IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO.

8, AUGUST 2002 1843

Maximum-Likelihood Source Localization and

Unknown Sensor Location Estimation for Wideband
Signals in the Near-Field
Joe C. Chen, Student Member, IEEE, Ralph E. Hudson, and Kung Yao, Fellow, IEEE

Abstract—In this paper, we derive the maximum-likelihood have proposed to use focusing matrices to transform the wide-
(ML) location estimator for wideband sources in the near field band signal subspaces into a predefined narrowband subspace
of the sensor array. The ML estimator is optimized in a single [7]–[9]. Then, the MUSIC algorithm can be applied afterwards
step, as opposed to other estimators that are optimized separately
in relative time-delay and source location estimations. For the as if it was the narrowband case. This class of methods is re-
multisource case, we propose and demonstrate an efficient alter- ferred to as the coherent signal subspace method (CSM). Other
nating projection procedure based on sequential iterative search techniques that evolved from the CSM method include the use of
on single-source parameters. The proposed algorithm is shown steered covariance matrix [10] and interpolated array [11]. How-
to yield superior performance over other suboptimal techniques, ever, the drawback of these methods is the preprocessing that
including the wideband MUSIC and the two-step least-squares
methods, and is efficient with respect to the derived Cramér–Rao must be performed beforehand. For instance, the CSM methods
bound (CRB). From the CRB analysis, we find that better source need to construct the focusing matrices that require focusing an-
location estimates can be obtained for high-frequency signals than gles that are not far from the true DOAs. The interpolated array
low-frequency signals. In addition, large range estimation error technique requires the construction of the interpolation matrix,
results when the source signal is unknown, but such unknown which is also a function of angle. Furthermore, these techniques
parameter does not have much impact on angle estimation. In
some applications, the locations of some sensors may be unknown have never been tested for the near-field case. Recently, an ex-
and must be estimated. The proposed method is extended to tension of the above methods, but without any preprocessing
estimate the range from a source to an unknown sensor location. step, has been studied for the near-field case [12]. It is a nat-
After a number of source-location frames, the location of the ural wideband extension of the conventional MUSIC algorithm;
uncalibrated sensor can be determined based on a least-squares thus, we refer to this method as the wideband MUSIC algorithm
unknown sensor location estimator.
and compare it with the proposed method of this paper. It is
Index Terms—Array shape calibration, beamforming, shown that wideband MUSIC yields poor estimation results (es-
Cramér–Rao bound, source localization. pecially in range) for a finite number of data samples (limited
due to moving source) and low SNR.
I. INTRODUCTION Another class of wideband source localization algorithms is
based on time delays (usually for a single source only). These

S OURCE localization using sensor arrays has been one of

the central problems in radar, sonar, navigation, geophysics,
and acoustic tracking. For the past few decades, a wide variety
techniques include two independent steps, namely, estimating
relative time delays between sensor data and then source loca-
tion based on the relative time-delay estimates. Closed-form so-
of techniques have been proposed for the estimation of the di-
lutions can be derived for the second step based on spherical
rection-of-arrival (DOA) of narrowband sources in the far-field
interpolation [13], [14], hyperbolic intersection [15], or linear
[1]–[5]. In [6], the authors provided an excellent review and
intersection [16]. However, these algorithms require the knowl-
comparison of the various narrowband techniques. Recently,
edge of the speed of propagation (e.g., the acoustic case). In
there has been an interest in locating wideband sources in the
[17] and [18], closed-form least-squares (LS) and constrained
near field, including both the range and the DOA. One obvious
least-squares (CLS) solutions are derived for the case of un-
application is to locate nearby acoustic sources using a micro-
known speed of propagation (e.g., the seismic case). The CLS
phone array.
method improves the performance from that of the LS method
For wideband signals, e.g., acoustic or seismic, the algorithms
by forcing an equality constraint on two components of the
need to be tailored to account for the large bandwidth. Many
unknowns. When the speed of propagation is known, the LS
method in [17] and [18], after removing such unknowns, be-
Manuscript received June 11, 2001; revised April 18, 2002. This work was comes the method independently derived later in [19]. In [19],
supported in part by the Defense Advanced Research Projects Agency SensIT the authors show that their method is mathematically equivalent
Program under Contract N66001-00-1-8937, NASA-Dryden under Grant to the spherical interpolation technique [13], [14] but with less
NCC2-374, an UC DIMI grant sponsored by STMicroelectronics, Inc., and a
Raytheon Doctoral Fellowship. The associate editor coordinating the review of computational complexity. We refer to the method in [19] as the
this paper and approving it for publication was Dr. Joseph Tabrikian. LS method of known speed of propagation and use it to com-
The authors are with the Electrical Engineering Department, University pare with the proposed method of this paper. In theory, these
of California, Los Angeles, CA 90045 USA (e-mail: [email protected];
[email protected]; [email protected]). two-step methods can work for multiple sources, as long as the
Publisher Item Identifier 10.1109/TSP.2002.800420. relative time delays can be estimated from the data. However,
1053-587X/02$17.00 © 2002 IEEE
1844 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 8, AUGUST 2002

