiley, 1983. K . A. Doksum, Mathematical Statistics. CA: P. J. Bickel and Holden-Day, 1977.
If the noise power can be estimated, the value of C can be chosen so that E{Z(&)} = 1S;I. This value of C is
C = ~ U ~ ( N - KN / K ,> l . 1) -K (42)
Harish M. Subbaram was born in Mysore, India, on November 8, 1957. He received the %.Tech. degree inelectronicsengineering from the Indian Institute of Technology, Madras, in 1980, and the M.S.E. and Ph.D. degrees in electrical engineering from the University of Pennsylvania, Philadelphia, in 1983 and 1986, respectively. He is currently a Postdoctoral fellow atthe Valley Forge Research Center, UniversityofPennsylvania. His research interests include signal and image processing, microwave and ultrasonic imaging, and spectrum analysis.
N - K= 1
(43d
N-K>l
(43b)
(44d
Wb)
var {l(Lii)}=o3, - K 5 2 N
= IS;I,
N-K>2.
1216
JEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION,VOL. AP-35, NO. 11, NOVEMBER 1981
Bernard D. Steinberg (S148-A5O-SM&I-F66) was born in Brooklyn, N Y , in 1924. He received the B.S. and M.S. degrees in electrical engineering
from the Massachusetts Institute of Technology, Cambridge, in 1949, andthe Ph.D. degree from the University of Pennsylvania, Philadelphia, in 1971. He worked in the Research Division of Philco through the middle 1950s. He was one of the founders of General Atronics Corporation in Philadelphia in 1956 and served as its Vice President and Technical Directorfor 15 years.Currently, he is Chairman of the Board and a founder of Interspec Inc., also in Philadelphia. He has been aProfessor of ElectricalEngineering at the University of Pennsylvania, Moore School of Electrical Engineering, since1971. He is the
Director of the University of Pennsylvanias Valley Forge Research Center. The primary work ti laboratory is the developmentof large, self-adaptive of hs microwave imaging systems based on the radio camera technology developed by the laboratory. His work has been inradar backscatter and in signal processing techniques and their applications to radar, HF communications, hydroacoustics and seismology. His most recent work is in seIf-adaptive signal processors, particularly in large antenna arrays. Dr.Steinberg is the author of hincipla of Aperture and Array System Design (Wiley, 1976). in which the random array is analyzed, and Microwave Imaging with Large Antenna Arrays (Wiley,1983), which descriks radio camera principles and techniques. For many years, he was a consultant to the Airborne Radar Branch of the Naval Research Laboratory, Washington, D.C. He is a member of U.S. Commissions B and C of the International Union of Radio Science.