Anda di halaman 1dari 17

Array Processing for Target DOA, Localization,

and Classication based on AML and SVM


Algorithms in Sensor Networks
L. Yip, K. Comanor, J. C. Chen, R. E. Hudson, K. Yao, and L. Vandenberghe
Electrical Engineering Department, University of California at Los Angeles,
Los Angeles, CA 90095-1594
Abstract. We propose to use the Approximate Maximum-Likelihood
(AML) method to estimate the direction-of-arrival (DOA) of multiple
targets from various spatially distributed sub-arrays, with each sub-array
having multiple acoustical/seismic sensors. Localization of the targets
can with possibly some ambiguity be obtained from the cross bearings
of the sub-arrays. Spectra from the AML-DOA estimation of the target
can be used for classication as well as possibly to resolve the ambi-
guity in the localization process. We use the Support Vector Machine
(SVM) supervised learning method to perform the target classication
based on the estimated target spectra. The SVM method extends in a
robust manner to the nonseparable data case. In the learning phase, clas-
sier hyperplanes are generated o-line via a primal-dual interior point
method using the training data of each target spectra obtained from a
single acoustical/seismic sensor. In the application phase, the classica-
tion process can be performed in real-time involving only a simple inner
product of the classier hyperplane with the AML-DOA estimated tar-
get spectra vector. Analysis based on Cramer-Rao bound (CRB) and
simulated and measured data is used to illustrate the eectiveness of
AML and SVM algorithms for wideband acoustical/seismic target DOA,
localization, and classication.
1 Introduction
Recent developments in integrated circuits have allowed the construction of low-
cost small sensor nodes with signal processing and wireless communication capa-
bilities that can form distributed wireless sensor network systems. These systems
can be used in diverse military, industrial, scientic, oce, and home applica-
tions [1], [2]. One of the central tasks of these systems is to localize the target of
interest by collaborative processing of the received sensing data. In this paper,
we consider source localization by cross bearing from dierent subarrays. In this
0
Submitted to the 2nd International Workshop on Information Processing in Sensor
Networks (IPSN 03), April 22-23, 2003 Palo Alto Research Center (PARC), Palo
Alto, California, USA.
approach, processing of the data is performed locally within each subarray and
no data communication is needed among the subarrays.
Source localization and DOA estimation using sensor arrays have received
much attention in the array signal processing eld for decades. Many high resolu-
tion algorithms such as Multiple Signal Classication(MUSIC)[3] and Maximum
Likelihood(ML)[4], [5] have been proposed for narrow band DOA estimation.
Recently there has been an interest in locating wideband sources, for example,
the tracking of multiple acoustic sources using a microphone array. The algo-
rithm development in wideband source localization can be categorized into two
classes. The rst class of algorithms involves two steps: estimation of the time
dierence of arrivals(TDOAs) among the sensors followed by a least square(LS)
t to obtain the source location [6]. However, this class of methods is usually
for a single source only. The second class of algorithms uses a ML parameter
estimation method [7] to perform DOA estimation of the sources for far-eld
scenarios, which is capable of estimating multiple source DOAs. The Approx-
imated ML(AML) algorithm discussed in this paper processes the data in the
frequency domain as this may be more attractive for acoustic signals due to their
wideband nature. Moreover, the estimated source spectrum can be obtained as
a byproduct of the source angle estimation algorithm , and this spectrum can
then be used for source identication and source classication.
We propose the use of the support vector machine (SVM) to perform the
source classication from the estimated source spectra. The SVM is a supervised
learning algorithm which attempts to linearly separate labeled training data by
a hyperplane. Since its introduction by Vapnik in 1992 [14], it has received
considerable attention and is widely used in a variety of applications [12]. The
SVM method not only includes the simple linearly separable training data case,
but extends in a robust manner to the nonseparable data case. In the learning
phase, classier hyperplanes are generated o-line via a primal-dual interior point
method using the training data of each target spectra obtained from a single
acoustical/seismic sensor. In the application phase, the classication process can
be performed in real-time involving only a simple inner product of the supporting
hyperplane with the AML beamforming estimated target spectra vector obtained
from a sub-array.
Besides the development of dierent estimation and classication algorithms
using a sensor array, we also derive the theoretical Cramer-Rao bound(CRB)
for both performance comparison and basic understanding purposes. The CRB
provides a common tool for all unbiased estimators. It has been shown that the
CRB can be asymptotically approached by an ML estimator when SNR and
sampling data length are suciently large [8]. The rst explicit formula for the
CRB on the covariance matix appeared in [8]. However, it was only restricted to
the narrowband DOA estimation case. In [9], a CRB for a single source case was
given for both wideband source localization and DOA estimation. In this paper,
we extend the result in [9] to a multiple sources case. The resulting formula shows
that the CRB for a particular source DOA is increased due to interference from
the other sources. Most of the CRB derivations were focused on the received noise
at each sensor. However, the time synchronization among each sensor is crucial
for coherent type array signal processing and this error should be minimized to
obtain good performance. However, analysis of this kind of error is rare in the
array signal processing literature. In this paper, we derive a CRB equation for
theoretical analysis of the time-synchronization error.
In section 2, we derive the theoretical Cramer-Rao bound for multiple source
DOA estimation. The AML method for DOA estimation is described in section
3. In section 4, we review the formulation of the SVM as a quadratic pro-
gramming (QP) problem and describe a primal-dual interior-point method for
solving it. We present experimental results in section 5, and nally, we draw
some conclusions in section 6.
2 CRB Analysis for Multiple Source DOA Estimation
2.1 Array Data Model for DOA Estimation
When the source is in the far-eld of the array, wavefront arriving at the array
is assumed to be planer and only the angle of arrival can be estimated. For
simplicity, we assume both the source and sensor array lie in the same plane(a
2-D scenario). Let there be M wideband sources, each at an angle
m
from
the array. The sensor array consists of P randomly distributed sensors, each at
position r
p
= [x
p
, y
p
]
T
. The sensors are assumed to be omni-directional and have
identical response. The array centroid position is given by r
c
=
1
P

