Layered Representation for Pedestrian Detection and Tracking in Infrared Imagery

This paper introduces a layered representation for infrared imagery and studies its application into pedestrian detection and tracking. We present a generalized EM algorithm to decompose infrared images into background and foreground layers and study the phenomenon of polarity switch. We propose a hybrid (shape+appearance) algorithm for pedestrian detection, in which shape cue is first used to eliminate non-pedestrian moving objects and appearance cue is then used to pin down the location of pedestrians. We also formulate the problem of shot segmentation and present a graph matching-based pedestrian tracking algorithm. Experimental results with OSU Thermal Pedestrian Database are reported to demonstrate the excellent performance of our algorithms.