I.

INTRODUCTION

Target tracking and target classification are critical

Joint Target Tracking and problems in land-based and airborne surveillance
systems. Although they are fundamentally linked,
Classification Using Radar and they are usually treated as separate problems, often
using a separate suite of sensors or sensor models and
ESM Sensors techniques to solve them. For example, target tracking
is usually performed using data from kinematic
sensors (e.g. radar) while target classification is
usually performed using data from identity or
attribute sensors (e.g. electronic support measure
SUBHASH CHALLA (ESM), inverse synthetic aperture radar (ISAR),
GRAHAM W. PULFORD high resolution radar). More recently, contrary to
CSSIP this general trend, the target classification problem
University of Melbourne using radar data based on observed target dynamics
was considered in the literature by Jacobs, et al. [13]
and Cutaia and O’Sullivan [10]. On the other hand,
Chang and Fung [9] have treated the problem of target
Bayesian target classification methods using radar and classification using target identity information from
electronic support measure (ESM) data are considered. A joint
treatment of target tracking and target classification problems is high resolution radar and ESM using Bayes’ inference
introduced. First, a method for target classification using radar nets. Although they consider the tracking problem in
data and class-dependent kinematic models is presented. Second, conjunction with the classification problems, there
a target classification method using ESM data is presented.
is no consideration, however, of the link between
Then, a Bayesian radar and ESM data fusion algorithm, refered
to as direct identity fusion (DIF), for target classification is the target state and the target class. More recently,
presented. This algorithm exploits the dependence of target state methods introduced in [13, 11] consider the link
on target class via the use of class dependent kinematic models between target state and target class, where they
but fails to exploit the dependence of target class on target state.
We then introduce a method, referred to as joint tracking and
exploit a one-way dependence of target class on
classification (JTC), for treating target tracking and classification target state by using class-dependent kinematic
problems jointly, by exploiting the dependence of target class models. However, we assert that there is a two-way
on target state via flight-envelope-dependent classes and the dependence between target state and class and that it
dependence of target state on target class via class dependent
kinematic models.
can be exploited by defining classes based on flight
A two-dimensional example demonstrates the relative merits envelopes, i.e., regions of state space occupied by
of these methods. It is shown that, while the incorporation of a target belonging to a particular class. Methods of
the two-way dependence between target state and class (i.e., integrating target tracking and classification into a
JTC) promises some benefits over the method that incorporates
only a one-way dependence (i.e., DIF), there are severe filter single framework were considered by Miller, et al.
implementation difficulties for the former. The results also in [16] and by Kastella and Zatezalo in [14], where
demonstrate that the fusion of information from radar and the high resolution radar information relating to the
ESM sensors via the DIF approach results in improvements over geometry and orientation of targets is harnessed for
classification methods based on either of the individual sensors.
classification purposes. Although some attempts have
been made at deducing target identity from kinematic
(radar) measurements [11] and using state constraints
(flight envelopes) in improving target state estimates
[3], there exists no unified framework for treating
these two problems jointly. Such a framework is
developed here.
Manuscript received June 15, 2000; revised May 7, 2001; released First, we present the method that assumes a
for publication May 7, 2001. one-way dependence of the target state and class
IEEE Log No. T-AES/37/3/08571. as introduced in [13, 11]. Second, a method for
incorporating the information from an identity sensor
Refereeing of this contribution was handled by P. K. Willett.
(e.g. ESM) as described in [19] is presented. Then
This work was carried out under DSTO Contract 482919. a method for modeling the two way dependence
Authors’ current addresses: S. Challa, CSSIP, Dept. of Electrical between target state and target class, by defining
and Electronic Engineering, University of Melbourne, Parkville, flight-envelope-based target classes, class dependent
Victoria 3010, Australia, E-mail: ([email protected]); G. kinematic models, and a joint target state-class
Pulford, Thompson Marconi Sonar, Sydney NSW, Australia. probability density function (pdf), is introduced.
The optimal filtering problem in this context, called
0018-9251/01/$10.00 °
c 2001 IEEE joint target tracking and classification (JTC) [6],

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001 1039
models the joint target state-class pdf and allows for
cross-coupled feedback between state and class. The
JTC method incorporates flight envelope information
for different target classes as a means of representing
the link from target state to class. It may also include
target class information from ESM or other identity
sensors as well as target behavioral information from
observed trajectory dynamics. The incorporation
of flight envelopes poses great challenges in filter Fig. 1. Illustration of flight envelopes of targets belonging to
different classes.
development and leads to computationally intensive
filter implmentation.
On the other hand the suboptimal filtering method, (angular dynamical) properties of aircraft can be
referred to here as direct identity fusion (DIF) of incorporated as extra states [10]. Let the static target
ESM and radar, is computationally tractable and is class ck = c belong to one of s known classes f1, : : : , sg
based on a technique described in [10, 19]. It uses an at any time k. Each target class is characterized
interacting multiple model (IMM) filter to estimate by a set of possible maneuvering inputs S c from a
target class probability based on observed target Markov chain with corresponding transition matrix
maneuvers and a stochastic ESM model to estimate P c . A linear dynamical model for target behavior
target class probability based on emitter declarations. corresponding to class i can then be represented by
These two independent target class probability
distributions are then fused by Bayesian classification xk+1 = Fxk + Guk + vk , uk » ¡ (i),
(1)
techniques. Importantly, there is no feedback of xk+1 , xk 2 ¥i , vk » N(0, Q)
target identity information into the tracking algorithm
in the suboptimal filter (while it is present in the where vk is a white independent identically distributed
optimal JTC filter). In addition, the details of the ESM (IID) sequence and N(0, Q) represents a multivariate
sensor model based on the work presented in [8] are Gaussian with zero mean and covariance Q (as
summarized here. described in [1] for example). The requirement
The paper is organized as follows. The dynamical xk 2 ¥i models the flight envelope constraints for
model for the target dynamics and flight envelope class i. In practice, the sets ¥i may only constrain
information is presented in Section II, followed some components (e.g. velocity and acceleration)
by a description of the radar measurement model of the target state. A representative flight envelope
(Section IIIA) and ESM sensor model (Section set for a one dimensional tracking problem with two
IIIB). Section IV details the computation of target possible classes is shown in Fig. 1. The union of flight
class probabilities based on the emitter declarations envelope sets f¥i g is partitioned into nonoverlapping
from the ESM sensor (Section IVA) and from target regions fªj g. The maneuver input at time k, uk =
maneuvers observed in the radar data (Section IVB). Mji (k), is assumed to be from a finite-state Markov
Section IVC describes a feed-forward algorithm (DIF) chain ¡ (i), i = 1, : : : , s defined
for the fusion of identity information obtained via
radar and ESM sensors. The JTC approach, which ¡ (i) = (S i , P i )
allows feedback of classification information into S i 2 fM1i , : : : , Mr(i)
i
g
the tracking system, is presented in Section IVD. (2)
This is followed by a presentation in Section V of P i = (Plji ) l, j = 1, : : : , r(i)
some simulated test scenarios in which the target
identification and tracking performance of the two Plji = Pr(uk+1 = Mji j uk = Mli )
algorithms are compared and discussed. The report
where fM1i , : : : , Mr(i)
i
g is the set of feasible maneuver
concludes in Section VI with a discussion of the work
accelerations for a target in class i including a
and suggestions for further study.
constant velocity model, and P i is the maneuver
transition matrix for class i. It is to be noted that the
II. DYNAMICAL MODELING WITH FLIGHT number of models r(i) in class i may be different for
ENVELOPES each class.
Different classes of aircraft (e.g. fighter plane, Remark. There is an underlying inconsistency
commercial passenger aircraft) differ in their in this framework: for any bounded initial state,
geometry, size, and flight envelope and in particular maneuvering inputs and the (Gaussian) process
in their maneuvering capabilities. The acceleration noise can drive the state outside of the bounds.
capabilities of various targets can most naturally be Set theoretic approaches (e.g. [3]) appeared as a
incorporated into the target’s discrete-time dynamical possible alternative. However, such approaches to
state equation as input terms [5], while the rotational state estimation set initial bounds on the state and

