Tracking Everything Around Me

We propose to develop and conduct initial evaluations of real time computer vision algorithms for detecting and tracking people and other movers from robot mounted video cameras. The rapid sensor motion and depth of field motivates the use of direct shape and appearance-based methods for detection and tracking, instead of detecting pixel changes against a background model. The detection and tracking algorithms need to manage frequent and sometimes severe occlusions arising from crowded conditions. Our approach has three processing levels. First, objects are detected using shape-based methods tuned to humans or other objects. Then multiple objects are tracked simultaneously using a novel multi-target particle filter that dynamically handles grouping behaviors. Finally, scene content is extracted and used to link tracks across long occlusions using a global optimization. In Phase 1, we will develop an initial end-to-end algorithm by enhancing our current techniques for each processing step. The feasibility of tracking under difficult conditions will be investigated by measuring tracking performance under various levels of inter-object occlusions, object size and camera motion.