Pedestrian detection in images by integrating heterogeneous detectors

Pedestrian Detection in still images is a key problem in computer vision. Traditional approaches design features for representing the holistic human body. Unfortunately, occlusions and articulations pose challenges and degrade their performances. Part-based representations have more potential to solve these problems. However, they tend to produce more false alarms than holistic approaches. This paper proposes a framework to integrate heterogeneous detectors (including holistic, part-based and face detectors) to boost pedestrian detection performance. Responses from heterogeneous detectors cast probability votes using Hough transform and considering geometric relationship of different detectors. Peaks of votes localize where pedestrians are. To avoid false alarms, cell models are learned in advance to evaluate local alignment and to reject wrong detections. Experiments on the INRIA dataset show that our framework provides a better performance than some state-of-the art methods.