Research On Moving Object Detection And Analysis Cross Multi-camera

With the rapid development of intelligent video surveillance technology in recent years, the video surveillance system which based on the multi-camera is more and more applied in a variety of monitoring sites, and caused a widespread concern of many scholars as the advantages of the widely monitoring range, the extensive observation angle and the overall capture information. The algorithms of single camera vision in target detection and matching areas are increasingly improved, however, as the different of light, color and location information in the view of multi-camera, these perfect algorithms are difficult to successfully migrate to multi-camera surveillance network. Multi-camera topology estimation can provides some necessary information for the analysis of the active of target in the multi-camera visual domain, so it has important theoretical significance and application value for multi-camera sight temporal correlation study.In this paper, some of the major issues in multi-camera surveillance network intelligent moving target detection and analysis of research activities are discussed and researched. The major work and conclusions of present paper are as follows:In the moving target detection of the view of signal camera, there are some the problems in the typical Gaussian mixture background modeling process:the background model initialization is susceptible to be impacted by the recurring target prospects, the existence of some unnecessary updates during the update detection model process and the false detection caused by the over convergence of variance false detection. This paper pertinently puts forward some improved methods and suggestions: using secondary screening method to get rid of the mean difference between the larger and the prospect of background interference background model initialization process, establish the conditions to ensure that the replacement delayed update coming Gaussian distribution is a stable background Gaussian distribution in the model update process, the compensatory threshold is introduced to avoid the false detection of the excessive convergence of the Gaussian distribution variance. The experimental results show that the proposed algorithm can effectively solve the problems of typical Gaussian mixture background modeling with a good robustness and detect moving targets accurately.In multi-camera network topology matching estimates and target areas, as a result of the activities of traditional target tracking based on the analysis, the inaccuracy of analysis of motion target is lead by the confusion of the trajectory tracking of moving targets. Based on the multi-camera visual spatial distance within the target and movement patterns similar similarity, the camera sight area is divided into several semantic, then the nuclear canonical correlation analysis (KCCA) is used for these regions correlation analysis, thus to estimate the of and temporal correlation and spatial correlation for each semantic area and camera. Target matching between multiple cameras use these correlation information of temporal and spatial to reduce the computation and eliminate these false targets to improve the matching accuracy. Experimental results show that the proposed method can accurately divide the semantic area and get the correlation matrix between them, turn right to establish a multi-camera network topology, and improve the matching accuracy of active targets cross multiple cameras.