Research On Multi-camera Collaborative Target Tracking

With the rapid development of computer technology, the application of sensor network technology and embedded technology are more and more extensive. The space of video sensor networks is also effectively expanded. Multi-camera collaborative tracking as an important branch has received widespread concern from many scholars. Collaborative multiple cameras can reduce resource redundancy, enhance the effectiveness of information,and increase the robustness of information through cooperative fusion. But in the actual scene,appeared node restrictions、target occlusion and other non-controllable factors, often make the multi-camera cooperative tracking face enormous challenges. Therefore, researching the algorithms of multiple cameras collaborative tracking is the focus of this thesis.Some basic knowledge of collaborative tracking target such as the technologies of detecting target, the technologies of tracking target and the technologies of collaborative tracking target are studied firstly. Secondly, three typical target detection technologies and the related research methods in the field of collaborative tracking target are summarized. Thirdly,a systematic analysis and summary of the problems of collaborative multi-camera tracking is conducted. Then, those hot issues of collaborative multi-views tracking are studied deeply.Based on the above analysis, the algorithms of the choice of scheduling and the information fusion of multi-camera were improved in this research as follows:(1) Most researches on camera network are analyzed based on the image’s information.Few researches on camera network were focus on the network structure of the deployment information. Therefore, this paper launches the research in two-dimensional space. Since the camera network node belongs to the directed node, a new method of collaborative location target is proposed for the directed sensor networks. This method is based on the neighborhood topology between nodes. First of all, the connection between nodes is established through the neighborhood topology. Secondly, the information of target by the node sensing is shared through the neighborhood topology. Then the cluster members can be determined according to each node’s capture gain size. The object eventually moving region is defined combined with the sensing area and the largest mobile theorem. Finally, the region centroid as the target location is calculated by using the grid method. The simulation results show that the proposed approach outperforms other similar methods in both tracking accuracy and stability.(2) The scheduling choosing problems of the multi-camera is studied in this part. Due to the problem of target tracking in video sensor networks, a new algorithm is proposed for cooperative monitoring and tracking of node selection based on particle filter. Firstly, the method obtains posterior distribution through particle filtering in tracking problem and gets the information entropy to evaluate the estimate uncertainty. Background modeling and the phase division are applied to extract target blob, then the number of pixels of the target blob is calculated to measure detection information. Finally, confidence measure of nodes is determined based on the two factors to realize the optimization of the selected node and then track by particle filter. Experiment results show that the proposed method can effectively improve the tracking accuracy in comparison with the similar method, which selects cameranodes to achieve.(3) Aiming at the collaborative tracking research based on multi-camera, The existing algorithms of multi-view target tracking based on the homography constraint is simulated.The following problems can be found from the analysis of simulation results. The corresponding relationship between the number of foot position and the number of targets will be blurred because of many uncertain factors such as the occlusion between multiple targets.An improved algorithm of the multi-camera collaborative tracking multiple targets based on modified tracking is proposed. First of all, the foot regions are identified based on the homography constraint and histogram feature. The errors of the homography constraint and histogram feature should be tolerable. Then according to the relationship between the number of location area and target number, combined with the results of detecting moving target and the foot region clustering algorithm, a revised strategy positioning design goals and complete occlusion more accurate target location environment. Finally, the targets can be tracked based on the optimal path. Experiment results show that the algorithm can effectively reduce the tracking error in comparison with the similar method, enhance the robustness of tracking.