this remains to be a challenging task in practice. Recently, the output an estimate of the range from a source to one or more sen-
use of higher order statistics to estimate the relative time delays sors of unknown locations. When several ranges from different
for multiple independent sources has been studied in [20]. source locations are accumulated, they can be combined by a
In this paper, we derive the “optimal” parametric ML solution least-squares calibration (LSC) method to estimate the unknown
to locate wideband sources in the near field. The wideband data sensor locations. This extended AML method shows promising
is transformed to the frequency domain, and the signal spectrum improvement over our previous method based on LS source lo-
can be represented by the narrowband model for each frequency calization [26]. Other issues such as calibrating the sensor gain,
bin. This allows a direct optimization for the source location phase, mutual coupling, and channel mismatch are also critical
under the assumption of i.i.d. noise instead of the two-step opti- in practice, but we ignore them for the purpose of this paper.
mization that involves relative time-delay estimation. However, The paper is organized as follows. In Section II, we derive
in practice, we apply the DFT, and thus, there are a few artifacts. the AML solution for source localization. Then, the CRB for
The circular shift property of the DFT introduces an edge effect source localization is derived in Section III. In Section IV, the
problem for the actual linear time shift, and this edge effect is LSC unknown sensor location estimation is introduced. Then,
not negligible for a small block of data. To remove the edge ef- simulation examples and systematic evaluations are provided to
fect, appropriate zero padding can be applied. However, it is also show the effectiveness of the proposed algorithms in Section V.
known that zero padding destroys the orthogonality of the DFT Finally, we give our conclusions.
transform, which makes the noise spectrum appear correlated
across frequency. As a result, an exact ML solution for data of II. MAXIMUM-LIKELIHOOD SOURCE LOCALIZATION
finite length does not exist. Instead, we ignore these finite effects
A. Derivation of the AML Solution
and derive the solution, which we refer to as the approximated
ML (AML) solution. Note a similar solution has been derived The derivation of the AML solution for real-valued signals
independently in [21] for the far-field case, but here, we con- generated by wideband sources in the near field is an exten-
sider the near-field case. sion of the classical maximum-likelihood DOA estimator for
In practice, the number of sources must be determined inde- narrowband signals. Due to the wideband nature of the signal,
pendent of any localization algorithm, but here, we assume it is the AML metric results in a combination of each subband. Since
known for the purpose of this paper. For the single-source case, we consider near-field sources, the signal strength at each sensor
we show the AML formulation is equivalent to maximizing the can be different due to nonuniform spatial loss in the near-field
sum of the weighted cross-correlation functions between time geometry. The sensors are assumed to be omnidirectional and
shifted sensor data. The optimization using all sensor pairs mit- have identical responses. For a randomly distributed array of
igates the ambiguity problem that often arises in the relative sensors, the data collected by the th sensor at time can be
time-delay estimation between two widely separated sensors for given by
the two-step LS methods. In the case of multiple sources, we
apply an efficient alternating projection (AP) procedure, which (1)
avoids the multidimensional search by sequentially estimating
the location of one source while fixing the estimates of other
source locations from the previous iteration. for , , and ,
Besides the development of the AML method, we also derive where
the theoretical Cramér–Rao bound (CRB) for both performance number of sources (assumed to be known and
comparison and basic understanding purposes. The CRB shows );
that the localization variance bound can be broken down into signal gain level of the th source at the th sensor
two separate parts: one that depends on the signal characteris- (assumed to be constant within the block of data);
tics and one that depends on the array geometry. The signal de- source signal;
pendent part shows that theoretically, the source location RMS fractional time-delay in samples (which is allowed
error is linearly proportional to the noise level and the speed of to be any real-valued number);
propagation and inversely proportional to the source spectrum zero mean i.i.d. Gaussian noise with variance
and frequency. Thus, better source location estimates can be ob- The time delay is defined by , where
tained for high-frequency signals than low-frequency signals. In th source location;
further sensitivity analysis, large range estimation error is found th sensor location;
when the source signal is unknown, but such an unknown pa- speed of propagation in length unit per sample.
rameter does not affect the angle estimation. Define the relative time-delay between the th and the th sen-
It is well known that the performance of source localization sors by
degrades when the sensor locations are in error. This problem A block of samples in each sensor data can be transformed to
arises in many cases where the sensors may be randomly de- the frequency domain by a DFT of length . As discussed ear-
ployed in an ad hoc network or even move to different positions lier, the DFT creates a circular time shift rather than the actual
in time. The estimation of unknown sensor locations, which are time shift. When , severe edge effect results for small
often called array shape calibration, has also drawn consider- , but it becomes a good approximation for large . We can
able attention in the literature [22]–[25]. In the later section of apply zero padding for small to remove such edge effect, i.e.,
this paper, we extend the AML source localization algorithm to , where is the maximum relative time-delay among
CHEN et al.: MAXIMUM-LIKELIHOOD SOURCE LOCALIZATION 1845