1040 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
make bounded noise assumptions; this addresses the signals by time of arrival, angle of arrival and radio
problem of the evolution of the state bounds, however frequency, deinterleaving the received pulse trains and
it does not address the issue of estimation when the finally, correlating the measured signal parameters
bounds on the state are fixed. As the inconsistency is like modulation format, pulse repetition frequency,
due to maneuvering inputs and noise, by redefining pulsewidth, beamwidth, scan type, scan rate, etc., with
them and making them state dependent, the pdf of the emitter library. This leads to the identification of
the state can always be made to lie inside the flight the likely source emitters on the target or targets in the
envelopes. The inclusion of the state dependent current angular look direction.
maneuvers and noise is, however, beyond the scope The signal processing that is carried out in
of this paper. It should be noted that the method of ESM sensors is complicated and there are many
deriving the JTC filter does not depend on the type sources of possible errors in the emitter identification
of maneuver and noise models used. Hence, for the process. We have chosen to adopt a simple three stage
sake of simplicity, we have used state-independent stochastic model of the ESM processing chain that
maneuvering inputs and Gaussian noise as defined in takes into account the following primary sources of
(1). error: emitters may be on or off (usage), emitters that
are on may not be detected, and detected emitters may
III. MEASUREMENT MODELS FOR RADAR AND ESM be confused with other emitters. Possible emitter types
SENSORS under consideration belong to the set

A detailed consideration of the ESM measurement −E = fE1 , E2 , : : : , EN g

model is presented in Section IIIB while the radar where N is the total number of emitter types in the
measurement process is considered in the next section. ESM sensor’s emitter library. We assume that each
In what follows, Yk = fykx , ykc g is the set of sensor type of target has a unique set of emitters:
measurements at time k and Yk = fY1 , : : : , Yk g is the
set of sensor measurements up to time k where ykx "i = fEj1 (i) , Ej2 (i) , : : : , Ejn (i) g (4)
i
represents a kinematic measurement obtained from
the radar at time k, and ykc represents a target class with elements in the set −E . The number of emitters
measurement obtained from the ESM sensor at time k. for target type i is denoted by ni .
The ESM measurement space, denoted −q ,
represents all possible emitter type combinations from
A. Radar Measurement Model −E including the empty set
For simplicity, we assume a standard linear, ¢
discrete-time radar measurement equation of the form −q =fQ0 , Q1 , : : : , Qm g m = 2N ¡ 1
[1]
ykx = Hxk + wk , wk » N(0, R) (3) where Qj is the jth measurement subset. Under this
definition the ESM measurement at time k can only
where xk is the target kinematic state (e.g., containing match one of the measurement subsets:
position and velocity components in one or more
dimensions of target motion) and wk is a zero ykc = Qj 2 −q , k = 0, 1, : : : : (5)
mean white Gaussian noise process with known The component emitters comprising a given
covariance R. In practice the measurement function H measurement subset Qj are written
is nonlinear and techniques such as extended Kalman
filtering can be used in constructing the maneuvering Qp = [Ei1 , Ei1 , : : : , Eip ], i1 , : : : , ip 2 f1, : : : , Ng (6)
target tracking filter such as the IMM algorithm.
with each emitter occurring at most once in a given
For the subsequent description of the JTC filter, the
Qj .
nonlinearity of the range/bearing measurement process
The emitter usage state indicates whether the
is not restrictive since the JTC framework allows for
emitter under consideration is switched on or off.
non-Gaussian probability densities.
Denote by ¯j (k) the usage indicator for emitter j at
time k. This is modeled as a binary random variable
B. ESM Sensor Measurement Model with possible values:
½
ESM sensors are passive directional receivers 0 emitter j is “off”
which scan the frequency range of interest to intercept ¯j (k) = :
1 emitter j is “on”
emitted electromagnetic signals from targets in the
surveillance region and identify the likely source Define bli (k) as the lth usage vector of the emitter set
emitters. Commercially available ESM systems "i at time k,
may have emitter libraries of 2048 emitter types
bli (k) = [¯1 (k), ¯2 (k), : : : , ¯ni (k)], ¯i (k) 2 f0, 1g
or more [18]. The typical operational sequence of
an ESM sensor includes [17] sorting the received i = 1, 2, : : : , s l = 1, 2, : : : , hi

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1041

where hi = 2ni ¡ 1. It should be noted that while ¯j (k) independent for each emitter, a simpler description
deals with the state of a single emitter, bli (k) deals with is possible for the confusion probabilities:
the state of a set of emitters.
Emitter switching behavior is described by cij = Pr(declare Ej j detect Ei ): (10)
defining an emitter usage Markov chain for each
The probability for the type combinations in (9) may
emitter on a target. In this case we model emitter
now be decomposed into a product of elementary
j as having a fixed-usage Markov chain regardless
confusion probabilities in the form:
of which target it is located on and independent of
the other emitters. Let Ej (k) denote the event that ¼pq = ci1 j1 ci2 j2 : : : cip jq (11)
emitter j is in use (i.e., “on”) at time k and Ēj (k) the
complementary event that it is not in use. The usage where by assumption, emitter subsets Qp and Qq have
transition probability matrix for emitter j, denoted by the same number of emitters.
Áj , then has elements The modeling parameters must be specified for a
" # given ESM sensor and would require detailed system
Pr(Ej (k + 1) j Ej (k)) Pr(Ēj (k + 1) j Ej (k)) knowledge and practice. The ESM model in (10)
Áj = :
Pr(Ej (k + 1) j Ēj (k)) Pr(Ēj (k + 1) j Ēj (k)) requires a relatively small number of parameters,
growing approximately as the square of the number of
(7) emitters. Specifically, for N emitters, 2N parameters
Note that the row sums of the usage Markov chain are required to specify the usage Markov chain and
matrix are unity. For target type i, the overall emitter N 2 for the confusion matrix. In addition, a further
usage transition probability matrix ªi (a matrix of N detection probabilities, one per emitter, may
dimension 2ni ) for emitter set "i , when the emitter be required. While this may appear excessive for
switching processes are independent, is given by the large emitter libraries, it is likely that many of the
following Kronecker product: confusion probabilities could be assigned to zero due
to processing constraints.
ªi = Áj1 (i) − Áj2 (i) − ¢ ¢ ¢ − Ájn (i) : (8)
i

The assumption that the emitters on board a given IV. BAYESIAN TARGET CLASSIFICATION
target have independent switching processes is not
true in general. However, this approach has the Let the target belong to one of s classes i.e.,
advantage that only 2 free parameters per emitter c 2 f1, : : : , sg. and let Pr(c = i) be the prior probability
are required to specify the overall usage transition that the target belongs to class i. Then the target
probability matrix. classification problem consists of finding the posterior
To account for the errors which may be caused probability Pr(c = i j Yk ) that the target belongs to
by the processing chain of the ESM receiver (i.e., the class i given the sensor measurement sequence Yk =
deinterleaving processing), we define a measurement fy1 , y2 , : : : , yk g. The Bayesian solution to finding the
confusion matrix as follows: posterior target class probability is given by
¼pq = Pr(declare yc (k) = Qq j Qp detected)
Pr(Yk j c = i) Pr(c = i)
p, q = 0, 1, : : : , m. (9) Pr(c = i j Yk ) = Ps k
j=1 Pr(Y j c = j) Pr(c = j)

