General Direction-of-Arrival Tracking with

Acoustic Nodes
Volkan Cevher, Student Member, IEEE, James H. McClellan, Fellow, IEEE,
Abstract
Traditionally in target tracking, much emphasis is put on the motion model that realistically represents
the targets movements. In this paper, we rst present the classical constant velocity model and then
introduce a new model that incorporates an acceleration component along the heading direction of the
target. We also show that the target motion parameters can be considered part of a more general feature
set for target tracking. This is exemplied by showing that target frequencies, which may be unrelated
to the target motion, can also be used to improve the tracking performance. In order to include the
frequency variable, a new array steering vector is presented for the direction-of-arrival (DOA) estimation
problems. The independent partition particle lter (IPPF) is used to compare the performances of the
two motion models by tracking multiple maneuvering targets using the acoustic sensor outputs directly.
The treatment is quite general since IPPF allows general type of noise models as opposed to Gaussianity
imposed by Kalman type of formulations. It is shown that by incorporating the acceleration into the state
vector, the tracking performance can be improved in certain cases as expected. Then, we demonstrate
a case in which the frequency variable improves the tracking and classication performance for targets
with close DOA tracks.
Index Terms
Motion dynamics, particle lter, Mont e-Carlo simulation methods, reference priors, importance sam-
pling, time-frequency analysis
I. INTRODUCTION
The direction-of-arrival (DOA) estimation problem has been extensively studied in the signal process-
ing literature [1]. Narrow-band solutions based on beamforming such as multiple signal classication
V. Cevher and J.H. McClellan are with the Center for Signal and Image Processing, School of ECE, Georgia Institute of
Technology, Atlanta GA 30332-0250.
Prepared through collaborative participation in the Advanced Sensors Consortium sponsored by the U. S. Army Research
Laboratory under the Collaborative Technology Alliance Program, Cooperative Agreement DAAD19-01-02-0008.
2
(MUSIC) [2], minimum variance beamforming, and Pisarenkos method suffer degraded performance
when the targets are moving relatively fast during the estimation batch (i.e., the snapshot period). The
performance loss in these high resolution algorithms can be attributed to the nonstationarity of the array
data caused by the rapid target motion. In order to remedy this problem, one can incorporate target
motion dynamics to jointly estimate the DOAs while tracking targets [3]. These rened DOA estimates
provide better performance in exchange for increased computational complexity.
In the case of wideband acoustic signals, the pioneering work by Wang and Kaveh [4] on coherent
subspace processing coherently integrates the array autocorrelation matrices corresponding to the multiple
frequencies of interest, so that signal-to-noise (SNR) and resolution gains can be achieved. The work
by Gershman and Amin [5] approximates the signals within the DOA snapshot period as chirps and
performs time-frequency MUSIC on the acoustic array outputs. The varying frequency approximation
enables these wideband methods to tackle more realistic estimation problems. However, these algorithms,
like the narrow-band techniques, produce snapshot DOA estimates; and hence, require heuristics for target
association.
Advances in large scale integration of computer systems have made Mont e-Carlo techniques a feasible
alternative to solve the target tracking problem. Conventionally, given a target dynamics model, the
underlying motion equations are simplied by linearization and Gaussian noise assumptions so that an
analytical solution can be obtained. The extended Kalman lter is such a method; it is also the best
minimum mean-squares linear estimator for the problem at hand. Mont e-Carlo techniques, on the other
hand, do not linearize or assume Gaussian noise; however, they approximate the posterior density of
interest (e.g., a density that describes the likelihood of a targets position in space) by particles that
represent a discrete version of the posterior. The idea is that if a sufcient number of effective particles
can be used, the estimation performance will be close to the theoretically optimal solution.
A solution to the multiple target tracking problem has been given for the narrow-band case by Orton
and Fitzgerald in [6] using an independent partition particle lter (IPPF). The independent partition
assumption of the IPPF solves the DOA association problem common to the multiple target tracking
algorithms. The implementation of the IPPF is shown in [6], given the target motion dynamics developed
in [3] to track constant velocity targets with Brownian disturbance acting on the target heading directions.
In this paper, the IPPF is again used, but along with a new target motion model that allows accelerations
along the heading direction of the targets. It is shown with simulations that since the new motion model
enables the IPPF to relax the constant velocity assumption, target tracking is improved when the targets
have high accelerations as intuitively expected.
3
In target tracking, it is usually a good idea to retain certain target features even if they are not related to
the motion parameters. This may, in turn, also help the data fusion problem with different type of sensors
since the extra algorithmic computational load is usually much less taxing on the energy budget of the
sensor than transmitting the raw data to a central processor. Hence, in addition to the motion parameters,
we also show how to incorporate new features such as time-varying frequency signatures of the targets
into the particle lter algorithm. It is assumed that a separate time-frequency tracker is tracking the
dominant instantaneous frequencies of the targets; however, the issues related to the frequency estimation
are not considered in this paper. It is demonstrated by simulations that in the case of multiple targets, the
introduction of the frequency variable improves the tracking performance of the IPPF when two targets
have similar motion parameters but different time-frequency signatures.
Observability is one of the main concerns in acoustic tracking problems because it must be possible to
determine the target states from the array data uniquely. Observability of a time-varying system depends
on both the observations and the state equation (i.e., the acoustic sensor outputs and the motion model)
[7]. However, when the observations are sufcient to determine the target DOAs, the observability test
reduces to a simple rank test on the Jacobian of the local representation of the motion model [3]. In this
paper, it is assumed that the observations are sufcient to determine the target DOAs and the system
observability is proved using the observability rank condition. We argue that with the same assumption
on the observations, it is possible to automatically generate the IPPF code to do target tracking for some
other target motion dynamics as long as the rank condition is satised. This can enable dynamic switching
of motion models appropriate for different types of targets, resulting in a more exible tracker.
The array model employed in the simulations has a special structure. We use a single node consisting
of a circular omnidirectional microphone array that will supply DOA and frequency information about
the targets. However, the derivations do not depend on the particular structure of the nodes. The reader
should be cognizant of the fact that tracking, in this paper, implies the temporal estimation of the target
DOAs as opposed to target positions. Spatial diversity of multiple nodes can be exploited in triangulating
the targets of interest, which also presents a data fusion problem. However, the data fusion from multiple
nodes can be facilitated by some of the variables in the target feature set such as the target orientation
and frequency. It should be noted that these two variables are invariant from the node positions when the
nodes have the same reference frame.
1
Hence, in a simple application such as triangularization, if the
DOA information coming from each node also includes these features, the data association for multiple
1
If the nodes have different reference directions, the orientation angles of targets can be found in another reference frame by
simple additions and subtractions of the reference angles of the nodes.
4
targets can be done optimally with minimal information exchange.
The organization of the paper is as follows. Section II introduces the models that account for the
target motion dynamics as well as the acoustic observations. Section III describes how to construct the
necessary probability density functions as the backbone of the particle lter solution. The IPPF algorithm
details are discussed in section IV before the simulation results, which are presented in section V.
II. DATA MODELS
Consider K far-eld targets coplanar with a sensor node consisting of P acoustic sensors. The sensor
node (or sensor array) is not assumed to possess any special structure. Two motion models will be rst
presented, differing in the assumptions of constant velocity versus constant acceleration.
A. State Model-I
The targets are assumed to have constant speeds with a Brownian disturbance acting on their heading
directions [3], [6]. With the introduction of the frequency variable, the model has the following state
vector:
x
k
(t)
_

