Proceeding of the 2004 American Control Conference FrM06.

1
Boston, Massachusetts June 30 - July 2, 2004

Tutorial on Multisensor Management and Fusion Algorithms

for Target Tracking
Michael K. Kalandros, Lidija Trailović, Lucy Y. Pao, and Yaakov Bar-Shalom

Abstract— This paper provides an introduction to sensor

fusion techniques for target tracking. It presents an overview Target m
of common filtering techniques that are effective for moving
targets as well as methods of overcoming problems specific
to target tracking, such as measurement-to-track association
and sensor registration. The computational demand of such
algorithms is discussed and various practices, including dis-
tributed processing of target tracks and sensor management,
are proposed to help reduce this demand. Final comments Target n
include a discussion of applications and implementation issues
specific to the presented scenarios.

I. I NTRODUCTION
Sensor A
Target tracking is the process of maintaining state es- :: The tracking system
timates of one or several objects over a period of time. : Sensor B
requires additional
processing of the
obtained information
These objects can be aircraft, ships, or ground-based targets. Centralized or Distributed
before making a decision
information processing
Mobile robots can track the location of landmarks in their
environment to maintain an accurate estimate of their loca-
Fig. 1. An illustration of the sensor fusion idea: the radars provide
tion, as well. Target tracking algorithms are basically state measurements of the surveillance region and the processing units (cen-
estimation algorithms, where the estimate of the state is tralized or distributed) gather data, perform a sensor fusion algorithm,
corrected by measurements from various sensors, which can and determine positions of targets. Sensor fusion algorithms are capable
of combining information from diverse sensing equipment, and improve
include radar, sonar, and CCD cameras, to name a few. An tracking performance, but at a cost of increased computational complexity.
illustration of tracking multiple aerial targets using sensors
such as radars is presented in Figure 1. The use of multiple
sensors can dramatically improve tracking accuracy in a
process known as sensor fusion. Section II discusses the requiring data association algorithms to associate measure-
extension of common state estimation and target tracking ments to target tracks. Data association is further compli-
algorithms, such as the Kalman filter [9], to include the cated by the possible presence of non-target or “clutter”
fusion of data from multiple sensors based on a centralized measurements (detections or returns from nearby objects,
processing architecture as shown in Figure 2. clouds, electromagnetic interference, acoustic anomalies,
One of the difficulties that sets target tracking apart from false alarms, etc.) that must be identified as not originating
other estimation tasks is the uncertainty in measurement with any targets of interest. The resulting uncertainty about
origin. Typically in state estimation tasks, the measurement measurement origin decreases the accuracy of the state
is clearly linked to the object about which state informa- estimate of each track and can ultimately lead to track loss.
tion is desired. In many tracking environments, however, One approach to data association is known as the Nearest
measurements are received from multiple objects, and it Neighbor (NN) filter, which simply selects the measurement
is not known which measurement originated from which that is closest to the predicted measurement, discarding all
object. In these environments (radar-based tracking, for the others ([8], p. 123). Another approach is Probabilistic
example) there is no clear linkage of measurement to object, Data Association (PDA), which creates a weighted average
of all the received measurements, based on the likelihood
This work was partially supported by the Office of Naval Research that each measurement came from the target of interest
(Grant N00014-02-1-0136). The authors thank Weerawat Khawsuk for his ([8], Section 3.4). Sensor fusion techniques for tracking
comments on Section 5, and Richard M. Powers for providing the data for
Figures 5, 6, and 7. environments requiring data association are discussed in
M. K. Kalandros is with Air Defense Systems Department, The Johns Section III.
Hopkins Applied Physics Lab, Johns Hopkins Road, Laurel, MD 20723- Another factor that affects the accuracy of sensor fusion is
6099, [email protected].
L. Trailović and L. Y. Pao are with the Department of Electrical and imperfect knowledge of the location of each sensor, leading
Computer Engineering, University of Colorado, Boulder, CO 80309-0425, to errors common to each measurement from that sensor.
[email protected], [email protected]. Techniques to reduce these errors, discussed in Section IV,
Y. Bar-Shalom is with the Department of Electrical and Sys-
tems Engineering, University of Connecticut, Storrs, CT 06269-1157, generally involve augmenting the state of the target track to
[email protected]. include estimates of the sensor’s relative location.
0-7803-8335-4/04/$17.00 ©2004 AACC 4734
 2

Sensor 1 observation. For example, let us consider a moving object

such as an aircraft, observed by both a pulsed radar and
Sensor 2 Tracking an infrared imaging sensor. The radar provides accuracy
System Target
Environment Tracks in determining the aircraft’s range, but a limited ability
(Data to determine the angular direction of the aircraft. On the
Association (State
& State Estimates) other hand, the infrared imaging sensor has accuracy in
Sensor Ns Estimation) determining the aircraft’s angular direction, but is unable to
measure range. If these two observations can be correctly
associated, the combination of the two sensor data would
Fig. 2. Tracking system using a centralized processing architecture. It provide improved estimates of the location (better than
includes a set of sensors that observe the environment, including the targets either of the two independent sensors), but at a cost of
to be tracked, and a central system for processing the outputs of those additional computation.
sensors.
Computational complexity studies of the tracking algo-
rithms are numerous and can be categorized as studies of the
filter complexity and studies of data association complexity.
While the fusion of multiple sensor measurements can For example, comparison of different filtering techniques
result in dramatically increased tracking performance, it in terms of handling nonlinearities were discussed in [2],
often leads to large increases in communication load and [62], [63], similar tracking scenarios were analyzed in [26],
computational complexity. One method of implementing [28], [36], [74], and comparison of several data association
such algorithms is to distribute the tracking task across algorithms were addressed in [60]. More on computational
multiple processors. For example, each sensor could have its issues in tracking algorithms will be presented in Sec-
own processor that uses single-sensor tracking algorithms. tion VII.
The tracks from each processor are either shared among Multisensor fusion was designed to combine sensor in-
the different processors or sent to a “global” processor formation from a variety of sensors and to improve target
for fusion. This approach has its own drawbacks, most state estimates (over those resulting from using a single
notable of which is that while measurements of a target sensor) in a surveillance region, or in other words, to resolve
from different sensors are generally uncorrelated, the tar- uncertainty in retrieved data sets. Significant activity has
get tracks created from the measurements are correlated, taken place in the research and development of sensor
requiring additional processing. However, since the number and data fusion systems for military applications such as
of target tracks is generally much lower than the number automatic target recognition, mission-critical systems (in
of individual measurements, the complexity seen by each battlefield assessment), and for smart weapons. More re-
processor will still be less than that required for a single cently, applications such as integration of intelligent systems
centralized processor system. Distributed and decentralized (automated plant management, computer vision, vehicle
tracking are discussed in Section V. teleoperation, intelligent robot navigation, etc.), atmospheric
As the number of targets and sensors increases, track- sciences data modeling (e.g., handling large state spaces
ing systems can very quickly become overloaded by the in remote sensing), medical diagnosis (e.g., imaging), and
incoming data. Furthermore, as the number of available smart buildings are beginning to apply methods that have
sensors and sensor modes increases, it is easy to overwhelm originated in multisensor fusion. A number of projects of
human operators, such as fighter pilots, as well. Sensor interdisciplinary nature for civilian applications have been
manager systems that balance tracking performance with carried out — Distributed Space Systems Technology, Earth
system resources are often required to control the flow Science Technology, to name a few — where some form of
of information. Although functions controlled by sensor uncertainty is built in the acquired data sets. Implementation
managers include a wide array of sensing activities such of multisensor fusion systems has become practical in the
as the selection of sensor modes or scanning an area last decade due to the advances in sensor manufacturing,
for unknown targets, we primarily overview methods for signal processing algorithms, VLSI technology, and high
determining how best to task sensors to illuminate specific performance computing and communication [29].
targets in Section VI. By setting an allowable threshold on It is worth noting the difference between sensor data
track accuracy, sensors can be assigned by the needs of each fusion and general data fusion (or frequently, information
target track, instead of attempting to apply all sensors to all fusion), according to terminology suggested in [29]. Mul-
targets. Alternatively, sensors can be selected in such a way tisensor fusion implies a process which generally employs
that their effectiveness is maximized, thus deriving the most both correlation and fusion processes to transform sensor
information from the available sensor resources. measurements into updated states and covariances for entity
In principle, fusion of multisensor data provides signif- tracking (also known in the literature as correlator-tracker
icant advantages over single source data. In addition to or level 1 fusion). Information fusion refers to acquir-
the statistical advantage gained by combining same-source ing, processing, and intelligently combining information
data (e.g., obtaining an improved estimate of a physical gathered by various knowledge sources and sensors to
phenomenon through redundant observations), the use of provide a better understanding of the phenomenon under
multiple types of sensors may increase the accuracy of the consideration. It implies the use of multisensor fusion as an
4735
 3

integral part of a more complex process. Information fusion The Kalman filter is based on the following models of
is capable of handling data diverse in nature and origin, re- the target and measurement systems ([7], Chap. 5):
quires access to a variety of resources (e.g., data/knowledge
x(k) = Fx(k − 1) + w(k − 1) (3)
base), and can be described as a process by which the
tracked entities are associated with environmental, doctrinal, z(k) = Hx(k) + v(k) (4)
and performance constraints, or presents a structured multi- where x(k) is the current state of the target; F, H are known
perspective assessment of entities and trends (situation system matrices; z(k) is the measurement of the target;
assessment and threat assessment, or level 2 and level w(k) is a variable representing process noise or higher-
3 fusion). Techniques for information fusion are complex order motion not modeled by F; and v(k) is a variable
and drawn from a wide range of areas including artificial representing measurement noise in the sensor. Both w(k)
intelligence, pattern recognition, statistical estimation, etc. and v(k) are assumed to have zero-mean, white, Gaussian
Though the terms sensor fusion and information fusion may probability distributions.
be used interchangeably in the literature, issues related to Since w(k) and v(k) are zero-mean noise processes, the
information fusion as described above are beyond the scope target states and measurements in the next time interval can
of this tutorial paper. be predicted by
x̂(k|k − 1) = Fx̂(k − 1|k − 1) (5)
II. F ILTERING ẑ(k|k − 1) = Hx̂(k|k − 1) (6)
From a Bayesian perspective, the general tracking prob- where x̂(k|k − 1) is the expected value of x(k), E[x(k)],
lem is to recursively calculate some degree of belief that given data up to the time period k − 1 and ẑ(k|k − 1) is the
the state x(k) at time k has some value given the data expected value of the measurement at time k. The quantity
z(1 : k) up to time k. Thus, it is required to construct ν(k) = z(k) − ẑ(k|k − 1) is known as the innovation. The
the pdf p(x(k)|z(1:k)). It is assumed that the initial pdf covariance of the state and the innovation predictions are
p(x(0)|z(0)) ≡ p(x(0)) of the state vector, also known P and S, respectively:
as the prior, is available (z(0) being the set of no mea-


