See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3357875

Taxonomy of multiple target tracking methods

Article in IEE Proceedings - Radar Sonar and Navigation · November 2005

DOI: 10.1049/ip-rsn:20045064 · Source: IEEE Xplore

CITATIONS READS

89 2,195

1 author:

Graham W Pulford
109 PUBLICATIONS 697 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Graham W Pulford
letting you access and read them immediately. Retrieved on: 24 September 2016
Taxonomy of multiple target tracking methods

G.W. Pulford

Abstract: A concise summary of techniques for multiple target tracking is provided and their main
characterics assessed qualitatively. The techniques have been catergorised into more than
35 different algorithmic types. A comparison chart is provided that lists each algorithm and
categorises the processing scheme, data association mechanism, complexity scaling (with number
of targets and with state dimension), overall complexity and a subjective performance figure.
Although some recent filtering theory developments have been omitted, the survey should serve to
demonstrate the large variety of ‘classical’ estimation theoretic algorithms already available for the
design of multiple target tracking systems. A number of areas deserving of further study are
identified in the concluding remarks.

1 Introduction systems used approximations to the Kalman filter called

Alpha-Beta filters.
It has been over a quarter of a century since Bar-Shalom’s Today it is increasingly the case that MTT approaches are
survey of multiple target tracking (MTT) methods was being applied in non-aerospace arenas. Examples of this
published in IEEE Transactions on Automatic Control [1]. trend include image processing [6 – 10], oceanography
At that time there were about half a dozen MTT algorithms [11 – 13], autonomous vehicles and robotics [14 – 17],
in evidence in the public domain. The lower visibility but remote sensing [18, 19] and biomedical research [20 – 22].
more recent survey report by Dézert [2] also lists fewer than Whereas in some of these areas the application of MTT
10 algorithms. Since that time many new approaches to techniques is already advanced, in others the full power of
MTT and the inseparable problem of data association have MTT methods is yet to be realised. In image processing and
been developed and it is expedient to take stock of what has computer vision, the MTT problem goes by the name of
transpired in the interim. This is especially true when we motion correspondence. The 1993 article by Cox [23]
consider that quite recent developments such as finite set reviews a number of MTT methods such as multi-
statistics and particle filtering, which, although not expli- hypothesis tracking (MHT) and joint probabilistic data
citly covered in this taxonomy, are poised to offer the next association (JPDA) in this context. Cox also adds to the list
generation of advanced tracking methods. In view of this of areas in which data association arises, viz. psychology,
potential ‘generational’ change in MTT algorithms, we biological vision, molecular dynamics and particle physics.
A popular account of MTT due to Uhlmann [24], including
provide in this article a fairly complete survey of what we
an efficient gating strategy, appeared in 1992 in connection
call ‘classical’ methods for the MTT problem. These
with the Strategic Defense Initiative in the United States.
methods are based largely on probability, stochastic
In order to maximise accessibility of the material we have
processes and estimation theory, which, when combined
adopted a somewhat informal, tutorial style of presentation
with systems theory and combinatorial optimisation, lead to in this article. In Section 2 we discuss the major aspects of
a plethora of approaches that can seem somewhat daunting the multiple target tracking problem. We then provide in
to the uninitiated. Although the recent book by Blackman Section 3 a brief overview of some of classical algorithms
and Popoli [3] contains a detailed coverage of approaches put forward as solutions to the MTT data association
for the MTT data association problem, the presentation is problem. We mention some more recent approaches
quite specialised and connections between the approaches (particle filering, finite set statistics) in Section 4 and clarify
may not always be apparent. The approach taken in this the scope of the article. In Section 5 we discuss the area of
article should be easier to digest if only because of its performance assessment for MTT algorithms. This is
brevity. followed in Section 6 by a taxonomy of the more than 35
Automatic target tracking, which is founded on Kalman different MTT approaches that are now in existence. By way
filtering [4], had its genesis in the early 1970s and was of justification for the lack of mathematical detail, we point
driven primarily by aerospace applications such as radar, out that an exposition of all these techniques would be
sonar, guidance, navigation, and air traffic control (see ambitious in the confines of a journal article.
chapter 8 of [5] and associated references). Most early
2 The multiple target tracking problem
q IEE, 2005 The MTT problem as understood in this paper concerns the
IEE Proceedings online no. 20045064 single-sensor, automatic tracking of multiple, independent
doi: 10.1049/ip-rsn:20045064 point targets. This is distinct from the area of extended
Paper first received 10th August 2004 and in revised form 15th March 2005 object tracking that arises, for instance, in image processing.
The author is with QinetiQ, Bldg A9, Cody Technology Park, Farnborough, A point target is modelled as having neither physical length
Hampshire, GU14 0LX, UK nor resolvable features. It exists purely in a dynamical
E-mail: [email protected] state-space usually consisting of position, velocity
IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005 291
and acceleration. The basic tracking problem concerns the measurements YðkÞ ¼ fy1 ðkÞ; . . . ; yMk ðkÞg is received at
estimation of the target state given noisy measurements each time. The cumulative set of measurement data is
from a sensor whose field of view includes the target of denoted Y k ¼ fYð1Þ; . . . ; YðkÞg. It is now not possible to
interest. write down the posterior PDF of the state pðxðkÞjY k Þ without
Multiple target tracking techniques are fundamentally defining a set of measurement-to-target data association
different from single target tracking techniques. The hypotheses fli ðkÞg under each of which there applies a
difference lies in the state-space model used. In a single system of the form:
target tracking algorithm, the state of only one target is
modelled; detections from other targets are assumed to be xðk þ 1Þ ¼ FðkÞ xðkÞ þ vðkÞ ð2Þ
false alarms and problems result when tracking closely-