P
p=1
r
p
=
[x
c
, y
c
]
T
. We use the array centroid as the reference point and dene a signal
model based on the relative time-delays from this position. The relative time-
delay of the mth source is given by t
(m)
cp
= t
(m)
c
t
(m)
p
= [(x
c
x
p
) sin
m
+
(y
c
y
p
) cos
m
]/v, in where t
(m)
c
and t
(m)
p
are the absolute time-delays from the
mth source to the centroid and the pth sensor, respectively, and v is the speed
of propagation. In a polar coordinate system, the above relative time delay can
also be expressed as:t
(m)
cp
= r
p
cos(
m

p
)/v, where r
p
and
p
are the range
and angle of the p sensor with respect to the array centroid. The data received
by the pth sensor at time n is then x
p
(n) =
M
m=1
S
(m)
c
(n t
(m)
cp
) + w
p
(n), for
n = 0, ..., N 1, p = 1, ..., P, and m = 1, ..., M, where S
(m)
c
is the mth source
signal arriving at the array centroid position, t
(m)
cp
is allowed to be any real-valued
number, and w
p
is the zero mean white Gaussian noise with variance
2
.
For the ease of derivation and analysis, the received wideband signal can be
transformed into the frequency domain via the DFT, where a narrowband model
can be given for each frequency bin. However, the circular shift property of the
DFT has an edge eect problem for the actual linear time shift. These nite
eects become negligible for a sucient long data. Here, we assume the data
length N is large enough to ignore the artifact caused by the nite data length.
For N-point DFT transformation, the array data model in the frequency domain
is given by
X(k) = D(k)S
c
(k) + (k), (1)
for k = 0, ..., N1, where the array data spectrum is X(k) = [X
1
(k), ..., X
R
(k)]
T
,
the steering matrix D(k) = [d
(1)
(k), ..., d
(M)
(k)], the steering vector is given by
d
(m)
(k) = [d
(m)
1
(k), ..., d
(m)
R
(k)]
T
, d
(m)
p
= e
j2kt
(m)
cp
/N
, and the source spec-
trum is given by S
c
(k) = [S
(1)
c
(k), ..., S
(m)
c
(k)]
T
. The noise spectrum vector
(k) is zero mean complex white Gaussian distributed with variance N
2
. Note,
due to the transformation to the frequency domain, (k) asymptotically ap-
proaches a Gaussian distribution by the central limit theorem even if the ac-
tual time-domain noise has an arbitrary i.i.d. distribution (with bounded vari-
ance). This asymptotic property in the frequency-domain provides a more reli-
able noise model than the time-domain model in some practical cases. A more
compact expression of (2) can be formed by stacking up the N/2 positive fre-
quency bins of X(k) into a single column, that is X = G() + , where
X = [X(1)
T
, ..., X(N/2)
T
]
T
, G() = [S(1)
T
, ..., S(N/2)
T
]
T
, S(k) = D(k)S
c
(k),
and R

