Variational Particle Filter for Multi-Object Tracking

Abstract

The paper proposes an edge-based multi-object tracking framework which deals with tracking multiple objects with occlusions using a variational particle filter. Object is modelled by a mixture of a non-parametric contour model and a non-parametric edge model using kernel density estimation. Visual tracking with a mixture model is formulated as a Bayesian incomplete data problem, where measurements in an image are associated with a generative model which is a mixture of mixture models including object models and a clutter model and unobservable associations of measurements to densities in the generative model are regarded as missing data. A likelihood for tracking multiple objects jointly with an exclusion principle is presented, where it is assumed that one measurement can only be generated from one density and one density can generate multiple measurements and it significantly reduces the complexity of enumerating all feasible events. To address the curse of dimensionality in tracking multiple objects jointly, a variational particle filter (VPF) is proposed for multi-object tracking, where the proposal distribution is based on the approximated posterior from variational inference rather than using the prior as the proposal distribution in sampling importance resampling (SIR) particle filter. With the variational particle filter, the number of particles needed for multi-object tracking can be significantly reduced. Experimental results in challenging sequences demonstrate the robust performance of the proposed method.