A new strategy based on spatiogram similarity association for multi-pedestrian tracking

Multiple pedestrian tracking is an active and challenging research topic that many different approaches have addressed it. Since Human stick changes over time and person usually moving in random way, the identity association remains a hard task. In this paper we propose a new method for coupling detections over all the frames of the video sequence in order to make a performant tracker. The pedestrian detection is ensured using the Dalal and Triggs´s human detector. In order to overcome the problem of missing detections caused by occlusion, we propose to use an interpolation process based on average speed. Then, all the previously detections are organized over a tree structure, where each frame represents a tree level. All detections in level `i´ are linked to the next level by an arc characterized by a cost representing the spatiogram similarity between these 2 detections. After trajectories refinement is done based on Euclidean distance to palliate the false detection association. An experimental study conducted in 2 datasets (CAVIAR and CWV) proves the good performance of our proposed method in term of tracker precision and tracker accuracy.