Research On Object Handoff Under Overlapping Views In Intelligent Video Surveillance

With the urban population increasing, serious social public security greatly promote the development of the security industry. Intelligent Video Surveillance System can replace people to guard important occasions and warn suspicious behavior in time, this made it become the main safeguard. As the key technology of Intelligent Video Surveillance System, object handoff fundamentally solves the requirement of big scene and multiple perspectives monitoring. And its essence is keeping same label of the object which appears in different cameras. This thesis focuses on the research of object handoff under overlapping views, it mainly contains object detection and tracking, object handoff under overlapping views and collaboration of multi-camera. Specific work is follows:(1) Object detection and tracking. The method of Gaussian Mixture Model is used to obtain the moving object. To reduce the influence of the adhesion on object handoff, a kind of adhesion processing method based on statistics is used. First, Determine whether adhesion by the statistics of target rectangle area, width and height, then split adhesion object averagely. For multiple targets tracking, a quick method based on multiple characteristics is used. Introduce multimode processing method, and split target prediction into two modes according to the objects centroid distance, then predict the state by the target position, color and movement information of the historical frame, the detected target and the predicted target is related by the NN method, and exception handling mechanism is introduced for entry of new target and left of old target. Experiments show that target tracking method is high reliability and real-time under sparse scenarios.(2) Object handoff under overlapping views. A method of handoff opportunity judgment was proposed which is combined with target visibility judgment and flags. Firstly, SIFT algorithm and RANSAC algorithm is used to obtain the matching point from different perspectives, with which the homograph matrix is computed, and the Field Of View(FOV) lines, obtained by the boundary points of images and the homograph transformation, were applied for defining the set of visibility judgment function, besides, the flags which respectively represented object visibility and handoff succeed are also defined. Projective transformation and distance judgment with tolerated error were then used to achieve object handoff. If more than one object meets the condition, the color information is used to compare further. Finally, update the succeed flag according to the result of object handoff. Experiments show that this method can exactly grasp the opportunity of handoff and the effect of handoff is good.(3) The collaboration of multi-camera. For soft-collaboration, An improved scheme of multiple cameras scheduling is adopted, which defines a weight function relying on object visibility, size, completion and camera load, by calculating the weights of the visibility cameras to the particular target at any time, the manager assigns the maximum weight camera to the target as the main tracking camera.