Context driven tracking using particle filters

This paper introduces a novel approach to robust tracking that combines Particle Filters (PFs) and estimation of physical constraints using Bayesian Networks (BNs). Heterogeneous Context Data (CD) describing the environment in which tracked objects move, is fused with the help of BNs. The resulting uncertain constraints are incorporated into the filtering process through a modification of the importance weights. Causal probabilistic models representing relations between the tracked objects and their environment are used to derive an updating rule that allows theoretically sound incorporation of uncertain constraints into PF. The approach allows incorporation of new types of CD without requiring any adaptation of the PF algorithm itself. The experimental results confirm that the presented method significantly improves the tracking accuracy in a relevant class of problems characterized by partial sensor coverage and low updating frequencies.