P(k|k − 1)=E (x(k) − x̂(k|k − 1))(x(k) − x̂(k|k − 1))
surements). In principle, the pdf p(x(k)|z(1 : k)) may be
obtained recursively in two stages: prediction and update. =F P(k − 1|k − 1)F
+ Q(k − 1) (7)
In the update stage the measurement z(k) is used to modify


the prior density to obtain the required posterior density of S(k) =E[(z(k) − ẑ(k)) (z(k) − ẑ(k)) ]
the current state. =H P(k|k − 1) H
+ R(k) (8)
To define the general problem of tracking, let us consider
the evolution of the state-space model of a target defined where Q(k) is the process noise covariance and R(k) is the
by a sequence {x(k)}, measurement noise covariance. The state estimate is then
updated with the new measurement:
x(k) = fk (x(k − 1), w(k − 1)), (1) x̂(k|k − 1) = x̂(k|k − 1) + K(k) (z(k) − H x̂(k|k − 1))
(9)
where fk (·) is a possibly time-varying nonlinear function where K(k) is the Kalman gain:
of the state x(k), w(k) is an i.i.d. process noise sequence,
and k is the discrete time index. The tracking objective is K(k) = P(k|k − 1) H
S−1 (k) . (10)
to recursively estimate x(k) based on measurements from The state covariance is updated by
sensor j
P(k|k) = (I − K(k)H) P(k|k − 1) . (11)
z(k) = hk (x(k), v(k)), (2)
The Kalman filter, as well as the other, more advanced
where hk (·) is a possibly time-varying nonlinear function of algorithms in this section, can be extended to include
the state, and v(k) is an i.i.d. measurement noise sequence. inputs from multiple sensors per scan in either a parallel
In particular, tracking is the calculation of filtered estimates or sequential fashion [81]. In the parallel implementation,
of x(k) based on the set of all available measurements the measurement vectors zs (k) = Hs x(k) are stacked into
z(1 : k) = {z(i), i = 1, . . . , k} up to time k. a single vector Z(k) = [z1 (k), z2 (k), . . . zNs ], where Ns
The recursive relations for prediction and update form the is the total number of sensors and s is the sensor index,
basis for the optimal Bayesian solution [3] but in general and then processed as a single sensor as described above.
it can not be determined analytically. Solutions exist in a However, extending the size of the measurement vector can
restrictive set of cases. In the special case when the state and dramatically increase the computational complexity of the
measurement equations are linear, the noise is Gaussian, and algorithm, particularly in the calculation of the inverse of
in the absence of clutter, the optimum analytical solution to the innovation covariance matrix for the Kalman gain in
the problem of recursively estimating the states x(k) based (10). Because of this, the parallel approach to multisensor
on the measurements z(1 : k) is given by the Kalman filter fusion is rarely used. In the sequential implementation a dif-
(KF). ferent KF is run for each sensor, with the a posteriori state
4736
 4

estimate and covariance from each sensor’s KF becoming for any nonlinearity. Methods that demand approximation
the a priori state estimate and covariance for the following and/or linearization (as does the EKF) modify the system
sensor’s KF. model itself in order to solve the filtering problem. The
A central operation performed in the KF is the propa- EKF achieves only first-order accuracy, but interestingly,
gation of a Gaussian random variable through the system the computational complexity of the UKF is the same order
dynamic equations (5) and (6). However, many real-world as that of the EKF.
systems, such as target tracking problems with clutter and/or The first step in the UKF is to sample the prior state
maneuvering targets, involve elements of nonlinearity and distribution, that is, generate the Sigma points:
non-Gaussianity, which preclude exact analytic solution.
χ0,k−1 = x̂k−1
Resolving nonlinearities in tracking problems motivated κ
the development of a number of suboptimal filtering al- W0 =
(n + κ)
gorithms. The extended Kalman filter (reviewed in [2],  
[78]) has become a standard technique used in a number χi,k−1 = x̂k−1 + (n + κ)Pk−1
i
of nonlinear estimation and machine learning applications. 1
These include estimating the state of nonlinear dynamic Wi =
2(n + κ)
systems, estimating parameters for nonlinear system identi-  
fication (e.g., learning the weights of a neural network), and χi+n,k−1 = x̂k−1 − (n + κ)Pk−1
i
dual estimation (e.g., the expectation maximization (EM) 1
algorithm) where both states and parameters are estimated Wi+n = (16)
2(n + κ)
simultaneously. In the extended Kalman filter (EKF), the
states are propagated using their nonlinear descriptions in where i = 1, 2, . . . , n, and n is the state dimension. Each
(1) and (2) without contributions from the noise compo- Sigma point (χi,k−1 ) has an associated weight (Wi,k−1 ).
nents, The parameter κ is a design parameter of the UKF which
affects the the distribution of the Sigma points. After the
x̂(k) = fk (x(k − 1)) (12) prior distribution is sampled to generate the Sigma points,
ẑ(k) = hk (x(k)), (13) they are propagated through the nonlinear model. This is
the prediction step of the UKF:
and then the state distribution is approximated by a Gaus-
sian random variable. The state estimate is then propagated χi,k = f (χi,k , uk−1 , (k − 1)) (17)
analytically through a first-order linearization of the non- 2n

linear system by replacing the state transition matrix F x̂−

i,k = Wi χi,k (18)
and the measurement matrix H with the partial derivatives i=0
2n
of the nonlinear functions of the state and measurement,  

P− = Wi χi,k − x̂(k)− χi,k − x̂(k)− .(19)
respectively, k
 i=0
∂(x(k))  In the UKF, the prediction step also requires determining the
F(k) = , (14)
∂x  x=x̂(k|k) output distribution by propagating the Sigma points through
the nonlinear output equation. Denoting the output Sigma
 points as χyi,k , the output Sigma points and the distribution
∂hk (x(k)) 
H(k) =  . (15) are
∂x x=x̂(k|k)  
χyi,k = h χyi,k−1 , uk−1 , (k − 1) (20)
The new F and H matrices are used in the rest of the
KF equations, which are unchanged. If the true density 2n

is non-Gaussian (e.g., bimodal or heavily skewed), this ŷk− = Wi χyi,k (21)
can introduce large errors in the true posterior mean and i=0
covariance of the transformed Gaussian random variable, 2n   

which may lead to sub-optimal performance and sometimes P−
yy,k = Wi χyi,k − ŷk− χyi,k − ŷk− (22)
i=0
divergence of the EKF. The unscented Kalman filter (UKF)
2n
   

addresses this problem by using a deterministic sampling
P−
xy,k = Wi χi,k − x̂−
k χyi,k − ŷk− . (23)
approach. The UKF has been proposed [32], [33] and
i=0
adopted in recent years as an alternative to the EKF. The
state distribution is again approximated by a Gaussian In the correction step, an alternative Kalman gain is cal-
random variable, but is now represented using a minimal culated and used to update the a priori estimate with the
set of carefully chosen sample points (called Sigma points) measurement,
 −1
instead of approximating the nonlinear system dynamics. Kk = P− −
xy,k Pyy,k + Rk . (24)
These sample points capture the true mean and covariance
completely, and when propagated through the true nonlinear The state and covariance updates are calculated as in the
system, capture the posterior mean and covariance accu- standard Kalman filter, using the alternative gain computa-
rately to the third order (of the Taylor series expansion) tion.
4737
 5

Another approach when resolving the nonlinear and non- supporting a particular target track can also vary due to
Gaussian nature of tracking is using Sequential Monte the target maneuvering mode or the type and resource
Carlo methods that have evolved from the general problem limitations of the individual sensors. This variability of
of estimation in nonlinear, non-Gaussian systems such as the target motions and the sensor systems complicates the
particle or bootstrap filters. In many estimation and tracking tracking process. A Kalman filter is often employed to
application areas it is important to include both nonlinear filter the position measurements for estimating the position,
or non-Gaussian elements in order to accurately model the velocity, and acceleration of a target. When designing a
dynamics of physical systems. To reduce storage costs and Kalman filter, the process (or acceleration) noise Q is
allow for rapid adaptation to changing signal characteristics, selected such that the 65 − 95% probability region contains
it is required to process data on-line in a sequential manner. the maximum acceleration level of the target [15]. When
Particle filters [2], [22] are sequential Monte Carlo meth- targets maneuver, the acceleration changes in a deterministic
ods based on point mass (or “particle”) representations of manner, the white noise assumption associated with the
probability densities, which can be applied to any (possibly process noise does not hold, and the filter develops a bias
nonlinear and non-Gaussian) system model, and which in the state estimate during maneuvers. If a “larger” Q
generalize traditional Kalman filtering methods. Because of is chosen, the bias in the state estimates is less during
the highly nonlinear dynamics and measurement models, maneuvers, but such a Q poorly characterizes the target
and non-Gaussian distributions in ground moving target motion when the target is not maneuvering, leading to filter
scenarios, particle filters have been applied for littoral performance that is far from optimal.
tracking (targets on land or in the sea near the boundary) The Interacting Multiple Model (IMM) algorithm can
and yielded good tracking performance [50]. be applied in such scenarios [15], [16]. When the motion
Consider a one-dimensional random process with pdf of the target is represented by multiple models which
p(x). If Np particles xi , i = 1, . . . , Np are drawn from are hypothesized to include the correct one, the IMM
the process, then a discrete approximation of the density estimation algorithm is an efficient approach for merging
p(x) is written as the different model hypotheses. In the IMM algorithm,
the state estimate is computed under each possible model

