Rao-Blackwellised Particle Filter for Tracking with Application in Visual Surveillance

Abstract

Particle filters have become popular tools for visual tracking since they do not require the modeling system to be Gaussian and linear. However, when applied to a high dimensional state-space, particle filters can be inefficient because a prohibitively large number of samples may be required in order to approximate the underlying density functions with desired accuracy. In this paper, by proposing a tracking algorithm based on Rao-Blackwellised particle filter (RBPF), we show how to exploit the analytical relationship between state variables to improve the efficiency and accuracy of a regular particle filter. Essentially, we estimate some of the state variables as in a regular particle filter, and the distributions of the remaining variables are updated analytically using an exact filter (Kalman filter in this paper). We discuss how the proposed method can be applied to facilitate the visual tracking task in typical surveillance applications. Experiments using both simulated data and real video sequences show that the proposed method results in more accurate and more efficient tracking than a regular particle filter.