= E[
H
] = N
2
I
NR/2
. Throughout this paper, we denote superscript
T
as the transpose, and
H
as the complex conjugate transpose.
2.2 CRB Derivation for Multiple Source DOA Estimation
The CRB for any unbiased estimator of parameters with an arbitrary distribu-
tion is in general given by the inverse of the Fisher Information matrix[10], that
is CRB(
i
) = F
1
[]
ii
. In the white Gaussian noise case, the Fisher information
matrix is given by
F = 2Re[H
H
R
1

H] = (2/N
2
)Re[H
H
H], (2)
where H =
G
()
T
. We rst assume the source signals are known and the unknown
parameter is = [
1
, ...,
M
]. Applying (2), the element of Fisher Information
matrix can be shown to be
F
ii
=
2
N
2

k
a
2
ip
(
2kr
p
v
)
2
sin
2
(
i

p
)|S
(i)
c
(k)|
2
(3)
F
ij
=
2
N
2

k
a
ip
a
jp
(
2kr
p
v
)
2
exp(j
2kr
p
[cos(
i

p
) cos(
j

p
]
v
)
sin(
i

p
) sin(
j

p
)S
(i)
c
(k)S
(j)
c
(k) (4)
The Fisher Information matrix is a M M Hermitian with real diagonal ele-
ments. The CRB can be obtained by the inverse of F. In order to make compar-
ison of the single source formula, assume the matrix F has the format of
F =
_
F
11
F
1x
F
x1
F
xx
_
, (5)
where F
11
is the scaler which is the same as the single source case, and x = 1.
By applying the block matrix inverse lemma, the CRB of the source 1 angle is
given by[F
1
]
11
= (F
11
F
1x
F
1
xx
F
x1
)
1
. The penalty term F
1x
F
1
xx
F
x1
is due
to the interference from the other sources. It can be shown that this term is
always nonnegative, therefore, the CRB of DOA of source 1 in a single source
case is always less than the multiple source case.
In practice, the source signals are usually unknown. In this case, the unknown
parameter for estimation is [, S
T
c
]. The H matrix is given by H = [
G

T
,
G
S
T
c
].
The Fisher information matrix can be shown as
F =
_
F

B
B
T
C
_
, (6)
where B =
2
N
2
(
G

T
)
H G
S
T
c
, C =
2
N
2
(
G
S
T
c
)
H G
S
T
c
. By the block matrix inversion
lemma, the inverse of the upper left MM submatrix is given by [F
1
,Sc
]
MM
=
(F
1

Z
Sc
)
1
, where Z
Sc
= B
T
C
1
B. It can be shown that the penalty matrix
Z
Sc
due to unknown source signal is non-negative denite. Therefore, the DOA
estimation error of the unknown signal case is always larger than that of the
known case.
2.3 CRB Derivation for the Time Synchronization Error
In this subsection, we evaluate the theoretical performance of DOA estimation
for the far-eld case by CRB analysis. For clear illustration of the eect of the
time synchronization error, we only consider the case of a single source here. From
the data model of the far-eld DOA case, when only the time synchronization
error is considered, the received waveform of the pth sensor at the k frequency
bin is given by
X
p
(k) = S
c
(k) exp
_
j2k(t
p

p
)
N
_
, (7)
where S
c
(k) is the received signal spectrum at the reference sensor, t
p
is the
relative time delay from the pth sensor to the reference sensor. For the far-eld
case, t
p
=
rp cos(sp)
v
.
p
is the time synchronization error and assumed to be
IID white Gaussian with zero mean and variance
2
r
. Taking natural logarithm
of both sides of (7), and rearranging terms, we obtain
Z
p
(k) = f
k

r
p
cos(
s

p
)
v
+
p
, (8)
for k = 1, . . . , K and p = 1, . . . , P, where Z
p
(k) =
N
2k
{ln(X
p
(k))}, f
k
=
N
2k
{ln(S
c
(k))}, and {} represents the imaginary part of a complex value. At
the frequency bin k, the P equations are stacked up to form a complete matrix.
We have the following real-valued white Gaussian data model X = G() + ,
where is the unknown parameter that we need to estimate, = [
s
, f
k
] in
our case, i.e., the source angle and the source spectrum, = [
1
. . .
P
]
T
and
G() = f
k
[1 . . . 1]
T

_
r1 cos(s1)
v
. . .
rP cos(sP )
v
_
T
. From (2), the Fisher in-
formation matrix of this white Gaussian model is given by
F =
1

