Evaluation of a Particle Filter to Track People for Visual Surveillance

Previously a particle filter has been proposed to detect colour objects in video [1]. In this work, the particle filter is adapted to track people in surveillance video. Detection is based on automated background modelling rather than a manually-generated object colour model. A labelling method is proposed that tracks objects through the scene rather than detecting them. A methodical comparison between the new method and two other multi-object trackers is presented on the PETS 2004 benchmark data set. The particle filter gives significantly fewer false alarms due to explicit modelling of the object birth and death processes, while maintaining a good detection rate.