k
(t)
Q
k
(t)

k
(t)
f
k
(t)
_

_
(1)
where
k
(t),
k
(t), and f
k
(t) are the DOA, the heading direction, and the instantaneous frequency
of the k
th
target. Q
k
(t) log q
k
(t) is a compound variable where q
k
(t) v
k
/r
k
(t). Target DOAs
are measured clockwise with respect to the y-axis whereas the target heading directions are measured
counter clockwise with respect to the x-axis. Figure 1 illustrates the geometry of the problem. The state
update equation can be derived by relating the DOAs of the target at times t and t + T using the
geometrical relation of position 1 at (r
k
(t) sin
k
(t), r
k
(t) cos
k
(t)) to position 2 at (r
k
(t) sin
k
(t) +
v
k
T cos
k
(t), r
k
(t) cos
k
(t) +v
k
T sin
k
(t)). Then, it is straightforward to obtain the following update
relations:
tan
k
(t + T) =
r
k
(t) sin
k
(t) + v
k
T cos
k
(t)
r
k
(t) cos
k
(t) + v
k
T sin
k
(t)
(2)
and
r
k
(t + T) =
_
r
2
k
(t) + 2r
k
(t)v
k
T sin (
k
(t) +
k
(t)) + v
2
k
T
2
(3)
5
1
2
v
k
T

k
(t)
r
k
(t)

k
(t)

k
(t + T)
r
k
(t + T)
Fig. 1. The k
th
target is at position 1 at time t and moves to position 2 in T seconds with a constant speed. The target is
assumed to be in the far-eld of the sensor array whose center coincides with the origin.
Equations (2) and (3) form a scalable system for the target motion dynamics at hand. To elaborate on
this, consider scaling the range and the speed of the k
th
target. It can be shown that this scaled target
has the same set of update equations as above since the scale factor can be cancelled out. This, in fact,
leads to the introduction of the compound variable q
k
(t) dened earlier. In the state update, however,
the logarithm of q
k
(t) is used since an additive noise component can be employed (as opposed to the
multiplicative noise when q
k
(t) is used). Note that the particle lter can cope with the general type noise
and that additivity of the noise is not required; however, the logarithm is used to decouple the noise
from the state update, which is given below. Also, note that the target maneuvers (i.e., the perpendicular
accelerations) are modelled in the state update through the state noise on
k
(t):
x
k
(t + T) =

f
I
(x
k
(t), u(t + T)) = f
I
(x
k
(t)) +u(t + T)
=
_

_
arctan
_
sin k(t)+e
Q
k
(t)
T cos k(t)
cos k(t)+e
Q
k
(t)
T sin k(t)
_
Q
k
(t) 1/2log[1 + 2e
Qk(t)
T sin (
k
(t) +
k
(t)) + (e
Qk(t)
T)
2
]

k
(t)
f
k
(t) + 2b
k
(t)T
_

_
+
_

_
u
,k
(t + T)
u
Q,k
(t + T)
u
,k
(t + T)
u
f,k
(t + T)
_

_
(4)
where u N(0, diag{
2

,
2
Q
,
2

,
2
f
}). When the target speed becomes negative due to acceleration, the
Q
k
term becomes complex due to the logarithm. This problem can be solved by updating only the real
part of Q
k
, but adding the imaginary part (which is when q
k
is negative) to the heading parameter
k
,
which, in effect, will reverse the direction of the target.
6
Implicit in (4) is a second order polynomial approximation done for the phase of the signals of interest.
Hence, for t [t
i
, t
i
+ T), the following relation is assumed on the k
th
signal:
s
k
(t)
k
(t
i
)e
j2[f0,k(ti)t+bk(ti)t
2
]
(5)
where
k
(t
i
) is the complex amplitude, f
k
(t) = f
0,k
(t
i
) + 2b
k
(t
i
)t is the instantaneous frequency; and
2b
k
(t
i
) is the rate of change of the instantaneous frequency of the k
th
signal at time t during the i
th
batch
period. The rest of the paper assumes that dominant narrow-band frequencies of each target are tracked
by a separate time-frequency lter. Some frequency tracking examples using Markov chains can be found
in [8][10]. Note that the state estimate of the IPPF also results in a distribution on the instantaneous
target frequencies, which can be exploited by the frequency tracker in determining b
k
(t
i+1
)s for the next
recursion.
The variances of the components of the state noise vector u are usually very small, which may lead
to sample impoverishment [11] (explained in Sec. V.A.3). In fact, if the process noise is zero, the state
variables can be treated as static variables in an estimation problem where using a particle lter may not
be appropriate. Techniques to prevent this sample impoverishment or degeneracy are discussed in [6].
Moreover, the state noise vector is chosen to be Gaussian due to its analytical tractability. Justications
of this model can also be found in [6].
B. State Model-II
In this second formulation, the targets are now assumed to have slowly varying accelerations along
their heading directions. The new state vector is the following:
x
k
(t)
_

k
(t)
Q
k
(t)