Np
hypothesis for the most recent sampling period with each
p(x) ≈ wi δ(x − xi ) , (25)
model using a different combination of previous model-
i=1
conditioned estimates. All the hypotheses are then merged
where δ(·) is the Dirac delta function and wi is the weight to compute the output state estimate and associated error
of the i-th particle. A particle filter operates by recursively covariance. The IMM algorithm consists of a filter for each
propagating a set of Np particles through the model (1)- model, a model probability evaluator, an estimate mixer at
(2). Under certain assumptions, it has been shown [22] the input of the filters, and an estimate combiner at the
that the discrete density approximation approaches the true output of the filters. With the assumption that the model
a posteriori density p(x(k)|z(1), . . . z(k)) as Np → ∞. switching is governed by an underlying Markov chain, the
Variations in particle filters differ in the details of the mixer uses the model probabilities and the model switching
numerical procedures used to recursively propagate the set probabilities to compute a mixed estimate for each filter.
of particles. Several versions of the algorithms can be found Define x̂q (k − 1|k − 1) as the a posteriori estimate for
in [2], [20], [22]. model q at time k − 1. Define ωq (k) as the probability that
Studies of advanced filtering algorithms are numerous the target is defined by model q at time k − 1. Now define
and, as a rule, include comparisons between Kalman filter, a “mixing probability”, ωr|q (k − 1|k − 1) as the probability
EKF, UKF, and particle filters, showing tradeoffs among that the model was defined by model r at k − 1 given that
these approaches in different applications. For example, a it is now defined by model q. This can be calculated using
nonlinear trajectory tracking problem in a one-dimensional Bayes’ rule
state space has been proposed as a benchmark example in
[2] where it was shown that classical Kalman or extended 1
ωr|q (k − 1|k − 1) = pqr ωr (k − 1),
Kalman filters do not perform as well as a particle filter. cq
Another example is tracking (from radar observations) a q, r = 1, . . . , Im (26)
ballistic object on reentry, considered a highly complex
problem in nonlinear filtering. In [63] the Cramer-Rao lower where Im is the number of models, pqr is the probability
bounds were derived for the variance of the estimation error of the target switching from model r to model q and cq is
and a comparison of several nonlinear filtering techniques a normalization constant.
was presented (Kalman filter, EKF, UKF, and the particle Now create the a priori state estimate for model q by
filter). Considering the computational and statistical perfor- using an average of the state estimates from each model,
mance, the unscented Kalman filter was shown to be the weighted by the mixing probabilities,
preferred choice for this particular application.
In many tracking systems and scenarios, targets may 
Im
x̂0q (k − 1|k − 1) = ωr|q (k − 1|k − 1) x̂r (k − 1|k − 1)
switch between a number of maneuvering and non- r=1
maneuvering modes and the number and type of sensors (27)
4738
 6


Im
specifying which target is being referenced [8], [11], [56],
P0q (k − 1|k − 1) = ωr|q (k − 1|k − 1) ·
[58]:
r=1
(Pr (k − 1|k − 1)+ xt (k + 1) = Ft (k)xt (k) + Gt (k)wt (k) . (32)
0q
[x̂ (k − 1|k − 1) − x̂ (k − 1|k − 1)] ·
r
The noise vectors wt (k) are independent Gaussian random
[x̂r (k − 1|k − 1) − x̂0q (k − 1|k − 1)]
. variables with zero mean and known covariances Qt (k).
(28) Measurements (also called reports or returns) from the
sensor are received by a central processor at discrete time
Note that these estimates will be different for each model, intervals. Each measurement can originate from at most
because the mixing probabilities will be different for each one target. Sensor may not provide measurements at every
model. Once the a priori estimates are calculated, each filter interval. Some of the measurements arise from targets,
is processed independently using the measurements from and some from clutter; some targets may not yield any
the current scan. The model probabilities are updated by measurements at all in a particular time interval. The
calculating the likelihood ∆q (k) that each model produced probability of detection is assumed to be constant across
the current target measurement and combining it with the a targets for a given sensor and will be denoted PD . False
priori probability of being the model currently that defines measurements (clutter) are uniformly distributed throughout
the target motion, the surveillance region with density λ.
Let Mk be the number of measurements from the sensor
1 
Im
ωq (k) = pqr ωr (k − 1) . (29) at the k-th time interval. Assuming a pre-correlation gating
c r=1 process is used to eliminate some of the returns [8], [58],
let mk denote the number of validated measurements from
This process is repeated for each scan. To extract an the sensor at time k. The volume of the gate at time k is
updated state and covariance estimate, the model outputs chosen such that with probability PG the target-originated
are merged by averaging the state estimates weighted by measurements, if there are any, fall into the gate of the
their model probabilities: sensor. The number of gated false measurements is modeled
by a Poisson distribution µF (mk ) given by

Im
x̂(k|k) = ωr (k)x̂r (k|k) (30) (λVk )mk
µF (mk ) = e−λVk . (33)
r=1 mk !

Im
P(k|k) = ωr (k) (Pr (k|k) + [x̂r (k|k)− x̂(k|k)] · The target-originated position measurements are determined
r=1
by
[x̂r (k|k) − x̂(k|k)]
) . (31) zt (k) = H(k)xt (k) + vt (k), (34)
where 1 ≤ t ≤ T and 1 ≤ $ ≤ Mk . The H(k) matrices
Improvements in the application of the IMM algorithm
are known, each vt (k) is a zero-mean Gaussian noise vector
for radar management (pointing, scheduling, and waveform
uncorrelated with all other noise vectors, and the covariance
selection) and highly maneuvering targets in the presence of
matrices R(k) of the noise vectors vt (k) are known. For a
false alarms and electronic counter measures are presented
given target t it is not known which measurement $ (1 ≤
in [44]. The IMM estimator was used in combination
$ ≤ Mk ) originates from the target. That is the problem of
with the probabilistic data association (PDA) technique to
data association whereby it is necessary to determine which
compare several benchmark scenarios of multiple sensors
measurements originate from which targets [8].
tracking a single target. The combined IMMPDAF tech- Let Z(k) denote the set of gated measurements at time
nique presents a unified framework for target tracking and k,
radar management (estimation results were used quanti- Z(k) = (z1 (k), . . . , zmk (k)) . (35)
tatively to select the operating conditions of the radar).
More recently, IMM was applied to improve algorithms The superscripts t are not indicated, since it is not known
for resolving multipath problems (due to reflections) which which measurements originated from which target. Finally,
occur in tracking low elevation targets near the sea-surface let Z k denote the sequence of the first k observations,
[69].
Z k = (Z(1), . . . , Z(k)). (36)
We generally assume that the initial states of the targets
III. DATA A SSOCIATION
are perfectly known, and each target is always well inside
In this section, we overview several approaches for the surveillance region. In practical KF-based data associ-
data association. A centralized processing architecture is ation algorithms, gating is usually required to limit com-
assumed, where all sensor measurements are received by putational burden. Assuming that the measurement noise
a central processor prior to data fusion. and model process noise are normally distributed, gating
Consider tracking T targets with a single sensor in is done by a Chi-squared test [8] to determine whether a
a cluttered environment. To distinguish between different given measurement lies within the minimum-volume hyper-
targets, augment the Kalman filter equations with notation ellipsoid which contains a set percentage (PG ) of the
4739
 7

statistical distance between the propagated track’s predicted

measurement and each measurement received during each
z2(k) scan.
^
zn(k) ^
zr(k) The Nearest Neighbor algorithm [8] selects the mea-
surement that is closest (minimizing dt (k)2 ) to the target
track at each scan as the “true” measurement for the
z1(k) z3(k) Kalman filter. However, when the clutter density is high, or
targets are closely-spaced, it is easy to associate the wrong
measurement with the target.
Measurements To reduce the effects of misassociation, the probabilistic
Predicted Measurements data association (PDA) algorithm [8], [11] and its extension
for multiple targets, joint probabilistic data association
(JPDA), use a weighted average of gated measurements
Fig. 3. Tracking multiple targets in the same environment may cause
overlapping of validation regions and create sharing of measurements.
rather than selecting a single measurement candidate. Given
The gating regions for targets n and r are centered around the predicted the gated (true and false) measurements, the JPDA track-
measurements ẑn (k) and ẑr (k). In this example, measurement z2 (k) is ing algorithm computes the probabilities of each of the
gated by both targets n and r.
measurements being the true measurement of target t. A
combined measurement is then calculated as a weighted
average of these gated measurements,
probability distribution of the predicted measurement. More
rigorously, for a given target t, a gated measurement $ must 
M

