IEEE SIGNAL PROCESSING MAGAZINE [46] SEPTEMBER 2010
Digital Object Identifier 10.1109/MSP.2010.937395
1053-5888/10/$26.00©2010IEEE
Bayesian Tracking
for Video Analytics
[
An overview
]
[
Alessio Dore, Mauricio Soto, and Carlo S. Regazzoni
]
isual tracking represents the basic process-
ing step for most video analytics applica-
tions where the aim is to automatically
understand the actions occurring in a
monitored scene. Consequently, the perfor-
mances of these applications are significantly depen-
dent on the accuracy and robustness of the tracking
algorithm. Bayesian state estimation and probabilistic
graphical models (PGMs) have proved to be very pow-
erful and appropriate mathematical tools to efficiently
solve the inference problem of motion estimation by
combining object dynamics and observations. In this
article, the impact of these signal processing techniques
on the development of recent tracking algorithms is shown
and a categorization of the most common approaches is pro-
posed. This categorization intends to logically organize different
concepts related to Bayesian visual tracking to give a global over-
view to the reader. Finally, general considerations on the design of
visual trackers for video analytics systems are discussed, focusing
on the tradeoff that is usually performed between the accuracy of
the target motion assumptions and the robustness of the object
appearance representation.
INTRODUCTION
The basic tracking task consists in estimating the trajectory of a
moving object by consistently assigning a label over frames con-
sidering noisy measurements. Therefore, tracking can be consid-
ered as a filtering process that eliminates the noise from
measurements to derive its motion. In this context, enhanced
tracking algorithms have been investigated for many applications
such as human computer interaction, automated surveillance,
traffic monitoring, and automotive preventive safety. For example
in [1]–[3], complex industrial surveillance systems are presented
pointing out the paramount importance of reliable tracking accu-
racy to ensure acceptable performances in the monitoring task.
Recently, researchers focused on developing accurate trackers
able to obtain a rich object description with the aim of accomplish-
ing advanced scene interpretation tasks. This trend is driven by the
growing interest in developing video analysis systems for helping
and supporting humans in many assignments. To achieve this, a
versatile, flexible, and expressive framework is required where the
differences between algorithms can be easily understood.
Additionally, coherent and consistent paradigms should be
employed to obtain robust, reliable, and efficient algorithms. In
this regard, the application of Bayesian filtering algorithms to
PGMs can provide useful tools for visual tracking.
PGMs are able to provide an appropriate theoretical framework
where object dynamics and appearance can be combined and the
motion estimation problem can be efficiently solved. As a matter
of fact, PGMs can represent, learn, and compute complex probabil-
ity distributions by explicitly defining statistical dependencies
between the elements of the probability model. Then by combin-
ing graph theory and probability theory, PGMs aim to deal with
© BRAND X PICTURES
V