spaced or crossing targets. Multiple target tracking algor- HðkÞ xðkÞ þ wðkÞ for target measurements
ithms, however, take the existence of more than one target yðkÞ ¼
clutter otherwise;
into account simultaneously in their measurement
association processes. In theory, MTT algorithms are where it should be understood that the measurement matrix
capable of tracking closely-spaced and crossing targets. H(k) is modified to account for those targets that are actually
Usually, only one measurement is assumed to be produced detected under the association hypothesis. Note that the
by each target at a given time and the targets are assumed to association hypotheses effectively partition the data set by
have independent dynamics. While these assumptions can defining disjoint events in which there is no uncertainty as to
be relaxed, it comes at a high price in terms of the origin of the measurements. Assuming prior Gaussia-
computational requirements. nity, under each possible li ðkÞ the posterior PDF of the joint
The target motion is modelled as a stochastic, or in some target state is Gaussian. However the correct association
cases deterministic, process involving a vector of state hypothesis is unknown and must be accounted for in a
variables x(k) where k denotes the discrete time. For probabilistic manner, resulting in a (Gaussian) mixture PDF
example, in a passive sonar system the state might consist of of the form
frequency, frequency-rate, bearing and bearing-rate for each
X
Nk
target, whereas in a 3-D active radar system the state pðxðkÞjY k Þ ¼ bi ðkÞ NfxðkÞ; x^ i ðkjkÞ; Pi ðkjkÞg ð3Þ
variables could include the Cartesian (x,y,z) position and i¼1
velocity for each target. For a deterministic model, the
motion is determined completely once the initial state is where Nk is the number of terms in the mixture, bi ðkÞ ¼
specified, whereas for a stochastic model one seeks to Prðli ðkÞjY k Þ is the association probability under hypothesis
characterise the probability density function (PDF) of the i, and Nfx; m; Pg is a multivariate Gaussian PDF in
target state. The sensor measurement model describes how the variable x with mean vector m and covariance matrix
the measurements y(k) are coupled to the target state x(k) P. The terms x^ i ðkjkÞ and Pi ðkjkÞ are none other than the state
and also models errors due to measurement noise. In many estimate and error covariance from a Kalman filter for the
applications it is possible to write the system and system in (2). Unfortunately the hypothesis tree, represent-
measurement equations as a linear, discrete-time Gauss- ing all possible measurement-to-target data associations,
Markov system: undergoes continual branching, resulting in a number of
association hypotheses Nk that increases exponentially with
xðk þ 1Þ ¼ FðkÞ xðkÞ þ vðkÞ time. The optimal estimators for the MTT problem are
yðkÞ ¼ HðkÞ xðkÞ þ wðkÞ ð1Þ therefore not implementable and many suboptimal
approaches have been developed. We briefly discuss some
where v(k) and w(k) are iid zero-mean Gaussian random of these in the following Section.
processes with known covariance matrices Q(k) and R(k)
respectively. Ignoring data association for the time being, 3 Classical solution approaches
the estimation theoretic (filtering) problem is then to
compute the PDF of the state given a sequence of noisy Different tracking methods can be distinguished on the basis
measurements Y k ¼ fyð1Þ; . . . ; yðkÞg. When the initial of the estimation criterion used. For instance, in a Bayesian
mean and covariance of the state vector are specified, the approach, the posterior PDF of the state pðxðkÞjY k Þ is
solution is given by the recursive discrete-time Kalman filter sought, from which the conditional mean (CM) or minimum
(KF) [25]. In practice it is necessary to use a combination of mean-square-error (MMSE) estimate of the state may be
raw data and prior knowledge to establish the initial state obtained by taking the expectation EðxðkÞjY k Þ. In a
and covariance – a process called track initiation [26]. maximum a posteriori probability (MAP) approach, one
If system nonlinearities are present, then extended Kalman wishes to find the state that maximises the posterior
filtering or one of its variants may be applied. Finite- probability: argmax pðxðkÞjY k Þ. In problems where
dimensional nonlinear filters also exist for certain specific a deterministic target motion model has been used, it is
types of nonlinear systems [27]. often required to find the maximum likelihood
In real applications spurious measurements called false (ML) estimate of the initial target state by maximising
alarms or clutter are also generally present so that there is pðY k jxð0ÞÞ with respect to x(0).
ambiguity in the origin of the measurements; in other words, MTT methods also differ in their processing scheme. Two
it is not clear a priori which measurement corresponds to the major variants are recursive and batch methods. In a
target of interest. This problem is aggravated by the recursive estimator, the state at time k is estimated based on
presence of multiple targets and the non-unity detection data at time k and the previous state estimate from time
probability of real sensors. All of these factors conspire to k
1 (and usually its error covariance). A batch method
produce data association ambiguity, which is largely to is one that uses a given sequence of data Y K ¼ fyð1Þ; . . . ;
blame for the complexity of the MTT problem (see [26, 28]). yðKÞg from which a sequence of state estimates f^xð1jKÞ;
When there is data association ambiguity, the model in x^ ð2jKÞ; . . . ; x^ ðkjKÞ; . . . ; x^ ðKjKÞg is produced. Another
(1) must be modified. Instead of a single measurement variant is referred to as fixed-lag smoothing in which a
vector, a scan consisting of a variable number Mk of look-ahead is set of, say, L
1 samples and the current state
292 IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005
estimate x^ ðkjk þ L
1Þ is generated using the measurement particle filtering may however be able to remedy this
set fyð1Þ; . . . ; yðk þ L
1Þg. Broadly speaking, there is an situation.)
inverse relationship between decision delay (L
1 in this Recognising that a track is basically a sequence of
case) and the estimation error variance. In commonsense measurements, an alternative approach to solving the MTT
terms this says that the longer we are prepared to wait, the problem is to map the measurement set onto a trellis and to
better the estimates will be (on average). A further scheme is seek the optimal measurement association sequence. These
multi-pass batch in which the entire data set is reprocessed approaches may either be ML or MAP, depending on the
several times to refine the state estimates. Typically these assumed form of the cost function. Conceptually such
approaches are based on the expectation-maximisation approaches provide state estimates as a by-product of the
(EM) algorithm [29, 30] and a sequence of initial state data association process, although the state estimates are
estimates is required for initialisation. usually required to evaluate the cost function. Candidate
The basic solution approaches for the MTT problem have techniques include the Viterbi algorithm [38 –40], the EM
already been adequately described elsewhere [23, 26]. algorithm [41, 42], multiple scan assignment (also known as
The simplest of these is the nearest-neighbour Kalman filter matching or set partitioning) [43 – 48], and network
(NNKF). This technique allows only the measurement theoretic algorithms [49]. The article by Pattipati et al.
closest in statistical distance to the predicted track to be used [50] provides an in-depth survey of assignment techniques
to update the target state estimate. A variant of this approach for the MTT problem.
is the strongest neighbour filter in which the SNR is Although algorithms that seek to solve the data
taken into account to resolve the association ambiguity. association problem directly are essentially taking a ‘hard
Such approaches ignore the other measurements in the decision’ approach, it is possible to hedge bets, so to speak,
data set although they can be surprisingly effective in by retaining not only the optimal solution, but the N best
practice [31]. solutions [50]. This kind of approach (called a list Viterbi
A common suboptimal Bayesian approach is known as algorithm) is also used in telecommunications where the
probabilistic data association (PDA) [32]. In this approach, N-best surviving paths in the trellis are retained [51] so as to
all measurements that are close to the predicted track minimise the risk of an incorrect decoding decision. For
location are used in a Bayesian update of the form (3). The MTT the N-best assignment algorithm of [52] provides an
overall Gaussian mixture PDF is then condensed to a single efficient means to achieve this [6, 53, 54]. The N-best
Gaussian by retaining only the first and second moments assignment formalism may also be generalised to higher-
(i.e. the mean and covariance). Both the NNKF and PDA are dimensional assignment problems.
recursive single-target tracking algorithms. Multi-scan assignment algorithms for MTT, which are
equivalent to integer programming problems of the kind
The simplest ‘true MTT’ algorithm is the global nearest
originally posed by Morefield [55], are known to be NP-
neighbour algorithm, also known as the 2-D assignment
hard problems in combinatorial optimisation. This means
algorithm [33]. As the name suggests, an assignment
that the worst-case complexity of the problem is exponential
problem is set up that accounts for the distances between
in the size of the problem (i.e. the number of targets and
all measurement and all tracks. The single best solution to
measurements). The currently preferred solution is via
this assignment problem allows each track to be updated
Lagrangian relaxation [45, 56], wherein an S-dimensional
with at most one measurement via Kalman filtering.
assignment problem (with S . 2) is solved to desired
The preceding approaches consider data association accuracy by solving a sequence of lower-dimensional
decisions one scan at a time, summarising previous data (preferably 2-D) assignment problems. In each stage of
by a single set of track estimates and covariances. More the process, a constraint is omitted or ‘relaxed’, being
complex approaches are possible that consider data included as a term in the cost function of a dual problem via
association across multiple scans or allow limited branching a Lagrange multiplier. The constraint is then reintroduced in
in the hypothesis tree. The multi-hypothesis tracking another lower-dimensional problem (the primal problem).
approach is of the latter type. MHT, which is essentially a This iterative solution process works because the dual
MAP estimator, can be formulated in two principal ways problem lower bounds the optimal cost while the primal
referred to as measurement-oriented [34] and track-oriented problem upper bounds it, with the so-called duality gap
[35], depending on the mechanism used to generate the decreasing at each iteration (see [43] for more detail).
association hypotheses. Implementations of MHT also differ An alternative approach in [57] starts with the integer
in respect of the strategies used to compute hypothesis programming formulation and, by relaxing the constraints
probabilities, cluster tracks, merge track histories and prune for an integer solution, arrives at a standard linear
low probability branches from the hypothesis tree (see [3]). programming (LP) problem. This makes the problem
A recent article by Blackman [36] discusses these issues in a amenable to efficient polynomial time algorithms developed
tutorial context and points to some areas for further study. for linear programming, such as interior point methods [58].
Some approaches to the MTT problem utilise the ‘hard Interestingly, since the interior point solution of an LP may
association’ approach characterised by (3) in connection sometimes be fractional, rather than integral, a probabilistic
with the optimal Bayesian estimator. Other approaches seek updating approach may be used. The fractional solutions
to mitigate this complexity by applying a ‘soft association’ can also be harnessed for track initiation. A linear
or non-enumerative approach. The latter approach ignores programming approach to MHT has recently been presented
the constraint that a point target may only produce a single in [59].
measurement from a single sensor at a given time. The problems of automatic detection and tracking are
Alternatively, numerical approximation of the multi-target closely linked: the former is one of declaring the presence or
PDF (obtained via Bayes’ rule) may be attempted [37]. absence of a target signal given noisy observations; the
In the latter approach, the complexity of the data latter is that of estimating the parameters of the target (e.g.
association problem, which requires a mixture PDF with its location and speed) given noisy observations. When the
an ever-increasing number of terms, is replaced by an signal-to-noise ratio is low or the false alarm rate high, the
equally undesirable numerical approximation of a multi- boundary between detection and tracking can become
dimensional PDF of arbitrary form, (current research in blurred. This is because it may not be reasonable to declare
IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005 293
the presence of a target unless a number of detections from the approximation error on the dimension of the state vector
this target have been observed that correlate in space and [72]. So-called Rao-Blackwellized particle filters have
time; but a number of detections from the same target also recently been applied directly to the MTT problem by
constitutes a track. Moreover the manner of correlating representing some components of the posterior state PDF,
detections from one time to the next clearly depends on the for instance the means and covariances of the terms in the
location and motion of the target, which are generally mixture PDF, in a parametric form [73]. This reduces the
unknown. Conventional tracking approaches implicitly number of particles required to approximate the PDF in
assume that some form of thresholding has been applied addition to reducing the estimation variance.
to the raw sensor data in order to reduce the amount of The FISST approach is primarily for multi-sensor multi-
measurement data for processing. Approaches that do not target estimation problems, in which the number of targets is
apply thresholding are referred to as track-before-detect also a random variable. Mahler, in a formidable body of
(TBD). Such approaches tend to use discrete or quantised work, presents a new differential and integral calculus as a
state and measurement spaces teamed with either hidden foundation for this class of problems. In FISST, random
Markov models (HMMs) or multi-frame detectors (for 2-D variables are set-valued, that is, they are random sets.
image data), and their complexity=performance is therefore Conventional probability-mass functions are replaced by
determined by the number of discrete states they require. belief-mass functions, integrals are replaced by set-inte-
The HMM technique relies on a finite-state Markov chain grals, and so on. The methodology also encompasses
model of the target dynamics accompanied by a (memory- Dempster-Schafer evidential reasoning. The interested
less) probabilistic mapping that represents the observation reader should consult [74, 75] and references therein.
process of the sensor. Under this formalism the problem is At the time of writing FISST cannot be viewed as a practical
amenable to computation of the smoothed posterior state alternative for Bayesian MTT owing to difficulties with
PDF pðxðkÞjY K Þ via the Forward-Backward (FB) algorithm its implementation, which, even in the approximate
given a batch of (quantised) measurement data Y K [60]. The ‘probability hypothesis density’ (PHD) case, may require
sequence of discretised MAP state estimates techniques such as particle filtering.
Finally we mention a number of other omissions from our
x~ ðkjKÞ ¼ arg max pðxðkÞjY K Þ; k ¼ 1; . . . ; K ð4Þ review. Image processing approaches, e.g. Hough Trans-
can then be computed. The sequential MAP state estimator forms [76 –78]) and 3-D matched filters for multi-frame
detection [79 – 83], have not been included in the tables
f^xð1jKÞ; . . . ; x^ ðKjKÞg ¼ argmax pðxð1Þ; . . . ; xðKÞjY K Þ since they are designed for sequences of 2-D images and
are not easily generalised to multidimensional state=
ð5Þ measurement data. We have deliberately left out artificial
may also be computed by applying the Viterbi algorithm neural networks: although some work relevant to MTT has
(VA) [61]. Note that the sequence of individually most been performed in this area, it tends to replicate known
probable states in (4) is not the same as the single most structures like JPDA (see, e.g. [84]) or replace parts of
probable state sequence in (5). Although mathematically known adaptive filters [85]. An important area that is
feasible, the HMM method is not practicable in the full beyond the purview of this article is that of manoeuvring
MTT case except in low dimensional state-spaces (e.g. target tracking where the system and=or noise covariance
passive sonar). matrices in (1) can change with time according to an
unobservable stochastic process. A large class of algorithms
has been developed for this problem including multiple
4 Recent developments model filters and filter banks [26] (chapter 4), interacting
multiple model (IMM) [86] and generalised pseudo-
As alluded to previously, some recent theoretical develop- Bayesian (GPB) algorithms [87]. Surveys of techniques
ments have not been covered in our survey. The reader for the manoeuvring target tracking problem may be found
should however be aware of these areas at least in name and in [26, 88, 89].
we suggest some starting references. The areas in question
are random set theory=finite set statistics (FISST) [62], 5 Performance assessment
unscented Kalman filters (UKFs) [63, 64], statistical
bootstrapping [65] and particle filters (PFs) [66 –69]. The Whatever the approach used to solve the MTT problem, it is
last area in particular is experiencing very rapid growth at necessary to devise measures of estimation performance so
present, as evidenced by two recent books on the subject that different algorithms may be compared or benchmarked.
[70, 71]. It should be pointed out that the UKF and PF For problems involving only parameter estimation (maxi-
methods arose in the context of estimation for nonlinear mum likelihood techniques) an absolute limit exists called
systems and are thus not explicitly MTT methods, although the Cramér-Rao lower bound. This is a lower bound on the
they are highly relevant since they can be used to replace the estimation variance for any unbiased estimator and it
traditional extended Kalman filter, which is widely used in requires knowledge of the true target state. Extensions of
MTT approaches. this result are available for MAP estimators (see [90] or
Particle filtering is a random sampling approach, that is, it [42], chapter 4), although the problem of missed detections
is non-parametric. As such, a set of sample vectors called and data association can greatly complicate the analysis
particles is taken to represent the PDF. Estimation statistics [91]. The evaluation of CRLBs, when possible, is often
may then be generated on the basis of these samples. The laborious.
attraction of the approach is that the actual filtering Tracker performance analysis is still an open research
procedure is conceptually much simpler than in conven- area: very few analytical performance results are available
tional MTT algorithms and is unhampered by system for even basic algorithms like the nearest neighbour Kalman
nonlinearities. On the other hand, it should be stressed that filter. It is therefore commonplace to perform Monte Carlo
there are many unanswered questions of practical import- trials in order to accumulate track error data from which
ance concerning PFs, such as the number of particles performance metrics such as average errors and false track
required for a given level of accuracy and the dependence of probabilities can be computed. A promising method is the
294 IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005
tracker operating characteristic (TOC) [92 – 94], which is finite sensor resolution. We confess that we have sided with
the tracking analogy of the receiver operating characteristic the majority of researchers in not covering finite resolution
in detection theory [95]. The TOC plots the true and false in this literature review, despite its obvious practical
track probabilities for given input data statistics (detection importance. The interested reader should consult [3]
and false alarm probabilities), and can be visualised as a set (chapter 6).
of surfaces. All algorithms we include herein relate to the multi-target
Owing to the many different types of possible tracking tracking problem with unassociated or unlabelled data, i.e.
errors, one must be careful to define what is understood by measurements are not tagged in terms of their origin. This
‘tracker performance’ [31, 96]. In MTT studies, compara- excludes multiple target localisation and tracking methods
tive performance is typically gauged on the basis of state for which the data association problem does not arise since
estimation errors. For instance the average position and the individual targets signals are identified by their angles of
velocity errors or the normalised-squared innovations [26] arrival at a receiver array [98 – 102]. We now go on to
may be computed for a given scenario. This interpretation of describe the scheme we have adopted to organise and
performance is due in part to the classical control-theoretic tabulate the various approaches.
viewpoint that is the legacy of the Kalman filtering era. First the processing scheme is stated. This may be either
Unfortunately such an approach fails to characterise recursive (sequential), single- or multi-pass batch. In a
adequately an algorithm’s performance in solving the data similar manner to the fast Fourier transform, most batch
association problem in an MTT environment. Thus such algorithms can be implemented in a sliding window or
quantities as track initiation delay, overshoot, label swaps, fixed-lag format that may or may not entail overlap of data
misassociation errors, track loss and coalescence, etc., are segments.
generally not taken into account. It is also rare to see Next the data association rule for the algorithm is
algorithm robustness characterised, that is, the sensitivity of described. This is a very important factor in determining
an algorithm to modelling errors and mistuning. These the performance=complexity trade-off of the algorithm. The
considerations, while more difficult to analyse in a methods can be loosely categorised as either enumerative or
theoretical context, are often paramount in the design of a non-enumerative. Much of the complexity of the data
multi-target tracking system for a real application. association problem stems from the enumeration of
In practice algorithm performance against an agreed set measurement-to-target association hypotheses. Enumera-
of criteria for a given application can usually only be tive methods include the NNKF (nearest neighbour); PDA
assessed after extensive testing on both simulated and real (all-neighbours); 1:1 single and multiple scan assignment;
data. A comprehensive set of simulation scenarios might 1:1 association along each branch of a hypothesis tree; 1:1
include: crossing tracks, parallel tracks and manoeuvres in association along each path of a trellis. Non-enumerative
light=heavy clutter and at high=low probability of detection. (soft-association) algorithms, such as the SME [103],
Factors such as finite sensor resolution, field of view and PMHT [104], event-averaged mean field [105] and Markov
target detection probability modelling should also be taken random field [106] approaches, do not require explicit
into account at least in the data generation stage. enumeration of data association hypotheses; instead, the
Furthermore the dimension of the measurement data can computations occur in a higher dimensional state space
have a dramatic impact on tracker performance for the same (avoiding data association) or some kind of average is used
algorithm. It is easy to be fooled by simulations on to take all associations into account in a statistical sense
multidimensional measurement data in which the effective (this can be described as many-to-one association).
volumetric clutter density is very low. The approximate scaling of computational complexity
Apart from performance, other important criteria used in with the number of targets (T) and with the state dimension
the evaluation of MTT algorithms include the compu- (nx ) is subsequently tabulated. This is described in order of
tational complexity (as a function of the number of targets, increasing complexity as linear, polynomial, combinatorial
measurements, and state dimension), implementational and exponential. Some licence has been used since actual
complexity (data structures, hypothesis management), algorithmic complexity depends strongly on the implemen-
whether automatic track initiation and maintenance are a tation (e.g. gating and hypothesis management strategies).
part of the algorithm, and the decision delay (also referred to The Kalman filter has a generic complexity of Oðn3x Þ.
as processing lag). The decision delay is important since the The scaling for a single KF is therefore quadratic in the state
theoretically optimal approach is to base the state estimation dimension. Many algorithms employ multiple KFs and so
on as much data as possible, delaying the actual production this scaling will be multiplied by the number of filters
of tracks. This must be weighed against the practical or required, which is a function of the number of targets and
operational consequences of not displaying the filtered track measurements in general. The (average) number of
data in a timely manner. Some algorithms like the VA and measurements in a single scan is denoted by M. The batch
MHT may make retrospective changes to the track history in length or history is denoted by K. Some recursive algorithms
view of new data. We mention some further outstanding like PDA have a history of 0; others like GPB consider
issues to do with performance comparisons in the association hypotheses over several scans of data.
Conclusion. Following this a measure of overall computational
complexity has been given as reflected in Table 1.
6 Taxonomy of MTT methods The reader should be mindful that this simplistic linear
characterisation of complexity spans many orders of
In the following tables the open literature algorithms for the magnitude. Simple algorithms such as the NNKF are
single-sensor, multiple target tracking problem have been classed as ‘low’ complexity whereas algorithms such as
summarised (although many of the approaches are also
applicable to the multi-sensor case). This list covers the
major journals, in addition to a number of conference Table 1: Scale used for both algorithmic complexity and
performance
proceedings and reports. Our categorisation of MTT
algorithms is motivated by a book review by Daum [97], low low=medium medium medium=high high
which deals with numerous MTT approaches and focuses on
IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005 295
the optimal Bayesian filter are classed as ‘high’ complexity, detailed in the literature. Clearly, however, the higher
although the latter cannot be implemented since its complexity algorithms also tend to have higher memory
computational requirements increase with time. Even for requirements, as do multiple scan and batch algorithms.
the same algorithm, complexity may vary widely depending Some algorithms, such as hidden Markov models, use a
on the implementation details. discretised state space (with N states) and their complexity
Memory requirements are not addressed since these are cannot be directly compared with the majority of
highly implementation dependent and are not usually approaches (which are based on continuous state spaces).

