Object Tracking In Cooperative Video Surveillance Under Complex And Crowded Environment

As the public security gets more and more attention in China and abroad, video surveillance systems have being widely used and become the major surveil-lance method. High-definition and intelligent video surveillance system is the most promising direction in contrast to the traditional video surveillance systems which are labor-consuming and can only monitor large areas at lower resolution. Object tracking is the basic problem in intelligent video surveillance. The main challenging factors for object tracking in crowded and complex environments are: 1) Background clutters,2) Frequent occlusion between objects,3) Non-rigid defor-mation of the target,4) The low resolution of the target which makes it difficult to distinguish it from other objects. This dissertation focuses on developing effective algorithms to solve the above problems. The main contributions and innovations are summarized as follow:1. Propose a single object tracking algorithm based on online shared dic-tionary learning. To solve the problem caused by background clutters, it learns three dictionaries:target dictionary, background dictionary and shared dictio-nary. The shared dictionary models the commonality between the object and background, and the difference is captured by the target and background specific dictionaries. In this way, the commonality and the difference between the object and background are seperated, which makes it much easier to distinguish the tar-get from the background. Experimental results demonstrate that:compared to state-of-the-art methods, the proposed method is more robust against background clutters, and achieves 10% improvement in terms of average overlap ratio.2. Propose a single object tracking algorithm based on part-based multi-graph ranking. To solve the problem caused by occlusion, it exploits the intrinsic manifold structure of data, and formulizes object tracking as a graph ranking problem. First, the target is divided into local parts, and graphs are constructed on these parts using multiple feature representations. Each graph is assigned a weight, then the weights of the graphs and the ranking vector are learned jointly in a regularization framework to reduce the effects of the occluded parts. Ex- perimental results demonstrate that:compared to state-of-the-art methods, the proposed method is more robust against occlusion, and achieves 8%improvement in terms of average overlap ratio.3. Propose a single object tracking algorithm based on multi-level represen-tation and hierarchical tree structural constraint. To solve the problem caused by occlusion and deformation, it represents the target at multiple levels and mod-els the structure relationships between parts by constructing a hierarchical tree. The optimal locations of the parts at all levels are obtained by quantifying the appearance similarity and the deformation cost, and optimized jointly in a uni-fied objective function. Thus, each level can benefit from other levels, and the overall performance of the tracker is enhanced. Experimental results demonstrate that:compared to state-of-the-art methods, the proposed method is more ro-bust against occlusion and deformation, and achieves 9% improvement in terms of average overlap ratio.4. Propose a cooperative tracking algorithm in dual-camera systems with bidirectional information fusion. In cooperative video surveillance, the goal of cooperative tracking is to obtain the high resolution images of the target and en-hance the tracking performance in the active camera. First, multiple observation models are adopted to make the tracking algorithm in each camera more stable. Moreover, a bidirectional information fusion strategy is proposed to enhance the tracking performance, which takes full use of the information from both cameras. The experimental results on realistic simulations and the implementation on a real surveillance system validate the effectiveness of the proposed algorithm.5. Propose a novel multi-target tracking algorithm based on heterogeneous camera fusion. In cooperative video surveillance, to reduce the ID switch errors caused by the low resolution images in the static camera, a discriminative model is learned from higher resolution images in the active camera. Information from the heterogeneous cameras is fused into a tracking-by-detection framework to improve the association accuray of the tracklets. Experimental results show that:com-pared to state-of-the-art multi-target trackers, the proposed method can reduce the ID switch errors and improve the multi-target tracking performance signifi-cantly.