P
p=1
r
2
p
sin
2
(sp)
v
2

P
p=1
rp sin(sp)
v

P
p=1
rp sin(sp)
v
P
_
. (9)
The CRB for the source angle can then be given by the rst diagonal element
of the inverse of the Fisher information matrix, which is given by
CRB =

2

P
p=1
r
2
p
sin
2
(sp)
v
2

1
P
_

P
p=1
rp sin(sp)
v
_
2
. (10)
Some observation can be made from the CRB formula 10. First, the numerator
of the CRB only depends on the variance of the time-synchronization error; while
the denominator of the CRB depends on the array geometry and source angle.
Therefore, the CRB is proportional to the time synchronization error. Further-
more, the array geometry also has eect on the CRB. Poor array geometry may
lead to a smaller denominator, which results in a larger estimation variance. It
is interesting to note that the geometric factor is the same as the CRB formula
for additive Gaussian noise at [9], which means the array geometry produces the
same eects on both kinds of errors. Second, although the derivation is limited
to one frequency bin, the resulting CRB formula is independent of that partic-
ular frequency bin. Therefore, unlike the CRB of AWGN, the CRB can not be
reduced by increasing the number of frequency bins. In other words, the time
synchronization error can not be reduced by increasing the data length of the
received signal.
2.4 Variance Lower Bound for Time Synchronization Error and
AWGN
By considering time synchronization error and AWGN together, the received
signal spectrum at the k frequency bin and the pth sensor, will be
X
p
(k) = S
c
(k) exp
_
j2k(t
p

p
)
N
_
+
p
(k), (11)
where the rst term is the same as (7),
pk
is the complex white Gaussian with
zero mean and N
2
n
is the variance. Exact CRB requires the derivation of the
probability density function (pdf) of the above data model, which may be a
formidable task.
Here, we provide a variance bound based on the independence assumption of

p
and
pk
. With this condition, the variance of the estimator will be the sum
of the variance induced by these errors independently. By using var(; , ) =
var(; ) +var(; ), CRB(; ) var(; ) and CRB(; ) var(; ), we obtain
var(; , ) CRB(; ) + CRB(; ). (12)
The CRB induced by AWGN is given by
CRB =
1

P
p=1
r
2
p
sin
2
(sp)
v
2

1
P
_

P
p=1
rp sin(sp)
v
_
2
_, (13)
where =
2
N
2
V
2

N/2
k=1
_
2kSc(k)
N
_
2
and the geometric factor is the same as the
time synchronization error. Although the variance bound may not be as tight
as the CRB, it can be shown to match well with the root-mean-square (RMS)
error of the simulation of AML algorithm. Furthermore, it oers a much simpler
and more ecient way to evaluate the variance lower bound.
3 AML Method for DOA Estimation
3.1 Derivation of the AML Algorithm
In contrast to the TDOA-CLS method where the data is processed in the time
domain, the AML estimator does the data processing in the frequency domain.
The ML metric results in a coherent combination of each subband. Therefore, the
AML approach can gain advantage where the signal is wideband, for example, the
acoustic signal. By assuming both the source angles and spectrums are unknown,
the array signal model dened in section 2.1 is given by X = G(, S
c
) + ,
and R

= E[
H
] = N
2
I
NR/2
. The log-likelihood of this complex Gaussian
noise vector , after ignoring irrelevant constant terms, is given by L(, S) =
||XG(, S)||
2
. The maximum-likelihood estimation of the source DOA and
source signals is given by the following optimization criterion
max
,S
L(, S) = min
,S
N/2

k=1
||X(k) D(k)S(k)||
2
, (14)
which is equivalent to a nonlinear least square problem. Using the technique of
separating variable[5], the AML DOA estimate can be obtained by solving the
following likelihood function
max

J() = max

N/2

k=1
||P(k, )X(k)||
2
= max

N/2

k=1
tr(P(k, )R(k)), (15)
where P(k, ) = D(k)D

(k), D

= (D(k)
H
D(k))
1
D(k)
H
is the pseudo-inverse
of the steering matrix D(k) and R(k) = X(k)X(k)
H
is the one snapshot co-
variance matrix. Once the AML estimate of is found, the estimated source
spectrum can be given by

