Abstract
This paper presents pedestrian detection embedded into a framework of level-set-based moving object detection and tracking. Most video surveillance systems use stationary cameras to watch scenes. On the premise that the targets are active in the scene, this paper propose to use motion cues to exclude the background positions from evaluation of histograms of oriented gradients (HOG). Discovering that it is the shape that has distinguishable characteristics to differentiate pedestrian from other categories, the boundary based tracking algorithm is adopted to detect and track moving objects. By using the level set method to implement the model, we encode the HOG descriptor only using the pixels located on a narrow band around the zero level set. Considering that an extracted contour may enclose pedestrians in a group, a method for segmenting individuals is also proposed. A set of experiments demonstrate the efficiency and performance of the proposed pedestrian detection system.