k
(t)

k
(t)
f
k
(t)
_

_
(6)
where
k
(t), Q
k
(t),
k
(t), and f
k
(t) are as dened above.
k
(t) is dened using the k
th
targets
acceleration a
,k
along its heading direction:
k
(t)
a,k
2rk(t)
. Figure 2 illustrates the geometry of the
problem used to derive the state update relations.
The state update equation can be derived similar to the State Model-I:
x
k
(t + T) =

f
II
(x
k
(t), u
k
(t + T)) = f
II
(x
k
(t)) +u
k
(t + T) (7)
7
1
2
v
k
(t)T

k
(t)
r
k
(t)

k
(t)

k
(t)
r
k
(t + T)

k
(t + T)
1
2
a
,k
T
2
d
k
(t)
Fig. 2. The k
th
target is at position 1 at time t and moves to position 2 in T seconds with constant acceleration. Notice that

k
(t) corresponds to
k
(t + T) in Fig. 1.
with

k
(t + T) =
k
(t) + tan
1
_
T
2

k
(t) cos(
k
(t) +
k
(t))
1 + T
2

k
(t) sin(
k
(t) +
k
(t))
_
+ u
,k
(t + T) (8)
where

k
(t) = tan
1
_
sin
k
(t) + Te
Qk(t)
cos
k
(t)
cos
k
(t) + Te
Qk(t)
sin
k
(t)
_
(9)
and

k
(t) =
a
,k
2d
k
(t)
(10)
Q
k
(t + T) = log{2T
k
(t)[1 + 2Te
Qk(t)
sin(
k
(t) +
k
(t)) + T
2
e
2Qk(t)
]
1/2
+ e
Qk(t)
}

1
2
log[1 + 2Te
Qk(t)
sin(
k
(t) +
k
(t)) + T
2
e
2Qk(t)
]

1
2
log[1 + 2T
2

k
(t) sin(
k
(t) +
k
(t)) + T
4

2
k
(t)] + u
Q,k
(t + T)
(11)

k
(t + T) =

k
(t)
[1 + 2T
2

k
(t) sin(
k
(t) +
k
(t)) + T
4

2
k
(t)]
1/2
+ u
,k
(t + T)

k
(t) =

k
(t)T
[1 + 2Te
Qk(t)
sin(
k
(t) +
k
(t)) + T
2
e
2Qk(t)
]
1/2
(12)

k
(t + T) =
k
(t) + u
,k
(t + T) (13)
f
k
(t + T) = f
k
(t) + 2b
k
(t)T + u
f,k
(t + T) (14)
The state noise is dened similarly as u N(0, diag{
2

,
2
Q
,
2

,
2

,
2
f
}). Intuitively, the noise variables
8
u

and u
Q
need to be correlated due to the dependence of the velocity on acceleration along the heading
direction. An expression for the correlation between these two noise variables can be derived to better
track the targets when the targets are actually not changing their directions. However, as the targets
undergo slow maneuvers between each time step, it is the authors observation that the correlation between
the acceleration along the heading direction at the beginning of the batch and the actual target speed
becomes much weaker. Hence, the noise variables on the velocity and the acceleration related variables
are modelled independently to better accommodate the maneuvering targets as well as for simplicity.
C. Observation Model
The sensor array in the simulations consists of P omnidirectional acoustic sensors situated uniformly
on a circle of radius R. A steering vector associated with the array denes the complex array response for
a source at DOA . The received signal at the p
th
sensor corresponding to the i
th
target is rst derived
using the same second order polynomial approximation on the phases of the signals shown in (5). For
an isotropic and non-dispersive medium, the signal received at the p
th
sensor can be written as
s
i
(t (
i
T
z
p
)) s
i
(t)e
j2[f0,i(t)(i
T
zp)2bi(t)t(i
T
zp)+bi(t)(i
T
zp)
2
]
s
i
(t)e
j2[bi(t)(i
T
zp)
2
fi(t)(i
T
zp)]
(15)
where i = 1, 2, , K, z
p
is the p
th
sensor position, and
i
(1/c)[
cos(
i
), sin(
i
)
]
T
is the i
th
slowness vector in cartesian coordinates. Equation (15) leads to the following array steering vector for
the i
th
source signal:
a(
i
) =
_

_
e
j2{bi(t)(i
T
z1)
2
fi(t)(i
T
z1)}
e
j2{bi(t)(i
T
z2)
2
fi(t)(i
T
z2)}
.
.
.
e
j2{bi(t)(i
T
zP )
2
fi(t)(i
T
zP )}
_

_
(16)
A similar derivation of the array steering vector for signals with constant narrow-band frequencies can
be found in [1]. Note that, in the second line of (15), the term b
i
(t)(
i
T
z
p
)
2
can be ignored for small
aperture sizes and slowly varying frequencies to obtain the same steering vector used in [5].
As mentioned above, the phase characteristics of the signals of interest are approximated with chirps
within a batch period T. We also require that each steering vector a(
i
) uniquely correspond to a signal
whose direction is the objective of the DOA estimation problem. Signals coming from multiple targets
are added to form the observations. For the IPPF, these acoustic observations are updated every = T/M
9
seconds where M denotes the number of batch samples. Then, the array outputs are written as follows:
y(t) = A((t))s(t) +n(t) t/ = 1, 2, , M (17)
In (17), y(t) is the noisy array output vector, n(t) is additive noise (e.g., n(t) N(0,
2
n
)), and A((t))
has the steering vectors in its columns. It should be noted that the sensor positions must be perfectly
known in order to dene A((t)) for this model [12].
For notational convenience and tractability, the data collected at each time increment during the batch
period T is stacked to form the data vector Y
t
: MP 1. The signal vector S
t
and the noise vector W
t
are formed in the same manner. Thus, the array data (or observation) model for the batch period can be
compactly written as:
Y
t
=

