Integrated Bayesian multi-cue tracker for objects observed from moving cameras

This paper describes an approach to tracking multiple independently moving objects observed from moving cameras. The method addresses difficulties typically associated with tracking, including changes in background, parallax in the scene, arbitrary camera motion, object occlusions, cross-overs, and appearance changes. Using a bottom up approach, independently moving objects are detected in images acquired from a camera in free motion. These object detection results are then used in a top down particle filter framework to generate and evaluate object hypotheses. Integrating bottom up and top down approaches leads to more robust object detection, an improved object representation, and more effective generation and evaluation of target hypotheses. We demonstrate the effectiveness of the approach on real image sequences taken from hand-held cameras and from PETS 2005 dataset.