Vehicle Tracking With Non-overlapping Views For Multi-camera Surveillance System

With the increasing emphasis on city security issue and the implementation of statepolicies and regulations, there has been a rapid development of intelligent videosurveillance technology. Many products in this field have become familiar to the public.Nowadays, with the rapid expansion and development of the city, traditional technologyfor single-camera-based video analysis has become insufficient for the applicationrequirements. By taking cost, maintenance and management into consideration, it isimpossible and unnecessary for the camera network to cover all area. To solve the aboveproblems, we designed a vehicle tracking method for multi-camera with non-overlappingviews, to achieve target tracking for wide-area video surveillance environments.There proposes a novel approach of data association using minimum cost andmaximum flow on the basis of single-camera video analysis in the previous works. Byassociating structured information of moving objects with the topology information ofcamera network, this approach is able to keep track of target vehicles using multi-cameranetwork. Besides, by taking video characteristics into consideration, we also present analgorithm for camera network topology partitioning, which makes it possible for eachsub-graph to track target independently and as well as in parallel. This not only reducesthe size of the data processing, but also improves the computational performance of thesystem in dealing with huge amounts of data.The presented experiment result demonstrates that the vehicle tracking system basedon non-overlapping camera network is able to perform target vehicle tracking analysis andtrajectory prediction with high efficiency and accuracy. Compared with the traditionalmaximum posterior probability algorithm based on Bayesian framework, the dataassociation algorithm improves the system performance of20%. Meanwhile, the proposedcamera network sub-graph partitioning algorithm meets the system requirements formassive data processing. System has a certain accuracy, efficiency and robustness.