all sensor pairs. However, the zero-padding removes the orthog- the source signal vector that yields the minimum residual at any
onality of the noise component across frequency. In practice, the source location is given by
size of is limited due to the nonstationarity of the source lo-
cation. Thus, in the following derivation, we assume that either (5)
is large enough or that the noise is almost uncorrelated across
frequency and claim such a formulation as an approximation where is the pseudo-in-
to the maximum-likelihood solution. Throughout this paper, we verse of the steering matrix . Define the orthogonal
denote superscript as the transpose, as the complex conju- projection and the complement
gate transpose, and as the complex conjugate operation. orthogonal projection By sub-
In the frequency domain, the array signal model is given by stituting (5) into (4), the minimization function becomes
. After substituting the estimate of
, the AML source locations estimate can be obtained by
(2) solving the following maximization problem:
for , where the array data spectrum is
given by , the steering matrix (6)
is given by , the steering
vector is given by ,
Note that the AML metric has an implicit form for the es-
, and the source spectrum is given
timation of , whereas the metric shows the explicit
by . The noise spectrum form. Once the AML estimate of is obtained, the AML esti-
vector is zero mean complex white Gaussian distributed mate of the source signals can be given by (5). It is interesting
with variance in each element. Note due to the transforma- that when zero-padding is applied, the covariance matrix is
tion to the frequency domain, asymptotically approaches no longer diagonal and is indeed singular; thus, an exact ML so-
a Gaussian distribution by the central limit theorem, even if lution cannot be derived without the inverse of . In the above
the actual time-domain noise has an arbitrary i.i.d. distribution formulation, we derive the AML solution using only a single
(with bounded variance) other than Gaussian. This asymptotic block. A different AML solution using multiple blocks could
property in the frequency domain provides a more reliable noise also be formed with some possible computational advantages.
model than the time-domain model in some practical cases. For When the speed of propagation is unknown, as in the case of
convenience of notation, we define . By seismic media, we may expand the unknown parameter space
stacking up the positive frequency bins (zero frequency to include it, i.e., .
bin is not important and the negative frequency bins are merely
mirror images) of the signal model in (2) into a single column, B. Single-Source Case
we can rewrite the sensor data into a space–tem-
poral frequency vector as , where In the single-source case, the AML metric in (6) be-
, and . comes , where
We assume, initially, that the unknown parameter space is is the beam-steered beamformer output in the
, where the source locations frequency-domain [27], is the normalized
are denoted by , and the source signal steering vector, and is the normalized
spectrum is denoted by signal gain level at the th sensor. It is interesting to note that in
The signal gain levels are assumed to be known in the near-field case, the AML beamformer output is the result of
this derivation, and the effect of not knowing will be forming a focused spot (or area) on the source location rather
discussed in a later section. The log-likelihood function of than a beam since range is also considered. In Appendix A,
the complex Gaussian noise vector , after ignoring irrelevant the AML criterion is shown to be equivalent to maximizing
constant terms, is given by . The the weighted cross correlations between sensor data, which is
maximum-likelihood estimation of the source locations and commonly used for estimating relative time delays.
source signals is given by the following optimization criterion: The source location can be estimated based on where
is maximized for a given set of locations. Define the normalized
metric
(3)

which is equivalent to finding for all bins, (7)

where
where , which is useful to
(4) verify estimated peak values. Without any prior information on
possible region of the source location, the AML metric should
The minima of with respect to the source signal vector be evaluated on a set of grid points. A nonuniform grid is sug-
must satisfy ; hence, the estimate of gested to reduce the number of grid points. For the 2-D case,
1846 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 8, AUGUST 2002

polar coordinates with nonuniform sampling of the range and much performance degradation for the single-source case when
uniform sampling of the angle can be transformed to Cartesian the actual gain is merely modeled by the spatial loss.
coordinates that are dense near the array and sparse away from
the array. When the crude estimate of the source location is ob- C. Multiple-Source Case
tained from the grid-point search, iterative methods can be ap- For the multiple-source case, the parameter estimation is a
plied to reach the global maximum (without running into local challenging task. Although iterative multidimensional param-
maxima given appropriate choice of grid points). In some cases, eter search methods, such as the Nelder–Mead direct search
grid-point search is not necessary since a good initial location method, can be applied to avoid an exhaustive multidimen-
estimate is available, e.g., from the estimate of the previous data sional grid search, finding the initial source location estimates
frame for a slowly moving source. One simple iterative gradient is not trivial. Since iterative solutions for the single-source case
algorithm can be formulated in the following: are more robust and the initial estimate is easier to find, we
extend the alternating projection method in [2] to the near-field
(8) problem. The alternating projection approach breaks the
multidimensional parameter search into a sequence of single
where is the step size that depends on the shape of and source parameter search and yields fast convergence rate. The
initial estimate following describes the alternating projection algorithm for the
two-source case, but it can be easily extended to the case of
(9) sources.
Alternating Projection Algorithm:
Step 1) Estimate the location of the stronger source on a
is the gradient of the normalized , single source grid
, and .
Other advanced and robust methods may be used without (10)
specifying a step size, e.g., the Nelder–Mead direct search
method [28]. Note that the efficiency and complexity of these Step 2) Estimate the location of the weaker source on a
iterative methods are beyond the scope of this paper and, thus, single source grid under the assumption of a two-
will not be discussed. source model while keeping the first source location
Note that in the derivation of the AML solution, we have estimate from step 1) constant:
assumed known sensor gain levels , or the normalized
sensor gains , for . However, in practice, (11)
they are also unknown. The signal model in the far-field case
usually ignores this scale factor since the SNR is assumed to For , repeat steps 3) and 4) until conver-
be uniform across the array when the source is far away (for gence.
omni-directional sensors). When the source is in the near field, Step 3) Do an iterative AML parameter search (direct search
the SNR at each sensor is different due to the different spatial or gradient) for the location of the first source while
loss from the source to each sensor. Thus, the maximization keeping the estimate of the second source location
on the AML metric is indeed based on a weighted from the previous iteration constant:
least-squares criterion, where each sensor data is weighted
by its signal strength in the overall metric. When the are
(12)
unknown, they can be replaced by their estimates. Two simple
but biased estimators using the and norms for the can
Step 4) Do an iterative AML parameter search (direct search
be given by
or gradient) for the location of the second source
and , respec- while keeping the estimate of the first source loca-
tively. In Appendix B, an extended AML solution is derived tion from step 3) constant:
to incorporate unknown for the single-source case. It is
shown that the extended AML metric involves finding the
(13)
principle singular value of a newly defined data spectrum
matrix. The resulting source location estimate using this metric
yields a weighted least-squares solution, where the are also Similar to the single-source case, the sensor gain levels
implicitly estimated at the same time. Although this new metric can be nonuniform and unknown when all the sources are in
can effectively estimate the sensor gain levels, we find that the near field. However, when there are multiple sources, the
the source location estimation degrades when we expand the gains are not easily determined (still an open issue). Thus, we
parameter space to include the . Based on our simulation may assume uniform gains, i.e., . In our simulations,
results, we find that it is better to use separately estimated or when the signal gains are modeled by the spatial loss, we find
simply assume uniform gains, i.e., , for the single source the location estimates of two sources to be off from the actual lo-
AML method. The uniform gain assumption makes the AML cations by quite a significant amount. Therefore, the gains
solution a least-squares one instead of a weighted least-squares are crucial parameters in the multisource case, as opposed to the
one. However, we find that such an assumption does not give single-source case.
CHEN et al.: MAXIMUM-LIKELIHOOD SOURCE LOCALIZATION 1847