S
c
ML
(k) = D

(k,

ML
)X(k). The AML algorithm
in eect performs signal separation by utilizing the physical separation of the
sources, and for each source signal, the SINR is maximized in the ML sense.
3.2 Multiple Snapshot Implementation of AML
In the previous formulation, we derive the AML algorithm using only a single
block data. A variant of the AML solution using multiple snapshots can also
be formed. In this approach, a block of N data samples are divided into N
s
snapshots, each snapshot contains N
t
samples, i.e. N = N
s
N
t
. The sample
covariance matrix R can then be obtained by averaging the N
s
time snapshots
R =
1
Ns

Ns
t=1
X(k)X(k)
H
. Since the multiple snapshots approach uses less sam-
ple data to perform FFT, the edge eect becomes more severe that the single
snapshot approach. Appropriate zero padding is necessary to reduce this artifact.
3.3 Alternating Projection of Multiple Source DOA Estimation
In the multiple source case, the computational complexity of the AML algorithm
requires multi-dimensional search, which is much higher than the MUSIC type
algorithm that requires only 1-D search. The alternating projection technique
breaks the multi-dimensional search into a sequence of 1-D search, and reduced
the computational burden greatly. The following describes the alternating pro-
jection algorithm for the case of M sources.
Step 1: Solve the problem for a single source,

(0)
1
= arg max
1
J(
1
).
Step 2: Solve the second source

2
(0)
= arg max
2
J(

(0)
1
,
2
). by assuming the
rst source is at
1
. Continuing in this fashion until all the initial values

1
(0)
, ...,

(0)
M
are computed.
For k=1,..., repeat Step 3 until it converges.
Step 3: At every iteration a maximization is performed with respect to a single
parameter while all the other parameter are held xed. Therefore, at (k+1) itera-
tion,

(k+1)
i
= arg max
i
J(

(k)
(i)
,
i
), where

(k)
(i)
denotes the (M1)1 vector of
the computed parameters at (k) iteration, i.e.

(k)
(i)
= [

(k)
1
, ...,

(k)
i1
,

(k)
i+1
, ...,

(k)
M
].
3.4 Source Localization by Cross Bearing
When two or more subarrays simultaneously detect the same source, the crossing
of the bearing lines can be used to estimate the source location. Without loss of
generality, let the centroid of the rst subarray be the origin of the coordinate
system. Denote r
ck
= [x
ck
, y
ck
]
T
as the centroid position of the kth subarray,
for k = 1, ..., K. Denote
k
as the DOA estimate (with respect to north) of
the kth subarray. From simple geometric relationship, we have Ay = b, where
A =
_