satisfy dt < γ, where z̃t (k) = βt (k) z (k) , (38)
=1
dt (k)2 = νt (k) S t (k)−1 νt (k) (37) where the weights, βt , are the computed probability that
where dt is known as the normalized distance, νt is the measurement $ originated from target t. Each combined
innovation between target track t and the $-th measurement measurement is then used to update the the target state esti-
during scan k, and S t (k) is the covariance of the innovation mate using the Kalman filter. The state covariance resulting
for each measurement. The threshold γ is set such that the from using the combined innovation is
true target track will be gated with probability PG . Pt (k|k) = Pt (k|k − 1) −
In each time interval where measurements are received, (1 − β0t )Kt (k)H(k)Pt (k|k − 1) + P̃t (k)
several measurements may be associated with (fall into the
gate of) each target as shown in Figure 3. The associated (39)
measurements give rise to a Gaussian mixture of compo- 
mk
P̃t (k) = Kt (k) βt (k)νt (k)νt (k)
− ν̃ t (k)ν̃ t (k)

nents representing several possible estimates for each target
=1
state. In the JPDA algorithm, each Gaussian mixture is
(Kt (k))
(40)
approximated by a single Gaussian mode with the same
mean and covariance as the original mixture [8], [25]. where β0t is the probability that none of the measurements
There are several approaches for constructing optimal and originated from target t.
sub-optimal estimators in cluttered environments and here In the multisensor case, these algorithms have been
we will briefly discuss two categories: Maximum A Pos- extended by typically using recursions in the number of
teriori (MAP) Probability and Maximum Likelihood (ML) sensors Ns . As with the filtering algorithms of the previous
filters. The Probabilistic Data Association (PDA) Filter [8] section, the two alternative implementations of multisensor
and Mixture Reduction (MR) filter [65] use a weighted algorithms are sequential and parallel, e.g., multisensor
average based on the total probability theorem. The Nearest JPDA (MSJPDA). In practice, sequential algorithms are
Neighbor (NN) filter [8] is an approximate MAP filter, not only superior in computational requirements over the
whereas the Mean-Field Event-Averaged Maximum Likeli- parallel implementation (in case of MSJPDA, [26], [58]),
hood Estimator (MFEAMLE) [38] is an ML estimator. The but in tracking performance as well. It was shown that
Probabilistic Multi-Hypothesis Tracking filter (PMHT) [72] computational complexity for the parallel implementation
is a hybrid of the two. MFEAMLE and PMHT are in- (which is not practical in reality due to the fact that multiple
herently multitarget algorithms, but there are multitarget sensors are hard to synchronize) grows exponentially with
extensions of the NN, PDA, and MR algorithms: the Global the number of sensors. A sequential implementation of the
Nearest Neighbor filter, Joint Probabilistic Data Association multisensor JPDA algorithm was presented in [26], where
(JPDA) [8], and Multi-Target Mixture Reduction [59]. it was shown that it has only linear growth in complexity
The challenge, of course is to correctly associate each with an increasing number of sensors, and that it results
target’s track with the measurements from that target, as in better performance (in terms of both root-mean squared
opposed to measurements from clutter or another target. position error (RMSE) and track lifetime metrics) for track-
This is done by evaluating the probability that each mea- ing in cluttered environments. While parallel and sequential
surement came from a given target, which is based on the implementations for pure Kalman filtering (i.e., no clutter
4740
 8

and hence no data association required) are equivalent in comes from target t at time k, and n is the iteration number.
terms of performance [81], this is not true for tracking in Having calculated the synthetic measurement data, apply a
cluttered environments. Kalman smoothing algorithm (a smoothing algorithm filters
The implicit assumption is that the probability density the incoming data moving both forward in backward in time
of the target track can be represented by a single Gaus- to create a more accurate, “smoothed” state estimate [7]).
sian distribution or “mode”. The Mixture Reduction Filter Repeat the calculation of the synthetic measurements using
relaxes this constraint and propagates multiple distribution the new, smoothed state estimates and repeat until some
modes for each target [65]. Each mode essentially represents stopping criterion is reached.
a separate target estimate. In multisensor centralized tracking systems, measure-
Allowing an infinite number of distribution modes, the ments are typically collected in “scans” or “frames” and
optimal maximum a posteriori estimator is, in the multiple then transmitted to a processing center. Communication
target case, Multiple Hypothesis Testing (MHT). This ap- network delays and varying pre-processing times at the sen-
proach results in an exponential increase in the number of sor platforms can result in out-of-sequence measurements
track modes over time [8], as a target track (mode) and an (OOSMs) [10], [48], [49]. The algorithms that account for
associated probability are created in each time step for all OOSMs can be classified in two types, OOSM filtering
possible combinations of the present set of tracks and the set algorithms and OOSM tracking algorithms. OOSM filtering
of new measurements. Associating every new measurement addresses the update of the state and covariance using
with every target mode is generally impractical for more the OOSMs. Research in multisensor multitarget OOSM
than a few iterations and hence, various suboptimal filters tracking involves data association, filtering, and hypothesis
have been proposed, namely pruning and combining algo- management. The resulting problem (namely, how to update
rithms. Pruning algorithms are characterized by eliminating the current state estimate with an “older” measurement) is
new modes with lower probabilities in each time step. In a nonstandard estimation problem. It was first solved for
combining algorithms, two or more modes are combined the case where the out-of-sequence measurement lag is less
into one. In ([5], Ch. 2) an optimization-based approach for than a sampling interval (the single-lag case). Subsequently,
the MHT using multidimensional assignment as opposed to the suboptimal algorithm was extended to the case of an
the conventional, enumerative approach is introduced. arbitrary (multistep) lag [10]. A single-model multiple-
MFEAMLE, a maximum likelihood algorithm, performs lag OOSM algorithm for data association, likelihood com-
data association using a mean-field approach to sum over putation, and hypothesis management for a multisensor
all associations [38]. The mean-field approach simplifies the multitarget multi-hypothesis tracking (MHT) system that
summation over all the associations and avoids assignment handles missed detections and clutter is presented in [49].
of a priori probabilities to the associations through the use More recently, particle filters have also been applied to
of a continuous variable, η whose integral converges to the OOSM problems [50].
Kronecker delta function. Estimation of the state x̂(k) is
done by maximizing a pseudo-log likelihood function based IV. S ENSOR R EGISTRATION
on the state and η. However, this function must be solved An added complication to the fusion of multiple sensor
using (computationally complex) numerical methods. outputs is the effect of relative errors on the target tracks.
PMHT is an empirical Bayesian algorithm [72] that Most sensors measure target location relative to their own
includes the simplifying assumption that different target-to- position (for example a radar might infer the distance from
measurement assignments are independent, which is gen- the radar antenna to the target from the time-of-flight of
erally not the case, but allows for a reduction in the an electromagnetic pulse). In many applications, only the
computational growth of algorithm the relative to those that relative distance from the platform to the target is important.
do not use this assumption (e.g., JPDA). Association of In those cases, the target position can be represented on
measurements with tracks is modeled as random variables a coordinate system centered on the sensor. However, if
estimated jointly with the target states, and the ML estimate the location of the target relative to some other location is
is found iteratively using the method of Expectation Maxi- important, then a more general, global coordinate system is
mization (EM). In practice this is done as a batch process required (as when an airborne radar sentry relays a target’s
(taking multiple scans of data at once) and replacing the location to a ground-based or sea-based tracking system). In
measurement and measurement covariance in the Kalman this case, the location of the sensor in the global coordinate
filter equation with a “synthetic” measurement z̃t (k) and a system becomes as important as that of the target. Any error
corresponding measurement covariance R̃t (k). For a set of in the estimate of the sensor’s position will become an error
measurements collected between time k1 and k2 , in the estimate of the target’s position. If multiple targets are
tracked by that sensor, they will all have a correlated error
mk
m=1 wm,n (k)zm (k)
t that is a function of the sensor position error. When fusing
z̃t (k) = mk , k = k1 , . . . , k2 (41) the outputs of multiple sensors, it will generally be the case
m=1 wm,n (k)
t
that these sensors are not all located in one position. Thus
R(k)
R̃t (k) = mk (42) a global coordinate system is required, as well as accurate
t
wm,n (k)
m=1 locations of each sensor. The process of eliminating such
t
where wm,n (k) is the probability that measurement m errors is known as sensor registration.
4741
 9

In its most general sense, sensor registration is another External Environment

state estimation problem, only in this case the state being
estimated is that of the sensor platform, regardless of
whether it is stationary or mobile. The most straightforward S S S S S S
approach is to measure the location of a target whose
position is very well known, and use that measurement to
correct the estimate of the sensor’s location. Of course for
P P P
mobile platforms this will involve repeated measurements of
the reference target as well as some sort of predictive filter
such as the Kalman filter to maintain an accurate estimate.
One way to accomplish registration for both stationary Communication Network
and moving platforms is by augmenting the target’s state
with that of the sensor platform’s (e.g., [53], [64]). Thus the
same tracking filter that is being used to locate the target Global Processor
will do the same for the sensor platform.
The addition of multiple targets tracked by each sensor
S = sensor P = local processor
increases the complexity of the sensor registration task.
Now the different target tracks must be correlated with one
another to determine which tracks correspond to the same Fig. 4. A distributed architecture divides the tracking task among multiple
processors to reduce communication burden and the computational demand
objects. However, if the tracks can be correctly associated, on individual processors.
the relative bias of the different sensors can be identified and
eliminated — relative registration. Several methods of track
correlation and bias estimation exist, including [46], [47], tracks i and j being fused at the global level, the state
[71]. The issues of sensor bias observability and absolute estimate becomes [19]
registration with only targets of opportunity are discussed 
in [46], [47]. x̂(k|k) = P(k|k) P−1 i (k|k)x̂i (k|k)
+P−1
j (k|k)x̂j (k|k)
V. D ISTRIBUTED VS . C ENTRALIZED T RACKING
−P−1
i (k|k − 1)x̂i (k|k − 1)
While the fusion of measurements from multiple sensors
−P−1
j (k|k − 1)x̂j (k|k − 1)
can dramatically improve tracking accuracy, there is a cost
in computational demand as well as communication band- +P−1 (k|k − 1)x̂(k|k − 1) (43)
width requirements. Each measurement must be transmitted