h((t), W
t
) = h((t)) +W
t
Y
t
= A
t
S
t
+W
t
(18)
where the steering matrix A
t
= diag{A((t)), A((t + )), , A((t + (M 1)))} implicity incor-
porates the DOA information of the targets.
D. Observability of the Motion Model
Even though the formulation of the problem as a state and observation model seems intuitive at rst,
it would be unwise to use this type of formulation if the new state model (7) is not observable given
the measurements (18). For the array model, it is assumed that the array can resolve the target DOA
parameters uniquely. This leaves only the examination of the motion model to determine the observability
of the system [3], [7]. It is veried in [3] that the state model described by (4) satises the strongly locally
observability rank condition. The strongly locally observability rank condition checks if the determinant
of the Jacobian of the state vector with respect to the target DOA is non-zero for observability [13].
It follows then that the new state vector x
k
(t) R
6
is observable if {
k
(t+m); m = 0, 1, ,
k
1}
and {b
k
(t); m = 0, 1, ,
k
1} are sufcient to determine x
k
(t) for a nite integer
k
. The Jacobian
of the state vector (7) with respect to the target DOAs is very tedious to calculate by hand. However, a
symbolic algebra manipulator can be programmed to prove that the Jacobian matrix is full rank for all
the regions of the system for the new state formulation, making the new state vector observable.
10
III. PDF CONSTRUCTIONS FOR THE IPPF
The particle lter is a convenient way of recursively updating a target posterior of interest. In this sec-
tion, we will show the derivations necessary for the operation of the particle lter, which are summarized
in Figure 3.
A. Data Likelihood
While formulating the state update equations in our problem, one encounters two nuisance parameters:
the signal vector S
t
and the noise variance for the additive Gaussian noise vector W
t
. For simplicity, we
will assume that the noise variance is approximately constant during the batch period [t, t+T). Following
the notation introduced in [6] , we will denote this noise variance as
2
w
(t) corresponding to the batch
period starting at time t. The noise has the complex Gaussian probability density function (pdf) described
by Goodman [14]. The data likelihood given the signal and noise vectors can be written as follows:
Relate the observed data
to the state variables:
p(Y
t
|A
t
, S
t
,
2
w
(t))
Construct the pdf for the
state transition:
p(X
t+T
|X
t
)
Use reference priors
to eliminate nuisance
parameters.
Noise model is crucial.
Need an analytical
relation.
Use Taylor series
expansion for the
importance function.
Ignore h.o.t. and
approximate with a
Gaussian.
l
y
(X
t
) log p(Y
t
|X
t
) l
x
(X
t
) log p(X
t
|X
tT
)
Obtain the importance function
(X
t
|X
0:tT
, Y
0:t
) N((X) +X, (X))
(X) =
_
l

x
(X) + l

y
(X)

1
,
(X) = (X)
_
l

x
(X) + l

y
(X)

.
Fig. 3. The mechanics of the necessary derivations needed by the particle lter. X is chosen to be the predicted state vector by
the state update relation without any noise.
L log p(Y
t
|A
t
, S
t
,
2
w
(t))
= MP log MP log
2
w
(t)
1

2
w
(t)
(Y
t
A
t
S
t
)
H
(Y
t
A
t
S
t
)
(19)
11
If the priors are known for the signals and noise variance given the state vector at time t, they can
be integrated out (a procedure also known as marginalization). If one desires to assume the least about
these parameters and let the observed data speak for itself, then the use of reference priors comes into
play
2
. Hence, even for moderate sample sizes, the information in the data dominates the prior information
because of the vague nature of the prior knowledge [16]. The intuitive choice of the prior is usually the
uniform prior on the natural space of the parameter. A good discussion of these issues can be found in
[16], [17], and [18].
The square root of the determinant of the Fisher information matrix is used as our reference prior
(a.k.a. Jeffreys prior). The resulting reference prior is not integrable (and hence, is improper) on the
entire unbounded space for the parameter vectors. This, in turn, stipulates compactness arguments on
the parameter space such as the ones used in [16], [19]. Assuming that the columns of A
t
are linearly
independent, the reference prior is given by [20]
p(S
t
|A
t
) |A
H
t
A
t
|
1/2
(20)
where |.| denotes the determinant of a matrix. At this point, we can use (20) to integrate out the signal
vector from our problem.
p(Y
t
|A
t
,
2
w
(t)) =
_
p(Y
t
|A
t
, S
t
,
2
w
(t))p(S
t
|A
t
)dS
t
p(Y
t
|A
t
,
2
w
(t)) exp
_

Y
H
t
(I A
t
(A
H
t
A
t
)
1
A
H
t
)Y
t

2
w
(t)
_
(21)
Finally, notice that (20) is not integrable if S
t
has innite multidimensional support. However, the
condition {S
t
: |[S
t
]
i
| <
i
} can be easily imposed on the i
th
signal component for some large
i
. This
makes the prior (20) integrable on the signal vector space and, in turn, the marginalization integrals
become proper. This condition is always satised in practice (e.g., when the signals of interest have nite
magnitudes at all times.)
B. State Likelihood
In the previous section, we omitted the motivation for constructing the pdf for the data and put the
emphasis on the use of reference priors. Now, it is necessary to elaborate on the reasons for constructing
2
Bernardo derives the reference prior using an estimation model based on communication channel with a source and data
[15]. The reference prior maximizes the mutual information between the source and the data.
12
the pdfs for the data and the state. The state and observation models (4), (7) and (18) form a hidden
Markov model (HMM), which can be compactly described by the following pdfs:
p(X
t
|X
tT
)
p(Y
t
|X
t
)
(22)
where X
t
= [x
T
1
(t), x
T
2
(t), , x
T
K
(t)]
T
and p(Y
t
|X
t
) = p(Y
t
|A
t
) or p(Y
t
|A
t
,
2
w
(t)) depending upon
whether or not we treat the noise variance as a known parameter. Here, we introduce a common notation
in the particle ltering literature, z
0:t
{z
0
, z
T
, . . . , z
t
}. The recursive update for the HMM model
described by (22) can be written as follows [11]:
p(X
0:t
|Y
0:t
) = p(X
0:tT
|Y
0:tT
)
p(Y
t
|X
t
)p(X
t
|X
tT
)
p(Y
t
|Y
0:tT
)
(23)
Hence, the recursive evaluation of p(X
0:t
|Y
0:t
) requires the pdfs shown in (22). The previous section
considered the construction of the second pdf in the model. This section will concentrate on the rst pdf
in (22).
The objective is to nd p(X
t
|X
tT
) given the state model. By inspection of (4) or (7), one can see that
X
t
is also normal with mean f
I,II
(X
tT
) and covariance equal to that of the additive noise. Therefore,
we can write the pdf for the state update as follows:
p(X
t
|X
tT
) N(f
I,II
(X
tT
), diag{
2