Table 2: Single-target Bayesian tracking algorithms: nearest neighbour (no branching); all-neighbour (single scan
branching and merging); optimal (branching with no merging)

Algorithm Nearest neighbour Kalman filter Probabilistic data association Optimal Bayesian filter

Processing recursive recursive recursive

History (scans)=batch length 0 0 unlimited
Association rule nearest neighbour all neighbour all paths=all neighbours
Scaling for number of targets – – –
Scaling for state dimension polynomial polynomial polynomial
Computational complexity low low high
Performance low low NA
Comments single target single target single target optimal MMSE
estimator; infeasible
References [26, 109] [32, 26, 43] [26, 1, 110, 111]

Table 3: Single target tracking algorithms (continued): the GPB filter approximates the OBF by allowing limited
branching followed by merging. The TSF is a non-Bayesian, multi-target branching filter with likelihood-based pruning

Track splitting filter

Algorithm Generalised pseudo-Bayesian filter Track splitting filter (optimal) (approximate)

Processing recursive (fixed lag) recursive recursive

History (scans)=batch length K .0 unlimited limited
Association type all paths=all neighbours 1:1 per branch 1:1 per branch
Scaling for number of targets – exponential combinatorial
Scaling for state dimension polynomial polynomial polynomial
Computational complexity medium high medium
Performance medium NA medium
Comments single target, K scan back; single target assumed in single target likelihood
also called multiple frame likelihood calculations; with pruning
data association. infeasible
References [111] [26, 1] [112, 113]

Table 4: Also known as ‘global nearest neighbour’, 2-DA is the simplest ‘true MTT’ algorithm. Its extension to multiple
scans is an NP-hard integer programming problem

Nearest neighbour
Algorithm 2-D assignment Multiple scan assignment Integer programming

Processing recursive batch batch

History (scans)=batch length 0 K >0 K >0
Association type 1:1 1:1 per scan 1:1 per scan
Scaling for number of targets polynomial polynomial polynomial
Scaling for state dimension polynomial polynomial polynomial
Computational complexity low=medium medium=high medium=high
Performance low=medium medium =high medium=high
Comments Complexity of 2-D 3-D assignment solved as MTT problem
assignment is OðN 3 Þ, sequence of 2-D problems; reformulated as
N ¼ minfM; T g, T targets, NP-hard for K > 0. constrained integer
M measurements programming problem
with likelihood cost function
References [33, 3, 114] [43–45, 47, 48, 20, 50, 57, 115–117] [1, 43, 3, 55]

296 IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005
The complexity of HMM-based approaches is usually experience. In many cases, however, no comparative
written in terms of the number of discrete states. performance results are available. Even though a benchmark
Penultimately a performance figure is given for each problem has been proposed for manoeuvring target tracking
algorithm. In assessing performance, rated on a scale of [107, 108], there is no agreed MTT benchmark problem,
‘low’ to ‘high’ as in Table 1, a rather subjective assessment which makes claims of algorithm superiority at times
has been made based on (i) the available public domain debatable. To complicate matters, algorithm performance is
simulation results; (ii) theoretical considerations such as strongly application- and data-dependent, although there
MAP or ML ‘optimality’; and (iii) the author’s own seems to be consensus in the tracking community that

Table 5: JPDA is the MTT generalisation of PDA. Its variants use simplified formulae for the data association
probabilities

