Object Detection And Activity Analysis Cross Multi-camera

Multi-camera surveillance system has the advantage of wide range ofmonitoring, broad angle of observing, capturing comprehensive information. It isapplied into a lot of places by many organizations. However, intensities of light indifferent camera views are not alike, and information of hue and saturation indifferent cameras are not unanimous. So even though the technique level ofdetection and tracking of objects within one camera view is very developed, it isworth to research and study on object detection and activity analysis crossmulti-camera using the algorithms used within one camera view to providespatio-temporal information of moving object.The research is on moving object detection and activity analysis cross cameras,the main work is as follow:Algorithms of common object detection are introduced, including optical flowfield method, temporal difference method, as well as background subtractionmethod. As to background subtraction method, the Gaussian Mixture Model(GMM)method, ViBe(Vision background extraction) and the optical flow field method aredescribed particularly. This thesis proposed a method of modeling backgroundwhich can effectively handle with complex scene of video of low resolution andframe frequency. Original video have been split into several overlapped blocks,each block comprise of frames having same counts. Each estimate background hasbeen produced by background modeling based on combination optimization, thanthese estimate backgrounds are formed a background model. As shown in theexperiment the method is stable in complex outdoor environment and robust withlight changing and shadow influence.Traditional methods of analysis object are based tacking trajectory, andusually are invalid due to large number of objects in a scene and confusion oftrajectories. In this proposed research, based on the similarity of moving modelsand the relationship of moving space, each vision field of camera network issegmented into semantic active regions automatically. Then a Cross KernelCanonical Correlation Analysis (xKCCA) is implemented to explore thecorrelations between these active regions and the topology of the camera network. This spatio-temporal information can improve the accuracy of objectsre-identification. The experiment results show that our approach performseffectively and efficiently in multi camera surveillance network.