A Pedestrian Multiple Hypothesis Tracker Fusing Head and Body Detections

Abstract

We present a multiple hypothesis pedestrian tracker for surveillance video that combines head and whole-body detections. The multiple hypothesis tracker deals with ambiguity in track-to-observation matching by maintaining the most likely valid data association hypotheses. Observations are head and body detections from HOG sliding window detectors. The head detector has a high probability of detection and high false alarm rate, whereas for the body detector these probabilities are lower. The two detection types are fused in a probabilistic framework to achieve robust pedestrian tracking in a crowded environment with clutter and partial occlusions. Experiments show that the use of head and body detections along with multiple hypothesis tracking can improve online track-by-detect methods.