to a central processor to be processed, representing a signif- P(k|k) = P−1 −1 −1
i (k|k) + Pj (k|k) − Pi (k|k − 1)
icant burden in high clutter or target-rich environments. An −1
alternative architecture is shown in Figure 4, where a single −P−1
j (k|k − 1) + P
−1
(k|k − 1)
processor has been replaced by a group of processors, each (44)
of which monitors a subset of the total number of sensors
where x̂i (k|k), x̂j (k|k), and x̂(k|k) are the updated state
and reports the results to a single, global processor. By
estimates of the tracks from processor/sensor i, j, and the
distributing the tracking tasks among a group of processors,
global processor, respectively, while x̂i (k|k−1), x̂j (k|k−1)
the computational demand on any one processor can be
and x̂(k|k − 1), are the a priori predictions of the state
reduced. Furthermore, if those processors can send only
from the previous scan, and Pi (k|·), Pj (k|·), and P(k|·)
the target tracks, rather than each measurement, then the
are the respective covariances. This approach is only exact
communication requirements would also be dramatically
when the updates from the processors are synchronized and
reduced (even more so if track information can be commu-
take place at every scan. If the updates from the sensor-
nicated at a reduced rate). Additionally, such an architecture
level processors are not passed to the global processor every
will result in a more robust tracking system as the failure
scan, then the propagation of the common process noise
of a single processor no longer represents a catastrophic
again causes correlation between the estimation errors. This
failure.
track-to-track fusion approach has not yet been extended to
The drawback to this approach is that because each pro-
account for clutter and data association as well as more
cessed track includes the effects of process noise originating
general classes of architectures.
with the target itself, the errors among the tracks are now
An alternative approach1 is known as decorrelation,
correlated [4]. There are several approaches to distributed
where the correlation between sensor-level tracks is re-
track fusion, including both maximum-likelihood (e.g., [6])
moved, allowing the global processor to treat the incoming
and minimum mean square error (MMSE) solutions. Here,
target tracks as uncorrelated measurements. Although orig-
we will overview two approaches to MMSE track fusion.
inally derived for pure filtering in [23], [27], it has recently
The first approach, direct track-to-track fusion, includes
the correlation effects directly in the fusion process. Apply- 1 This is really an algebraic rearrangement of (43)–(44); see Eq. (8.6.1-9)
ing this approach to the linear case, with two sensor-level in [8].
4742
 10

been extended to address tracking targets in clutter [40]. In

this approach, decorrelation is achieved by subtracting the Sensor Fi
a priori component of each state estimate Manager

y(k) = x̂(k|k) − C(k)x̂(k|k − 1) (45) P(k|k-1)

C(k) = P(k|k)P−1 (k|k − 1) (46)
Y(k) −1
= P(k|k) − P(k|k)P (k|k − 1)P(k|k) Sensor 1
Tracking {Z(k)}

...
...
(47)
System
where x̂(k|k − 1) is the a priori target state estimate,
P(k|k − 1) is the a priori covariance of that state estimate, Sensor Ns
y(k) is the decorrelated state estimate, and Y(k) is the
covariance of that estimate. As a “pseudomeasurement”,
y(k), along with its covariance, are sent from each local
processor for each track to a global processor. The mea- Fig. 5. Block diagram of a multisensor tracking system with a sensor
surement matrix for the decorrelated estimate becomes manager. The a priori covariance from the target tracker is passed to the
sensor manager which then selects a subset Φi of the sensor suite to
B(k) = I − C(k) (48) use during the following scans. Sensors can be chosen to maximize the
information gain from each sensor [66], achieve a desired track covariance
= I − P(k|k)P−1 (k|k − 1) (49) [36], or to optimize some other tracking metric.

where I is the identity matrix. Because the y(k) se-

quences are decorrelated, the global processor can use well-
across all targets or the amount of information from the
established ‘centralized’ fusion algorithms for processing
sensors. This approach can generally be defined by
the y(k) to form global estimates of the target states.
This technique has been further extended for multisensor Φi = arg mini f (P(k|k − 1), Pi (·|k)) (50)
multitarget tracking of targets in clutter [40], [41], including
feeding back information from global to local processors where Φi is sensor combination i, P(k|k − 1) is the a
[42]. priori covariance of the target track, and Pi (·|k) is the a
posteriori covariance (possibly propagated forward in time)
VI. S ENSOR M ANAGEMENT after the application of sensor combination i. The possible
cost functions include one consisting of the trace of each
Sensor managers are a general class of systems that
target’s state estimate covariance weighted by target priority
generate sensing actions, then prioritize and schedule those
(which presumably increases with importance) [54],
actions [52]. Sensing actions can include the tasking of
sensors to illuminate a target, the selection of sensor modes, f (P(k|k − 1), Pi (k + 1|k)) = priority
or scanning an area for unknown targets. These actions will
×T r(Pi (k + 1|k))
be selected to achieve various goals such as maintaining a
target track, optimizing the chance of detecting new targets, (51)
identifying detected targets, and minimizing electromag-
as well as one that uses a spectral decomposition of the
netic emissions to reduce the chance of detection by the
covariance matrix to select a sensor that will reduce the
enemy. Since sensing resources are generally limited, sensor
largest eigenvalue of that matrix [57].
managers must solve an optimization problem that balances
The 1990’s brought the application of information theory
those resources with tracking goals. A discussion of the
to the sensor management field, which has dominated it
various sensor manager approaches becomes essentially a
ever since. Here, instead of minimizing a cost function as in
discussion on how to formulate the optimization problem.
(51), the goal is to maximize the information gained by each
Although much work has been done on the management
sensor. Information gain is defined by taking the natural log
of individual sensors, especially radar (where the sampling
of the quotient formed by dividing the determinant of the
or revisit rate [44], [13], [21], [43], [79], [80], or emitted
a priori covariance by the determinant of the a posteriori
waveforms [39] can be modified to control the level of
covariance for each sensor combination [66]:
uncertainty in the target track), since this tutorial focuses
on multisensor fusion, only the selection of sensors from a 1 |P(k|k − 1)|
I (P(k|k − 1), Pi (k|k)) = ln
larger group of sensors will be discussed. 2 |Pi (k|k)|
Figure 5 shows how a sensor manager is incorporated into (52)
a target tracking system. The target state estimate covariance
is used to select the appropriate set of sensors to use on a This amounts to a measure of the uncertainty volume reduc-
given target. Those sensors then produce measurements that tion. Other, similar definitions of information gain have also
will be used to update the state estimate of that target. The been proposed [31], [68]. The use of the information metric
sensors are selected based on their ability to optimize some has been extended to include multitarget tracking scenarios
function of either the resulting target estimate covariances requiring data association [67].
4743
 11

5
The information metric is particularly useful in that it 10
MR-10
can be applied to a range of sensing activities, allowing
the sensor manager to compare the utility of a variety PDAF
of activities under one metric. For example, information PMHT

Average Track LIfetime

4
10 NN
resulting from a target search can be calculated based on
the difference in the probability of a given cell containing MFEAMLE
a target depending on whether the sensor will or will not 3
detect a target there. When combined with the information 10
metric of [66], the tasks of target search and tracking can
be combined into a single sensor manager [51].
2
A relatively new approach to sensor management is 10
known as Covariance Control. This approach defines a
desired target estimate covariance, Pd , for each track and
attempts to achieve that covariance. The desired covariance 1
10 -4 -3 -2
can be a function of the target’s priority or of an external 10 10 10
goal such as achieving enough estimate accuracy before Normalized Clutter Density
applying weapons to that target. A variety of approaches
based on these metrics exist [36]. One metric, known Fig. 6. Average track lifetime as a function of normalized clutter density
as the Eigenvalue/Minimum Sensors Algorithm, attempts
to achieve the desired covariance goal while minimizing
the number of sensors used. The fact that achieving Pd
The performance measures were: track lifetime (Figure 6),
will result in the difference Pd − Pi having all positive
root-mean squared position error (RMSE) (Figure 7), and
eigenvalues, gives this algorithm its name. More rigorously,
computational complexity (Figure 8), while the variable
the algorithm selects a sensor combination iev with the
parameters were normalized target acceleration and normal-
fewest sensors that achieves the covariance goal:
ized clutter density. Figures 6, 7, and 8 present results for
Φev = {Φi : Pd − Pi > 0} (53) tracking a single target until it is lost.
iev = arg min |Φi |, Φi ∈ Φev (54) The definition of track loss used is that proposed in [65]:
i
a track is determined to be lost when either the estimate is
where |Φi | = Nsi is the number of sensors in combination sufficiently far from the true track (the “norm test”) or the
i. This approach has been extended to include the influence true measurement falls outside of the gate (the “acceptance
of the PDAF algorithm on the covariance [34]. Addition- test”) for five consecutive time steps. Track lifetime is the
ally, algorithms for selecting a desired covariance and the difference between the time step of track loss and the time
appropriate sensors have been derived to reduce the effect step of track initiation (assumed to be at time zero, since
of closely-spaced targets on probabilistic data association track initiation was not considered in [60]). Track lifetimes
methods [35]. for a given normalized clutter density were averaged across
all trials.
VII. C OMPUTATIONAL C OMPLEXITY AND R EAL -T IME The reported results show some expected trends, e.g., the
I MPLEMENTATION I SSUES nearest neighbor algorithm has acceptable performance for
The view of the computational complexity for different low and medium clutter, has low computational burden, but
tracking algorithms is affected by their intended application. tracks poorly in high clutter; MR with 10 modes performs
An understanding of the tradeoffs between the performance well in terms of track lifetime, but at the cost of higher
and computational complexity of the applied algorithms computational complexity compared to other algorithms,
may provide guidance for users to select the best algorithms whereas the PDA algorithm yields a “middle of the road”
for their purposes. solution; PMHT shows superior performance in the RMSE
One such study presents the computational requirements metric at the cost of additional computational burden across
and the performance of a parallel implementation of the the clutter range; and the MFEAMLE algorithm shows high
multisensor JPDA algorithm [26], [58]. A sequential imple- computational demand due to inherent numerical issues.
mentation of the multisensor JPDA algorithm was presented The rise in computational complexity with the number
in [26], and it was shown that it has linear growth in of sensors is an issue for sensor managers as well. A
complexity with an increasing number of sensors, and that global search over all possible sensor combinations grows
it results in better performance (in terms of both root-mean exponentially with the number of sensors, often outstripping
squared position error (RMSE) and track lifetime metrics) the computational cost of simply using all of the sensors
for tracking in cluttered environments. (other limitations, such a maximum number of sensor
In [60], five data association algorithms were compared dwells, may make sensor management necessary even if the
in simulation using a common linear single-target motion computational demand is large). There are however, several
test bed. The comparison included the NN, PMHT, PDA, search methods that can reduce this growth. The most
MR, and MFEAMLE algorithms described in Section III). straightforward is the “Greedy” algorithm, which selects
4744
 12

