Abstract—The probability hypothesis density (PHD) filter is
a practical alternative to the optimal Bayesian multiple targets
filter based on random finite sets (RFS). It propagates the
posterior intensity of the random sets of targets. In this paper,
we apply the Gaussian Mixture (GM) PHD filter to track a
random number of moving targets in visual sequences. To
obtain the PHD of visual objects, we propose a method to
approximate the posterior intensity using the feature
measurement. Monte Carlo technology is adopted to obtain
the feature measurement random set by sample particles with
the integer label. And we adopt an adaptive weight to fuse the
color and edge features to improve the represent ability of
tracking targets. The experimental results have demonstrated
the effectiveness of our method.
Keywords: probability hypothesis density filter, feature
measurement, visual tracking,Monte Carlo technology
I. INTRODUCTION
isual objects tracking is an important task in
many applications, such as road traffic control,
medical image sequence analysis, video
compression, security, and surveillance systems.
The data association problem in multiple targets
tracking usually involves ensuring that the correct
measurement is given to each stochastic filter so that the
trajectories of each target can be accurately estimated.
This is referred to as measurement-to-track association.
The three classical approaches to this are the nearest
neighbor (NN) method [1], the Joint Probabilistic Data
Association Filter (JPDAF) [1], and the Multiple
Hypothesis Tracking filter (MHT) [2].
The NN filter simple takes the nearest validated
measurement to the predicted measurement to update
each of the target states. This can result in problems
since the nearest validated measurement may be the
same for two different targets. The JPDA method takes
into account the fact that a measurement may fall inside
the intersection of two or more validation gates of
several different targets and so could have originated
from any of these targets or form clutter. The ideal MHT
filter maintains probabilities of all possible associations
at each time step. This does not just consider the
probabilities from the previous time step, which allows
for backtracking and also track initiation. In practice, it is
not feasible to keep track of all possible hypotheses, as
the computational complexity grows exponentially [3].
An alternative solution to the multiple targets tracking
problem is to view the set of observations collectively,
and try to estimate the set of target states directly, where
the correct report-to-track association is considered
unobservable [4]. The disadvantage of this approach is
that the continuity of the individual target tracks is not
kept. One such method uses Finite Set Statistics for
multiple target tracking [5], with an approach analogous
to the recursion used in Bayesian filtering by
constructing multiple target posterior distributions.
This paper is organized as follows. Section 2 provides
an overview of probability hypothesis density filter and
the Gaussian Mixture implementation. Extracting feature
measurement set and selected feature are described in
section 3. Section 4 provides the experiment results and
performance evaluation. Finally, Section 5 concludes the
paper.
II. BACKGROUND
Multi-sensor multi-target tracking is a class of dynamic
state estimation problems in which the entity of interest is a
set that is random in the number of elements as well as the
values of individual elements [6]. Finite random sets are
therefore natural and intuitive representations of multi-target
states and multi-target measurements. The modelling of
multi-target dynamic using random sets naturally leads to
tracking algorithms which incorporate track initiation, a
procedure that has mostly been performed separately in
traditional tracking algorithms. Bayesian multi-target
tracking (in the random finite set framework) propagates the
multi-target posterior density recursively in time. This
involves the evaluation of multiple (set) integrals and the
computational intractability is more severe than its
single-target counterpart [7]. Detailed description on the
multi-target Bayes recursion is presented in literature [8].
A. PHD filter
The PHD filter is an approximation developed to
alleviate the computational intractability in the
multi-target Bayes filter. Instead of propagating the
multi-target posterior density in time, the PHD filter
propagates the posterior intensity, a first-order statistical
A novel probability hypothesis density filter for tracking visual
targets
Xiaofeng LU
1*
, Xinhong Hei
1
, Lei Wang
1
, Haiyan Jin
1
1
The School of Computer Science and Engineering
Xi’an University of Technology