Distributed Multi-view Target Tracking, Statistical Inference Methods And Achieve

Visual object tracking is under intensive research in the field of computer vison and has a wide range of application, including video monitoring, video retrieving, vehicle navigating, and etc. Multi-camera video can provide us more redundant information which can help to improve tracking robustness, thus multi-camera tracking system is gaining more and more attention lately. Kalman filter and Sequential Monte Carlo Particle Filter are both Bayesian inference methods. Kalman filter is a traditional method of state estimation. With the development of computing capacity, Particle Filter is implemented more and more common in real-time video object tracking systems. A novel algorithm is proposed to perform object tracking with multiple cameras in the Bayesian Inference framework including the Kalman filter based adaptive prediction of initial searching point algorithm, the data fusion method based on Sequential Monte Carlo Particle Filter. Finally, the algorithm is optimized and implemented real-time on TI OMAP3530 platform.Redundant information provided by multi-camera system can help to improve tracking robustness. However, traditional template matching based multi-camera object tracking suffers from tracking failure brought by different views and computation requirement is high. To solve this problem, a new algorithm is proposed, which uses the geometrical constraints to communicate object location between different camera views so as to predict the initial searching point of the object in the current view. Besides, by calculating the Kalman Filter model parameters online and adaptively adjust the Kalman Gain to limit the error passing during the interaction process. Experiments show that the proposed algorithm can greatly reduce computation cost as well as improving tracking performance.Besides, this paper explores Bayesian network to fuse spatial-temporal position and template object feature in multiple cameras. Firstly, Bayesian network is used to model the system of multiple static cameras'tracking system. Then, the high-dimensional joint posterior is propagated spatiotemporally. Finally, the estimation of the target location in each camera view is achieved by using an efficient message passing mechanism with sequential Monte Carlo Approximation (Particle Filter) of the joint posterior. By taking full advantage of image data and position data from multiple cameras, the tracking algorithm is quitey robust to occlusion in some cameras of the system. Both qualitative and quantitative experiments have demonstrated the effectiveness and robustness of the proposed algorithm.Finally, an efficient tracking algorithm suited to real-time implementation is proposed and implemented on OMAP 3530 platform. Our system is composed of multiple cameras and the corresponding processing modules. The cameras collect data and perform tracking distributedly. In order to reduce the load of network transmission, only the necessary arguments are transmitted among cameras. Our robust tracking system efficiently fuses information from different views and is capable of dealing with partial and full occlusion. In order to achieve these goals, we designed a distributed object tracking software framework and transmission protocols. The system fully utilizes the powerful processing and network transmission ability of OMAP3530 with optimized allocation of calculation between the ARM and DSP processing core. After optimization in program level, the target tracking program can run robustly and efficiently in real time, which are proved by considerable results.