The element ¼ij of the measurement confusion matrix (12)

¦ is defined as the probability of declaring the emitter
Since Pr(c = i) is assumed known, (12) reduces
set measurement as Qj , when the emitter set Qi was
to finding the measurement likelihood Pr(Yk j c = i).
actually detected.
The evaluation of Pr(Yk j c = i), depends on the
EXAMPLE Consider an ESM sensor able to identify probabilistic relationship between the sensor
emitters −E = fE1 , E2 , E3 g, then s = 3 and −q consists measurements Yk and the target classes. This
of a total of 8 measurement subsets, namely relationship may be modeled in various ways
depending on the sensor in question. In the case of
−q = fØ, fE1 g, fE1 , E2 g, fE1 , E3 g, fE2 g,
an ESM measurement, the observed emitter sequence
7
fE2 , E3 g, fE3 g, fE1 , E2 , E3 gg = µj=0 may be used for target classification. For radar
measurements, the observed dynamical behavior is
Also assume that the emitter sets "1 = fE1 , E3 g
used for target classification. For ESM and radar
and "2 = fE2 , E3 g represent target type T1 and T2
measurements, both the observed emitter sequence
respectively.
and the target’s dynamical behavior may be used for
For realistic numbers of emitters N the size of target classification.
the confusion matrix becomes prohibitively large. A key point to note in the Bayesian approach to
If we assume that the emitter confusion process is target classification is that it assumes the availability

1042 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
of the measurement likelihood function for every where ¼j1 is derived from the confusion matrix
sensor under consideration. For a radar sensor, elements as in (11).
this likelihood function is obtained from the linear The second term in (14) is the probability of the
transformation of the measurement noise density predicted usage vector for emitter set "i which can be
and for the ESM sensor, the likelihood function is derived as follows
obtained from the confusion and the emitter usage
transition matrices. The details are examined in the Pr(bli (k) j Yck¡1 )
following sections. hi
X
= Pr(bli (k) j bpi (k ¡ 1)) Pr(bpi (k ¡ 1) j Yck¡1 )
A. Target Classification Using ESM Data p=0

hi
X
As presented in Section IIIB, the ESM data i
= Ãpl Pr(bpi (k ¡ 1) j Yck¡1 ) (15)
comprises a sequence of emitter identity declarations.
p=0
It is assumed that the targets to be classified have a
unique sets of emitters on board. Hence, ESM data where, Ãpl i
is the (p, l) element of the transition
may be used to distinguish the targets. A Bayesian probability Ãi defined in (8).
approach to this problem is presented below. The term Pr(bpi (k ¡ 1) j "i , Yck¡1 ) is the prior
Given the ESM measurement sequence Yck = probability of the emitter usage vector for emitter set
fyc (0), yc (1), : : : , yc (k)g up to time k, the problem "i . Its posterior probability is given by
is to find the posterior probability of target class:
Pr(c = i j Yck ), i = 1, : : : , s. Applying Bayes’ rule, we Pr(bli (k) j Yck ) = Pr(bli (k) j yc (k), Yck¡1 )
have
Pr(yc (k) j bli (k)) Pr(bli (k) j Yck¡1 )
Pr(c = i j Yck ) = Phi
i i k¡1 )
p=0 Pr(yc (k) j bp (k)) Pr(bp (k) j Yc
Pr(yc (k) = Qj j c = i, Yck¡1 ) Pr(c = i j Yck¡1 ) (16)
= Ps k¡1 ) Pr(c = j j Y k¡1 )
j=1 Pr(yc (k) = Qj j c = j, Yc c where Pr(bli (k) j Yck¡1 ) is obtained from (15).
(13)
where we assume Pr(c = i) has a uniform prior B. Target Classification Using Radar Data
distribution at time k = 0.
The term Pr(yc (k) = Qj j c = i, Yck¡1 ) in (13) is the The dynamical behavior of the target is embedded
likelihood that a measurement yc (k) will be observed in the observed radar data and extraction of this
at time k for target type i, given that the sequence information is a basis for target classification. In terms
of measurements Yck¡1 up to time k ¡ 1 has been of the model in Section II, the dynamical behavior
observed. Applying the Total Probability Theorem of interest is the sequence of maneuvering inputs
with respect to all possible emitter usage states of the initiated by the target and the region of state space
set "i at time k, we have it occupies (the flight envelope). These are the two
distinctive dynamical properties that differ between
Pr(yc (k) = Qj j c = i, Yck¡1 ) target classes. The use of information relating to the
hi flight envelopes is considered in Section IVD. Here,
X
= Pr(yc (k) = Qj j bli (k)) Pr(bli (k) j Yck¡1 ): we consider a method for target classification using
l=0
only the information obtained from maneuvering input
(14) probabilities. This method was proposed by Cutaia
and O’Sullivan [10].
The first term Pr(yc (k) = Qj j bli (k))
in the above The problem of target classification using radar
expression is the probability that the emitter subset Qj measurements may be stated as follows.
is declared under the hypothesis that the usage vector Given the measurement sequence Yxk = fyx (0),
of emitter set "i for type i is bli (k). This conditional yx (1), : : : , yx (k)g up to time k, the problem is to find
probability can be evaluated from the elements of the posterior probability of target class: Pr(c = i j Yxk ),
confusion matrix elements defined in (9). The usage i = f1, : : : , sg. The Bayesian classification solution is
emitter vector, a string of ones and zeros, can be given by
mapped on a one-to-one basis to the emitter set space.
For example, if bli (k) = f1, 0, 0g, the corresponding Pr(Yxk j c = i) Pr(c = i)
Pr(c = i j Yxk ) = Ps
emitter set is Q1 = fE1 g. This mapping can then be k
j=1 Pr(Yx j c = j) Pr(c = j)
used in evaluating Pr(yc (k) = Qj j bli (k)) in (14) by
subsituting where Pr(c = i) is the prior probability distribution
assigned to various target classes. The class-
Pr(yc (k) = Qj j bli (k)) = Pr(yc (k) = Qj j Q1 ) = ¼j1 conditioned measurement likelihood can be

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1043

expanded as mutually independent:
Pr(Yxk j c = i) = Pr(yx (k), Yxk¡1 j c = i) Pr(Yxk , Yck j c = i) = Pr(Yxk j c = i)Pr(Yck j c = i)
= Pr(yx (k) j Yxk¡1 , c = i) Pr(Yxk¡1 j c = i) i = 1, 2, : : : , s. (18)
k
Y Splitting the measurements sequences into present and
= p(yx (m) j Yxm¡1 , c = i): past components, it follows that
m=1