III. CRAMÉR–RAO BOUND FOR SOURCE LOCALIZATION where the penalty matrix due to unknown speed of propagation
is defined by . The matrix
The CRB is most often used as a theoretical lower bound for
any unbiased estimator [29]. Most of the derivations of the CRB is non-negative definite; therefore, the source localization error
for wideband source localization found in the literature are in of the unknown speed of propagation case is always larger than
terms of relative time-delay estimation error. In this section, we that of the known case.
derive a more general CRB directly from the signal model. By For the third case, when the source signal is also un-
developing a theoretical lower bound in terms of signal charac- known, i.e., , the matrix is
given by , where
teristics and array geometry, we not only bypass the involvement
of the intermediate time-delay estimator but also offer useful in- , and and are the magni-
sights into the physics of the problem. tude and phase part of , respectively. The Fisher information
We consider the following three cases: matrix can then be explicitly given by
1) known signal and known speed of propagation;
2) known signal but unknown speed of propagation; (18)
3) known speed of propagation but unknown signal.
The comparison of the three conditions (see Section V) provides where and are not explicitly given since they are not needed
a sensitivity analysis that explains the fundamental differences in the final expression. By applying the block matrix inversion
of different problems, e.g., unknown speed of propagation for lemma, the leading submatrix of the inverse Fisher in-
seismic sensing and known signal for some applications. The formation block matrix can be given by
case of unknown signal and unknown speed of propagation is
not too different from the case of unknown signal and known (19)
speed of propagation; thus, it is not considered. For all cases,
we can construct the Fisher information matrix [29] from the where the penalty matrix due to unknown source signal is de-
signal model defined in Section II by fined by

Re Re (14)
(20)
where for the first case, assuming that is the
only unknown in the single-source case. In this case, ,
where is the scale The CRB with unknown source signal is always larger than that
factor that is proportional to the total power in the derivative with known source signal, as discussed later. This can be easily
of the source signal shown since the penalty matrix is non-negative definite.
The matrix acts as a penalty term since it is the average
(15) of the square of weighted vectors. The estimation variance
is larger when the source is far away since the vectors are
similar in directions to generate a larger penalty matrix, i.e.,
is the array matrix, and is the unit vectors add up. When the source is inside the convex hull of
vector indicating the direction of the source from the th sensor. the sensor array, the estimation variance is smaller since
The matrix provides a measure of geometric relations be- approaches the zero matrix, i.e., vectors cancel each other.
tween the source and the sensor array. Poor array geometry may For the 2-D case, the CRB for the distance error of the estimated
lead to degeneration in the rank of matrix . It is clear from the location from the true source location can be given by
scale factor , as will be shown later, that the performance does
not solely depend on the SNR but also the signal bandwidth and (21)
spectral density. Thus, source localization performance is better
for signals with more energy in the high frequencies. where . By further expanding
For the second case, when the speed of propagation is also un- the parameter space, the CRB for multiple source localizations
known, i.e., , the matrix for this case is given by can also be derived, but its analytical expression is much more
. The Fisher information block matrix complicated. Note that when both the source signal and sensor
is given by gains are unknown, it is not possible to determine the values of
the source signal and the sensor gains (they can only be esti-
(16) mated up to a scaled constant).
By transforming to the polar coordinate system in the 2-D
where , diag , and case, the CRB for the source range and DOA can also be
. By applying the well-known block matrix given. Denote as the source range from a
inversion lemma (see [30, App.]), the leading submatrix reference position such as the array centroid. The DOA can be
of the inverse Fisher information block matrix can be given by given by with respect to the -axis. The
time delay from the source to the th sensor is then given by
(17) , where is
1848 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 8, AUGUST 2002

the polar position of the th sensor. We can form a polar array

matrix , where

(22)

Similarly, the polar penalty matrix due to unknown source signal

can be given by

(23)

The leading 2 2 submatrix of the inverse polar Fisher infor-

mation block matrix can be given by

(24)

Then, the lower bound of range estimation variance can be given

by and the lower bound of DOA estima- Fig. 1. Geometry of three source locations and one unknown sensor location.

tion variance can be given by The polar

array matrix shows that a good DOA estimate can be obtained at may be of distinct ones or one traveling source recorded at dif-
the broadside of a linear array, and a poor DOA estimate results ferent times. Then, the source locations and the distances
at the endfire of a linear array. A two-dimensional (2-D) array is are used in the LSC method to estimate the unknown sensor
required for better range and DOA estimations, e.g., a circular location.
array. The LSC unknown sensor location estimator is derived by
expanding and manipulating the circle relations shown in Fig. 1.
IV. UNKNOWN SENSOR LOCATION ESTIMATION Consider the 2-D case, where , and
. The equation of a circle from the th source with
In this section, we propose a method for estimating the un- radius is given by
known sensor location using an extended AML method fol-
lowed by the LSC method. Although the following formulation (25)
is only for a single sensor of unknown location, it can be ex-
tended to estimate other unknown sensor locations in sequential By expanding the circle equation in (25) and some manipula-
steps one at a time. We assume the locations of sensors are tions, we can write
known a priori, but the calibrating source is an unknown signal
at an unknown location (i.e., online calibration). The AML es- (26)
timator is extended to include the parameters ,
where is the distance from the source to the and in matrix form, it becomes
unknown sensor location . The source location is estimated
using the sensors of known locations, and the distance is
estimated using the sensor of unknown location. Only the dis- (27)
tance is added to the parameter space since an ambiguity ex-
ists in a circle around the source. The extended AML metric where
can be given by , where
, ,
is the data spectrum of the sensor of unknown location, and .. .. ..
. . . (28)
is the normalized sensor gain level at this sensor.
For the 2-D case, the unknown sensor location can be
obtained by the intersection of the circles drawn from three or
more source locations, as depicted in Fig. 1. The estimation and . Note that (27) is also applicable for the
then becomes a two-step process. The extended AML method 3-D case by defining and as three-dimensional (3-D)
is used to sequentially estimate sets of parameters, i.e., vectors. An overdetermined solution to the unknown sensor lo-
, where is the number of cation and can be given by , where
sources. Note that the estimation of these parameters can be is the pseudoinverse of . However, when the
done either separately or jointly for all sources, but we only source locations form a straight line, degenerates, and the
consider the separate case for simplicity. The source locations above formulation does not apply.
CHEN et al.: MAXIMUM-LIKELIHOOD SOURCE LOCALIZATION 1849

