Coping with uncertain dynamics in visual tracking : redundant state models and discrete search methods

Abstract

A model of the world dynamics is a vital part of any tracking algorithm. The observed world can exhibit multiple complex dynamics at different spatio-temporal scales. Faithfully modeling all motion constraints in a computationally efficient manner may be too complicated or completely impossible. Resorting to use of approximate motion models complicates tracking by making it less robust to unmodeled noise and increasing running times. We propose two complimentary approaches to tracking with approximate dynamic models in a probabilistic setting. The Redundant State Multi-Chain Model formalism described in the first part of the thesis allows combining multiple weak motion models, each representing a particular aspect of overall dynamic, in a cooperative manner to improve state estimates. This is applicable, in particular, to hierarchical machine vision systems that combine trackers at several spatio-temporal scales. In the second part of the dissertation, we propose supplementing exploration of the continuous likelihood surface with the discrete search in a fixed set of points distributed through the state space. We demonstrate the utility of these approaches on a range of machine vision problems: adaptive background subtraction, structure from motion estimation, and articulated body tracking.