0
Average RMS Position Estimate Error 10 resources so that better designs get more simulation time).
MFEAMLE The developed method is applied to a general tracking
NN problem of Ns sensors tracking T targets using a sequential
PDAF multi-sensor data fusion tracking algorithm (MSJPDA). The
optimization consists of finding the order of processing
MR-10
sensor information that results in the smallest variance of
PMHT
the position error. Results obtained with high confidence
levels and in reduced simulation times show that processing
the best available sensor the last in sequential MSJPDA
Equivalent KF Error
has the smallest RMS error, e.g., the order of processing
sensor information affects tracking precision in the pres-
ence of clutter (i.e., when data association algorithm is
necessary). The presented method can be applied to any
-1 ranking and selection problem where variance (or standard
10
10
-4
10
-3 -2
10 deviation) is the performance metric. Further improvement
Normalized Clutter Density in comparison of tracking algorithm was achieved when the
distribution of position error was modeled with Gaussian
mixture pdf [74], [75]. Incorporating the improved statistical
Fig. 7. Average RMS position estimate error as a function of normalized
clutter density
modeling of the position error variance in computing budget
allocation algorithm [74], simulation time is reduced by
an order of magnitude compared to long Monte Carlo
Avg Flops / Track / Timestep / Expected Number of Measurements

8
10 experiment.
NN One application of the developed variance ranking al-
PDAF gorithm is comparing sensor-order designs in sequential
MR-10
7 PMHT MSJPDA as a function of clutter density λ and RMSE as
10 MFEAMLE performance metric. A design is an ordered set of sensor
noise covariance parameters, θj = {r1 , r2 , . . . , rNs }, with
senor noise covariance Rs = rs I in (8). A better sensor is
6
10 characterized with smaller sensor noise parameter r. In [75],
the sequential MSJPDA algorithm with Ns = 3 sensors
tracking T = 4 targets are considered. The compared de-
5
signs are defined as ordered sets of sensor noise parameters
10
(two identical sensors with r = 0.1 and one better sensor
with r = 0.01). The objective is to find the best design in
terms of the RMSE for various clutter densities λ. The three
4
10 -4 -3 -2
designs are
10 10 10
Normalized Clutter Density
θ1 = {0.1, 0.1, 0.01}, θ2 = {0.1, 0.01, 0.1},

Fig. 8. Computational burden as a function of normalized clutter density θ3 = {0.01, 0.1, 0.1}. (55)
The variance ranking tools [73], [75] were used in the
optimization algorithm to yield the results in Fig. 9 and
sensors that maximize the change in the cost function,
Table I. For low clutter densities λ, the order of sensor
driving the farthest towards its maximum or minimum,
processing does not affect the tracking performance. Since
depending on the formulation, which reduces the compu-
the RMSE’s do not differ significantly, the best ranking
tational demand to a polynomial growth with the number
design θb is selected with a small confidence level when
of sensors [36], [82]. Other search techniques add a random
the maximum number of simulation runs Nmax = 100
component to the search to further improve the performance
is exhausted. For higher clutter λ, the order of sensor
of the sensor selection algorithm [37].
processing affects the tracking more significantly. The best
Comparison of different tracking algorithms or parameter ranking design is consistently the one where the best sensor
sets within one algorithm relies on time-consuming and is processed last.
computationally demanding Monte Carlo simulations. In
[76] a method is presented to minimize simulation time,
yet to achieve a desirable confidence of recognizing better VIII. A PPLICATIONS
designs. It is achieved by applying ordinal optimization Multisensor fusion was developed to combine sensor in-
ideas (i.e., determining rank order instead of quantifying formation and to improve target state estimates (over using
the performance of the best design) and computing budget a single sensor) in a surveillance region, or in other words,
allocation techniques [73] (i.e., allocating computational to resolve uncertainty in diverse data sets retrieved from
4745
 13

RMSE
of this paper.

0.25
Q1= {0.1, 0.1, 0.01}
A. Tracking of Ground Targets
Q2= {0.1, 0.01, 0.1}
The tracking of aerial or sea-based targets usually rely on
0.2 Q3= {0.01, 0.1, 0.1} the assumptions that the motion of each target is constrained
only by the motion models, which may include types of
0.15 motion as well as limits to acceleration and velocity; that
the target will generally not be obscured (may not hold in
0.1 optical tracking); and that target clutter will be uniformly
distributed. These assumptions are far less likely to hold
when tracking ground targets (see e.g., [24] and [70],
0.05
p. 423). In these applications, the motion of the target may
be constrained by terrain, targets may often be obscured
0 by foliage or other terrain, and clutter distribution is more
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
clutter density l likely to deviate from uniform. These difficulties may hold
for air- and sea-based tracking as well when targets are
Fig. 9. Root-mean squared position error (RMSE) of the sequential flying close to the earth or are in littoral areas. One approach
MSJPDA algorithm as a function of clutter density λ for three designs. for tracking ground targets on-road and off-road has been
presented in [45].
λ 0.01 0.1 0.5 1.0 1.4 2.0 There are several ways of dealing with the increased
θb θ2 θ1 θ1 θ1 θ1 θ1 tracking difficulties described above. One is to use moving
Nruns 100 100 93 85 21 15
target indicators or doppler radar to distinguish moving
conf. level 0.43 0.86 0.98 0.98 0.98 0.99
TABLE I targets from stationary objects ([70], p. 104). While this
T HE SELECTED BEST RANKING DESIGN θb , THE TOTAL NUMBER OF will reduce the amount of clutter, it does not alleviate ob-
SIMULATION RUNS Nruns , AND THE CONFIDENCE LEVEL AS scuration (in this case, a stopped vehicle can be considered
FUNCTIONS OF CLUTTER DENSITY λ. T HE MAXIMUM ALLOWED obscured) or constraints on motion. Obscuration by foliage
NUMBER OF RUNS IS Nmax = 100. can be reduced through the use of lower frequency radars
that can see through the tree canopy. The inclusion of
terrain information can be used in conjunction with sensor
measurements to improve the target state estimates [1], [24],
different sensors. Early applications were mostly military, [30]. Roadway intersections lead naturally to the use of
such as automatic target recognition, battlefield assessment, MHT or IMM trackers, as well.
and smart weapons. However, techniques originally de- Investigation efforts in the development of high-accuracy
veloped in multisensor fusion have now found their way multisensor vehicle state estimation schemes using the UKF
in applications and projects of interdisciplinary nature for is presented in [77]. The model is able to cope with vehicle
civilian use (space technology [12], [18] and earth sciences slip using multisensor data from internal sensors, odometry,
[61]) where some form of uncertainty is built in the acquired and Differential Global Positioning System (D-GPS).
data sets.
Sensor fusion algorithms have also been validated with IX. S UMMARY AND A REAS OF F UTURE W ORK
biomedical data [17]. Though the methods for sensor fu- Multisensor data fusion seeks to combine information
sion in biomedical applications originate in information from multiple sensors and sources to improve inferences
processing, artificial intelligence, and classification algo- about a surveillance region that are not achievable from a
rithms, the applied estimation methods combine statistical single sensor or source. Applications include military appli-
and control theory algorithms (e.g., weighted least squares, cations such as automatic target recognition, automated situ-
maximum a posteriori probability, minimum a posteriori ation assessment, and for smart weapons, as well as civilian
variance estimation, etc.). Estimates are recursively updated applications such as automated plant management, mobile
via Kalman or adaptive Kalman filters. Clinical testing robot navigation, and environmental modeling. Thus, an ex-
of software for detecting respiratory phases has shown tensive legacy exists including: process models, algorithms,
that the complex techniques of sensor fusion have become evolving tool kits, and systems engineering methodology
accepted by physicians to aide diagnostics only after the (system design and algorithm selection).
introduction of user-friendly interfaces. Another step for- There are still a number of growing challenges for data
ward has been the implementation of soft computing on a fusion, at lower levels of inferences: understanding sensor
symbol level through which the user can integrate his or her processing and sensor limitations, multitarget tracking and
own experience and terminology. More complex medical identification, and at higher levels of inferences: automated
applications of multisensor fusion include improvement of reasoning for situation assessment, development of alter-
image processing and pattern recognition, and hence fall in native hypotheses for threat assessment, monitoring, and
the information fusion category, which is beyond the scope control of fusion processes.
4746
 14

