Occluded Pedestrian Tracking Using Body-Part Tracklets

Abstract

Detection of pedestrians under occlusion has been addressed previously with body-part-based approaches, in particular using the generalised Hough transform. Tracking is usually addressed by first detecting pedestrians in each frame independently and then tracking the detections over time. This paper presents a novel variation on the generalised Hough approach: tracking is performed first, and detection second. Robust features on a pedestrian are tracked over short time-frames to form tracklets. Not only do tracklets reduce false alarms due to unstable features, but they provide temporal correspondence information in Hough space. Consequently tracking can be posed as optimal path finding in Hough space and efficiently solved using the Viterbi algorithm. The paper also presents an improvement to the random Hough forest training method by using multi-objective optimisation.