Algorithm Joint PDA Cheap JPDA Suboptimal=fast JPDA

Processing recursive recursive recursive

History (scans)=batch length 0 0 0
Association type all neighbour all neighbour all neighbour
Scaling for number of targets combinatorial linear polynomial
Scaling for state dimension polynomial polynomial polynomial
Computational complexity medium low=medium low=medium
Performance medium low=medium low=medium
Comments Complicated by association Greatly simplified association Simplified association
probability calculations probability calculations probability calculations
References [26, 43, 118–122] [123] [124, 125]

Table 6: Further variants of JPDA use, e.g. assignment algorithms, to obtain the single best or N-best joint association
hypotheses. The SME filter avoids data association by nonlinear filtering in a higher dimensional space

Symmetric measurement
Algorithm Nearest neighbour JPDA N-best JPDA equation (SME) filter

Processing recursive recursive recursive

History (scans)=batch length 0 0 0
Association type 1:1 all neighbour non-enumerative
Scaling for number of targets linear polynomial linear
Scaling for state dimension polynomial polynomial polynomial
Computational complexity low=medium medium medium
Performance low=medium low=medium Medium
Comments Better for M and T approximately Sequence of 2-D assignment EKF on symmetric
equal; single joint association problems to solve. combinations of
event selected. position-only measurements.
References [123, 31] [126, 127, 31] [128, 103]

Table 7: JPDA hypotheses can be extended across multiple scans. Smoothing (retrodiction) of state estimates may also
be introduced

Algorithm JPDA smoothing Multi-scan JPDA Bayesian MHT (GCDA_SMRN)

Processing recursive recursive (fixed lag) recursive (fixed lag)

History (scans)=batch length K >0 K >0 mþnþ1
Association type multiple scan, all multiple scan, all 1:1 per branch with PDA
neighbours neighbours combining
Scaling for number of targets combinatorial combinatorial combinatorial
Scaling for state dimension polynomial polynomial polynomial
Computational complexity medium=high medium=high medium=high
Performance medium medium medium
Comments State vector concatenation Association probabilities Global combining data association
and Kalman smoothing updated with K scans suspended m frames and
of data: multiple retrodicted n frames: multiple
observation combining tracks are combined.
References [129] [3, 57, 130] [3, 131, 132]

IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005 297
Table 8: Mixture reduction is a structured approach to reducing the number of components in a mixture PDF for single
or multiple target states. ML-PDA uses a PDA likelihood to set up an optimisation problem for batch estimation

Algorithm Clustering algorithm filter Multi-target mixture reduction ML-PDA

Processing recursive recursive multi-pass batch

History (scans)=batch length 0 0 K >0
Association type all neighbour all neighbour all neighbour (in cost function)
Scaling for number of targets –- combinatorial combinatorial
Scaling for state dimension polynomial polynomial exponential
Computational complexity medium medium=high medium=high
Performance medium NA NA
Comments single-target Gaussian Extension of Gaussian Requires solution of
mixture reduction algorithm. mixture reduction non-linear multi-dimensional
Successive combining until algorithm to multi-target optimisation problem
only B branches case (e.g. quasi-Newton)
References [3, 133, 134] [3, 135] [136, 137]

Table 9: MHT is a Bayesian MAP estimator with unbounded computational requirements. Various types of clustering,
hypothesis selection, pruning and merging of state estimates may be applied in an implementation

Algorithm MHT (optimal) MHT (approximate) N-best MHT

Processing recursive recursive (fixed lag) recursive

History (scans)=batch length unlimited K >0 K >0
Association type 1:1 per branch 1:1 per branch 1:1 per branch
Scaling for number of combinatorial combinatorial polynomial
targets
Scaling for state dimension polynomial polynomial polynomial
Computational complexity high medium=high medium
Performance NA medium=high medium=high
Comments Optimal branching MAP Clustering, history Sequence of 2-D
estimator is infeasible; combining, pruning assignment problems
Track initiation and to reduce complexity to solve at each step
termination included
References [34] [33, 138 –144, 35] [53, 6, 145–147, 54, 52]

Table 10: PMHT takes a soft association approach in the EM algorithm to produce smoothed state estimates. HMM
approaches use a discretised state space coupled with the Viterbi or FB algorithm to produce state estimates across a
batch

Algorithm PMHT Single-target HMM HMM (decoupled)

Processing multi-pass batch batch batch

History (scans)=batch length K >0 K >0 K >0
Association type non-enumerative many:1 NN or PDA NN or PDA
Scaling for number of targets linear – –-
Scaling for state dimension polynomial exponential, base N exponential, base n
Computational complexity medium OðN 2 K Þ
Oðn2D K Þ OðnDþ2 K Þ
Performance low=medium medium=high medium
Comments EM algorithm with Kalman Single target, N discrete Single target, n discrete states
smoother; approximate sum states; forward backward per dimension; D-dimensional
over all association sequences or Viterbi algorithm, state; Component transitions
both OðN 2 K Þ separable
References [148–153, 104] [154–164] [165–167]

the MHT and multi-scan assignment approaches provide The reader should bear in mind that some of the
the best currently available performance (note however that algorithms are only for single target tracking but can be
there are a large number of variants of both of these applied to the MTT problem by performing computations
methods). independently for each target; thus their complexity scaling
298 IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005
Table 11: HMM approaches may be applied to unthresholded data for low SNR single targets. The Bayes-Markov
approach is a discretised, recursive MAP estimator. [172] applies to 2-D image data only

Algorithm HMM-TBD Single target Bayes-Markov TBD Multi-target Bayes-Markov TBD

Processing batch recursive recursive

History (scans)=batch length K >0 0 0
Association type observation PDF observation PDF observation PDF
Scaling for number of – – linear
targets
Scaling for state dimension exponential exponential exponential
Computational complexity OðN 2 K Þ for N states O(N) for N states O(NT) for N states, T targets
Performance NA NA NA
Comments unthresholded data; designed unthresholded data; discrete unthresholded data
for very low SNR tracking; state. forward recursion of discrete state; non-crossing
Viterbi algorithm FB algorithm tracks assumption
References [168– 170] [160, 171, 172] [173, 174, 172]

Table 12: Multi-target HMM requires a Cartesian product state space that is computationally prohibitive. Simplifications
use a ‘mixed state’ representation that ignores target identity or a mutual exclusivity assumption

Algorithm Multi-target HMM (full state) Multi-target HMM (mixed state) Multi-target HMM-FB

Processing batch batch batch

History (scans)=batch length K >0 K >0 K >0
Association type NN or PDA NN or PDA all neighbour (PDA)
Scaling for number of exponential combinatorial linear
targets
Scaling for state dimension exponential exponential exponential
 T 
2T
Computational complexity OðN K Þ for N states, O ðNT ! Þ2 K OðN 2 K Þ for N states
T targets
Performance NA NA NA
Comments discrete state; infeasible discrete permutation state; discrete state; forward backward
for most applications requires a hypothesis testing algorithm; assumes separability
stage to separate mixed track of tracks in state space
References [175, 176] [177–179] [180, 181]

Table 13: VDA harnesses the Viterbi algorithm to perform data association on a trellis. The EM algorithm may also be
applied to estimate either the association sequence or the target state (parameters)

Expectation maximisation Maximum likelihood

Algorithm Viterbi data association data association data association

Processing recursive (fixed lag) multi-pass batch multi-pass batch

History (scans) = batch length K >0 K >0 K >0
Association type 1:1 per path 1:1 per path all neighbour
Scaling for number of targets – – combinatorial
Scaling for state dimension polynomial polynomial polynomial
Computational complexity low=medium medium medium
Performance medium NA medium =high
Comments single target; Viterbi Single target MAP estimator Maximum likelihood deterministic
algorithm on M-state for association sequence; estimator for target parameters;
trellis VA þ Kalman smoother per pass Based on EM algorithm
References [38, 39] [41, 42] [182]

is linear in the number of targets. For single An additional space has been provided for general
target algorithms, the performance figure relates only to comments as to the particulars of the algorithm in question.
single target tracking: these algorithms are not intended for Note that a dash in the table indicates that a given term does
closely-spaced multiple targets. not apply to a given algorithm. When there is insufficient
IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005 299
Table 14: A number of operations research algorithms have been applied to solve the MTT problem on a trellis

Algorithm Multi-target Viterbi Multi-target Viterbi network flow Viterbi MHT

Processing batch batch recursive (fixed lag)

History (scans) = batch length K >0 K >0 K >0
Association type 1:1 per path 1:1 per path 1:1 per path
Scaling for number of targets combinatorial polynomial combinatorial
Scaling for state dimension polynomial polynomial polynomial
Computational complexity medium=high medium medium =high
Performance NA NA NA
Comments non-intersecting tracks; non-intersecting tracks; list VA on trellis with
Markov assumption; Markov assumption; permutation states
VA þ assignment algorithm relaxation algorithm
References [40] [49] [183]

Table 15: Ideas from statistical thermodynamics have been applied to approximate the data association process. These
include mean field theory and MRFs. Direct updating of the multi-target state PDF via Bayes’ rule may be attempted
using numerical methods

Mean field theory

Algorithm (event averaged MLE) Markov random field Numerical Bayes

Processing recursive multi-pass batch recursive

History (scans) = batch length 0 K >0 0
Association type non-enumerative many:1 non-enumerative many:1 non-enumerative
(joint state space)
Scaling for number of targets polynomial linear exponential
Scaling for state dimension polynomial polynomial exponential
Computational complexity medium medium medium =high
Performance NA NA NA
Comments thermodynamic approximation; MRF model for association Grid-based method;
2nd order numerical optimisation probabilities; EM Requires numerical integration
required at each step algorithm used for Markov prediction
and Bayes update
References [105, 184] [106] [37, 185]

Table 16: Abbreviations used in the paper

2-DA 2-dimensional assignment MRF Markov random field

CRLB Cramér-Rao lower bound MTT Multiple target tracking
FB Forward Backward (algorithm) NA Not assessed
EKF Extended Kalman filter NN Nearest neighbour
EM Expectation-maximisation OBF Optimal Bayesian filter
HMM Hidden Markov model PDA Probabilistic data association
IMM Interacting multiple model PDF Probability density function
FISST Finite set statistics PMHT Probabilistic multi-hypothesis tracking
JPDA Joint probabilistic data association SME Symmetric measurement equation
ML Maximum likelihood SNR Signal-to-noise ratio
MAP Maximum a posteriori TBD Track-before-detect
MFT Mean field theory TOC Tracker operating characteristic
MHT Multiple hypothesis track{er=ing} TSF Track-splitting filter
MLE Maximum likelihood estimat{ion=or} UKF Unscented Kalman filter
MMSE Minimum mean-square error VA Viterbi algorithm

300 IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005
information in the literature on which to assess perform- 8 Acknowledgments
ance, the abbreviation ‘NA’ (not assessed) has been used
Table 2 – 15. Further abbreviations have been listed in This work was carried out in part under Thales Underwater
Table 16 at the end of the paper. Systems self-funded R&D contract number 2002-P-GSS-
06.1-R-RY.