,
2
Q
,
2

,
2

,
2
f
}) (24)
We have two important remarks on the construction of the pdfs for our problem. The rst one is that
we generally need the analytical expressions for the pdfs to make use of the particle lter, which does
not assume a Gaussian model in general. The second remark is about the model order of the HMM. The
motion equations describe a rst order HMM model and hence the update equations (23) depend only
on the previous state. If more complicated motion equations are formulated in the state model so that
the HMM model order increases, then a new recursive update formulation becomes necessary.
C. The Importance Function
An appropriate choice of the importance function (.) may reduce the variance of the simulation
errors.
3
However, it was shown analytically in [22] that the importance weights have increasing variance
3
i.e., if we choose the exact posterior as the importance function then due to the nature of the data generating process the
variance of the estimator is inversely proportional to the number of particles N [21].
13
with time, which leads to increasing estimation errors (or simulation errors, will be used interchangeably).
Here, we restate an important result: the unconditional variance of the importance weights, i.e. with the
observations Y
0:t
being interpreted as random variables, increases over time. This fact is also known as
the degeneracy phenomenon: after a few iterations, all but one of the normalized importance weights will
be very close to zero [22].
It is crucial to note that the optimal importance function (X
t
|X
(i)
0:tT
, Y
0:t
) is proportional to p(Y
t
|X
t
)
p(X
t
|X
(i)
tT
) with the proportionality independent of X
t
. We have previously derived the analytical
relations for p(Y
t
|X
t
) and p(X
t
|X
tT
). Moreover, dene
l
y
(X
t
) log p(Y
t
|X
t
)
l
x
(X
t
) log p(X
t
|X
tT
)
(25)
(X) = [l

x
(X) + l

y
(X)]
1
(X) = (X)[l

x
(X) + l

y
(X)]
(26)
then, a suboptimal importance function which minimizes the variance of the importance weights is given
by the following [6], [22]:
(X
t
|X
0:tT
, Y
0:t
) N((X) +X, (X)) (27)
X is judiciously chosen to be the mode of p(X
t
|X
tT
, Y
t
) so that (X) 0 [22].
IV. ALGORITHM DETAILS
In this section, we will give the details of our modications to the independent partition particle
ltering algorithm by Orton and Fitzgerald. The outline of the IPPF is given in [6] and is repeated for
completeness.
A. Partitioning and Data Association
Each particle in the IPPF consists of multiple state vectors, e.g., for three targets, a particle consists
of three different state vectors corresponding to each targets motion parameters and frequency. The
target association problem is solved by the independence assumption on these partitions. To elaborate,
the independence assumption results in partition importance functions
k
(.)s. In order to generate the
partition importance functions, (26) is calculated for the whole particle; however, only the block diagonal
portions of (X) are used for each partition, which results in
k
(x) N((x)+x,
k
(X)) with
k
(X)
14
Pseudo Code for the IPPF
i. At time t, for each particle i (i = 1, . . . , N) and its partition k (k = 1, . . . , K), x
(i)
t
(k):
sample from the partition importance function: x
k
(x
t
(k)|x
(i)
tT
(k), Y
t
)
set x
(i)
t
(k) = x and calculate the partition weight q
(i)
k
(x)
normalize the partition weights across all particles for each partition:
q
(i)
k
=
q
(i)
k

i
q
(i)
k
for each partition, resample with replacement across the particles using the distribution generated
by q
(i)
k
and generate a new set of particles X
(i)
t
with their respective q
(i)
k
for each partition. Also,
reindex X
(i)
tT
accordingly
ii. For each particle i (i = 1, . . . , N), X
(i)
t
:
calculate the importance weights using
w
(i)
t
= w
(i)
tT
p(Y
t
|X
(i)
t
)p(X
(i)
t
|X
(i)
tT
)
(X
t
|X
(i)
tT
, Y
t
)

k
q
(i)
k
normalize the importance weights w
(i)
t
across the particles
iii. Resample the particles X
(i)
t
using a Metropolis-Hastings scheme keeping the reversibility of the chain

corresponding to the k
th
(dim{x} dim{x}) block diagonal matrix entry of (X). In particular, the
off-diagonal matrices in the particle Hessian corresponding to the cross partitions are ignored. Note that
after the particle is formed, the discrepancies generated by this method are augmented by the importance
weights, which are calculated using the the full Hessians generated from the new particle.
When the target DOAs cross, previous target states help distinguish the next state through the state
update probability. The important thing to remember is that unless one target is moving in tandem with
the other target, the partitions for two targets will be different from each other by the other elements of
the state vector (e.g, frequency, heading direction, and so on), which will be emphasized by the partition
probability q
k
(x). Then, the partition cross sampling is used to help the data association by generating
particles having high probability across all partitions. This method of generating particles helps the IPPF
to propagate particles with good predictive states and automatically handles the data association.
One modication is the use of the state transition probability (24) for the weighted resampling functions
q
k
(x). This choice alone seems to constrain the particles by the state update equation and hence is expected
to have poor performance for maneuvering targets. However, this choice of the weighted resampling
function makes sure that the created particles form a cloud around the expected mode of the target
state. The maneuvering target cases, on the other hand, are handled by the absolutely critical Mont e-
15
Carlo Markov chain (MCMC) resampling step. A classical Metropolis-Hastings scheme, which keeps the
reversibility of the Markov chain can be used to resample the state space using the data likelihood. In the
test cases run, the algorithm seems to handle the maneuvers better as the number of iterations increase
in the MCMC step. Details of an MCMC resampling scheme are given in [23].
When the targets maneuver, the expected mode of the next state predicted by the state update equation
changes. At the resampling state, the particles that are closer to this changed mean survive while the
particles around the predicted mean diminish. Hence, the resampling step, in effect, not only makes the
particles span most of the state space, but also compensates for the effects of the maneuver. It should be
noted that maneuvering has more impact on the heading direction than the other state variables. Hence,
a slight modication exploiting this fact in the resampling step may also improve the performance of the
algorithm for a given number of particles.
B. Effects of the Frequency Variable
The new state vectors include new motion variables, but the most important extension comes from the
frequency variable in the form of deriving new gradients and Hessians (26) for the linearization of the
optimal importance function. We will concentrate on l

y
and l

y
since it is necessary to approximate l

y
by
a positive denite matrix
4
and l

y
is used in setting up the equations. Derivations of l

x
and l

x
due to the
new motion parameters are straightforward. The notation in this section closely follows [3]. Dene
J(t) Y
H
t
(I A
t
(A
H
t
A
t
)
1
A
H
t
)Y
t
=
M1

m=0
|y(t + m) P
A
(t + m)y(t + m)|
2
=
M1

m=0
trace{P

A
(t + m)

R
y
(t + m)} =
M1

m=0
J
m
(t)
(28)
where P
A
and P

A
I P
A
are the projection matrices onto the column and the null spaces of A
and A
H
, respectively.