When the source locations form a straight line, e.g., a source

traveling in a linear path, the circles drawn from the source lo-
cations will intersect at two locations and result in an ambiguity.
However, it can be resolved if we know a priori that one of
the intersections is a more probable candidate for the unknown
sensor location. In this case, the proposed estimator needs to be
slightly modified. By taking the difference of (26) for the th
source from that of the first source, we obtain

(29)

where . Let without loss of gen-

erality, e.g., source locations form a straight line on the -axis,
we obtain the estimate of the -position of the unknown sensor
by , where , and
. Fig. 2. Three-dimensional plot of J (r ) for a source inside the convex hull
The estimate is then given by the circle equations in (25) of the array. Square: sensor locations.
with , which becomes , for
. The estimate is given by

(30)

If is positive, two solutions for with equal magnitude

but opposite signs are given, thus providing two ambiguous un-
known sensor location estimates. By having prior knowledge of
which side the unknown sensor is located, the ambiguity can be
resolved.

V. SIMULATION EXAMPLES AND EVALUATIONS

For all simulation examples and evaluations, we only con-
sider acoustic sources. In principle, the AML method can be
applied to signals of other nature, e.g., seismic, but they will
not be considered for the purpose of this paper. In all cases, the
sampling frequency is set to be 1 KHz. The speed of propagation Fig. 3. Image plot of J (r ) for a distant source. Square: sensor locations.
is 345 m/s. A prerecorded tracked vehicle signal, with signifi- 2
Circle: actual source location. : source location estimate.
cant spectral content of about 50 Hz bandwidth centered about
a dominant frequency at 100 Hz, is considered. For an arbitrary ence), as depicted in Fig. 4. In this case, we consider the spatial
array of five sensors, we simulate the array data (with uniform loss that is a function of the distance from the source location to
gain ) using this tracked vehicle signal with appropriate time each sensor location; thus, the gains are no longer uniform.
delays. The data length (which corresponds to 0.2 s), For each frame of samples, we simulate the array data
the DFT size (zero-padding), and all positive fre- for one of the 12 source locations. We apply the AML method,
quency bins are considered in the AML metric. In Fig. 2, the assuming unity , and estimate the source location at each
normalized in (7) is evaluated at different near-field po- frame. A 2-D polar grid point system is being used, where the
sitions and plotted for a source inside the convex hull of the array angle is uniformly sampled, and the range is uniformly sampled
(overlaid on top) under 20 dB SNR. A high peak shows up at in log-scale (to be physically meaningful). This 2-D grid-point
the source location, thus indicating good estimation. When the search provides the initial estimate of the direct search method
source moves away from the array, the peak broadens and re- [28]; thus, global maximum is guaranteed. In addition, we apply
sults in more range estimation error since it is more sensitive to the LS method [19] and the wideband MUSIC method [12] and
noise. As depicted in Fig. 3 (for the same setting), the image plot compute the CRB using (21) for comparison purposes. Note the
of evaluated at different positions show that the range es- complexities of these methods are not compared in this paper.
timation error is likely to occur in the source direction when the The wideband MUSIC divides the 200 samples into four subin-
source moves away from the array. However, the angle estima- tervals of 50 samples each for the computation of the sample
tion is not greatly impacted. correlation matrix and uses the same grid-point search and direct
In the following, we consider a single traveling source sce- search method as the AML case. As depicted in Fig. 5 [(a) for
nario for a circular array (uniformly spaced on the circumfer- and (b) for ], both the AML and LS methods ap-
1850 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 8, AUGUST 2002

Fig. 4. Single traveling source scenario.

Fig. 6. (a) Source localization and (b) DOA estimations for unknown speed of
propagation performance comparison (R = 7).

Fig. 5. Source localization performance comparison for known speed of

=
propagation case. (a) R 5. (b) R = 7. Fig. 7. CRB comparison for the traveling source scenario (R = 7). (a)
Localization bound. (b) DOA bound.
proach the CRB when the source is near the array (high SNR and
short range), but the AML outperforms the LS method. How- we compare the CRB for the known source signal and speed
ever, the wideband MUSIC yields much worse estimates than of propagation, unknown speed of propagation, and unknown
those of the LS and AML methods, especially when the source is source signal cases using seven sensors for the same single trav-
far from the array. The wideband MUSIC requires large number eling source scenario. As depicted in Fig. 7, the unknown source
of samples for good performance, but in this case, is limited signal is shown to have a much more significant degrading ef-
by the moving source. Note that when the source is far from the fect than the unknown speed of propagation in source location
array (low SNR and long range), the CRB is too idealistic since estimation. However, these parameters are not significant in the
it does not take into account other factors such as the edge effect DOA estimations. Furthermore, the theoretical performance of
or correlated noise across frequency. Thus, the reader should not the estimator is analyzed for different signal characteristics. For
get the impression that the AML performs poorly in this region. the same setting as the one shown in Fig. 3, we replace the
When we assume that the speed of propagation is unknown, tracked vehicle signal by a Gaussian signal after bandpass filter
we apply the extended AML method with unknown parameters (with fixed SNR of 20 dB). The center frequency of the source
. In this case, the CLS method in [18] is compared signal is first set to 250 Hz, and then, we let the bandwidth of this
with the extended AML method. In Fig. 6, we show the AML signal vary from 50 to 450 Hz. As shown in Fig. 8(a), the theo-
method yields better overall performance in both source location retical performance improves as the bandwidth increases. Then,
and DOA estimations for . To compare the theoretical for a fixed bandwidth of 100 Hz, we vary the center frequency
performance of source localization under different conditions, from 50 to 450 Hz and observe the performance improvement
CHEN et al.: MAXIMUM-LIKELIHOOD SOURCE LOCALIZATION 1851