7 Concluding remarks 9 References

Motivated by Daum’s 1996 book review [97], we 1 Bar-Shalom, Y.: ‘Tracking methods in a multitarget environment’, IEEE
Trans. Autom. Control, 1978, AC-23, (4), pp. 618–626
have provided a comprehensive survey of single and 2 Dézert, J.: ‘Poursuite multi-cibles mono-senseur: Analyse des principales
multiple target tracking methods in an estimation theoretic approches développées dans le domaine’ (ONERA, Châtillon, France,
1988
framework. This survey updates earlier reviews by Bar- 3 Blackman, S., and Popoli, R.: ‘Design and analysis of modern tracking
Shalom and Dézert. A taxonomy of the methods was systems’ (Artech House, 1999)
presented that comprised over 35 different algorithms, 4 Gelb, A. (Ed.): ‘Applied optimal estimation’ (MIT Press, 1974)
5 Skolnik, M.L. (Ed.): ‘Radar handbook’ (McGraw-Hill, 1990, 2nd edn.)
classified according to (i) the central algorithmic idea; (ii) 6 Cox, I.J., and Hingorani, S.L.: ‘An efficient implementation of Reid’s
the processing strategy; (iii) the history or batch length; (iv) multiple hypothesis tracking algorithm and its evaluation for the purpose
of visual tracking’, IEEE Trans. Pattern Anal. Mach. Intell, 1996, 18, (2),
the scaling as a function of the number of targets; (v) the pp. 138 –150
scaling as a function of the state dimension; and (vi) the 7 Cox, I.J., Rehg, J.M., and Hingorani, S.L.: ‘A Bayesian multiple
computational complexity. A subjective performance hypothesis approach to contour segmentation’, Int. J. Comput. Vis., 1993,
11, (1), pp. 5–24
measure was also provided. In some cases the theoretically 8 Lane, D.M., Chantler, M.J., and Dai, D.: ‘Robust tracking of multiple
optimal version of an algorithm was distinguished from its objects in sector-scan sonar image sequences using optical flow motion
estimation’, IEEE J. Ocean. Eng., 1998, 23, (1), pp. 31–46
practical implementation (as done in [97]). We reiterate that 9 MacCormick, J., and Blake, A.: ‘A probabilistic exclusion principle for
no coverage was given of finite sensor resolution, target tracking multiple Objects’. Proc. 7th Int. Conf. Computer Vision, 1999,
pp. 572 –587
manoeuvres or multi-sensor tracking. 10 Rasmussen, C., and Hager, G.D.: ‘Probabilistic data association
Even without considering new approaches like finite-set methods for tracking complex visual objects’, IEEE Trans. Pattern
statistics and particle filters, it is clear that the current range Anal. Mach. Intell., 2001, 23, (6), pp. 560–576
11 Moran, B.A., Leonard, J.J., and Chryssostomidis, C.: ‘Curved shape
of MTT algorithms is indeed large. It is also apparent that reconstruction using multiple hypothesis tracking’, IEEE J. Ocean.
more needs to be done to establish practically meaningful Eng., 1997, 22, (4), pp. 625–638
12 Kocak, D.M., da Vitoria Lobo, N., and Widder, E.A.: ‘Computer vision
performance benchmarks, using freely available measure- techniques for quantifying, tracking, and identifying bioluminescent
ment and truth data, so that the relative merits of the plankton’, IEEE J. Ocean. Eng., 1999, 24, (1), pp. 81–95
individual algorithms can be better understood. 13 Xie, Y.: ‘A range-dependent echo-association algorithm and its
application in split-beam sonar tracking of migratory salmon in the
In comparing MTT algorithms regard should be paid to Fraser River watershed’, IEEE J. Ocean. Eng., 2000, 25, (3),
the computational complexity of the approach (which is pp. 387–398
14 Cox, I.J., and Leonard, J.J.: ‘Probabilistic data association for dynamic
defined differently for continuous and discrete state spaces) world modeling: A multiple hypothesis approach’. Int. Conf. Advanced
and to whether the approach is a recursive filter as opposed Robotics, June 1991, vol. 2, pp. 1287– 1294
15 Jensfelt, P., and Kristensen, S.: ‘Active global localization for mobile
to a smoother=batch algorithm. There are many cases in robot using multiple hypothesis tracking’, IEEE Trans. Robot. Autom.,
which smoothers have been, in the author’s opinion, 2001, 17, (5), pp. 748 –760
unfairly compared with filters, when a more appropriate 16 Leonard, J.J., and Durrant-Whyte, H.F.: ‘Application of multi-target
tracking to sonar based mobile robot navigation’. Proc. 29th Conf. Dec.
yardstick might have been a smoother with associated Control., Hawaii, Dec. 1990, pp. 3118–3123
(labelled) data. Another potentially misleading statistic that 17 Leonard, J.J., Moran, B.A., Cox, I.J., and Miller, M.L.: ‘Underwater
sonar data fusion using an efficient multiple hypothesis
arises in the tracking literature is that of performance algorithm’. Proc. Int. Conf. Robotics and Automation, 1995, vol. 3,
metrics, such as true and false track probabilities: these pp. 2995–3002
depend strongly on track initiation and maintenance logic, 18 Nicoli, M., Rampa, V., and Spagnolini, U.: ‘Hidden Markov model for
multidimensional wavefront tracking’, IEEE Trans. Geosci. Remote
which is often defined differently by different authors. Sens., 2002, 40, (3), pp. 651–662
A further example is that of track loss where algorithms, 19 Spagnolini, U., and Rampa, V.: ‘Multitarget detection/tracking
for monostatic ground penetrating radar: application to pavement
such as PDA, that provide good track continuity for true profiling’, IEEE Trans. Geosci. Remote Sens., 1999, 37, (1),
tracks also tend to have high false track rates. Given the pp. 383 –394
conflicting nature of many tracking performance metrics, 20 Kirubarajan, T., Bar-Shalom, Y., and Pattipati, K.R.: ‘Multiassignment
for tracking a large number of overlapping objects’, IEEE Trans.
end-users stand to benefit from studies that highlight the Aerosp. Electron. Syst., 2001, 37, (1), pp. 2– 21
numerous trade-offs that are involved. 21 Goobic, A.P., Welser, M.E., Acton, S.T., and Ley, K.: ‘Biomedical
application of target tracking in clutter’ Dept. EE, Virginia Univ.
Whilst considerable effort has been expended by the Charlottesville, 2001
research community on approaches like JPDA, MHT and 22 Hammarberg, B., Forster, C., and Torebjörk, E.: ‘Parameter estimation
of human nerve C-fibers using matched filtering and multiple
HMMs for state estimation, relatively little attention has hypothesis tracking’, IEEE Trans. Biomed. Eng., 2002, 49, (4),
been paid to approaches that concentrate on the data pp. 329–336
association problem. Such techniques as Viterbi and EM 23 Cox, I.J.: ‘A review of statistical data association techniques for motion
correspondence’, Int. J. Comput. Vis., 1993, 10, (1), pp. 53–66
algorithms show potential that merits further study in this 24 Uhlmann, J.K.: ‘Algorithms for multiple target tracking’, Am. Sci, 1992,
context. In contrast to the large number of papers on 80, pp. 128–141
25 Anderson, B.D.O., and Moore, J.B.: ‘Optimal filtering’ (Prentice-Hall,
multiple-scan assignment algorithms, other techniques NJ, 1979)
from operations research and combinatorial optimisation 26 Bar-Shalom, Y., and Fortmann, T.E.: ‘Tracking and data Association’
(e.g. network theory, integer and linear programming) (Academic Press, 1988)
27 Daum, F.E.: ‘Exact finite-dimensional nonlinear filters’, IEEE Trans.
are somewhat underrepresented in the MTT literature. Autom. Control., 1986, 31, (7), pp. 616–622
Fundamental questions remain as to the effectiveness of 28 Sittler, R.: ‘An optimal data association problem in surveillance
theory’, IEEE Trans. Mil. Electron, 1964, MIL-8, pp. 125–139
so-called non-enumerative or ‘soft association’ approaches 29 Baum, L.E., Petrie, T., Soules, G., and Weiss, N.: ‘A maximization
(as used in PMHT, MFT and MRF models) when compared technique occurring in the statistical analysis probabilistic functions of
Markov chains’, Ann. Math. Stat., 1970, 41, pp. 164 –171
with conventional ‘hard association’ methods that use a 30 Titterington, D.M., Smith, A.F.M., and Makov, U.E.: ‘Statistical
single measurement per track assumption. analysis of finite mixture distributions’ (John Wiley, NY, 1985)

IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005 301
31 Pulford, G.W.: ‘Relative performance of single-scan algorithms for 65 Gordon, N.J., Salmond, D.J., and Smith, A.F.M.: ‘Novel approach to
passive sonar tracking’. Proc. Aust. Acoustical Soc. Ann. Conf, nonlinear/non-Gaussian Bayesian state estimation’, IEE Proc. F, Radar
Adelaide, Nov. 2002, pp. 212–221 Signal Process., 1993, 140, (2), pp. 107–113
32 Bar-Shalom, Y., and Tse, E.: ‘Tracking in a cluttered environment with 66 Avitzour, D.: ‘Stochastic simulation Bayesian approach to
probabilistic data Association’, Automatica, 1975, 11, pp. 451–460 multitarget tracking’, IEE Proc., Radar Sonar Navig., 1995, 142, (2),
33 Blackman, S.: ‘Multiple target tracking with radar applications’ (Artech pp. 41–44
House, MA, 1986) 67 Arulampalam, S., Maskell, S., Gordon, N., and Clapp, T.: ‘A tutorial on
34 Reid, D.B.: ‘An algorithm for tracking multiple targets’, IEEE Trans. particle filters for online non-linear/non-Gaussian Bayesian tracking’,
Autom. Control, 1979, pp. 843–854 IEEE Trans. Signal Process, 2002, 50, (2), pp. 174–188
35 Kurien, T.: ‘Issues in the design of practical multitarget tracking 68 Hue, C., Le Cadre, J.P., and Perez, P.: ‘Sequential Monte Carlo methods
Algorithms’, in Bar-Shalom, Y. (Ed.): ‘Multitarget-multisensor track- for multiple target tracking and data fusion’, IEEE Trans. Signal
ing: advanced applications’ (Artech House, 1990), pp. 43–83 Process, 2002, 50, (2), pp. 309–325
36 Blackman, S.: ‘Multiple hypothesis tracking for multiple target 69 Musick, S., Greenwald, J., Kreucher, C., and Kastella, K.: ‘Performance
tracking’, IEEE Trans. Aerosp. Electron. Syst. Mag. Tutor, 2004, 19, comparison of particle method and finite difference nonlinear filters for
(1), pp. 5–18 low SNR target tracking’. Proc. 4th Int. Conf. on Information Fusion,
37 Stone, L.D.: ‘A Bayesian approach to multiple target tracking’, in Hall, Montreal, 2001
D.L., and Llinas, J. (Eds.): ‘Handbook of multisensor data fusion’ (CRC 70 Doucet, A., de Freitas, N., and Gordon, N. (Eds.): ‘Sequential Monte
Press, 2001), pp. 101 –1031 Carlo methods in practice’ (Springer-Verlag, 2001)
38 La Scala, B., and Pulford, G.W.: ‘A Viterbi algorithm for data 71 Ristic, B., Arulampalam, S., and Gordon, N.: ‘Beyond the Kalman
association’. Proc. Int. Radar Symp Munich, Sept. 1998, vol. 3, filter: particle filters for tracking applications’ (Artech House, Boston,
pp. 1155–1164 2004)
39 Quach, T., and Farooq, M.: ‘Maximum likelihood track formation with 72 Daum, F.E., and Huang, J.: ‘Curse of dimensionality and particle
the Viterbi algorithm’. Proc. 33rd IEEE Conf. on Decision and Control, filters’. Proc. Aerospace Conf., Mar. 2003, 4, pp. 4-1979–4-1993
Orlando, FL, Dec. 1994, pp. 271–276 73 Särkkä, S., Vehtari, A., and Lampinen, J.: ‘Rao-Blackwellized Monte
40 Wolf, J.K., Viterbi, A.J., and Dixon, G.S.: ‘Finding the best set of K Carlo data association for multiple target tracking’. Proc. 7th Int. Conf.
paths through a trellis with application to multitarget tracking’, IEEE Info. Fusion, June 2004, pp. 583– 590
Trans. Aerosp. Electron. Syst., 1989, pp. 287 –296 74 Mahler, R.P.S.: ‘Multitarget Bayes filtering via first-order multitarget
41 Pulford, G.W., and Logothetis, A.: ‘An expectation-maximisation moments’, IEEE Trans. Aerosp. Electron. Syst., 2003, 39, (4),
tracker for multiple observations of a single target in clutter’. Proc. 36th pp. 1152–1178
Conf. Decision and Control, San Diego, 1997, pp. 4997–5003 75 Mahler, R.P.S.: ‘Statistics 101 for multisensor, multitarget data fusion’,
42 Bergman, N.: ‘Linköping Studies in Science & Technology Disser- IEEE Trans. Aerosp. Electron. Syst. Mag. Tutor., 2004, 19, (1),
tations No. 579’. 1999 pp. 53– 64
43 Bar-Shalom, Y., and Li, X.-R.: ‘Multitarget-multisensor tracking: 76 Carlson, B.D., Evans, E.D., and Wilson, S.L.: ‘Search radar detection
principles and techniques’ (YBS Publishing, 1995) and track with the Hough transform Part I: system concept’, IEEE
44 Kirubarajan, T., Bar-Shalom, Y., and Pattipati, K.R.: ‘Multiassignment Trans. Aerosp. Electron. Syst., 1994, 30, (1), pp. 102–108
for tracking a large number of overlapping objects’, in Bar-Shalom, Y., 77 Hu, Z., Leung, H., and Blanchette, M.: ‘Statistical performance analysis
and Blair, W.D. (Eds.): ‘Multitarget-multisensor tracking: applications of track initiation techniques’, IEEE Trans. Signal Process., 1997, 45,
and advances, volume 3’ (Artech House, 2000), pp. 199–231 (2), pp. 445–456
45 Pattipati, K.R., Deb, S., Bar-Shalom, Y., and Washburn, R.B.: ‘A new 78 Liou, R.-J., and Azimi-Sadjadi, M.R.: ‘Multiple target detection using
relaxation algorithm and passive sensor data association’, IEEE Trans. modified high order correlations’, IEEE Trans. Aerosp. Electron. Syst.,
Autom. Control., 1992, 37, (2), pp. 198 –213 1998, 34, (2), pp. 553–568
46 Pattipati, K.R., Deb, S., Bar-Shalom, Y., and Washburn, R.B.: ‘Passive 79 Im, H., and Kim, T.: ‘Optimization of multiframe target detection
multisensor data association using a new relaxation algorithm’, in schemes’, IEEE Trans. Aerosp. Electron. Syst., 1999, 35, (1),
Bar-Shalom, Y. (Eds.): ‘Multitarget-multisensor tracking: advanced pp. 176– 187
applications’ (Artech House, 1990) 80 Porat, B., and Friedlander, B.: ‘A frequency domain algorithm for
47 Poore, A.B.: ‘Multidimensional assignment formulation of data multiframe detection and estimation of dim targets’, IEEE Trans.
association problems arising from multitarget and multisensor Pattern Anal. Mach. Intell., 1990, 12, (4), pp. 398 –401
tracking’, Comput. Optim. Appl., 1994, 3, pp. 27–57 81 Reed, I.S., Gagliardi, R.M., and Stotts, L.B.: ‘Optical moving target
48 Deb, S., Yeddanapudi, M., Pattipati, K.R., and Bar-Shalom, Y.: detection with 3-D matched filtering’, IEEE Trans. Aerosp. Electron.
‘A generalized S-D assignment algorithm for multisensor-multitarget Syst., 1988, 24, (4), pp. 327 –336
state estimation’, IEEE Trans. Aerosp. Electron. Syst., 1997, 33, (2), 82 Reed, I., Gagliardi, R., and Stotts, L.: ‘A recursive moving-target-
pp. 523–538 indication algorithm for optical image sequences’, IEEE Trans. Aerosp.
49 Castanon, D.A.: ‘Efficient algorithms for finding the K best paths Electron. Syst., 1990, 26, pp. 434–440
through a trellis’, IEEE Trans. Aerosp. Electron. Syst., 1990, 26 83 Reed, I., Gagliardi, R., and Shao, H.M.: ‘Application of three-
pp. 405–409 dimensional filtering to moving target detection’, IEEE Trans. Aerosp.
50 Pattipati, K.R., Popp, R.L., and Kirubarajan, T.: ‘Survey of assignment Electron. Syst., 1983, 19, pp. 898–905
techniques for multitarget tracking’, in Bar- Shalom, Y., and Blair, 84 Silven, S.: ‘Neural approach to the assignment algorithm for
W.D. (Eds.): ‘Multitarget-multisensor tracking: applications and multiple-target tracking’, IEEE J. Ocean. Eng., 1992, 17, (4),
advances’ (Artech House, 2000) vol. 3, pp. 77–159 pp. 326–332
51 Seshadri, N., and Sundberg, C.E.W.: ‘List Viterbi decoding algorithms 85 Chin, L.: ‘Application of neural networks in target tracking data fusion’,
with applications’, IEEE Trans. Commun., 1994, 42, pp. 313–323 IEEE Trans. Aerosp. Electron. Syst., 1994, 30, (1), pp. 281– 287
52 Murty, K.G.: ‘An algorithm for ranking all the assignments in order of 86 Blom, H.P., and Bar-Shalom, Y.: ‘The interacting multiple model
increasing cost’, Oper. Res., 1968, 16, pp. 682–687 algorithm for systems with Markovian switching coefficients’, IEEE
53 Brogan, W.: ‘Algorithm for ranked assignments with applications to Trans. Autom. Control., 1988, 33, (8), pp. 780 –783
multiobject tracking’, J. Guid., 1989, 12, (3), pp. 357 –364 87 Bar-Shalom, Y., and Li, X.-R.: ‘Estimation and tracking: principles,
54 Miller, M.L., Stone, H.S., and Cox, I.J.: ‘Optimizing Murty’s ranked techniques, and software’ (Artech House, MA, 1993)
assignment method’, IEEE Trans. Aerosp. Electron. Syst., 1997, 33, (3), 88 Mazor, E., Averbuch, A., Bar-Shalom, Y., and Dayan, J.: ‘Interacting
pp. 851–861 multiple model methods in target tracking: a survey’, IEEE Trans.
55 Morefield, C.L.: ‘Application of 0-1 integer programming for multi- Aerosp. Electron. Syst., 1998, 34, (1), pp. 103–123
target tracking problems’, IEEE Trans. Autom. Control, 1977, 22, 89 Pulford, G.W., and La Scala, B.F.: ‘A survey of manoeuvring target
pp. 302–312 tracking methods and their applicability to over-the-horizon radar’,
56 Poore, A.B., and Rijavec, N.: ‘A Lagrangian relaxation algorithm for CSSIP Report No. 14/96 to High Frequency Radar Division’, July
multidimensional assignment problems arising from multitarget 1996, http://www.ee.mu.oz.au/research/cssip/publications/9798/
tracking’, SIAM J. Optim., 1993, 3, (3), pp. 544–563 manoeuvre_survey.pdf
57 Li, X., Luo, Z.Q., Wong, K.M., and Bosse, E.: ‘An interior point linear 90 Kerr, T.H.: ‘Status of CR-like lower bounds for nonlinear
programming approach to two-scan data association’, IEEE Trans. filtering’, IEEE Trans. Aerosp. Electron. Syst., 1989, 25, (5),
Aerosp. Electron. Syst., 1999, 35, (2), pp. 474–490 pp. 590–601
58 Karmarkar, N.K.: ‘A new polynomial time algorithm for linear 91 Farina, A., Ristic, B., and Timmoneri, L.: ‘Cramer-Rao bound for
programming’, Combinatorica, 1984, 4, pp. 163– 174 nonlinear filtering with Pd , 1 and its application to target tracking’,
59 Coraluppi, S., and Carthel, C.: ‘Multi-hypothesis sonar tracking’. Proc. IEEE Trans. Signal Process, 2002, 50, (8), pp. 1916–1924
7th Int. Conf. Info. Fusion, June 2004, pp. 33–40 92 Fortmann, T.E., Bar-Shalom, Y., Scheffe, M., and Gelfand, S.:
60 Rabiner, L.R., and Juang, B.H.: ‘An introduction to hidden Markov ‘Detection thresholds for tracking in clutter – a connection between
models’, IEEE ASSP Mag., 1986, 3, pp. 4–16 estimation and signal processing’, IEEE Trans. Autom. Control., 1985,
61 Forney, Jr, G.D.: ‘The Viterbi algorithm’, Proc. IEEE., 1973, 61, (3), AC-30, (3), pp. 221 –229
pp. 268–278 93 Bar-Shalom, Y., Campo, L.J., and Luh, P.B.: ‘Receiver operating
62 Mahler, R.: ‘Random set theory for target tracking and identification’, characteristic to system operating characteristic: evaluation of a track
in Hall, D.L., and Llinas, J., (Eds.): ‘Handbook of Multisensor data formation system’, IEEE Trans. Autom. Control, 1990, 35, (2),
Fusion’ (CRC Press, 2001), pp. 14-1–14-33 pp. 172– 179
63 Julier, S., and Uhlmann, J.K.: ‘Data Fusion in Nonlinear Systems’, 94 Pulford, G.W.: ‘Markov chain analysis of the sequential probability
in Hall, D.L., and Llinas, J. (Eds.): ‘Handbook of Multisensor data ratio test for automatic track maintenance’. Proc. 6th Conf. Info. Fusion,
fusion’ (CRC Press, 2001), pp. 13-1–13-22 Cairns, July 2003, pp. 1258–1265
64 Julier, S.J., and Uhlmann, J.K.: ‘Unscented filtering and nonlinear 95 Van Trees, H.L.: ‘Detection, estimation, and modulation theory. Part 1’
estimation’, Proc. IEEE, 2004, 92, (3), pp. 401 –422 (John Wiley, New York, 1968)