R
y
(t + m) = y(t + m)y
H
(t + m) is the one-sample estimate of the array
covariance matrix at batch time indexed by m. Note that l
y
= MJ/
2
w
, hence the gradients and the
4
l

y
represents the local covariance of the particles around their modes and is required to be positive denite; however, this
positive deniteness is not always available and modications are required to guarantee that l

y
remains positive denite at each
iteration of the lter. See [3] for more discussion.
16
Hessians of J(t) are linearly related to l

y
and l

y
. Dene the gradient of J(t) as G:
G
J(t)
X
t
= vec
_
1
M
M1

m=0
_
V
m
(t)diag(

J
m
) +
m
(t)diag(
f
J
m
)
_
_
(29)
where
V
m
(t) V (t + m) =
_
v
1
(t + m), v
2
(t + m), , v
K
(t + m)
_

m
(t) (t + m) =
_

1
(t + m),
2
(t + m), ,
K
(t + m)
_
(30)
with v
k
(t + m)
k(t+m)
xk(t)
and
k
(t + m)
fk(t+m)
xk(t)
. Moreover,

J
m
=
_
J
m
/
1
(t + m), J
m
/
2
(t + m), , J
m
/
K
(t + m)
_

f
J
m
=
_
J
m
/f
1
(t + m), J
m
/f
2
(t + m), , J
m
/f
K
(t + m)
_
(31)
Equation (29) follows from the chain rule, where target frequency and its DOA are assumed in-
dependent from each other. Dene
i,m
(t) and
i,m
(t) as the Hessian of
i
(t + m) and f
i
(t +
m) with respect to X
t
, respectively; and form
m
(t) = diag[
1,m
(t), . . . ,
K,m
(t)] and
m
(t) =
diag[
1,m
(t), . . . ,
K,m
(t)]. Then, the Hessian H

2
J(t)
XtXt
is given by
H = H

+ H
ff
+ H
f
(32)
where
H

=
1
M
M1

m=0
_
[
2

J
m
1] [vecV
m
(t)vec
H
V
m
(t)] + [diag(

J
m
I)]
m
(t)
_
(33)
H
ff
=
1
M
M1

m=0
_
[
2
ff
J
m
1] [vec
m
(t)vec
H

m
(t)] + [diag(
f
J
m
I)]
m
(t)
_
(34)
H
f
=
1
M
M1

m=0
_
[
2
f
J
m
1] [vecV
m
(t)vec
H

m
(t) + vec
m
(t)vec
H
V
m
(t)]
_
(35)
vec stands for the concatenation of the columns of a matrix; and denote the Kronecker and Schur
products, respectively. Moreover, 1 and I denote a matrix of all ones and the identity matrix of dim{x}
dim{x}.
In order to guarantee the positive deniteness of l

y
, the terms containing
m
(t) and
m
(t) on (33)
and (34) are ignored while calculating the Hessian in (32) as discussed in [3]. Dene A
i
(t + m) =
A(t + m)/
i
(t), C
m
(t + m) = A(t + m)/f
m
(t), and
i
(t + m) =
P

A
(t+m)Xt+m
i(t)
. The
17
following derivatives need to be approximated while calculating (33) and (34):

2
J
m

i
(t)
j
(t)
2Re{
H
i
(t + m)
H
j
(t + m)}

2
J
m
f
i
(t)f
j
(t)
2Re{tr[A
H
(t + m)C
H
i
(t + m)P

A
(t + m)C
H
j
(t + m)A
H
(t + m)]}
(36)
H
f
is not guaranteed to be negative denite and can be ignored; however, the authors found that with
the approximation below, it almost never affects the deniteness of the Hessian and can be used in the
calculation of the Hessian (32):

2
J
m

i
(t)f
i
(t)
2Re{tr[A
H
(t + m)A
H
i
(t + m)P

A
(t + m)C
H
j
(t + m)A
H
(t + m)]}, (37)
V. SIMULATION RESULTS
In the simulations, our objectives are the following: (i) compare the effectiveness of the two state
formulations for tracking targets; (ii) show the effect of the frequency variable on tracking. Comparison
of State Model-I with the extended Kalman lter also can be found in [6] and hence is not repeated here.
Issues related to initialization of the lters can be found in [23].
A. Single Target Tracking
A circular sensor array of 15 omnidirectional sensors is used to track a single target. The radius of
the array is such that the inter-element spacing is equal to 0.45 times the wavelength () of the target
of interest. Figure 4 shows the track and the temporal speed evolution of the target. The simulation
parameters are given in Table I.
TABLE I
SIMULATION PARAMETERS (A)
Number of Particles, N 100
noise

0.1

Q noise
Q
0.1
noise

0.0001
noise

Signal to Noise Ratio, SNR 7dB

Target Narrow-band Frequency, f
0
200Hz
Number of Batch Samples, M 8
In Table I, some explanation of M, the number of batch samples, is necessary. If the DOA tracking
was done by a snapshot algorithm, a much higher number of batch samples would be necessary to
18
0 200 400 600 800 1000 1200 1400 1600 1800
600
400
200
0
200
400
600
X
x in [m]
y

i
n

[
m
]

3
0
s
(a)
5 10 15 20 25 30 35 40 45 50
0
20
40
60
80
100
120
140
t in [s]
T
a
r
g
e
t