The joint likelihood of the observations is therefore Pr(c = i j Yxk , Yck ) = Pr(c = i j y(k), yc (k), Yxk¡1 , Yck¡1 )
the product of the likelihood functions of the filter 1
conditioned on target type up to time k. Furthermore, = p(y(k), yc (k) j c = i, Yxk¡1 , Yck¡1 )
±
Pr(yx (k) j Yxk¡1 , c = i) £ Pr(c = i j Yxk¡1 , Yck¡1 ) (19)
r(i)
X where ± is a normalizing constant.
= p(yx (k) j u(k) = Mj(i) (k), Yxk¡1 , c = i)
By independence of the two measurement
j=1
sequences, (19) can be further simplified as
£ Pr(u(k) = Mj(i) (k) j Yxk¡1 , c = i):
Pr(c = i j Yxk , Yck )
By noting that 1
= p(y(k) j c = i, Yxk¡1 ) p(yc (k) j c = i, Yck¡1 )
¢
¹ij (k) = Pr(u(k) = Mj(q) (k) j Yxk¡1 , c = i) ±
£ Pr(c = i j Yxk¡1 , Yck¡1 ) (20)
and
¤ij (k) = p(yx (k) j Mj(i) (k), Yxk¡1 , c = i) where ± is given by
s
X
we have
±= p(y(k) j c = j, Yxk¡1 ) p(yc (k) j c = j, Yck¡1 )
r(i)
X j=1
p(yx (k) j Yk¡1 , c = i) = ¹ij (k)¤ij (k)
j=1 £ Pr(c = j j Yxk¡1 , Yck¡1 ): (21)
so that the Bayesian solution to target classification Equation (20) implies that with independent data
based on radar measurements is given by from two sources, the probability of target type is a
rescaled product of the likelihoods contributed by the
Pr(c = i j Yxk )
hP i two measurement sequences. Furthermore, this result
Qk r(i) i i also applies to the case of independent data from
m=1 j=1 ¹j (m)¤j (m) Pr(c = i)
=P nQ hP i o more than two sources as long as their likelihoods
s k r(l) l l
l=1 m=1 ¹
j=1 j (m)¤j (m) Pr(c = l) are available. The inherent philosophy of the DIF
approach is illustrated in Fig. 2.
(17)
in which ¹ij (k), and ¤ij (k) can be obtained from
D. Joint Target Tracking and Classification
a any multiple model filter. This approach was
first considered in [11] where the IMM was used
The uncertainty about the true state and class of a
to evaluate ¹ij (k), and ¤ij (k). Incorporating ESM
target is captured by the posterior joint state-class pdf
measurements into this framework was considered
recently in [19] and is presented in the next section. p(xk , c = i j Yk ) i 2 f1, : : : , sg (22)

which satisfies the normalisation requirement

C. Direct Identity Fusion
s Z
X
Assume that time—aligned and associated data from p(xk , c = i j Yk )dxk = 1: (23)
both radar and ESM sensors are available and that i=1 xk 2¥i

both target maneuver models (in Section IIIA) and

the emitter set on board the target (as described in In this instance the measurement set Yk to time k
Section IIIB) reflect the same target types. With the consists of both radar measurements Yxk (1) and
radar position measurement sequence denoted as Yxk = target class measurements Yck from an ESM sensor as
fyx (0), yx (1), : : : , yx (k)g and the ESM measurement modeled in Section IIIB. In the following formulation
sequence denoted as Yck = fyc (0), yc (1), : : : , yc (k)g, of the JTC filter, the target class probability Pr(c = i j
the posterior probability of each target type is to be yc (k)) determined from the ESM measurement yc (k) is
estimated. We assume that data from two sources are simply denoted by p´ (ykc ¡ i). In (1) the process noise

1044 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
Fig. 2. DIF. Multiple Model tracker gives state and maneuver estimates and hence target identity probabilities which are combined with
ESM classifications.

pdf is denoted pv (¢) and the measurement noise pdf in jth mode probability for class i is defined by
(3) is written pw (¢). These are both assumed to be IID
¢
white noise processes in what follows. ¹ij (k) = Pr(uk¡1 = Mji j c = i, Yk ),
The theorem below provides the integral recursions
for the JTC pdf, collectively called the JTC filter j = 1, : : : , r(i), i = 1, : : : , s (27)
[6]. Note that the assumption of static (unknown
deterministic) target class ck = c 8k has been made. is updated recursively as
Pr(i)
THEOREM IV.1 (JTC Filter). Given the state dynamics ¤ij (k) i i
l=1 Plj ¹l (k ¡ 1)
as in (1) and (2), measurement equations as in (3) ¹ij (k) = Pr(i) P r(i) i i
: (28)
i
and (4), the joint target state-class pdf (22) satisfies m=1 ¤m (k) l=1 Pml ¹l (k ¡ 1)
the following integral recursion for i = 1, : : : , s and for The mode-conditioned likelihood ¤im (k) (for class i,
all k, conditioned on maneuver m) is defined by
1
p(xk , c = i j Yk ) = pw (ykx ¡ Hxk )p´ (ykc ¡ i) ¢
±k ¤im (k) = Pr(yk j c = i, uk¡1 = Mmi , Yk¡1 ),
Z
£ p(xk j xk¡1 , c = i, Yk¡1 ) m = 1, : : : , r(i), i = 1, : : : , s (29)
¥i

£ p(xk¡1 , c = i j Yk¡1 )dxk¡1 (24) and may be evaluated recursively via

Z
where the normalizing factor ±k is given by
¤m (k) = p´ (yk ¡ i) pw (ykx ¡ Hxk )
i c
s Z
X ¥i
±k = pw (ykx ¡ H»k )p´ (ykc ¡ l)
¥l £ p(xk j c = i, Mmi (k ¡ 1), Yk¡1 )dxk (30)
l=1
Z
where the mode-conditioned state prediction pdf is
£ p(»k j »k¡1 , c = i, Yk¡1 )
¥i
evaluated as

£ p(»k¡1 , c = i j Yk¡1 )d»k¡1 d»k (25) p(xk j c = i, Mji (k ¡ 1), Yk¡1 )

Z
where the class- and state-conditioned state prediction
pdf is expressible as = pv (xk ¡ F i xk¡1 ¡ GMji (k ¡ 1))
¥i
X
r(i)

p(xk j xk¡1 , c = i, Yk¡1 ) = pv (xk ¡ F i xk¡1 ¡ GMji (k ¡ 1)) £ p(xk¡1 j c = i, Yk¡1 )dxk¡1 : (31)
j=1
Finally, the class-conditioned prior state pdf can be
X
r(i)
reexpressed as
£ Plji ¹il (k ¡ 1) (26)
p(xk¡1 , c = i j Yk¡1 )
l=1 p(xk¡1 j c = i, Yk¡1 ) = R k¡1 )dx
:
¥i p(xk¡1 , c = i j Y k¡1
with Mji (k ¡ 1) denoting that maneuver input u(k ¡ 1) =
Mji is in effect from time k ¡ 1 to time k and where the (32)

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1045

 Fig. 3. JTC view of target tracking and classification: tracker outputs may influence classification and vice versa.

