Vision-based pedestrian monitoring at intersections including behavior & crossing count

This work presents a tracking system which delivers count and behavior analysis of pedestrians by leveraging existing traffic camera infrastructure. The proposed system is able to detect either stationary or moving pedestrians through contextual fusion of appearance and motion cues. Pedestrian recognition performance is improved through cooperating tracking algorithms. Greedy bipartite graph matching is used to initialize newly detected pedestrians and optical flow is then utilized to handle tracking through partial occlusions. Experimental results including system evaluation and behavior analyses of pedestrians show the efficiency of the system to count and assess pedestrians' waiting time and crossing speeds. Additionally, heat-maps which indicate the waiting and moving locations of pedestrians at the intersection are provided to better understand usage patterns.