ACCEPTED FOR PUBLICATION IN IEEE-TAES. ©IEEE. 1
Relationship between Finite Set Statistics and the
Multiple Hypothesis Tracker
Edmund Brekke, Member, IEEE, and Mandar Chitre, Senior Member, IEEE
Abstract—The multiple hypothesis tracker (MHT) and finite
set statistics (FISST) are two approaches to multitarget tracking
which both have been heralded as optimal. In this paper we
show that the multitarget Bayes filter with basis in FISST
can be expressed in terms the MHT formalism, consisting
of association hypotheses with corresponding probabilities and
hypothesis-conditional densities of the targets. Furthermore, we
show that the resulting MHT-like method under appropriate
assumptions (Poisson clutter and birth models, no target-death,
linear-Gaussian Markov target kinematics) only differs from
Reid’s MHT with regard to the birth process.
Index Terms—Multitarget tracking, Data association, MHT,
FISST, Random Finite Sets
I. I NTRODUCTION
T
He standard multitarget tracking problem can be decom-
posed into the two subproblems of filtering and data
association. Filtering concerns the estimation of a target’s
kinematic state from a time series of measurements. Data
association concerns how one determines which measurements
originate from which targets, and which measurements are to
be considered useless clutter. Many of the tasks carried out in
a surveillance system are most appropriately understood in the
context of data association. For example, establishing tracks
on new targets is fundamentally a data association problem,
since a decision to establish a new track amounts to a decision
regarding the origin of a sequence of measurements.
A Bayesian formalism is natural in target tracking since
prior knowledge about target kinematics is easily quantified
by a prior distribution. For filtering without measurement
origin uncertainty one can evaluate the posterior probability
density function (pdf) of the state given the sequence of
measurements so far observed. If the filtering problem is
Gaussian as well as linear, then the posterior is a Gaussian,
whose sufficient statistics are found by the Kalman filter (KF)
[1]. For non-linear and non-Gaussian problems, no closed-
form representation of the posterior can be found in general,
but the posterior can still be approximated by techniques such
as sequential Monte-Carlo methods (SMC) [2].
When data association is involved things get more com-
plicated, and it is not immediately clear that a well-defined
posterior exists. Consequently, the concept of optimality be-
comes problematic for data association problems. Truly Bayes-
optimal solutions, i.e., solutions which minimize a posteriori
E. Brekke is with the Department of Engineering Cybernetics, Norwegian
University of Science and Technology, O. S. Bragstads plass 2D, 7034
Trondheim, Norway. e-mail: edmund.brekke@ntnu.no.
M. Chitre is with the Acoustic Research Lab, Tropical Marine Science
Institute, National University of Singapore, 14 Kent Ridge Road, Singapore
119222, e-mail: mandar@arl.nus.edu.sg.
Manuscript received October 21, 2016; revised December 15, 2017.
expected loss, can only be achieved if the posterior is available.
From an intuitive perspective, it seems fairly obvious that
Bayes-optimal solutions to the standard multitarget tracking
problem must, in one way or another, consider all feasible
data association hypotheses.
A milestone was reached in 1979 when Donald Reid
proposed the multiple hypothesis tracker (MHT) [3], which
evaluates posterior probabilities of all feasible data association
hypotheses. Several variants of the MHT exist. In this paper
we are solely concerned with Reid’s original algorithm. In
contrast to many of the later developments, it has two attributes
which are very important here. First, Reid’s MHT is a recursive
method, which updates hypothesis probabilities according to a
recursive formula. Second, although one typically would want
to find the hypothesis with the highest posterior probability,
maximum a posteriori (MAP) estimation is never mentioned
in [3]. The output of Reid’s MHT as described in [3] is
simply a collection of possible hypotheses, with corresponding
probabilities. In practice, implementations of MHT-methods
rely on pruning, simplifications and sliding window techniques
to mitigate its exponential complexity. While a practical MHT
therefore will be suboptimal, it is commonly believed that
the ideal MHT without approximations is optimal in some
unspecified sense.
A different approach known as finite set statistics (FISST)
was developed by Ronald Mahler [4, 5, 6] in the 2000’s. FISST
is a reformulation of point process theory tailored to multi-
target tracking [7]. In FISST, both targets and measurements
are generally treated as random finite sets, i.e., as set-valued
random variables. This allows one to express a Bayes-optimal
solution to the full tracking problem using a single prediction
equation and a single update equation. This optimal recursion,
commonly known as the multitarget Bayes filter, is just as
intractable as the MHT, but Mahler and coworkers have devel-
oped several approximative solutions such as the probability
hypothesis density (PHD) filter, the cardinalized probability
hypothesis density (CPHD) filter and the multitarget multi-
Bernoulli (MeMBer) filter [6, 8].
The relationship between FISST and previously established
tracking methods is a controversial topic which has been
explored in several papers, although never fully resolved. The
integrated probabilistic data association (IPDA) and the joint
IPDA (JIPDA) have been linked to the FISST formalism in [9]
and [10]. The Set JPDA employs concepts based on FISST
to improve the JPDAF [11]. It was shown in [12] that the
CPHD filter is equivalent to MHT when maximally one target
is present. Several classical tracking methods were discussed
from the perspective of point process theory in [13]. The
relationship between MHT and FISST has been explored in