Fig. 8. CRB comparison for different signal characteristics (for scenario in Fig. 10. RMS error of the two-source location estimates in each alternating
Fig. 3). (a) Bandwidth. (b) Center frequency. projection iteration. (a) Unit sensor gains case. (b) Variable sensor gains case.

Fig. 9. Joint two-source location estimation using the AML alternating Fig. 11. Three single-source location estimations and unknown sensor
projection of five iterations (unit sensor gain case). location estimation.

for higher center frequency in Fig. 8(b). The theoretical perfor- the sensor gains are modeled by the spatial loss and we as-
mance analysis applies to the AML estimator since the AML sume they are uniform in the AML metric, the source location
estimator approaches the CRB asymptotically. estimates are slightly biased, i.e., converging to an offset loca-
For two coherent sources of equal power near the array [with tion, as depicted in Fig. 10(b).
uniform gains and ], we apply the alternating projection To demonstrate the effectiveness of the unknown sensor lo-
method (10)–(13) of only five iterations for the AML method cation estimation method, we consider a similar scenario to the
and show converging solution in Fig. 9. The same grid-point first example with one additional sensor of unknown location.
search is applied to obtain the initial estimate of one source As depicted in Fig. 11, the expanded parameters
and then the initial estimate of another source in the second for are estimated separately using the extended AML
step. The direct search method is applied in the following steps method. From the estimated source locations and distances, the
to reach the converging solution. In this case, the wideband LSC method, i.e., (27), is applied, and a good estimate of the
MUSIC yields a false peak (not shown) since the two sources unknown sensor location is shown. Systematic evaluations via
are coherent (which is a well-known problem with MUSIC-type computer simulations are performed for the same scenario by
algorithms). The RMS error of the two-source AML estimates placing sources uniformly within a 4 4 m2 area around the
for the first ten alternating projection iterations are plotted in array. The performance of the LSC using the extended AML
Fig. 10(a) after an average of 100 realizations. The alternating method (which we refer to as AML-LSC) is compared with our
projection converges in only a few iterations. However, when previously proposed method using LS source location estimates
1852 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 8, AUGUST 2002

terion. The weighted cross-correlation function of the th and

the th sensors is defined by

(31)

where the weight , is the

normalized signal gain level at the th sensor, and
is the weighted cross-power spectral density.
The maximum weighted cross-correlation criterion maximizes
the sum of the weighted cross-correlation functions with the rel-
Fig. 12. Unknown sensor location estimation RMS error versus (a) number of
sources and (b) SNR. ative time delays for parameter , which can be given by

(which we refer to as LS-LSC) [26]. In Fig. 12(a), the unknown

sensor location estimation RMS error is plotted with respect a
different number of sources or source locations under 20 dB
SNR. The proposed AML-LSC method yields improved esti-
mates that are closer to the theoretical lower bound (derived in
Appendix C). In Fig. 12(b), the RMS error is plotted with respect
to the SNR for ten sources or source locations. The AML-LSC
result is again better than the LS-LSC result.

VI. CONCLUSIONS
In this paper, the AML source localization solution is de- (32)
rived for wideband sources in the near field. In the single-source
case, the AML solution maximizes the weighted cross-correla-
Since the zero frequency bin only produces a constant term
tion functions of the array signal and yields the beam-steered
that is independent of and the negative frequency bins are
beamformer output. We have shown its superior performance
complex conjugated versions of the positive frequency bins,
over suboptimal techniques and its efficiency with respect to
is equivalent to the AML metric.
the CRB. For the two-source case, we have shown good results
using the alternating projection procedure of only a few itera-
APPENDIX B
tions. We have shown that the different signal gain levels due to
EXTENDED SINGLE-SOURCE AML SOLUTION FOR UNKNOWN
spatial loss are not sensitive parameters to source localization
SIGNAL GAIN LEVEL
in the single-source case but are more significant in the case of
multiple sources. When combining the extended AML method When the are unknown, we can expand the pa-
with the LSC unknown sensor location estimator, significant rameter space to , where
improvement is observed over other previous time-delay-based and In this
methods. The derived CRB also proves the physical observa- case, the frequency-domain data (positive frequency bins only)
tions that favor high energy in the higher frequency components is grouped into a data spectrum matrix with the
of a signal. The sensitivity of source localization to different un- following elements:
known parameters has also been analyzed. It has been shown
that an unknown source signal results in a much larger error (33)
than that of unknown speed of propagation. The CRB supports
the physical understanding of the performance in terms of the where is the element of the phase matrix
signal nature and source/array geometry. , is the element of the noise matrix , ,
and . By multiplying the conjugate of the phase
APPENDIX A term on both sides of (33), we obtain the elements of the trans-
WEIGHTED CROSS-CORRELATION CRITERION formed data spectrum matrix as
As discussed in Section II-B, the AML criterion can be shown
to be equivalent to the maximum weighted cross-correlation cri- (34)
CHEN et al.: MAXIMUM-LIKELIHOOD SOURCE LOCALIZATION 1853