S
p
e
e
d

i
n

[
m
p
h
]
(b)
Fig. 4. (a) One target initially heading in the positive x-direction with speed approximately equal to 7 mph, starts to maneuver
at t = 30s. The sensor node is situated at the origin. (b) The target has almost constant acceleration between t = 0 and t = 35.
The speed of the target is almost constant during the maneuver.
get high resolution DOAs. However, inclusion of the motion dynamics reduces the number of batch
samples required to estimate the target DOAs. Ideally, a higher number of batch samples also helps
the IPPF estimate the DOAs better: the gradient and Hessian terms that form the mean and covariance
matrix of the approximate importance function incorporate more data and hence are expected to improve.
Interestingly, the authors determined on synthetic data that even M = 2 makes a good approximation to
these parameters when the target accelerations are small and the algorithm is initialized close to the true
values. Additional improvement in DOA estimation performance by increasing M is empirically found
to quickly reach the point of diminishing returns. When State Models I and II are run for the same target
track in Fig. 4 with M = 2, both models perform the same. However, as M is increased, State Model-II
starts to perform better.
Figure 5 illustrates the estimation performance of the IPPF. The DOA estimation of both lters is
almost identical as shown in Fig. 5 (a). However, if the true target tracks were estimated using the IPPF
state estimates in conjunction with the correct target initial range and speeds (which are not available to
the IPPF), the resulting tracks would be quite different as illustrated in Fig. 5 (b). Even if the DOAs
are the same, State Model-I explains the increase in target speed with a change in the heading direction,
but leads to an incorrect heading direction estimate, which is a crucial parameter in data fusion. Figure
6 shows the estimated and Q parameters of the target. As one can observe, different values of these
parameters can lead to the same DOA track even if they do not correspond to the true physical target
track.
19
5 10 15 20 25 30 35 40 45 50
10
20
30
40
50
60
70
80
90
100
t in [s]
D
O
A

T
r
a
c
k
s

i
n

[

]
State ModelII
True DOA
State ModelI
(a)
0 200 400 600 800 1000 1200 1400 1600 1800
600
400
200
0
200
400
600
X
x in [m]
y

i
n

[
m
]

5
s

1
0
s

1
5
s

2
0
s

2
5
s

3
0
s

3
5
s

4
0
s

4
5
s

5
0
s
(b)
Fig. 5. (a) The diamonds are the true DOA track. Dashed line is the estimate from State Model-I while the solid line is the
estimate from State Model-II. The two estimates are nearly identical in tracking the target DOAs. (b) Small errors in the DOAs
are accentuated by the range. State Model-II obtains a good estimate of the true target track because it uses acceleration.
5 10 15 20 25 30 35 40 45 50
45
40
35
30
25
20
15
10
5
0
5
t in [s]
T
a
r
g
e
t

H
e
a
d
i
n
g

i
n

[

]
State ModelII
True (t)
State ModelI
(a)
5 10 15 20 25 30 35 40 45 50
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
t in [s]
e
Q(t)
State ModelII
True v(t)/r(t)
State ModelI
(b)
Fig. 6. (a) State Model-I tries to explain target accelerations through the heading parameter; hence, it fails to capture the true
target parameters. The true target heading direction has additive Gaussian noise with

= 4

, which is marked with diamond.

Solid line is the State Model-II estimate. (b) Note that e
Q(t)
corresponds to v(t)/r(t) for the target. State Model-II has a better
Q(t) estimate since the target headings are close to the true headings.
B. Multiple Target Tracking with Varying Narrow-band Frequencies
It is challenging to track two narrow-band targets whose DOA tracks are closer than the Rayleigh
resolution. As the reader has seen so far, the IPPF produces high resolution DOA estimates; however,
it can fail when the target DOAs as well as the movement parameters are very close. The objective
in this section is to show that by incorporating a frequency variable, it is sometimes possible to still
track targets even in this difcult case. In this example, three target tracks are simulated with different
20
frequency scenarios (Fig. 7). Table II summarizes the simulation parameters. Figure 8 shows the tracking
50 0 50 100 150 200
150
100
50
0
50
100
X
x in [m]
y

i
n

[
m
]
1
2 3
Fig. 7. Three targets move with constant speed with some Brownian disturbance acting on their heading directions. State Model-I
is sufcient for the tracking in this case.
TABLE II
SIMULATION PARAMETERS (B)
Number of Particles, N 100
noise

0.1

Q noise
Q
0.01
noise

Frequency noise
f
0.001
Signal to Noise Ratio, SNR 7dB
Number of Batch Samples, M 8
performance when the targets have the same time-frequency signatures. Targets marked with the diamonds
(1) and the circles (2) have very similar DOA tracks after time t = 15s. Initially, the IPPF does a good
job in tracking; however, as targets get closer, it fails.
Figure 9 simulates the same problem, but, in this case, targets 1 and 2 have different time-frequency
signatures. Note that the IPPF had problems resolving the DOAs of these targets since they had close
motion parameters and had the same time-frequency signature. Due to the way the particles are generated
(by the partition importance functions
k
(.) as explained in section IV), the DOA tracking of target 3
is also affected as shown in Fig. 8. Hence, the change in the time-frequency signature of target 2 helps
21
the independence assumption on the partitions and improves the DOA tracking performance of the lter
as illustrated in Fig. 9 (b). Moreover, the partition weights q
k
(x) for each partition not only depends on
the motion parameters, but also the frequency. Hence, when the time-frequency signatures of the targets
differ, the IPPF can create better predictive states, resulting in overall better tracking.
5 10 15 20 25 30 35 40
20
22
24
26
28
30
32
34
t in [s]
F
r
e
q
u
e
n
c
y

T
r
a
c
k
s

i
n

[
H
z
]
(a)
5 10 15 20 25 30 35 40
20
40
60
80
100
120
t in [s]
D
O
A

T
r
a
c
k
s

i
n

[

]
2
1
3
(b)
Fig. 8. (a) Initial target frequencies are 19.50Hz, 19.60Hz, and 19.60Hz. (b) As the targets 1 and 2 get close to each other, their
motion parameters are not sufcient to distinguish their DOAs; hence, the IPPFs DOA tracking perfomance deteriorates.
5 10 15 20 25 30 35 40
5
10
15
20
25
30
t in [s]
F
r
e
q
u
e
n
c
y

