An Improved Particle Filter Approach for Real-time Pedestrian Tracking in Surveillance Video

This paper presents a method for pedestrian tracking in surveillance video, and the method is based on an improved particle filter. In our algorithm, the dynamics is modeled as a second-order autoregressive process. And for the observation model, color histogram features are used for likelihood measure. The proposed color histogram method is operated on a sub-region of the target region and we explore how the background subtraction process affects the color histogram model. We further adopt rectangle filters and pixel-difference cues in the observation model to overcome the limitation of individual cue. Experiments show that the method yields better tracking performance with the improved observation model.