Detecting walking pedestrians from leg motion in driving video

Pedestrian detection in the driving video is an important function for accident avoidance. Different from the detection method based on human shape analysis, this paper introduces a new method to detect walking people from their motion in the driving video. Motion profiles of the driving video are acquired where we found walking people showing their leg moving trajectories as twisted chains. These chains are very different from the moving traces of background and other vehicles appearing as smooth curves according to the vehicle motion mechanism. Thus we design a method to recognize chains uniquely at leg crossing by using HOG features and confirmed with template matching. This method can detect a person in two walking steps. The results show a promising detection rate in the reduced data dimension of video.