PROOF The recursions are obtained by applying elsewhere. The current radar and ESM measurement
Bayes’ rule to (22) along with standard methods from likelihoods are given, respectively, by
the IMM filter [2].
The JTC pdf is defined by p(xk , c = i j Yk ) p(ykx j xk , c = i, Yk¡1 ) = pw (ykx ¡ Hxk )
where xk 2 ¥c , c 2 f1, : : : , sg and Yk = fY1 , Y2 , : : : , Yk g. p(ykc j xk , c = i, Yk¡1 ) = p´ (ykc ¡ i)
Applying Bayes’ rule we have
p(xk , c = i j Yk ) where pw (¢) is the pdf of the measurement noise
in (3) and p´ (¢) is the probability mass function of
= p(xk , c = i j Yk , Yk¡1 ) ESM measurements. The latter may be computed as
described in Section IVA. Noting that a target class
1
= p(Yk j xk , c = i, Yk¡1 )p(xk , c = i j Yk¡1 ) c = i is characterized by a set of maneuvering inputs
±k fMji gr(i) i i
j=1 with transition matrix P = (Pjl ), we have
1
= p(ykx , ykc j xk , c = i, Yk¡1 )p(xk , c = i j Yk¡1 ) p(xk j xk¡1 , c = i, Yk¡1 )
±k
r(i)
X
1
= p(ykx , ykc j xk , c = i, Yk¡1 ) = p(xk j xk¡1 , Mji (k ¡ 1), Yk¡1 )
±k j=1
Z
£ p(xk , xk¡1 , c = i j Yk¡1 )dxk¡1 £ p(Mji (k ¡ 1) j xk¡1 , c = i, Yk¡1 )
¥c
r(i)
X
1 pv (xk ¡ F i xk¡1 ¡ GMji (k ¡ 1))
= p(ykx , ykc j xk , c = i, Yk¡1 ) =
±k j=1
Z
£ p(xk j xk¡1 , c = i, Yk¡1 ) £ p(Mji (k ¡ 1) j c = i, Yk¡1 )
¥c
r(i)
X
£ p(xk¡1 , c = i j Yk¡1 )dxk¡1 = pv (xk ¡ F i xk¡1 ¡ GMji (k ¡ 1))
1 j=1
= p(ykx j xk , c = i, Yk¡1 )p(ykc j xk , c = i, Yk¡1 )
±k r(i)
X
Z £ p(Mji (k ¡ 1), Mli (k ¡ 2) j c = i, Yk¡1 )
£ p(xk j xk¡1 , c = i, Yk¡1 ) l=1
¥c
r(i)
X
£ p(xk¡1 , c = i j Yk¡1 )dxk¡1 = pv (xk ¡ F i xk¡1 ¡ GMji (k ¡ 1))
j=1
where ±k is a normalization factor given by
Z X s r(i)
X
±k = p(Yk j xk , c = i, Y k¡1
) £ p(Mji (k ¡ 1) j Mli (k ¡ 2), Yk¡1 )
<n i=1 l=1
Z
£ p(xk j xk¡1 , c = i, Yk¡1 )p(xk¡1 , c = i j Y k¡1 )dxk¡1 :
£ p(Mli (k ¡ 2) j c = i, Yk¡1 )
¥i r(i) r(i)
X X
We implicitly assume that the class pdfs are restricted = pv (xk ¡ F i xk¡1 ¡ GMji ) Plji ¹il (k ¡ 1)
to the appropriate flight envelope sets and are zero j=1 l=1

1046 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
where pv (x) = Nfx; 0, Qg is the pdf of the Gaussian Remarks.
process noise and the posterior maneuver probability 1) The JTC filter recursions are initialized by
¢
¹im (k) = p(Mmi (k ¡ 1) j c = i, Yk ) can be expressed defining the initial state-class pdfs p(x0 , c = i),
recursively as i = 1, : : : , s and the initial class-conditioned mode
probabilities ¹ij (0), j = 1, : : : , r(i) for all classes i.
¹im (k) = p(Mmi (k ¡ 1) j c = i, Yk ) A practicable method consists of assuming that the
state-class pdfs are Gaussian for each class subject to
= p(Mmi (k ¡ 1) j c = i, Yk , Yk¡1 ) the restriction to the appropriate flight envelopes. The
1 initial state-class pdf must satisfy the normalization
= p(Y j c = i, Mmi (k ¡ 1), Yk¡1 ) requirement (23).
¢ k
2) The JTC filter pdf is in general not finite
£ p(Mmi (k ¡ 1) j c = i, Yk¡1 ) dimensional since at each iteration for each class a)
r(i) the state pdf is restricted to the flight envelope, and
¤im (k) X
= p(Mmi (k ¡ 1), Mni (k ¡ 2) j c = i, Yk¡1 ) b) there is a summation over terms for the various
¢ maneuvers. Thus, even if the initial state-class pdf
n=1
r(i) is a finite component mixture density, the number
¤i (k) X of components will increase at each iteration, but
= m p(Mmi (k ¡ 1) j Mni (k ¡ 2))
¢ moreover, the restriction to the flight envelope and the
n=1
subsequent renormalization over all classes (equation
£ p(Mni (k ¡ 2) j c = i, Yk¡1 ) (22)) will change the functional form of the mixture
Pr(i) components.
¤im (k) i i
n=1 Pnm ¹n (k ¡ 1)
= Pr(i) P r(i) 3) The JTC filter reduces to the optimal Bayesian
i i i
q=1 ¤q (k) n=1 Pnq ¹n (k ¡ 1) filter for a maneuvering target [15] when there is only
one target class with an unrestricted flight envelope
where ¢ is the normalization factor appearing in the (i.e., ¥1 = Rnx ).
last line and the likelihood ¤im (k), defined below, is 4) The JTC filter reduces to the DIF approach
evaluated as under the following conditions: a) the flight envelopes
¢ are unrestricted for all classes, b) the IMM filter is
¤im (k) = p(Yk j c = i, Mmi (k ¡ 1), Yk¡1 ) used to provide state estimates.
Z Since a closed form solution for the joint state-
= p(Yk j Mmi (k ¡ 1), xk , Yk¡1 ) class pdf will not be obtainable in general, it is
¥i necessary to resort to numerical procedures for
£ p(xk j c = i, Mmi (k ¡ 1), Yk¡1 )dxk representation of the pdfs and approximation of
Z the associated integrals at each time k. Possible
= p´ (yk ¡ i) p(ykx ¡ Hxk )
c approaches to the numerical evaluation of the integrals
¥i include trapezoidal integration, Romburg integration,
£ p(xk j c = i, Mmi (k ¡ 1), Yk¡1 )dxk : and adaptive recursive Simpson’s rule. Concerning
computational requirements, the entire state-class
The mode-conditioned state prediction pdf p(xk j c = i, pdf p(xk , c j Yk ) must be stored at each time. In
Mji (k ¡ 1), Yk¡1 ) in the likelihood calculation above is contrast to [6], it was found that the mode-conditioned
given by state pdfs p(xk j Mji (k), Yk ) are not required to be
stored. In addition the computationally and memory
p(xk j c = i, Mji (k ¡ 1), Yk¡1 ) intensive operation of integration of the product
Z of transition and prior pdfs must be performed
= p(xk j xk¡1 , Mji (k ¡ 1), Yk¡1 ) at each iteration. The computational and memory
¥i requirements rapidly increase with the dimension
£ p(xk¡1 j c = i, Mji (k ¡ 1), Yk¡1 )dxk¡1 of the target state. This is in contrast to a standard
Z IMM filter where only the means and covariances of
= pv (xk ¡ F i xk¡1 ¡ GMji ) the mode-conditioned state pdfs must be stored, and
¥i where Kalman filter updating of the conditional pdfs
is performed. The computation of the likelihood terms
£ p(xk¡1 j c = i, Yk¡1 )dxk¡1
Z is particularly critical since numerical integration
errors cannot be “normalized out.” As in the IMM
= pv (xk ¡ F i xk¡1 ¡ GMji )
¥i algorithm, the response of the mode probabilities
to observed target maneuvers is a good indicator
p(xk¡1 , c = i j Yk¡1 ) of the correct functioning of the algorithm (see
£R k¡1 )d»
dxk¡1 :
¥i p(»k¡1 , c = i j Y k¡1 Section V).

