Research Of Tracking Algorithms For Objects Across Multiple Cameras

Using computer vision and multi-camera network to achieve the intelligence video surveillance is one of the future direction of video surveillance system. In multi-camera intelligence video surveillance, many problems have to be solved. Among them are two critical problems. On the one hand, we have to solve the problems in single camera tracking. On the other hand, we have to deal with the new problems aroused in multi-camera system. For example, how to coordinate multiple cameras in a surveillance system to fulfil reliable tracking when a person moves from the Field of View (FOV) of one camera to the FOV of another camera.Based on the application background of the intelligence video surveillance, this paper studies non-overlapping multi-camera based object tracking method, and the algorithm for cooperative multi-camera surveillance system. The research includes following aspects.(1) An algorithm for tracking objects across multiple cameras with disjoint views is presented. On the basis of single camera object tracking, using multi-camera data fusion method which combines estimating change in appearance of objects with establishing path model, it can accomplish object tracking in such situations that complete site models or calibrated cameras are not available, even in the light changing environment. Compared to other multi-camera tracking algorithms, this method is of 6% higher tracking accuracy.(2) A real-time surveillance system based on non-overlapping cameras in indoor or outdoor environment is designed. The system adopts non-overlapped multi-camera to enlarge field of view, achieves the detection, tracking and classification of moving objects at the end of each single camera, achieves the tracking of objects across cameras using data fusion method mentioned above at the end of multiple cameras. Experiment shows that the tracking accuracy of the system reaches 98%.(3) An algorithm for cooperative multi-camera surveillance is proposed. The algorithm distributes cameras to objects depending on task priority, distance between objects and cameras and visibility of objects, which guarantees that the visible object of high priority and nearest distance to camera can be distributed camera first. Qos function has been introduced to describe the system performance. This method is one time higher in system performance in comparison to existent coordination algorithm.