One potential area of future improvement in sensor [22] A. Doucet, N. de Freitas, and N. Gordon (editors), Sequential Monte
fusion algorithms is incorporating the continuously emerg- Carlo Methods in Practice, Springer-Verlag Inc., 2001.
[23] O. E. Drummond, “Feedback in Track Fusion Without Process
ing advanced filtering techniques developed to handle the Noise,” Proc. SPIE Conf. Signal and Data Processing of Small
nonlinear nature of tracking while reducing computational Targets, 2561, pp. 369-383, 1995.
burden. Another area is resolving issues in data fusion [24] R. H. Enders, “Fundamentals of on-road tracking,” Acquisition,
Tracking, and Pointing, Orlando, FL, pp. 334-342, 1999.
over distributed sensor networks with data communication [25] T. E. Fortmann, Y. Bar-Shalom, and M. Scheffe, “Sonar Tracking of
delays [55]. Multiple Targets Using Joint Probabilistic Data Association,” IEEE
J. Oceanic Engineering, Vol. 8, No. 3, pp. 173-183, July 1983.
[26] C. W. Frei and L. Y. Pao, “Alternatives to Monte-Carlo Simulation
R EFERENCES Evaluations of Two Multi-sensor Fusion Algorithms,” Automatica,
Vol. 34, No. 1, pp. 103-110, Jan. 1998.
[1] D. Andrisani, M. F. Tenorio, J. Lu, and F. P. Kuhl, “A Hierarchical [27] G. Frenkel, “Multisensor Tracking of Ballistic Targets,” Proc. SPIE
Target Tracker Using Image Data,” Proc. American Control Conf., Conf. Signal and Data Processing of Small Targets, 2561, pp. 337-
San Diego, CA, pp. 1004-1008, June 1990. 346, 1995.
[2] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A Tuto-
[28] L. Frenkel and M. Feder, “Recursive Estimate-Maximize (EM)
rial on Particle Filters for Online Nonlinear/Non-Gaussian Bayesian
Algorithms for Time Varying Parameters with Applications to Mul-
Tracking,” IEEE Trans. Sig. Proc., Vol. 50, No. 2, pp. 174-188,
tiple Target Tracking,” Int. Conf. on Acoustics, Speech, and Signal
Feb. 2002.
Processing, ICASSP-95, Vol. 3, pp. 2068-2071, 1995.
[3] D. Avitzour, “A Stochastic Simulation Bayesian Approach to Mul-
[29] D. L. Hall and J. Llinas, “An Introduction to Multisensor Data
titarget Tracking,” IEE Proc. on Radar, Sonar and Navigation,
Fusion,” Proc. IEEE, Vol. 85, No. 1, pp. 6-23, 1997.
Vol. 142, 1995.
[30] J. G. Herrero, J. A. Besada Portas, and J .R. Casar Corredera, “Use
[4] Y. Bar-Shalom, “On the Track-to-Track Correlation Problem,” IEEE
of Map Information for Tracking Targets on Airport Surface,” IEEE
Trans. Automatic Control, Vol. 26, No. 2, pp. 571-572, Apr. 1981.
Trans. Aerospace and Electronic Systems, Vol. 39, No. 2, pp. 675-
[5] Y. Bar-Shalom and W. Dale Blair (editors), Multitarget-Multisensor
693, April 2003.
Tracking: Applications and Advances, Vol. III, Artech House, 2000.
[6] Y. Bar-Shalom and L. Campo, “The Effect of the Common Pro- [31] K. J. Hintz and E. S. McVey, “Multi-Process Constrained Estima-
cess Noise on the Two-Sensor Fused-Track Covariance,” IEEE tion,” IEEE Trans. Systems, Man, and Cybernetics, Vol. 21, pp. 237-
Trans. Aerospace and Electronic Systems, Vol. 22, No. 6, pp. 803- 244, 1991.
805, Nov. 1986. [32] S. J. Julier and J. K. Uhlmann, “A New Extension of the Kalman
[7] Y. Bar-Shalom and X. Li, Estimation and Tracking: Principles, Filter to Nonlinear Systems,” Proc. AeroSense: The 11th Int. Symp.
Techniques, and Software, Artech House, Boston, 1993. on Aerospace/Defense Sensing, Simulation and Control, 1997.
[8] Y. Bar-Shalom and X. R. Li, Multitarget-Multisensor Tracking: [33] S. J. Julier, J. K. Uhlmann, and H. D. Durrant-Whyte, “A new
Principles and Techniques, YBS Publishing, 1995. approach for filtering nonlinear systems,” Proc. American Control
[9] Y. Bar-Shalom, X. R. Li and T.‘Kirubarajan, Estimation with Ap- Conf., pp. 1628-1632, June 1995.
plications to Tracking and Navigation: Algorithms and Software for [34] M. Kalandros and L. Y. Pao, “The Effects of Data Association on
Information Extraction, J. Wiley and Sons, 2001. Sensor Manager Systems,” Proc. AIAA Guidance, Navigation, and
[10] Y. Bar-Shalom, M. Mallick, H. Chen, and R. Washburn, “One-Step Control Conf., Denver, CO, 2000.
Solution for the General Out-of-Sequence-Measurement Problem in [35] M. Kalandros and L. Y. Pao, “Sensor Management for Tracking Inter-
Tracking,” Proc. IEEE Aerospace Conf., Vol. 4, pp. 1551-1559, acting Targets,” Proc. Workshop on Estimation, Tracking, Fusion, and
March 2002. Control: A Tribute to Yaakov Bar-Shalom, Monterey, CA, pp. 221-
[11] Y. Bar-Shalom and E. Tse, “Tracking in a Cluttered Environment 248, May 2001.
with Probabilistic Data Association,” Automatica, Vol. 11, No. 5, [36] M. Kalandros and L. Y. Pao, ”Covariance Control for Multisensor
pp. 451-460, Sept. 1975. Systems,” IEEE Trans. Aerospace and Electronic Systems, Vol. 38,
[12] E. T. Baumgartner and S. B. Skaar, “An Autonomous Vision-Based No. 4, pp. 1138-1157, Oct. 2002.
Mobile Robot,” IEEE Trans. Automatic Control, Vol. 39, No. 3, [37] M. Kalandros, L. Y. Pao, and Y. C. Ho, “Randomization and Super-
pp. 493-502, March 1994. Heuristics in Choosing Sensor Sets for Target Tracking Applica-
[13] S. S. Blackman, M. T. Busch, and R. F. Popoli, “IMM/MHT tions,” Proc. IEEE Conf. on Decision and Control, Phoenix, AZ,
Tracking and Data Association for Benchmark Tracking Problem,” pp. 1803-1808, 1999.
Proc. American Control Conf., Seattle, WA, pp. 2606-2610, 1995. [38] K. Kastella, “Event-Averaged Maximum Likelihood Estimation and
[14] S. Blackman and R. Popoli, Design and Analysis of Modern Tracking Mean-Field Theory in Multitarget Tracking,” IEEE Trans. Automatic
Systems, Artech House, 1999. Control, Vol. 40, No. 6, pp. 1070-1074, 1995.
[15] W. D. Blair and Y. Bar-Shalom, “Tracking Maneuvering Targets With [39] D. J. Kershaw and R. J. Evans, “Optimal Waveform Selection
Multiple Sensors: Does More Data Always Mean Better Estimates?” for Tracking Systems,” IEEE Trans. Information Theory, vol. 40,
IEEE Trans. Aerospace and Electronic Sytems, Vol. 32, No. 1, pp. 1536-1550, 1994.
pp. 450-456, Jan. 1996. [40] W. Khawsuk and L. Y. Pao, “Decorrelated State Estimation for Dis-
[16] H. A. P. Blom and Y. Bar-Shalom, “The Interacting Multiple Model tributed Tracking of Interacting Targets in Cluttered Environments,”
Algorithm for Systems with Markovian Switching Coefficients,” Proc. American Control Conf., Anchorage, AK, pp. 899-904, May
IEEE Trans. Automatic Control, Vol. 33, No. 8, pp. 780-783, 2002.
Aug. 1988. [41] W. Khawsuk and L. Y. Pao, “Decorrelated State Estimation for
[17] B. R. Bracio, W. Horn, and D. P. F. Moller, “Sensor Fusion in Distributed Tracking Using Multiple Sensors in Cluttered Environ-
Biomedical Systems,” Proc. 19th Annual Intern. Conf. of the IEEE ments,” Proc. American Control Conf., Denver, CO, pp. 3208-3214,
Society of Eng. in Medicine and Biology society, Vol. 3, pp. 1387- June 2003.
1390, 1997. [42] W. Khawsuk and L. Y. Pao, “Distributed Multi-sensor Multi-
[18] S. Brunke and M. Campbell, “Estimation architecture for future target with Feedback,” Proc. American Control Conf., Boston, MA,
autonomous vehicles,” Proc. American Control Conf., pp. 1108-1114, June 2004.
May 2002. [43] T. Kirubarajan, Y. Bar-Shalom, and E. Daeipour, “Adaptive Beam
[19] K. C. Chang, R. K. Saha, Y. Bar-Shalom, “On Optimal Track- Pointing Control of a Phased Array Radar in the Presence of ECM
to-Track-Fusion,” IEEE Trans. Aerospace and Electronic Systems, and False Alarms Using IMMPDAF,” Proc. American Control Conf.,
Vol. 33, No. 4, Oct. 1997. Seattle, WA, pp. 2616-2620, 1995.
[20] D. Crisan and A. Doucet, “A Survey of Convergence Results on [44] T. Kirubarajan, Y. Bar-Shalom, W. D. Blair, and G. A. Watson,
Particle Filtering Methods for Practitioners,” IEEE Trans. Signal “IMMPDAF for Radar Management and Tracking Benchmark with
Processing, Vol. 50, No. 3, pp. 736-746, March 2002. ECM,” IEEE Trans. Aerospace and Electronic Systems, Vol. 34,
[21] E. Daeipour, Y. Bar-Shalom, and X. Li, “Adaptive Beam Point- No. 4, pp. 1115-1134, 1998.
ing Control of a Phased Array Radar Using an IMM Estimator,” [45] T. Kirubarajan, Y. Bar-Shalom, K.R. Pattipati and I. Kadar, “Ground
Proc. American Control Conf., Baltimore, MD, pp. 2093-2097, 1994. Target Tracking with Topography-Based Variable Structure IMM
4747
 15