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1047

V. COMPARATIVE NUMERICAL SIMULATIONS

To demonstrate the operation of the JTC filter,

and compare its performance with the more standard
DIF approach (based on an IMM filter), we have
provided in this section a simple maneuvering target
example for a one-dimensional tracking problem
having a two-dimensional state (x-position and
x-velocity). The target is assumed to belong to one
of two classes representing either a highly agile
maneuvering target (e.g., a fighter aircraft) or a lower
speed target with limited maneuvering capability
(e.g., a commercial aircraft). The problem consists
of tracking the unknown target and identifying
the class to which it belongs using the kinematic
measurements received from a radar and the identity
measurements received from ESM. We assume that
Fig. 4. Position and velocity estimates for IMM and JTC
the ESM measurements are synchronized with the
algorithm from simulated maneuvering target data. In this
radar sampling rate correctly associated (by virtue simulation, the measurement noise standard deviation for position
of their bearing) with the radar measurements. was 50 m and a filter process noise scaling factor of 100 was
The zero-clutter, unity detection probability case used for the JTC filter. The IMM filter used a maneuver transition
is assumed so as to focus on the state and class parameter of 0.9 and process noise scaling of 1.
estimation performance of the two algorithms. The
incorporation of the identity measurements into the
JTC filtering framework is straightforward as long Measurements were generated assuming that the target
as multisensor multitarget association issues such as belongs to class 2 (i.e., the velocity is restricted to lie
ghosting and other ambiguities are ignored. in (150, 400) and the maneuvering inputs belong to
S 2 = f0, +5g, ¡5gg). A +5g maneuver was initiated
at the 6th sampling instant and terminated at the 9th
A. Maneuvering Target Models sampling instant. The true initial target state was
assumed to be
The prior information available about the two · ¸ · ¸
x1 8000m
classes is used to define the acceleration inputs to the = (34)
state dynamical equation and also to specify the flight x2 0 200m/s
envelope constraints. In this case the flight envelope and the process noise vk for data generation was
constrains only the target velocity. Both classes are assumed to be zero.
assumed to satisfy the following dynamical equation The radar position measurement at time k is given
where the target moves on the x1 -axis according to by
· ¸ · ¸· ¸ · ¸ · ¸ · ¸
x1 1 T x1 0 0 x1
= + uk + vk (33) yk = [1 0] + wk (35)
x2 k+1 0 1 x2 k 1 1 x2 k

where x1 and x2 are the x-position and x-velocity, where wk is zero mean white Gaussian measurement
respectively. The sampling time T was assumed to noise with standard deviation 50 m. The true target
be 1 s. The flight envelope constrains the x-velocity trajectory components are shown in Fig. 4 with
to lie in the interval (100, 250)m/s for class 1 and position-time in the upper plot and velocity-time in
(150, 400)m/s for class 2. It should be noted that the lower plot.
there is an overlap region in the velocity domain of
(150, 250)m/s. There is no constraint on the position B. Target Emitter and ESM Models
state variable for either class, although since a grid
approximation is used for numerical quadrature in For this simple scenario, only three possible
the JTC filter, the state-class pdf is represented on a emitter types were considered. A class 1 target was
finite number of position cells. The input acceleration assumed to possess emitters fE1 , E3 g while a class 2
uk for class i is assumed to be restricted to the set target was assumed to possess emitters fE2 , E3 g. The
of maneuvering inputs S i , i 2 f1, 2g. This a priori set of all emitters is therefore −E = fE1 , E2 , E3 g) as in
information is specified as follows. Section IIIB. ESM measurement sequence generation
S 1 = f0, +g, ¡gg is based on the given emitter usage Markov chains,
detection probabilities and confusion matrix. The
S 2 = f0, +5g, ¡5gg: usage process for each emitter is assumed to be

1048 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
independent with usage transition probability matrices numerical values used in the simulation presented here
given by were taken as
· ¸ · ¸ 2 3
0:7 0:3 0:8 0:2 0:9 0:05 0:05
Á1 = Á2 = 6 7
0:4 0:6 0:1 0:9 P i = 4 0:1 0:9 0 5 (36)
· ¸
0:8 0:2 0:1 0 0:9
Á3 =
0:3 0:7 for both classes, corresponding to a transition matrix
where state 1 is “emitter on” and state 2 if “emitter parameter of 0.9. Note that this stochastic maneuver
off.” All emitters were assumed to be off at the initial model does not coincide with the true target trajectory
time. which assumed a deterministic +5g target maneuver
Unlike the model described in Section IIIB and from time 6 to 9. The process noise scaling factor
used in [19] a simpler emitter confusion process was for the IMM filter was taken as q = 1 with the
assumed. In this model the confusion process for each measurement noise covariance matched to that of the
emitter was assumed to be independent of whether radar sensor.
or not other emitters were also detected. Thus with Focusing on the JTC filter implementation, the
3 emitters, a 6 free parameter model specifying the initial JTC pdfs for both classes were assumed
following quantities suffices: to be truncated Gaussian densities. Correct filter
initialization was assumed for both the IMM filter and
Pr(declare E1 j detect E1 ) = 0:8 the mean vector of the Gaussian used in constructing
the initial JTC pdf. The covariance of the Gaussian
Pr(declare E1 j detect E2 ) = 0:1
was based on the radar measurement covariance (as
Pr(declare E2 j detect E1 ) = 0:1 described for instance in [1]). These initial Gaussian
pdfs are then truncated so that the resulting density
Pr(declare E2 j detect E2 ) = 0:8
lies within the specified flight envelope for each class.
Pr(declare E3 j detect E1 ) = 0:1 Since the densities for the JTC filter are defined
numerically on a grid rather than as continuous
Pr(declare E3 j detect E2 ) = 0:1:
functions (characterized by statistics such as means
Note that in this model it is not possible to declare and covariances), the cell-size for the grid and the
the presence of all three emitters simultaneously, state space domain need to be chosen carefully. In
and given that an emitter in use has been detected, a this simulation we chose a state space domain large
single corresponding emitter type is always declared. enough to contain the significant support of the class
Thus the number of emitters declared is equal to the 2 pdf for the entire length 20 simulation. Selection
number of in-use emitters detected. Emitter detection of the grid cell-size is particularly important as it
probabilities of PD1 = 0:9, PD2 = 0:9, PD3 = 0:9 were affects the accuracy of approximation of the JTC
assumed for data generation. The ESM classifier pdf. The approach taken here is to implement a
for both the IMM-based DIF and for JTC assumed numerical Kalman filter for the radar measurements
unity emitter detection probabilities. This was done alone and to check the approximation accuracy of
to ensure that the generated ESM data did not the pdf against the analytical solution, known to be
exactly match the model parameters assumed in the Gaussian. For this scenario a cell-size of 50 m for
estimators. Target classes were assumed equiprobable position and 25 m/s for velocity (for both classes)
a priori for both methods. was found to be sufficient. The state space domain
Following is a typical length 20 ESM declaration chosen for position was taken as (7000 m, 14000 m)
sequence corresponding to a type 2 target as generated for both classes. Different domains for velocity were
using Matlab used to reflect the different flight envelopes of each
class, and these were matched to the flight envelopes
Yck = fØ, E3, E3, Ø, Ø, E2, Ø, Ø, Ø, Ø, E2, Ø, previously stated. Numerical integration for the
Ø, E2, Ø, E1E2, E2E3, E2, Ø, Øg: JTC filter was effected using the trapezoidal rule as
implemented in MatlabTM 5.2. State estimates for
the JTC filter were computed by taking the mean of
C. Numerical Filter Implementations the pdf for each class (corresponding to a minimum
mean-square error approach rather than a maximum
The DIF method uses an IMM filter to provide a posteriori approach). A process noise scaling factor
target state estimates, covariances, and maneuver of 100 was used in the JTC filter to provide a wide
(mode) probabilities. For both the IMM and JTC support for the product integrals. Simulations were
implementation it was assumed that the target performed for both algorithms with radar sensor
maneuvers for class i are governed by a Markov chain position measurement errors of 50 m. A further set
with known transition probability matrix P i . The of simulations was performed for the JTC filter using

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1049

a position measurement error of 150 m in order to
exhibit the shape of the pdfs obtained.

