Vehicle Trajectory Prediction across Non-overlapping Camera Networks

Using camera networks to monitor the trajectory of moving vehicles plays important role in many applications, such as video surveillance, intelligent traffic system, and social security management. Most of the previous works tried to track the moving vehicle by using either appearance matching or spatial and temporal information. However, we realized that the moving of vehicles should follow some underlying social tendency. By using training data for tendency learning, we proposed a new idea to predict the vehicle trajectory, which is a quite different viewpoint in contrast with previous works. In detail, we regarded trajectory prediction as a recommendation problem. By giving partial and fragmental observations of vehicle locations on the map, the proposed system attempted to predict or recommend the possible vehicle moving trajectory. Three types of algorithms for recommendation were evaluated, including a user-based method, an item-based method, and a latent-based method. The experimental results show the tendency learning could be used as useful prior information for trajectory prediction. Furthermore, the tendency learning could be combined with previous works without conflict.