Estimator,” IEEE Trans. Aerospace and Electronic Systems, Vol. 36, [71] L. D. Stone, M. L. Williams, and T. M. Tran, “Track-to-Track
No. 1, pp. 26-46, Jan. 2000. Association and Bias Removal,” Proc. SPIE Conf. Signal Processing
[46] X. D. Lin, Y. Bar-Shalom and T. Kirubarajan, “Exact Multisensor of Small Targets, Vol. 4728, pp. 315-329, Orlando, FL, 2002.
Dynamic Bias Estimation with Local Tracks,” Proc. 6th Int. Conf. In- [72] R. Streit and T. Luginbuhl, “Maximum Likelihood Method for
formation Fusion, Cairns, Australia, July 2003. Probabilistic Multi-Hypothesis Tracking,” Proc SPIE Conf. Signal
[47] X. D. Lin, Y. Bar-Shalom and T. Kirubarajan, “Multisensor Bias Esti- and Data Processing of Small Targets, Vol. 2335, pp. 394-405, 1994.
mation Using Local Tracks without A Priori Association”, Proc. SPIE [73] L. Trailović and L. Y. Pao, “Computing budget allocation for
Conf. Signal and Data Processing of Small Targets, San Diego, CA, optimization of sensor processing order in sequential multi-sensor
August 2003. fusion algorithms,” Proc. American Control Conf., Arlington, VA,
[48] M. Mallick, S. Coraluppi, and C. Carthel, “Advances in Asyn- pp. 1841-1847, June 2001.
chronous and Decentralized Estimation,” Proc. IEEE Aerospace [74] L. Trailović and L. Y. Pao, “Variance Estimation and Ranking for
Conf., pp. 1873-1888, March 2001. Gaussian Mixture Distributions in Target Tracking Applications,”
[49] M. Mallick, J. Krant, and Y. Bar-Shalom, “Multi-Sensor Multi- Proc. IEEE Conf. Decision and Control, Las Vegas, NV, pp. 2195-
Target Tracking Using Out-of-Sequence Measurements,” Proc. 5th 2201, Dec. 2002.
Int. Conf. Information Fusion, pp. 135-142, July 2002. [75] L. Trailović and L. Y. Pao, “Position error modeling using gaussian
[50] M. Mallick, S. Maskell, T. Kirubarajan, and N. Gordon, “Littoral mixture distributions with application to comparison of tracking
Tracking Using Particle Filter,” Proc. 5th Int. Conf. Information algorithms,” Proc. American Control Conf., Denver, CO, pp. 1272-
Fusion, pp. 935–942, July 2002. 1277, June 2003.
[51] G. A. McIntyre and K. J. Hintz, “An Information Theoretic Approach [76] L. Trailović and L. Y. Pao, “Computing Budget Allocation for
to Sensor Scheduling,” Proc. SPIE, Vol. 2755, pp. 304-312, 1996. Efficient Ranking and Selection of Variances with Application to Tar-
[52] S. Musick and R. Malhotra, “Chasing the Elusive Sensor Manager,” get Tracking Algorithms,” IEEE Trans. Automatic Control, Vol. 49,
Proc. IEEE 1994 NAECON, Vol. 1, pp. 606-613, Dayton, OH, IEEE: No. 1, pp. 58-67, Jan. 2004.
New York, NY, 1994. [77] M. Wada, K. S. Yoon, and H. Hashimoto, “Nonlinear Filter Road
[53] N. Nabaa and R. H. Bishop, “Solution to a Multisensor Tracking Vehicle Model Development,” Proc. IEEE Conf. Intelligent Trans-
Problem with Sensor Registration Errors,” IEEE Trans. Aerospace portation Systems, pp. 734-739, Aug. 2001.
and Electronic Systems, Vol. 35, No. 1, pp. 354-363, Jan. 1999 [78] E. A. Wan and R. Van Der Merwe, “The Unscented Kalman filter
[54] J. Nash, “Optimal Allocation of Tracking Resources,” Proc. IEEE for Nonlinear Estimation,” IEEE Symp. Adaptive Systems for Signal
Conf. Decision and Control, New Orleans, LA, pp. 1177-1180, 1977. Processing, Communications, and Control, AS-SPCC 2000, pp. 153-
[55] D. Nicholson, C. M. Lloyd, S. J. Julier, and J. K. Uhlmann, “Scalable 158, Oct. 2000.
Distributed Data Fusion,” Proc. 5th Int. Conf. Information Fusion, [79] G. A. Watson and W. D. Blair, “Revisit Control of a Phased Array
pp. 630-635, July 2002. Radar for Tracking Maneuvering Targets when Supported by a
[56] S. D. O’Neil and L. Y. Pao, “Multisensor Fusion Algorithms Precision ESM Sensor,” Proc. SPIE Conf. Signal Processing of Small
for Tracking,” Proc. American Control Conf., San Francisco, CA, Targets, Orlando, FL, Vol. 2235, pp. 448-459, 1994.
pp. 859-863, June 1993. [80] G. A. Watson, W. D. Blair, and T. R. Rice, “Enhanced Electron-
ically Scanned Array Resource Management through Multisensor
[57] Y. Oshman, “Optimal Sensor Selection Strategy for Discrete-Time
Integration,” Proc. SPIE Conf. Signal Processing of Small Targets,
State Estimators,” IEEE Trans. Aerospace and Electronic Systems,
Vol. 3163, Orlando, FL, pp. 329-340, 1997.
Vol. 30, No. 2, pp. 307-314, April 1994.
[81] D. Willner, C. B. Chang, and K. P. Dunn, “Kalman Filter Algorithms
[58] L. Y. Pao, “Centralized Multisensor Fusion Algorithms for Tracking
for a Multi-Sensor System,” Proc. IEEE Conf. Decision and Control,
Applications,” Control Engineering Practice, Vol. 2, No. 5, pp. 875-
pp. 570-574, Nov. 1976.
887, Oct. 1994.
[82] Z. Zhang and K. J. Hintz, “OGUPSA Sensor Scheduling Architecture
[59] L. Y. Pao, “Multisensor Multitarget Mixture Reduction Algorithms and Algorithm,” Proc. SPIE Conf. Signal Processing, Sensor Fusion,
for Tracking,” J. Guidance, Control, and Dynamics, Vol. 17, No. 6, and Target Recognition V, Vol. 2755, pp. 296-303, 1996.
pp. 1205-1211, Nov.-Dec. 1994.
[60] L. Y. Pao and R. M. Powers, “A Comparison of Several Different Ap-
proaches for Target Tracking with Clutter,” Proc. American Control
Conf., Denver, CO, pp. 3919–3924, June 2003.
[61] C. Pohl and H. Touron, “Operational Applications in Multi-sensor
Image Fusion,” Int. Archives of Photogrammetry and Remote Sens-
ing, Vol. 32, Part 7-4-3, Valladolid, Spain, June 1999.
[62] C. Rago, P. Willett, and R. L. Streit, “A Comparison of the JPDA
and PMHT Tracking Algorithms,” Proc. Int. Conf. Acoustics, Speech
and Signal Processing, Vol. 5, pp. 3571-3574, May 1995.
[63] B. Ristić, A. Farina, D. Benvenuti, and M. S. Arulampalam, “Per-
formance Bounds and Comparison of Nonlinear Filters for Tracking
a Ballistic Object on Re-entry,” IEE Proc. Radar, Sonar and Navi-
gation, Vol. 150, No. 2, pp. 65-70, April 2003.
[64] S. I. Roumeliotis and G. A. Bekey, “Distributed Multirobot Localiza-
tion,” IEEE Trans. Robotics and Automation, Vol. 18, No. 5, pp. 781-
795, Oct. 2002.
[65] D. J. Salmond, “Mixture Reduction Algorithms for Target Tracking
in Clutter,” Proc. SPIE Conf. Signal and Data Processing of Small
Targets, Vol. 1305, pp. 434-445, 1990.
[66] W. Schmaedeke, “Information-based Sensor Management,” Proc.
SPIE, Vol. 1955, 1993.
[67] W. Schmaedeke and K. Kastella, “Event-averaged Maximum Likeli-
hood Estimation and Information Based Sensor Management,” Proc.
SPIE, Vol. 2232, pp. 91-96, 1994.
[68] W. Schmaedeke and K. Kastella, “Information Based Sensor Man-
agement and IMMKF,” Proc. SPIE Conf. Signal Processing of Small
Targets, Vol. 3373, Orlando, FL, pp. 390-401, 1998.
[69] A. Sinha, Y. Bar-Shalom, D. W. Blair, and T. Kirubarajan, “Radar
Measurement Extraction in the Presence of Sea-Surface Multipath,”
IEEE Trans. Aerospace and Electronic Systems, Vol. 39, No. 2,
pp. 550-567, 2003.
[70] M. I. Skolnik, Introduction to Radar Systems, Third Ed. McGraw-
Hill, Boston, MA, 2001.
4748