_
cos(
1
) sin(
1
)
.
.
.
.
.
.
cos(
K
) sin(
K
)
_

_, y =
_
x
s
y
s
_
, and b =
_

_
x
c1
cos(
1
) y
c1
sin(
1
)
.
.
.
x
cK
cos(
K
) ycK sin(
K
)
_

_.
The source location estimate can be given by the least square (LS) solution of
the above equation using normal equation pseudo-inverse y = (A
T
A)
1
A
T
b,
or Moore-Penrose pseudo-inverse. The residual of the LS solution can also be
given by res = A yb. In the multiple sources case, two or more DOA can be
estimated from each subarray. For example, there are eight permutations in the
case where three subarrays uield six DOA estimates (assuming each subarray
yield two DOA estimates). The two source location estimates can be chosen by
the two lowest residuals of the LS solutions.
4 Support Vector Machine
4.1 Standard QP Formulation
Suppose we are given training data in the form of N vectors x
i
R
n
and
N binary labels y
i
{1, 1}. A support vector classier is based on an ane
decision function a
T
x b, where a R
n
and b R are the solution of the
following quadratic program
minimize (1/2)a
2
2
+ 1
T
u,
subject to Y (Xa 1b) 1 u,
u 0.
(16)
The variables are a R
n
, b R, and u R
N
. The matrix X R
Nn
has rows
x
T
i
, Y = diag(y
1
, . . . , y
N
), and 1 denotes a vector in R
N
with all its components
equal to one. The coecient > 0 is a parameter set by the user.
The constraints in (16) have the following interpretation. The training vector
x
i
is considered correctly classied by the decision function f(x) = a
T
x b, if
a
T
x
i
b 1, if y
i
= 1, or a
T
x
i
b 1, if y
i
= 1. The variable u is the slack
vector in these inequalities, i.e., measures the amount of constraint violation:
u
i
= 0 if the point is correctly classied and u
i
> 0, otherwise.
The cost function is a weighted sum of two objectives. The second term 1
T
u
is the total slack, i.e., total constraint violation. The rst term penalizes large
a, and has a very intuitive geometrical meaning. It can be shown that 2/a
2
is the distance between the hyperplanes a
T
x b = 1 and a
T
x b = 1. This
distance is a good measure of the robustness of the classier. By minimizing
a
2
we maximize the margin between the two hyperplanes. In the QP (16), we
control the trade-o between classication error (as measured by the total slack
violation) and robustness (inversely proportional to the a
2
) by the parameter
.
We can also use the QP formulation (16) to solve nonlinear classication
problems. We dene a nonlinear decision function f : R
p
R of the form
f(v) = a
T
F(v) b, where F : R
p
R
n
consists of a set of specied basis
functions (e.g., all monomials of a certain maximum degree). Given a set of
N training points v
i
R
p
, y
i
{1, 1}, we can then dene x
i
= F(v
i
), and
compute a and b by solving (16).
4.2 Solution via Primal-Dual Interior Point Method
Most SVM training methods solve the QP (16) via the dual problem
maximize (1/2)z
T
Qz +1
T
z,
subject to 0 z 1,
y
T
z = 0,
(17)
where the variable is z R
N
, and the matrix Q is dened as Q = Y XX
T
Y, i.e.,
Q
ij
= y
i
y
j
x
T
i
x
j
, i, j = 1, . . . , N. The dual problem has fewer variables (i.e., N)
than the primal problem (which has n+1+N variables). In nonlinear separation
problems we usually have n N, and the large dimension n aects the dual
problem only through the length of the inner products Q
ij
= y
i
y
j
F(v
i
)
T
F(v
j
).
While a naive implementation would require O(n) operations per inner product,
for many common basis functions F the inner products can be computed much
more eciently, in O(p) operations. This is referred to as the kernel trick in the
SVM literature. Most SVM implementations therefore solve the dual problem
using general-purpose interior-point solvers (such as LOQO [13] or MOSEK [11]).
However, similar savings can be achieved in any primal or primal-dual interior-
point method directly applied to (16). We will explain this for Mehrotras primal-
dual predictor-corrector method, one of the most popular and ecient interior-
point algorithms.
Skipping details, each iteration of the primal-dual method requires the solu-
tion of a set of linear equations
_

_
diag(z) diag(s) 0 0 0 0
0 0 diag(u) 0 0 diag()
I 0 0 Y X y I
0 I I 0 0 0
0 y
T
0 0 0 0
0 X
T
Y 0 I 0 0
_

_
_

_
s
z

a
b
u
_

_
=
_

_
r
1
r
2
r
3
r
4
r
5
r
6
_

_
. (18)
where s = Y (Xa 1b) +u1, = 1z, and a, b, u, z are the current primal
and dual iterates. These equations are obtained by linearizing the optimality
conditions (KKT conditions)
z
i
s
i
= 0, u
i

i
= 0 i = 1, . . . , N,
Y (Xa 1b) 1 + u s = 0, u 0, s 0,
1 z = , y
T
z = 0, z 0, 0,
a = X
T
Y z,
(19)
and provide search directions a, b, u, z for the primal and dual updates.
An ecient implementation of the primal-dual method requires solving the
equation (18) fast. A rst key observation is that if the starting values of a and
z satisfy a = X
T
Y z, then r
6
= 0 in the righthand side of (18). Therefore, the
steps a and z satisfy a = X
T
Y z, and the inequality a = X
T
Y z holds
throughout the algorithm. As a result, the large-dimensional variable a or its
update a is never needed. Instead, they are dened implicitly by the smaller
dimensional variables z and z.
A second important observation is that by straightforward elimination of
a, s, u, and , we obtain an equivalent system
_
Q + D y
y
T
0
_ _
z
b
_
=
_
r
7
r
5
_
, (20)
where D is a positive diagonal matrix and r
7
is computed from the righthand
sides of (18). The coecient matrix of (20) can be constructed by adding a
diagonal matrix to Q (evaluated eciently via the kernel trick). The solution
is obtained from the Cholesky factorization of Q + D. Given z and b, the
remaining search directions readily follow from (18). Additional savings are pos-
sible by using approximate search directions obtained by replacing Q with an
approximation (e.g., diagonal plus low rank), or solving (20) iteratively.
5 Simulation and Experiment Results
5.1 Simulation Examples of CRB Analysis
In this subsection, we compare the derived CRB of DOA estimation with the
RMS error of AML in several simulations. In the rst simulation, multiple CRB of
is calculated in a two-source scenario. The two far-eld sources are located at 45

