Moving Objects Detection And Tracking In Non-overlapping Surveillance Camera Network

With the high-speed development of visual surveillance technology, the research about multi-camera information fusion is gradually paid more attention to. As one of the key techniques, objects tracking across non-overlapping multi-camera becomes an active research topic. This is a rising direction of machine vision, which crosses several subjects, and has a wide spectrum of promising applications.In an overlapping multi-camera surveillance system, to track the objects persistently, the overlapping fields of cameras are used to get the correspondence of observed objects. But it doesn't work in a non-overlapping multi-camera surveillance system, the spatial and temporal information got by each camera are used to get the correspondence, the main technologies include moving objects detection and tracking in a single camera, objects association among non-overlapping multi-camera, generation and visualization of objects'trajectories. Each technology is researched, the application environments are outdoor vehicle surveillance and indoor human surveillance, the main contribution consists of four parts as follows:Moving objects detection in a single camera: (1) A moving objects detection algorithm based on MAP (Maximum A Posteriori) of pixel blocks is proposed. The detection unit swithes from pixels to pixel blocks. The continuity of pixeles is enhanced. The prior energy defined by spatial-temperal clique enhances the continuity of pixel blocks. The method greatly improves the integrality of detected objects; (2) Another efficient moving objects detection algorithm based on entropy power and GGD(Generalized Gaussian Distribution) is proposed. The entropy theory is used to get segmentation threshold. GGD is used to compute the local thresholds adaptively. The method has good detection results on the videos with cluttered backgrounds.Moving objects tracking in a single camera: (1) For the colour feature is sensitive to illumination change, a feature named LICH is proposed, which is robust to the variation of environment illumination; (2) The tracking is done based on the LICH feature with an occlusion prediction mehtod. The positions of objects'minium bounding boxes are predicted. Occlusions are predicted to happen on objects with overlapping bounding boxes, and are confirmed by the final matching results. Sub-block matching is used to handle the occlusion. The occlusion prediction method can greatly decrease the calculation.The main technologies of objects association among non-overlapping multi-camera include association feature extraction, camera topology estimation and data association. In the paper, the camera topology is built through prior knowledge, the other two topics are mainly researched. (1) New features used in association are presented. For viechles, the ILICS feature is used in association, which can deal with the posture variation; For human, the anthropometric dimensions of human body, such as stature, shoulder breadth and so on, are used in association; (2) A data association algorithm based on minium cost network flow is proposed. The association result is got by calculating the maximum utility of the network with limited association numbers. Compared with the conventional Bayes method, the proposed method greatly improves the association results in an open environment; (3) Describes an objects association strategy for large area camera network. The whole camera network is devided into many individual units, each unit just care about the association problem inside. The working mode has good adaptability to surveillance camera networks with different structures.Generation and visualization of objects trajectories: A method of generating the objects trajectories between non-overlapping cameras is proposed to fill the blanks of trajectories in blind regions, the continuous and integrated trajectories can be got on the whole surveillance network. Different visualization methods are proposed respectively for large area surveillance of vehicles and surveillance of human in multi-storey buildings.