TRACKING POORLY MODELLED MOTION USING PARTICLE FILTERS WITH ITERATED LIKELIHOOD WEIGHTING

Abstract

Human motion in cluttered scenes is often tracked using particle filtering. However, poorly modelled inter-frame motion is not uncommon, resulting in poor priors for the filtering step. Alternatives to the Condensation algorithm in the form of an Auxiliary Particle Filter (APF) and Iterated Likelihood Weighting (ILW) are described. Experimental results comparing these filters' accuracy and consistency are presented for a scenario in which a person is tracked in an overhead view using an ellipse model with a likelihood based on colour and gradient cues. ILW is not intended to give unbiased estimates of a posterior but rather to reduce approximation error. It is shown to outperform both Condensation and the APF on sequences from this scenario.