and 60

from the sensor array. The sensor array conguration is a uniform square
with four acoustic sensors, each spacing 0.305 meter apart. The two sources are
a prerecorded motorcycle and a car signal respectively. The sampling frequency
is set to be 5KHz and the speed of propagation is 345m/sec. The RMS error
of AML is computed via 100 Monte Carlo runs. Fig.1 shows the resulted RMS
error of the rst source as a function of the SNR. It can be seen that both the
AML estimation error and CRB decrease as the SNR increases. The somehow
saturation behavior of the AML RMS error may be due to the quartization error
of the angle sampling.
In the second simulation, we consider the time synchronization eect on the
AML DOA estimation. The derived variance bound of (12) is compared with
the RMS error of AML. Only a single source arrive to the sensor array from
45

in this case. The RMS error of AML and the derived variance bound are
plotted as a function of SNR for various
2
t
. It can be seen that the performance
of AML matches well with the variance bound. Furthermore, the performance
of AML is limited even at high SNR region for x
2
t
at g.2. This is due to the
time synchronization error becoming dominant at that region, which results in
a error oor eect. The above theoretical as well as simulation analysis of the
time synchronization error shows that it is crucial to obtain accurate time syn-
chronization among sensors in order to yield a good performance of the coherent
array signal processing.
0 5 10 15 20 25 30 35
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
SNR in dB
R
M
S


e
r
r
o
r

i
n

r
a
d
i
a
n
AML RMS error
CRB
Fig. 1. RMS error comparison of CRB and AML as a function of SNR.
Fig. 2. RMS error comparison of variance bound and AML.
5.2 Experiment Results of Source Localization by Cross Bearing
Several acoustic experiments were conducted in PARC, Palo Alto, CA. In the
rst outdoor experiment (outside of the Xerox PARC building), three widely
separated linear subarrays, each with four microphones, were used. A white
Gaussian signal was played through the loud speaker placed at the two locations
shown in Fig. 5. In this case, each subarray estimated the DOA of the source
independently using the AML method, and the bearing crossing from the three
subarrays provided an estimate of the source location. An RMS error of 32cm
was reported for the rst location, and an RMS error of 97cm was reported for
the second location. This shows the AML crossing bearing method can locate
the wideband source eectively.
In a dierent outdoor conguration, two linear subarrays of four microphones
were placed at the opposite sides of the road and two omni-directional loud
speakers were placed between them, as depicted in Fig 6. The two loud speakers
play two independent pre-recorded sounds of light wheeled vehicles of dierent
kinds. By using the alternating projection steps on the AML metric, the DOAs of
the two sources were jointly estimated for each subarray under 11dB SNR (with
respect to the bottom array). Then, the crossing yielded the location estimates
of the two sources. An RMS error of 37cm was observed for the source 1 and an
RMS error of 45cm was observed for source 2. We note, the AML DOA subarray
angular resolution for multiple targets is signicantly better than the classical
Fourier limited resolution of the same subarray. This advantage is similar to the
so-called superresolution eect of various parametric spectral estimators (e.g.,
MUSIC; subspaced-based method) as compared to classical DFT Fourier limited
spectral estimator.
5.3 Simulation Results of SVM Classication
We attempt to discriminate between two sources (a Harley motorcycle and a car
from microphone data collected by BAE, Austin, TX, in Oct. 2002) based on
the magnitude spectrum of their acoustic signals. A total of 382 training vectors
were acquired. For each time signal, sampled at 5 kHz, we used a window of
1000 samples and moved the window by 100 samples until the end of the signal
was reached. For each time window, we took the magnitude of the FFT, saving
only the rst 200 frequency bins to make up each training vector.
Shown in gure (5) is the solution of (16) for various s. We see that for
suciently large , the margin becomes small enough such that 1
T
u is reduced
to 0, i.e., the training vectors are separable.
Figure (6) shows the result of testing our SVM classier on source spectra
estimated via the AML algorithm. The plot shows the fraction of misclassied
testing vectors as a function of the source SNR for set to 1,10, and 100. As
expected, we see that for the larger values of , which corresponds to larger
a and hence smaller margin width, the classier performs slightly worse at
low SNRs and slightly better at high SNRs. This phenomenon illustrates the
tradeo between robustness and misclassication.
5 0 5 10
0
5
10
15
A B
C
Xaxis (meter)
Y

a
x
i
s

(
m
e
t
e
r
)
Sensor locations
Actual source location
Source location estimates
5 0 5 10
0
5
10
15
A B
C
Xaxis (meter)
Y

