Neighboring vehicles modeling for tracking across nonoverlapping cameras

Tracking vehicles across nonoverlapping cameras is required by video-based intelligent transportation systems (ITS) to efficiently calculate traffic parameters; such as link travel times and origin/destination counts. In traffic monitoring applications, cameras are usually mounted far from each other to cover wide areas. As such, object features (i.e., color information, shape, and direction) change significantly from one camera to another. These space-time differences raise serious challenges on efficient tracking. In this paper, we have presented a probabilistic model to solve the multicamera tracking task in a network of disjoint view cameras, with attention paid on estimating the density function of different features such as space-time, appearance, and especially neighboring vehicles' relationships. As in highways each group of vehicles usually tend to keep their distances, using the similarity of neighboring vehicles plays an important role in finding the correspondent vehicles. A graph-based approach is used to solve the assignment problem. Experimental results show the efficiency of the proposed tracking method.