D. Simulation Results
Figs. 4—13 depict the results of the comparative
simulation scenario described in Sections VA to
VC. This one-dimensional scenario comprised a
single target moving at constant velocity from time
0 to 6 with a +5g maneuver from time 6 to 9, and
constant velocity from time 9 to 20. Radar and ESM
measurement sequences were generated for this
scenario, which assumed a class 2 target. Target truth
(solid line) for position and velocity versus time
are depicted in Fig 4 along with the filtered state
estimates for the IMM algorithm (dashed line) and
JTC algorithm (dash-dot line). Performance for the
Fig. 5. Evolution of class probabilities for IMM-based and
IMM algorithm was not greatly affected by choice of ESM-based classifiers, together with fused classification
the process noise scaling factor for values between probability for DIF and JTC methods.
0.1 and 100, and a value of 1 was fixed for Fig.
4. The maneuver transition matrix parameter value
was fixed at 0.9 as in Section VC, with acceptable
performance obtained over an approximate range of
0.8 - 0.95. The JTC filter assumed the same transition
matrix parameter as the IMM, but it was found that a
higher process noise value was required for acceptable
numerical integration performance. For this relatively
low measurement noise case, the filtered positions do
not differ appreciably from the target truth as shown
in Fig. 4. The x-velocity estimation performance
of the IMM filter showed a slightly more rapid
response to the maneuver onset and consequently
larger overshoot following the maneuver termination.
The JTC filter response was more cautious with less
overshoot despite the higher process noise value used.
Fig. 5 illustrates the evolution of target
classification probability for classes 1 and 2 for
the 1) JTC filter (closely spaced dots), 2) the DIF
Fig. 6. Maneuver (mode) probabilities for IMM and JTC
method (dashed line), 3) ESM-only classification
algorithms for class 1 maneuvering target model.
(dash-dot line), and 4) maneuver-only classification
from the IMM filter (dotted line). From these results
it can be seen that the maneuver-only classifier has decreased (e.g., around time 16). The JTC filter had
its highest class 2 probability over the interval in more erratic performance (suspected to be due to
which it detected the target maneuver (refer to mode numerical integration errors that adversely affect the
probability curve in Fig. 7); in the remainder of the mode likelihood calculations) but nonetheless achieved
time (i.e., during the constant velocity phase) the IMM faster classification than the DIF method, reaching
tended to favor class 1 as this is the model with the near-certainty at time 9 and remaining at a very high
lowest acceleration components. The ESM-based probability value. It should be noted that the DIF
classifier had good discriminative power as expected method in its current form cannot use the target flight
due to the close matching of ESM data generation envelope information, and this explains the better
and filtering parameters. The fusion of ESM and classification result of the JTC filter in this case.
radar identity information obtained via the DIF Figs. 6 and 7 show the maneuver (mode)
method clearly demonstrates the trade-off of fusing probabilities for both the IMM and JTC filters for
the more ambiguous radar-only identity with the class 1 and class 2, respectively. The IMM mode
more conclusive ESM identity information, attaining probabilities function as expected for class 1 and 2
a near-certainty value for class 2 at around time with no numerical problems encountered. The class
10, but dipping occasionally from this high value 1 mode probabilities for class 1 indicate that a +1g
of confidence when the IMM-based probability maneuver occurred over the interval 8 to 18. This

1050 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
 for the JTC filter in Fig. 7 are more acceptable,
although not as definitive as for the IMM, with a +5g
maneuver detected from time 9 to time 15. We stress
that the numerical integration errors due to the grid
representation of the pdfs are an important factor in
these results.
The remaining Figs. 8—13 depict various stages
of the joint state-class pdf for the JTC filter. Figs.
8—10 are for the JTC filter for a class 1 target and
show the posterior state pdf in position-velocity space
at times 3, 7, and 12, respectively. It can be seen
that the class 1 pdf has been truncated by the flight
envelope for class 1, with the majority of the pdf
having moved outside the flight envelope by time
12. The absence of a significant probability mass in
the state region the numerical integration routine to
perform poorly and this is believed to be responsible
Fig. 7. Maneuver (mode) probabilities for IMM and JTC for the failure to detect the target maneuver. Figs.
algorithms for class 2 maneuvering target model.
11—13 are for the JTC filter for a class 2 target and
the true multimodal structure of the state pdf is in
overestimate of the maneuver duration is due to the evidence in these pictures. Recall that with 3 possible
mismatching of the modeled acceleration (+1g instead maneuver modes (including constant velocity),
of +5g), with the IMM compensating for the deficit there are 3k possible maneuver sequences after k
in acceleration by extending the maneuver time. The time samples, thus the Bayesian state pdf contains
IMM performance when the class 2 model was used is contributions from an exponentially increasing number
much better, with the mode probabilities indicating a of maneuver sequence components as time increases.
+5g maneuver from time 8 to 12, i.e., and maneuver Notably, not all of the components result in significant
estimation delay of 2—3 samples. The JTC mode probability mass, nor do all components give rise
probabilities for class 1 (mismatched model) do to peaks that lie within the given flight envelope.
not indicate any mode switching. This may be due Components corresponding to sequences of +5g
to the limited flight envelope for the class 1 target maneuvers yield greater displacements in x-position
forcing the §5g maneuver-conditioned densities out than sequences with a majority of ¡5g maneuvers; the
of the restricted state domain, hence reducing their pdf is therefore somewhat skewed from left to right
contribution to the likelihood. The mode probabilities (see the contour plot of Fig. 11). By time 12 (Fig.

Fig. 8. Posterior pdf in position-velocity space for JTC filter assuming a class 1 target at time 3. The multimodal nature of the
state-pdf is clearly visible. In this simulation, the measurement noise standard deviation for position was 150 and a filter process noise
scaling factor of 100 was used. A class 2 target maneuver of +5g was effected from time 6 to 9.

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1051

Fig. 9. Posterior pdf in position-velocity space for the JTC class 1 filter at time 7. As expected for this class 2 target, a large portion
of the pdf lies outside the flight envelope for class 1. The pdf has also been displaced to the right compared with Fig. 11, indicative of
the target’s increasing x-position.

Fig. 10. Posterior pdf in position-velocity space for the JTC class 1 filter at time 12. Following the class 2 target maneuver from time
6—9, the pdf has almost completely moved outside the class 1 flight envelope.

13) the JTC pdf for class 2 shows a clear peak, which VI. DISCUSSION AND CONCLUSIONS
when correlated with the mode probability in Fig. 7
is seen to correspond with a +5g target maneuver. We have presented a comparison of two algorithms
During program execution it is easy to visualize for tracking and classification based on associated
the JTC algorithm steps corresponding to state pdf radar and ESM measurements. The ESM sensor
prediction followed by correction (updating) with the model, as reported in Caromicoli and Kurien [8],
radar measurements. The net effect of the correction models the emitter usage, detection, and confusion
step is to cull components from the prediction pdf processes in a stochastic manner. Target classification
that do not have significant alignment with the is based on both observed target maneuvers and
measurement pdf. on ESM emitter declarations. In the DIF approach,

1052 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
 Fig. 11. Posterior pdf in position-velocity space for the JTC class 2 filter at time 3. The non-Gaussian nature of the posterior pdf is
apparent. Modes corresponding to hypothetical sequences of +5g maneuvers resulting in velocities above 400 m/s have been truncated
by the class 2 flight envelope.