a
x
i
s

(
m
e
t
e
r
)
Fig. 3. Source localization of white Gaussian signal using AML DOA cross bearing in
an outdoor environment.
5 0 5 10
0
2
4
6
8
10
12
14
16
Source 1
Source 2
A
C
Xaxis (meter)
Y

a
x
i
s

(
m
e
t
e
r
)
Sensor locations
Actual source locations
Source location estimates
Fig. 4. Two-source localization using AML DOA cross bearing with alternating pro-
jection in an outdoor environment.
0 50 100 150 200 250 300
0
0.1
0.2
0.3
=1
=10 =100
PSfrag replacements
a
2
2
1
T
u
/
N
Fig. 5. Tradeo curve between a2 and 1
T
u.
0 5 10 15 20 25 30 35
0
0.2
0.4
0.6
SNR(dB)
f
r
a
c
t
i
o
n

m
i
s
c
l
a
s
s
i
f
i
e
d
= 1
= 10
= 100
Fig. 6. Fraction of misclassied testing vectors from estimated source spectra as a
function of SNR.
6 Conclusions
In this paper, an AML algorithm is derived for multiple sources DOA and spec-
tra estimation. The source location can then be obtained via cross bearing from
several widely separated arrays. Furthermore, the proposed SVM algorithm can
be applied to the target classication based on the estimated source spectra.
Theoretical analysis based on the CRB formula is also derived. Simulation and
experimental results demonstrated the eectiveness of the AML and SVM algo-
rithms.
7 Acknowledgments
We wish to thank J. Reich, F. Zhao, and P. Cheung of PARC, Palo Alto, CA,
for supplying the data used in section 5.2 and S. Beck and J. Reynolds of BAE,
Austin, TX, for supplying the data used in section 5.3. This work is partially
supported by DARPA SensIT program under contract AFRL/IFG 315 330-1865
and AROD-MURI PSU contract S0126.
References
1. J. Agre and L. Clare, An integraed arcitecture for cooperative sensing and net-
works, Computer, vol. 33, pp. 106-108, May 2000.
2. G.J. Pottie and W.J. Kaiser, Wireless integrated network sensors, Comm. of the
ACM, vol. 43, pp. 51-58, May 2000.
3. R.O. Schmidt, Multiple emitter location and signal parameter estimation, IEEE
Trans. on Antenna and Propagation, vol. AP-34, no. 3, pp. 276-80, Mar. 1986.
4. F.C. Schweppe, Sensor array data processing for multiple signal sources, IEEE
Trans. Information Theory, vol. IT-14, pp. 294-305, 1968.
5. I. Ziskine ad M. Wax, Maximum likelihood localization of multiple sources by
alternating projection, IEEE Trans. Acoust., Speech, and Signal Processing, vol.
ASSP-36, no. 10, pp. 1553-60, Oct. 1988.
6. T.L. Tung, K. Yao, C.W. Reed, R.E. Hudson, D. Chen, and J.C. Chen, Source
localization and time delay estimation using constrained least squares and best
path smoothing, Proc.SPIE, vol. 3807, Jul 1999, pp. 220-23.
7. J.C. Chen, R.E. Hudson, and K. Yao, Maximum-likelihood source localization
and unknown sensor location estimation for wideband signals in the near-eld,
IEEE Trans. on Signal Processing, vol. 50, no. 8, pp. 1843-54, Aug. 2002.
8. P. Stocia and A. Nehorai, MUSIC, Maximum Likelihood, and Cramer-Rao
Bound, IEEE Trans. Acoust. Speech, and Signal Processing, vol. 37. no. 5, pp.
720-41, May 1989.
9. L. Yip, J.C. Chen, R.E. Hudson, and K. Yao, Cramer-Rao bound analysis of
wideband source localization and DOA estimation, in Proc. SPIE, vol. 4971, Jul.
7-11, 2002
10. S.M. Kay, Fundamentals of Statistical signal Processing:Estimation Theory,
Prentice-Hall, New Jersey, 1993.
11. MOSEK v2.0 Users manual, 2002. Available from http://www.mosek.com.
12. B. Scholkopf and A. Smola. Learning with Kernels. MIT Press, Cambridge, MA,
2002.
13. R.J. Vanderbei. LOQO: An interior point code for quadratic programming. Opti-
mization Methods and Software, 11:451484, 1999.
14. V. N. Vapnik. The Nature of Statistical Learning Theory. Springer, New York,
NY, 1995.

Anda mungkin juga menyukai