T
r
a
c
k
s

i
n

[
H
z
]
2
1,3
(a)
5 10 15 20 25 30 35 40
20
40
60
80
100
120
t in [s]
D
O
A

T
r
a
c
k
s

i
n

[

]
2
1
3
(b)
Fig. 9. (a) Only the frequency track corresponding to the target marked with circle is changed from the previous example. (b)
The tracking is improved due now to the difference in the frequency of the targets that move close to each other.
22
VI. CONCLUSIONS
The state and observation equation set is the heart of any tracking model. In this paper, two new state
models and a new observation model are demonstrated using the IPPF. Given an observable state update,
it is relatively easy to generate the IPPF equations automatically for an acoustic observation model that
can resolve target DOAs. When all the targets have either zero acceleration or small accelerations, both
state models (I and II) have nearly identical tracking performance. Since the computational complexity
of state model II is about twice that of state model I, this trade-off favors model I. In one of the examples
shown, two targets with the same DOA track have quite different tracks in x-y space depending on the
acceleration model. The ghost track estimated by model I, due to the incorrect heading direction estimate,
is undesirable in the data association process by the multiple nodes; hence, model II is preferred since it
included acceleration in its state.
A derivation for the array steering matrix in the case where the target signal phases are locally chirps
was presented. In addition, it is demonstrated that if the state is also augmented to include the frequency
of individual targets, then it is possible to track targets with similar motion parameters when the targets
have different time-frequency signatures. This is an important result, since it enables the acoustic trackers
using these new state models to track multiple targets moving close to each other such as in a convoy.
VII. ACKNOWLEDGEMENTS
The authors would like to thank Matthew Orton for his invaluable comments on the intuition of the
IPPF.
REFERENCES
[1] D.H. Johnson and D.E. Dudgeon, Array Signal Processing: Concepts and Techniques, Prentice Hall Signal Processing
Series, 1993.
[2] P. Stoica and A. Nehorai, Music, maximum likelihood, and cram er-rao bound, IEEE Transactions on Acoustics, Speech,
and Signal Processing, vol. 37, no. 5, pp. 720741, May 1989.
[3] Y. Zhou, P.C. Yip, and H. Leung, Tracking the direction-of-arrival of multiple moving targets by passive arrays: Algorithm,
IEEE Transactions on Signal Processing, vol. 47, no. 10, pp. 26552666, October 1999.
[4] H. Wang and M. Kaveh, On the performance of signal-subspace processing-part i: Narrow-band systems, IEEE
Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-34, pp. 12011209, October 1986.
[5] A.B. Gershman and M.G. Amin, Wideband direction-of-arrival estimation of multiple chirp signals using spatial time-
frequency distributions, IEEE Signal Processing Letters, vol. 7, no. 6, pp. 152155, June 2000.
[6] M. Orton and W. Fitzgerald, A bayesian approach to tracking multiple targets using sensor arrays and particle lters,
IEEE Transactions on Signal Processing, vol. 50, no. 2, pp. 216223, February 2002.
23
[7] W.L. Brogan, Modern Control Theory, Prentice Hall, 1991.
[8] J.F. Giovannelli, J. Idier, R. Boubertakh, and A. Herment, Unsupervised frequency tracking beyond the nyquist frequency
using markov chains, IEEE Transactions on Signal Processing, vol. 50, no. 12, pp. 29052914, December 2002.
[9] R.F. Barrett and D.A. Holdsworth, A frequency tracking using hidden markov models with amplitude and phase
information, IEEE Transactions on Signal Processing, vol. 41, no. 10, pp. 29652976, October 1993.
[10] R.F. Barrett and D.A. Holdsworth, Frequency line tracking using hidden markov models, IEEE Transactions on Acoustics,
Speech, and Signal Processing, vol. 38, no. 4, pp. 586598, April 1990.
[11] M.S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, A tutorial on particle lters for online nonlinear/non-gaussian
bayesian tracking, IEEE Transactions on Signal Processing, vol. 50, no. 2, pp. 174188, 2002.
[12] V. Cevher and J.H. McClellan, 2-d sensor position perturbation analysis: Equivalence to awgn on array outputs,
Washington DC., 2-4 August 2002.
[13] E.D. Sontag, On the observability of autonomous discrete time nonlinear system, Int. J. Contr., vol. 36, pp. 867874,
1982.
[14] N.R. Goodman, Statistical analysis based on a certain multivariate complex gaussian distribution (an introduction),
Annals of Mathematical Statistics, pp. 152177, March 1963.
[15] J.M. Bernardo, Reference posterior distributions for bayesian inference, J.R. Statist. Soc. B, vol. 41, pp. 113147, 1979.
[16] J.M. Bernardo and J.M. Ramon, An introduction to bayesian reference analysis: Inference on the ratio of multinomial
parameters, Bayesian Statistics, vol. 47, pp. 101135, 1998.
[17] J.O. Berger and J.M. Bernardo, On the development of reference priors, Bayesian Statistics, vol. 4, pp. 3560, 1992.
[18] A.D. Lanterman, Schwarz, wallace, and rissanen: Intertwining themes in theories of model selection, International
Statistical Review, vol. 69, pp. 185212, 2001.
[19] J.J. Rissanen, Fisher information and stochastic complexity, IEEE Transactions on Information Theory, vol. 42, no. 1,
pp. 4047, January 1996.
[20] G.E. Box and G.C. Tiao, On sequential simulation-based methods for bayesian ltering, Tech. Rep. CUED/F-
INFENG/TR.310, Department of Engineering, University of Cambridge, 2001.
[21] A.D. Ackerberg, A new use of importance sampling to reduce computational burden in simulation estimation, under
revision at Review of Economic Studies, Available http://www.econ.ucla.edu/ackerber/smoonber2.pdf.
[22] A. Doucet, N. Freitas, and N. Gordon, Eds., Sequential Monte Carlo Methods in Practice, Springer-Verlag, 2001.
[23] M. Orton and W. Fitzgerald, A bayesian approach to tracking multiple targets using sensor arrays and particle lters,
Tech. Rep. CUED/F-INFENG/TR.403, Department of Engineering, University of Cambridge, 2001.