Fig. 12. Posterior pdf in position-velocity space for the JTC class 2 filter at time 7.

an IMM algorithm provides state estimation and A Bayesian recursion was developed for the JTC
maneuver detection and ESM and radar data are state-class pdf and numerical representation of this
processed independently and then the classification pdf was performed using a grid-based approximation
probabilities are fused. There is no feedback of with integration effected using a trapezoidal rule. This
classifier output to the tracking in the DIF approach. is a very computationally intensive approach both in
On the other hand, the JTC filter allows for the terms of memory and processing requirements and
interaction between classification and tracking. This is certain numerical problems remain to be resolved. In
effected by using flight envelopes for different target particular, the computation of JTC mode likelihoods
classes as state constraints and by employing, as in the was problematic due to inaccuracies in the numerical
DIF approach, different maneuver models for different integration of the pdfs that could not be “normalized
target classes. out.” Based on the somewhat preliminary results

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1053

 Fig. 13. Posterior pdf in position-velocity space for JTC class 2 filter at time 12. The JTC pdf indicates that the most probable
velocity is around 325 m/s.

obtained so far with the current implementation for the development of efficient “particle filtering”
of the JTC filter, it is fair to conclude that while methods for the representation and propagation of
there is a modest improvement in classification non-Gaussian pdfs such as those arising in the JTC
performance of the JTC filter over the DIF approach, problem.
we have not observed a significant improvement
in maneuver tracking performance of JTC over ACKNOWLEDGMENT
IMM-based methods. We point out however that
there are significant numerical problems in obtaining The authors are grateful to Prof. Rob Evans for
accurate evaluation of the likelihood integrals in helpful suggestions and to Mr. X. Wang for assistance
JTC which are central to the mode-probability and in computer simulations.
filtering calculations. In fact, the improvement in REFERENCES
classification performance for JTC can be attributed
[1] Bar-Shalom, Y., and Fortman, T. E. (1988)
to the inclusion of flight envelope information in
Tracking and Data Association.
the filter, whereas the standard IMM filter does Academic Press, 1988.
not allow for state-constraint information. From a [2] Bar-Shalom, Y., and Li, X.-R. (1993)
theoretical viewpoint, there is great potential in the Estimation and Tracking: Principles, Techniques, and
JTC approach however, since it can in principle Software.
Boston: Artech House, 1993.
capture the complexity of the multimodal posterior [3] Best, R. A. (1996)
Bayes pdf and this conclusion is unaltered in the Integrated tracking and guidance.
presence of false measurements (clutter). The success Ph.D. dissertation, School of Electronic and Electrical
of the JTC filter performance will depend to a Engineering, University of Birmingham, Oct. 1996.
large degree on the effectiveness of the numerical [4] Box, G. E. P., and Rabinowtiz, P. (1973)
Numerical Integration.
integration. Waltham, MA: Blaisdell, 1973.
Several schemes for improving the numerical [5] Bogler, P. L. (1987)
robustness of the JTC mode likelihood computations Tracking a maneuvering target using input estimation.
are currently under study and these include Gaussian IEEE Transactions on Aerospace and Electronics
Engineering, AES-23, (May 1987), 298—310.
and Gaussian mixture approximations of the prior
[6] Challa, S., and Pulford, G. W. (1999)
and predicted JTC densities. Mechanisms are also Joint target tracking and classification.
being sought for the inclusion of flight envelope In Proceedings of Radar’99 Conference, Brest, France,
information into standard tracking filters like IMM May 1999.
to improve target tracking performance through the [7] Challa, S., and Pulford, G. W. (1999)
Joint target tracking and classification of non-cooperative
feedback of more accurate target identity information. targets.
The field of Monte Carlo integration and resampling Presented at the American Control Conference, ACC’99,
techniques has been identified as a promising area San Diago, CA, June 1999.

1054 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001
 [8] Caromicoli, A., and Kurien, T. (1989) [14] Kastella, K., and Zatezalo, A. (1998)
Multitarget identification in airborne surveillance. Nonlinear filtering for tracking low elevation targets in
In Proceedings of SPIE on Aerospace Sensing. multipath.
Orlando, FL, Mar. 27-31, 1989, 1—16. SPIE, 3373, (Apr. 1998), 452—459.
[9] Chang, K. C., and Fung, R. (1997) [15] Kenefic, R. (1981)
Target identification with Bayesian networks in multiple Optimum tracking of maneuvering target in clutter.
hypothesis tracking system. IEEE Transactions on Automatic Control, AC-26, (June
Optics Engineering, 36, (Mar. 1997), 684—691. 1981), 750—753.
[10] Cutaia, N. J., and O’Sullivan, J. A. (1995) [16] Miller, M. I., Srivastava, A., and Grenander, U. (1995)
Identification of maneuvering aircraft using class Conditional-mean estimation via jump-diffusion processes
dependent kinematic models. in multiple target tracking/recognition.
Research monograph ESSRL-95-13, Electronic Signals IEEE Transactions on Signal Processing, 43, 11 (Nov.
and Systems Research Laboratory, Department of 1995).
Electrical Eng., Washington University, St. Louis, MO, [17] Stimson, G. W. (1998)
May 1995. Introduction to Airborne Radar.
[11] Cutaia, N. J. (1996) NJ: SciTech Publishing, 1998.
Performance of automatic target-recognition algorithms [18] Streetly, M. (Ed.) (1998)
using kinematic priors. Jane’s Radar and Electronic Warfare Systems (10th ed.).
Ph.D. dissertation, Washington University, St. Louis, MO, Surrey, UK: Jane’s Information Group.
May 1996. [19] Wang, X., Challa, S., and Pulford, G. W. (1999)
[12] Doucet, A. (1998) Target tracking and classification using radar and ESM
On sequential simulation-based methods for Bayesian sensors.
filtering. In Proceedings of International Aerosense Conference,
Technical report CUED/FINFENG/TR.310, Signal IAC’99, Adelaide, Scotland, 1999.
Processing Group, Department of Cambridge, Cambridge,
UK 1998.
[13] Jacobs, S. P., and O’Sullivan, J. A. (1997)
High resolution radar models for joint tracking and
recognition.
Presented at IEEE International Radar Conference,
Syracuse, NY, May 1997.

Subhash Challa received the B. Tech degree from Jawaharlal Nehru

Technological University, Hyderabad, India, in 1993, and the Ph.D. degree from
Queensland University of Technology (QUT), Brisbane, Australia, in 1998.
He was teaching in various capacities at QUT from 1994—1996 while studying
for the Ph.D. degree before taking up the appointment as a visiting fellow at
Harvard Robotics Lab, Harvard University, Boston, MA, and an internship at
Scientific Systems Company, Inc., Boston, in 1996. He is currently a senior
research fellow at the Cooperative Research Centre for Sensor Signal and
Information Processing (CSSIP) in the Department of Electrical and Electronics
Engineering, the University of Melbourne, Parkville, Australia. He has been
working in the areas of nonlinear filtering with applications to Bayesian target
tracking and identification since 1994. His current research interests include
particle filtering approaches to nonlinear filter design, situation/threat assessment,
sensor fusion with time-delayed and out-of-sequence measurements and data
incest problems in sensor networks.

Graham W. Pulford received the B.E. degree in electrical engineering and the
B.Sc. (Hons) degree in applied mathematics from the University of New South
Wales, Australia, in 1987. In 1992 he received his Ph.D. in telecommunications
(equalisation) from the Department of Systems Engineering at the Australian
National University.
He undertook post-doctoral studies in biological system identification in the
Department of Chemistry at the Australian National University during 1992.
In 1993 he joined the Department of Electrical and Electronic Engineering,
University of Melbourne, as a researcher in the Cooperative Research Centre
for Sensor Signal and Information Processing. He was responsible for several
commercial research contracts on radar tracker development in connection with
the Jindalee over-the-horizon radar including multipath data association and
Viterbi algorithm trackers, and holds a patent in this area.
During 2000 he was a consultant on the Jindalee Operation Radar Network
project in the areas of multitarget multipath track clustering and oblique ionogram
trace extraction. In 2001 Dr. Pulford joined the General Sonar Studies Group
at Thompson Marconi Sonar in Sydney, where he is currently engaged in the
development of target tracking and data fusion techniques for sonar and related
systems.

CHALLA & PULFORD: JOINT TARGET TRACKING AND CLASSIFICATION 1055