Multi-object Tracking Using Particle Swarm Optimization on Target Interactions - Advances in Heuristic Signal Processing and Applications - page 63

states together [ 23 ], or inferring this joint data association problem by estimating all
possible associations between the targets and the observations [ 17 , 24 ]. In contrast
to the above mentioned approaches, in order to achieve multi-target tracking, the
multiple parallel filters where a single filter per target has its own state space were
proposed in [ 9 ]. However, when the interactions among the moving targets take
place, difficulties in maintaining the correct object identities might arise. Therefore,
modeling the interactions among targets and occlusion reasoning play an incredi-
bly important role in multi-target tracking. Khan et al. [ 17 ] use a Markov Random
Field (MRF) motion prior to modeling the interactions among targets. Andriyenko
et al. [ 2 ] propose a model for global occlusion reasoning. In an approach that is
based on particle swarm optimization [ 30 ], the object interactions are modeled as
species competition and repulsion. Particle Swarm Optimization (PSO) is a popu-
lation based stochastic optimization technique [ 16 ] which shares many similarities
with evolutionary computation techniques. It has been shown to perform well on
many nonlinear and multimodal optimization problems.
Visual object tracking is an important ingredient of any multi-object tracking
algorithm. Particle filters [ 13 ] are one of the most efficient techniques for object
tracking. They were successfully applied in many visual tracking applications [ 28 ],
including multi-object tracking [ 8 , 23 ]. The task of object tracking can be considered
as a numerical optimization problem, where a local optimization is used to track the
local mode of the similarity measure in a parameter space of translation, rotation,
and scale. In [ 29 ], it was shown that, in tasks consisting in tracking a face or a
human, a particle swarm optimization-based tracker outperforms a tracker built on
a particle filter in terms of accuracy.
Visual object tracking using particle swarm optimization has been an active re-
search area for several years [ 18 , 19 ]. Recently, particle swarm optimization was
proposed to achieve full body motion tracking [ 14 , 20 , 31 ]. The particle swarm
optimization, which is a population-based searching technique, has high search effi-
ciency by combining a local search (using self-experience) and a global one (using
neighbor experience). In particular, a few simple rules result in high effectiveness of
exploration of a high-dimensional search space. In contrast, in a particle filter, the
samples do not exchange information and do not communicate with each other, and
thus they have reduced capability of exploring huge search spaces.
In this work, we present a PSO based algorithm for multi-target tracking. At
the beginning of each frame, the targets are tracked individually using highly dis-
criminative appearance models among different targets. Each of them is tracked on
the basis of separate particle swarm optimizations. The target locations and veloci-
ties that are determined by independent trackers are further employed in a particle
swarm optimization based algorithm which refines the trajectories extracted in the
first phase. Afterwards, a conjugate method is used in the final optimization. At this
stage, we utilize a complex energy function which represents the presence, move-
ment, and interaction of all targets within a temporal window consisting of the recent
frames.