Abstract

Bayesian particle filters have become popular for tracking human motion in cluttered scenes. The most commonly used filters suffer from two drawbacks. First, the prior used for the filtering step is often poor due to relatively large, poorly modelled inter-frame motion. Second, the use of the prior as an importance function results in inefficient sampling of the posterior. The use of the auxiliary particle filter (APF) and the novel iterated likelihood weighting filter (ILW) are proposed here in order to help address these problems. Experimental results comparing the filters’ accuracy and consistency are presented for a scenario in which a person is tracked in an overhead view using an ellipse model. A likelihood model based on combined region (colour) and boundary (gradient) cues is motivated and used. The ILW filter is shown to outperform both Condensation and the APF on typical sequences from this scenario.