Research On Algorithm Of Target Relay Tracking For Multi-Camera Video Surveillance

With the rapid development of video surveillance technology,the information fusion of multi-camera monitoring network has been paid more and more attention.As one of the core technologies,multi-camera target relay tracking in non-overlapping field of view has also become the focus of research.This is a new,multidisciplinary research direction in the field of machine vision research,and it has a wide application prospect.In the multi-camera surveillance environment with overlapping views,it is possible to obtain the correspondence between objects by overlapping parts of different camera scene,so as to achieve the goal of continuous tracking.However,this method can not be used in the multi-camera surveillance environment in non-overlapping field of view,and the target correspondence between different cameras can only be achieved by using the target spatial and temporal information which is obtained by a single camera.The main technologies include single-camera moving target detection and tracking,multi-camera space-time topology building,target retrieval and matching.In this thesis,set the application environment for the outdoor pedestrian detection,the specific content is as follows:Firstly,for single camera moving object detection,this thesis use the ViBe algorithm to detect the target,and optimize the result to get the motion foreground.The bounding box is used to track moving objects,and Kalman filter algorithm is used to track them,and the occlusion problem is analyzed and processed.Second,because the color feature itself does not contain spatial distribution information,different targets with the same color characteristics will be mis-matched.In this thesis,a method of feature extraction is proposed in LAB color space.this thesis classify pixel regions with similar color,divide the pedestrian image into the largest color stable region,and use these regions' center,color mean and Variance to describe pedestrians.Furthermore,this thesis propose a combinatorial feature based on segmentation,and establish a more robust model.Finally,this thesis combines the spatiotemporal topology information with the apparent model to compute the probability of single transition between fragments,and forms the segment transfer graph.The retrieval of the fragment to be retrieved is based on the transfer graph.According to the retrieval result,the fragments generated by the motion of the target in the whole multi-camera surveillance network are connected in series.The final result of this thesis is chain of fragments generated by the motion of the target in the whole multi-camera surveillance network.