Enhanced video surveillance using a multiple model particle filter

This paper describes a new visual target tracking algorithm which can be applied to intelligent video surveillance systems. We model the target under track as a nonlinear switching dynamic system, which is often referred as a jump Markov process. More specifically, we assume the target operates according to one dynamic model from a finite set of hypothetical models, known as regimes. The probability of switching from one model to another is governed by a predefined regime transition matrix. Then a particle filter is applied to each dynamic model to estimate the target location based on current measurement cues. The term particle filtering is a nickname given to the sequential Monte Carlo importance sampling technique for approximating a target distribution by a set of weighted samples. As shown from the experimental results, the multiple-model method is able to render a robust tracking of a target in the presence of strong background clutters compared to standard condensation method.