Note that the transformed noise matrix is still white is the time-delay from the th source to this sensor, and
Gaussian distributed with the same mean and vari- is the complex white Gaussian noise vector with
ance. The transformed data spectrum matrix is given by variance in each element. By including the pos-
By the singular value decomposition, itive frequency bins and sources, we can write the
can be decomposed to space–temporal frequency vector as
, where ,
, and
(35) . The Fisher information matrix
can be given by (14) with , where
where is the singular value, and and are the . The resulting leading submatrix
th column of the left and right singular matrices, respec- of the inverse Fisher information matrix is then given by
tively. The Frobenius norm square of can be given by
, where
(39)
is the sum of the square of the noise
subspace singular values. At the true source location, the square
of the principle singular value corresponds to the total array where is the
signal power plus the noise power in the signal subspace. In
scale factor for the th source, is
this transformed formulation, the AML solution is given by the
the source geometry matrix, ,
parameter set that minimizes the metric
is the penalty matrix due to un-
(36) known source locations, and .
The CRB for the 2-D unknown sensor location estimation
with the constraint that is a rank-one matrix (i.e., outer variance can be given by
product). At any source location , the optimal estimate of the
rank-one outer product of the source spectrum vector and sensor (40)
gain vector is given by
and the distance error of the unknown sensor location to its true
(37) location can be given by

where and are the principle left and right singular (41)
vectors of , respectively. By substituting (37) into (36),
the resulting AML metric for the source location is given by where .

ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers for
their valuable comments. J. C. Chen would like to thank Dr. A.
(38) G. Jaffer of Raytheon Systems for his valuable suggestions.

Since is a constant term, we can define an equivalent REFERENCES

metric , and the source location is [1] F. C. Schweppe, “Sensor array data processing for multiple signal
estimated by finding the maximum of . The normalized sources,” IEEE Trans. Inform. Theory, vol. IT-14, pp. 294–305, 1968.
[2] I. Ziskind and M. Wax, “Maximum likelihood localization of multiple
sensor gain vector can be estimated by . sources by alternating projection,” IEEE Trans. Acoust., Speech, Signal
Note that the actual sensor gain vector cannot be estimated Processing, vol. 36, pp. 1553–1560, Oct. 1988.
since is also unknown. [3] J. Capon, “High-resolution frequency-wavenumber spectrum analysis,”
Proc. IEEE, vol. 57, pp. 2408–2418, Aug. 1969.
[4] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
APPENDIX C tion,” IEEE Trans. Antennas Propagat., vol. AP-34, pp. 276–280, Mar.
CRB FOR UNKNOWN SENSOR LOCATION ESTIMATION 1986.
[5] S. S. Reddi, “Multiple source location—A digital approach,” IEEE
Without loss of generality, let the sensor of unknown Trans. Aerosp. Electron. Syst., vol. AES-15, Jan. 1979.
[6] J. Krim and M. Viberg, “Two decades of array signal processing re-
location be the th sensor in addition to the search: The parametric approach,” IEEE Signal Processing Mag., vol.
sensors of known locations. The array signal vector 13, pp. 67–94, July 1996.
for the th source at the th bin can be given by [7] H. Wang and M. Kaveh, “Coherent signal-subspace processing for
the detection and estimation of angles of arrival of multiple wide-
, band sources,” IEEE Trans. Acoust., Speech, Signal Processing, vol.
where , ASSP-33, pp. 823–831, Aug. 1985.
[8] H. Hung and M. Kaveh, “Focusing matrices for coherent signal-sub-
, is the signal gain level space processing,” IEEE Trans. Acoust., Speech, Signal Processing, vol.
at the sensor of unknown location, 36, pp. 1272–1281, Aug. 1988.
1854 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 8, AUGUST 2002

