Improvement of optical flow in pedestrian detection based on pictorial structure

Non-rigid characteristics of the human body and the diversification of human articulations are the two challenging problems in pedestrian detection, especially in cluttered scenes that commonly involve multiple people, such as surveillance cameras. Moreover occlusion and body changes also increase the difficulty of the people detecting. The general pictorial structure can divide human body into some parts, then use the appearance descriptor and the position information between parts to detect pedestrian in the case of greatly change in human body. This paper adds the part´s optical flow into the pictorial structure. We take advantage of the different limb movement direction and its speed to further increase the detection precise. In the experiment, we take comparison between our results and the results without using optical flow in the pedestrian movement database, which confirms that our improvement can increase the accuracy.