302 IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005
96 Manson, K., and O’Kane, P.A.: ‘Taxonomic performance evaluation for 125 Fisher, J., and Casasent, D.: ‘Fast JPDA multitarget tracking
multitarget tracking systems’, IEEE Trans. Aerosp. Electron. Syst., algorithm’, Appl. Opt., 1989, 28, pp. 371–376
1992, 28, (3), pp. 775– 787 126 Roecker, J.A.: ‘A class of near optimal JPDA algorithms’, IEEE
97 Daum, F.E.: ‘Review of “Multitarget-multisensor tracking: principles Trans. Aerosp. Electron. Syst., 1994, 30, (2), pp. 504 –510
and techniques”’, IEEE Aerosp. Electron. Syst. Syst. Mag., 1996, 127 Zhou, B., and Bose, N.K.: ‘Multitarget tracking in clutter: fast
pp. 41 –44 algorithms for data association’, IEEE Trans. Aerosp. Electron. Syst.,
98 Frenkel, L., and Feder, M.: ‘Recursive expectation-maximization (EM) 1993, 29, (2), pp. 352 –363
algorithms for time-varying parameters with applications to 128 Kamen, E.W., and Sastry, C.R.: ‘Multiple target tracking using
multiple target tracking’, IEEE Trans. Signal Process, 1999, 47, (2), products of position measurements’, IEEE Trans. Aerosp. Electron.
pp. 306 –320 Syst., 1993, 29, (2), pp. 476–493
99 Affes, S., Gazor, S., and Grenier, Y.: ‘An algorithm for multisource 129 Mahalanabis, A.K., Zhou, B., and Bose, N.K.: ‘Improved multi-target
beamforming and multitarget tracking’, IEEE Trans. Signal Process, tracking in clutter by PDA smoothing’, IEEE Trans. Aerosp. Electron.
1996, 44, (6), pp. 1512– 1522 Syst., 1990, 26, (1), pp. 113–121
100 Satish, A., and Kashyap, R.L.: ‘Multiple target tracking using 130 Roecker, J.A.: ‘Multiple scan joint probabilistic data
maximum likelihood principle’, IEEE Trans. Signal Process, 1995, association’, IEEE Trans. Aerosp. Electron. Syst., 1995, 31, (3),
43, (7) pp. 1677–1695 pp. 1204–1210
101 Rao, C.R., and Zhou, B.: ‘Tracking the direction of arrival of multiple 131 Drummond, O.E.: ‘Multiple target tracking with multiple
moving targets’, IEEE Trans. Signal Process, 1994, 42, (5), frame, probabilistic data association’, Proc. SPIE, 1993, 1954, pp.
pp. 1133–1144 394-408
102 Sword, C.K., Simaan, M., and Kamen, E.W.: ‘Multiple target angle 132 Koch, W.: ‘Experimental results on Bayesian MHT for maneuvering
tracking using sensor array outputs’, IEEE Trans. Aerosp. Electron. closely-spaced objects in a densely cluttered environment’. Proc.
Syst., 1990, 26, (2), pp. 367–372 Radar 97, Oct. 1997, pp. 729–733
103 Kamen, E.W.: ‘Multiple target tracking based on symmetric 133 Salmond, D.J.: ‘Mixture reduction algorithms for target tracking in
measurement equations’, IEEE Trans. Autom. Control, 1992, 37, clutter’, Proc. SPIE, 1990, 1305, pp. 434– 445
(3), pp. 371– 374 134 Williams, J.L., and Maybek, P.S.: ‘Cost-function-based Gaussian
104 Streit, R.L., and Luginbuhl, T.E.: ‘Probabilistic multi-hypothesis mixture reduction for target tracking’. Proc. 6th Conf. Info. Fusion,
tracking’, NUWC-NPT Technical Report 10,428, Naval Undersea Cairns, 2003, pp. 1047– 1054
Warfare Center, Newport, Rhode Island, Feb. 1995 135 Pao, L.Y.: ‘Multisensor multitarget mixture reduction algorithms for
105 Kastella, K.: ‘Event-averaged maximum likelihood estimation and tracking’, J. Guid. Control Dyn., 1994, 17, (6), pp. 1205– 1211
mean-field theory in multitarget tracking’, IEEE Trans. Autom. 136 Kirubarajan, T., and Bar-Shalom, Y.: ‘Target tracking using probabil-
Control, 1995, 40, (6), pp. 1070– 1074 istic data association-based techniques with applications to sonar, radar,
106 Molnar, K.J., and Modestino, J.W.: ‘Application of the EM algorithm and EO sensors’, in Hall, D.L., and Llinas, J. (Eds.): ‘Handbook of
for the multitarget/multisensor tracking problem’, IEEE Trans. Signal Multisensor Data Fusion’ (CRC Press, 2001), pp. 81–838
Process., 1998, 46, (1), pp. 115 –128 137 Shertukde, H.M., and Bar-Shalom, Y.: ‘Detection and estimation for
107 Blair, W.D., Watson, G.A., and Hoffman, S.A.: ‘Benchmark problem multiple targets with two omnidirectional sensors in the presence of
for beam pointing control of phased array radar against maneuvering false alarms’, IEEE Trans. Acoust. Speech Signal Process., 1990, 38,
targets’. Proc. Am. Control Conf., Baltimore, MD, 1994, vol. 2, (5), pp. 749 –763
pp. 2071–2075 138 Blackman, S., Dempster, R., and Broida, T.: ‘Multiple hypothesis
108 Watson, G.A., and Blair, W.D.: ‘Solution to second benchmark track confirmation for infrared surveillance systems’, IEEE Trans
problem for tracking manoeuvring targets in the presence of false Aerosp. Electron. Syst., 1993, 29, (3), pp. 810 –823
alarms and ECM’, Proc. SPIE, 1995, 2561, pp. 263 –274 139 Blostein, S.D., and Richardson, H.S.: ‘A sequential detection approach
109 Singer, R.A., Sea, R.G., and Housewright, K.B.: ‘New results in to target tracking’, IEEE Trans. Aerosp. Electron. Syst., 1994, 30, (1),
optimizing surveillance system tracking and data correlation perform- pp. 197–211
ance in dense multitarget environments’, IEEE Trans. Autom. 140 Chang, K.-C., Mori, S., and Chong, C.-Y.: ‘Performance evaluation of
Control., 1973, AC-18, (6), pp. 571 –582 a multiple-hypothesis multi-target tracking algorithm’. IEEE Conf.
110 Emre, E., and Seo, J.: ‘A unifying approach to multitarget tracking’, Dec. and Control, Hawaii, 1990, vol. TP-9, pp. 2258– 2263
IEEE Trans. Aerosp. Electron. Syst., 1989, 25, (4), pp. 520 –528 141 Koch, W., and Van Keuk, G.: ‘Multiple hypothesis track maintenance
111 Singer, R.A., Sea, R.G., and Housewright, K.B.: ‘Derivation and with possibly unresolved measurements’, IEEE Trans. Aerosp.
evaluation of improved tracking filters for use in dense multitarget Electron. Syst., 1997, 33, (3), pp. 883–892
environments’, IEEE Trans. Inf. Theory, 1974, IT-20, (4), 142 Quach, T., and Farooq, M.: ‘Application of the MHT to track
pp. 423 –432 formation of a single maneuvering acoustic source in clutter’. Proc.
112 Fraser, E.C., and Meier, L.: ‘Mathematical models and optimum 36th Midwestern Symp. on Circuits and Syst., 1993, vol. 1,
computation for computer-aided active sonar systems’, U.S. Navy pp. 684–687
Electronic Lab., SRI Final Report, San Diego, Mar. 1967 143 Torelli, R., Graziano, A., and Farina, A.: ‘IM3HT algorithm: a joint
113 Smith, P.L., and Buechler, G.: ‘A branching algorithm for discrimi- formulation of IMM and MHT for multitarget tracking’, Eur.
nating and tracking multiple objects’, IEEE Trans. Autom. Control, J. Control, 1999, 5, (1), pp. 46– 53
1975, AC-20, pp. 101–104 144 Van Keuk, G.: ‘Multiple hypothesis tracking with electronically
114 Drummond, O.E.: ‘Comparison of 2-D assignment algorithms for scanned radar’, IEEE Trans. Aerosp. Electron. Syst., 1995, 31, (3),
sparse, rectangular, floating point, cost matrices’, J. SDI Panels on pp. 916–927
Tracking, 1990, pp. 481–497, Iss. 4/1990 145 Cox, I.J., and Miller, M.L.: ‘On finding ranked assignments with
115 Chummun, M.R., Kirubarajan, T., Pattipati, K.R., and Bar-Shlom, Y.: application to multitarget tracking and motion correspondence’, IEEE
‘Fast data association using multidimensional assignment with Trans. Aerosp. Electron. Syst., 1995, 31, (1), pp. 486 –489
clustering’, IEEE Trans. Aerosp. Electron. Syst., 2001, 37, (3), 146 Cox, I.J., Miller, M.L., Danchick, R., and Newnam, G.E.:
pp. 898 –913 ‘A comparison of two algorithms for determining ranked assignments
116 Kirubarajan, T., Wang, H., Bar-Shalom, Y., and Pattipati, K.R.: with application to multitarget tracking and motion correspondence’,
‘Efficient multisensor fusion using multidimensional data association’, IEEE Trans. Aerosp. Electron. Syst., 1997, 33, (1), pp. 295 –301
IEEE Trans. Aerosp. Electron. Syst., 2001, 37, (2), pp. 386 –400 147 Danchick, R., and Newnam, G.E.: ‘A fast method for finding the exact
117 Popp, R.L., Pattipati, K.R., and Bar-Shalom, Y.: ‘Dynamically N-best hypotheses for multitarget tracking’, IEEE Trans. Aerosp.
adaptable M-best 2-D assignment algorithm and multilevel paralleli- Electron. Syst., 1993, 29, (2), pp. 555–560
zation’, IEEE Trans. Aerosp. Electron. Syst., 1999, 35, (4), 148 Gauvrit, H., Le Cadre, J.P., and Jauffret, C.: ‘A formulaton of
pp. 1145–1160 multitarget tracking as an incomplete data problem’, IEEE Trans.
118 Chang, K., and Bar-Shalom, Y.: ‘Joint probabilistic data association Aerosp. Electron. Syst., 1997, 33, (4), pp. 1242–1257
for multitarget tracking with possibly unresolved measurements 149 Malyutov, M., Nikiforov, A., and Protassov, R.: ‘Multitrajectory
and maneuvers’, IEEE Trans. Autom. Control, 1984, 29, (7), estimation in noise and clutter’. Proc. 4th Conf. Info. Fusion,
pp. 585 –594 Montréal, Aug. 2001, pp. 17–25
119 Colegrove, S.B.: ‘Multi-target tracking in a cluttered environment’. 150 Pulford, G.W.: ‘Is the probabilistic mulithypothesis tracking algorithm
Proc. ISSPA’87, Adelaide, Aug. 1987, pp. 307 –314 optimal?’, University of Melbourne, Dept. Electrical & Electronic
120 Dézert, J., and Bar-Shalom, Y.: ‘Joint probabilistic data association for Eng. Technical Note, March 1998
autonomous navigation’, IEEE Trans. Aerosp. Electron. Syst., 1993, 151 Streit, R.L., and Luginbuhl, T.E.: ‘Maximum likelihood method for
29, (4), pp. 1275–1285 probabilistic multi-hypothesis tracking’, Proc. SPIE, 1994, 2335, pp.
121 Fortmann, T.E., Bar-Shalom, Y., and Scheffe, M.: ‘Sonar tracking of 394 –405
multiple targets using joint probabilistic data association’, IEEE 152 Rago, C., Willett, P., and Streit, R.L.: ‘A comparison of the JPDAF
J. Ocean. Eng., 1983, OE-8, (3), pp. 173– 184 and PMHT tracking algorithms’. Proc. ICASSP, Detroit, May 1995,
122 Zhou, B., and Bose, N.K.: ‘An efficient algorithm for data association vol. 5, pp. 3571–3574
in multitarget tracking’, IEEE Trans. Aerosp. Electron. Syst., 1995, 31, 153 Willett, P., Ruan, Y., and Streit, R.: ‘PMHT: problems and some
(1), pp. 458– 468 solutions’, IEEE Trans. Aerosp. Electron. Syst., 2002, 38, (3),
123 Fitzgerald, R.J.: ‘Development of practical PDA logic for multitarget pp. 738–754
tracking by microprocessor’, in Bar-Shalom, Y. (Ed.): ‘Multitarget- 154 Arnold, J., Shaw, S., and Pasternack, H.: ‘Efficient target tracking
multisensor tracking: advanced applications’ (Artech House, 1990), using dynamic programming’, IEEE Trans. Aerosp. Electron. Syst.,
pp. 1 –23 1993, 29, pp. 44–56
124 Roecker, J.A., and Phillis, G.L.: ‘Suboptimal joint probabilistic data 155 Barrett, R.F., and Holdsworth, D.A.: ‘Frequency tracking using hidden
association’, IEEE Trans. Aerosp. Electron. Syst., 1993, 29, (2), Markov models with amplitude and phase information’, IEEE Trans.
pp. 510 –517 Signal Process., 1993, 41, (10), pp. 2965– 2976

IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005 303
156 Demirbaş, K.: ‘Manoeuvring-target tracking with the Viterbi algor- 171 Bethel, R.E., and Rahikka, R.G.: ‘An optimum first-order time delay
ithm in the presence of interference’, IEE Proc. F, Radar Signal tracker’, IEEE Trans Aerosp. Electron. Syst., 1987, 23, (6), pp. 718 –725
Process., 1989, 136, pp. 262 –268 172 Bruno, M.G.S., and Moura, J.M.F.: ‘Multiframe detector/tracker:
157 Martin, T., and Jauffret, C.: ‘Poursuite de cible manoeuvrante en optimal performance’, IEEE Trans. Aerosp. Electron. Syst., 2001, 37,
présence de fausses alarmes pour un système sonar actif’. Proc. 16th (3), pp. 925–945
GRETSI, Grenoble, France, 1997, pp. 801 –804 173 Bethel, R.E., and Rahikka, R.G.: ‘Multisignal time delay detection and
158 Martinerie, F., and Forster, P.: ‘Data association and tracking using tracking’, IEEE Trans Aerosp. Electron. Syst., 1992, 28, (3),
hidden markov models and dynamic programming’. IEEE Int. Conf. pp. 675 –696
Acoust. Speech Signal Process, 1992, II, pp. 449–452 174 Bethel, R.E., and Paras, G.J.: ‘A PDF multitarget tracker’, IEEE Trans
159 Martinerie, F.: ‘New data fusion and tracking approaches in multiple Aerosp. Electron. Syst., 1994, 30, (2), pp. 386–403
targets / distributed sensors network contexts’. IEEE Int. Conf. Acoust. 175 Pulford, G.W., and Evans, R.J.: ‘A survey of hidden Markov model
Speech Signal Process, 1993, I, pp. 249 –252 tracking with emphasis on OTHR - first report to High Frequency
160 Paris, S., Jauffret, C., and Goullet, G.: ‘Frequency line tracking in Radar Division’, CSSIP Report No. 7/95, Adelaide, May 1995, http://
passive sonar system’. Proc. GDR-PCR ISIS, Nov. 1998 www.ee.mu.oz.au/research/cssip/publications/9798/hmm_survey.pdf
161 Paris, S., and Jauffret, C.: ‘Frequency line tracking using HMM-based 176 Xie, X., and Evans, R.J.: ‘Frequency wavenumber tracking using
schemes’, IEEE Trans. Aerosp. Electron. Syst., 2003, 39, (2), hidden Markov models’, IEEE Trans. Signal Proc., 1993, 41, (3),
pp. 439–449 pp. 1391–1394
162 Sitbon, S.: ‘Comparative study between a new HMM tracker and 177 Chen, B., and Willett, P.: ‘Superimposed HMM transient detection via
conventional passive tracking algorithms’. Proc. Undersea Defence target tracking ideas’, IEEE Trans. Aerosp. Electron. Syst., 2001, 37,
Technology, 1994, pp. 441 –445 (3), pp. 946–956
163 Sitbon, S., and Passerieux, J.M.: ‘New efficient target tracking based 178 Xie, X., and Evans, R.J.: ‘Multiple target tracking and multiple
upon hidden markov model and probabilistic data association’. Proc. frequency line tracking using hidden Markov models’, IEEE Trans.
29th Asilomar Conf, Pacific Grove, USA, Nov. 1995, vol. 2, Signal Procss., 1991, 39, (12), pp. 2659–2676
pp. 849–854 179 Xie, X., and Evans, R.J.: ‘Multiple frequency line tracking with hidden
164 Streit, R.L., and Barrett, R.F.: ‘Frequency line tracking using hidden Markov models – further results’, IEEE Trans. Signal Process., 1993,
Markov models’, IEEE Trans. Acoust. Speech Signal Process, 1990, 41, (1), pp. 334 –343
38, pp. 586 –598 180 Paris, S., Jauffret, C., and Goullet, G.: ‘Extraction et détection
165 Logothetis, A., Evans, R.J., and Sciacca, L.J.: ‘Bearings-only tracking automatique de pistes fréquentielles en sonar passif’. Proc. 17th
using hidden Markov models’. Proc. 33rd Conf. Decision & Control, GRETSI, Vannes, France, Sept. 1999, pp. 535– 538
Orlando, FL, Dec. 1994, vol. 4, pp. 3301–3302 181 Paris, S., and Jauffret, C.: ‘A new tracker for multiple frequency
166 Van Cappel, D., and Alinat, P.: ‘Frequency line extractor using lines’. Proc. IEEE Aerospace Conf., Mar. 2001, vol. 4,
multiple hidden Markov models’. Proc. IEEE Ocean. Eng. Soc. pp. 4/1771–4/1782
OCEANS, Sept. 1998, vol. 3, pp. 1481–1485 182 Avitzour, D.: ‘A maximum likelihood approach to data association’,
167 Van Cappel, D., and Alinat, P.: ‘HMM-based frequency line extracting IEEE Trans. Aerosp. Electron. Syst., 1992, 28, (2), pp. 560 –565
in lofargrams’. Proc. IMDEX ASIA, 1999, pp. 81–86 183 Perry, R., Vaddiraju, A., and Buckley, K.: ‘Multitarget list Viterbi
168 Barniv, Y.: ‘Dynamic programming solution for detecting dim moving tracking algorithm’. Proc. 32nd Asilomar Conf. on Signals, Systems &
targets’, IEEE Trans. Aerosp. Electron. Syst. Aerosp. Electron. Syst., Computers, Nov. 1998, vol. 1, pp. 436–440
1985, AES-21, (1), pp. 144– 156 184 Kastella, K.: ‘A Maximum likelihood estimator for report-to-track
169 Barniv, Y., and Kella, O.: ‘Dynamic programming solution for association’, Proc. SPIE, 1993, 1954, pp. 386–393
detecting dim moving targets part II: analysis’, IEEE Trans. Aerosp. 185 Towers, J.J., and Chan, Y.T.: ‘Tracking with Doppler and bearing
Electron. Syst., 1987, 23, pp. 776–788 measurements using conditional PDF’, in Moura, and Lourtie (Eds.):
170 Barniv, Y.: ‘Dynamic programming solution for detecting dim moving ‘Acoustic Signal Processing for Ocean Exploration’ (Kluwer,
targets’, ‘Multitarget-multisensor tracking: advanced applications’ Boston, 1993), NATO ASI Series C: Math. and Phys. Sciences,
(Artech House, MA, 1990), pp. 85–154 pp. 301 –308

304 IEE Proc.-Radar Sonar Navig., Vol. 152, No. 5, October 2005