[9] S. Valaee and P. Kabal, “Wideband array processing using a two-sided Joe C. Chen (S’00) was born in Taipei, Taiwan,
correlation transformation,” IEEE Trans. Signal Processing, vol. 43, pp. R.O.C., in 1975. He received the B.S. (with honors),
160–172, Jan. 1995. M.S., and Ph.D. degrees in electrical engineering
[10] J. Krolik and M. Eizenman, “Minimum variance spectral estimation for from the University of California, Los Angeles
broadband source location using steered covariance matrices,” in Proc. (UCLA), in 1997 and 1998, and 2002, respectively.
IEEE ICASSP, vol. 5, Apr. 1988, pp. 2841–2844. Since 1997, he has been with the Sensors and Elec-
[11] B. Friedlander and A. J. Weiss, “Direction finding for wide-band signals tronics Systems Group of Raytheon Systems Com-
using an interpolated array,” IEEE Trans. Signal Processing, vol. 41, pp. pany (formerly Hughes Aircraft), El Segundo, CA.
1618–1634, Apr. 1993. From 1998 to 2002, he was a Teaching Assistant at
[12] T. L. Tung, K. Yao, D. Chen, R. E. Hudson, and C. W. Reed, “Source UCLA. He is currently also a part-time post-doctoral
localization and spatial filtering using wideband MUSIC and maximum researcher at UCLA. His research interests include
power beamforming for multimedia applications,” in Proc. IEEE SiPS, estimation theory and statistical signal processing as applied to sensor array
Oct. 1999, pp. 625–634. systems and radar systems.
[13] H. C. Schau and A. Z. Robinson, “Passive source localization Dr. Chen is a member of the Tau Beta Pi and Eta Kappa Nu honor societies.
employing intersecting spherical surfaces from time-of-arrival differ-
ences,” IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-35,
pp. 1223–1225, Aug. 1987.
[14] J. O. Smith and J. S. Abel, “Closed-form least-squares source location
estimation from range-difference measurements,” IEEE Trans. Acoust.,
Speech, Signal Processing, vol. ASSP-35, pp. 1661–1669, Dec. 1987. Ralph E. Hudson received the B.S. degree in elec-
[15] Y. T. Chan and K. C. Ho, “A simple and efficient estimator for hyperbolic trical engineering from the University of California,
location,” IEEE Trans. Signal Processing, vol. 42, pp. 1905–1915, Aug. Berkeley, in 1960 and the Ph.D. degree from the U.S.
1994. Naval Postgraduate School, Monterey, CA, in 1969.
[16] M. S. Brandstein, J. E. Adcock, and H. F. Silverman, “A closed-form In the U.S. Navy, he attained the rank of Lieutenant
location estimator for use with room environment microphone arrays,” Commander and served with the Office of Naval Re-
IEEE Trans. Speech Audio Processing, vol. 5, pp. 45–50, Jan. 1997. search and the Naval Air Systems Command. From
[17] K. Yao, R. E. Hudson, C. W. Reed, D. Chen, and F. Lorenzelli, “Blind 1973 to 1993, he was with Hughes Aircraft Com-
beamforming on a randomly distributed sensor array system,” IEEE J. pany and, since then, has been a Research Associate
Select. Areas Commun., vol. 16, pp. 1555–1567, Oct. 1998. with the Electrical Engineering Department, Univer-
[18] J. C. Chen, K. Yao, R. E. Hudson, T. L. Tung, C. W. Reed, and D. Chen, sity of California, Los Angeles. His research interests
“Source localization of a wideband source using a randomly distributed include signal and acoustic and seismic array processing, wireless radio, and
beamforming sensor array,” in Proc. ISIF, Aug. 2001, pp. TuC1: 11–18. radar systems.
[19] Y. Huang, J. Benesty, and G. W. Elko, “Passive acoustic source local- Dr. Hudson received the Legion of Merit and Air Medal and the Hyland Patent
ization for video camera steering,” in Proc. IEEE ICASSP, vol. 2, June Award in 1992.
2000, pp. 909–912.
[20] B. Emile, P. Common, and J. Le Roux, “Estimation of time-delays with
fewer sensors than sources,” IEEE Trans. Signal Processing, vol. 46, pp.
2012–2015, July 1998.
[21] P. J. Chung, M. L. Jost, and J. F. Böhme, “Estimation of seismic wave
parameters and signal detection using maximum likelihood methods,” Kung Yao (S’59–M’65–SM’91–F’94) received the
Comput. Geosci., vol. 27, pp. 147–156, Mar. 2001. B.S.E., M.A., and Ph.D. degrees in electrical engi-
[22] Y. Rockah and P. M. Schultheiss, “Array shape calibration using sources neering from Princeton University, Princeton, NJ.
in unknown locations—I: Far-field sources,” IEEE Trans. Acoust., He has worked at the Princeton-Penn Accelerator,
Speech, Signal Processing, vol. ASSP-35, pp. 286–299, Mar. 1987. Brookhaven National Laboratory, Brookhaven, NY,
[23] , “Array shape calibration using sources in unknown locations—II: and the Bell Telephone Labs, Murray Hill, NJ. He
Near-field sources and estimator implementation,” IEEE Trans. Acoust., was a NASA-NRC Post-Doctoral Research Fellow
Speech, Signal Processing, vol. ASSP-35, pp. 724–735, June 1987. at the University of California, Berkeley. He was a
[24] C. M. S. See and B. K. Poh, “Parametric sensor array calibration using Visiting Assistant Professor at the Massachusetts In-
measured steering vectors of uncertain locations,” IEEE Trans. Signal stitute of Technology, Cambridge, and a Visiting As-
Processing, vol. 47, pp. 1133–1137, Apr. 1999. sociate Professor at the Eindhoven Technical Univer-
[25] S. Gazor, S. Affes, and Y. Grenier, “Wideband multi-source beam- sity, Eindhoven, The Netherlands. From 1985 to 1988, he was an Assistant Dean
forming with adaptive array location calibration and direction finding,” of the School of Engineering and Applied Science at the University of Cali-
in Proc. IEEE ICASSP, vol. 3, May 1995, pp. 1904–1907. fornia, Los Angeles (UCLA). Presently, he is a Professor with the Electrical En-
[26] J. C. Chen, T. L. Tung, R. E. Hudson, and K. Yao, “Randomly dis- gineering Department, UCLA. His research interests include sensor array sys-
tributed sensor self-calibration using least squares methods,” in Proc. tems, digital communication theory and systems, wireless radio systems, chaos
Adv. Sensor Consort.: ARL Fed. Lab.: , Mar. 2000, pp. 105–109. communications and system theory, and digital and array signal processing. He
[27] D. H. Johnson and D. E. Dudgeon, Array Signal Pro- has published more than 250 papers. He is the co-editor of High Performance
cessing. Englewood Cliffs, NJ: Prentice-Hall, 1993. VLSI Signal Processing (New York: IEEE Press, 1997).
[28] J. A. Nelder and R. Mead, “A simplex method for function minimiza- Dr. Yao received the IEEE Signal Processing Society’s 1993 Senior Award
tion,” Comput. J., vol. 7, pp. 308–313, 1965. in VLSI Signal Processing. He was on the IEEE Information Theory Society’s
[29] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Board of Governors and is a member of the Signal Processing System Tech-
Theory. Englewood Cliffs, NJ: Prentice-Hall, 1993. nical Committee of the IEEE Signal Processing Society. He has been on the
[30] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estima- editorial boards of various IEEE Transactions, with the most recent being IEEE
tion. Englewood Cliffs, NJ: Prentice-Hall, 2000. COMMUNICATIONS LETTERS.