Study On Moving Object Detection And Tracking In Video Surveillance

Video surveillance system is the core sub-system of a security system, which captures video frames with surveillance equipment, then process the video frames using pattern recognition and computer vision technology, thereby achieving the purpose of detection and tracking. However, due to interference of external environment, the changes in the attitude of the target, there are many problems in moving object detecting and tracking. In this thesis, moving vehicle is taken as research object, a segmentation method based on the skeleton and corner is proposed to solve the problem of adhesions in moving object detection, and then it put forward some corresponding measures to overcome the defects in the CT (Real-Time Compressive Tracking) algorithm.In most cases, the images of vehicles will overlap each other in the frame of surveillance video obtained by the camera; at the same time, the expansion of the outside contour will lead to adhesion of the two cars in close proximity when we do some necessary operations on binary image such as erosion and dilation. Both of the two problems mentioned above have a serious impact on the follow-up work, such as tracking vehicle, traffic detection and so on. To solve the problems, a segmentation method based on the skeleton vehicle has been proposed. The skeleton of the adhesive area is extracted by "fire method", and the corners of the skeleton are obtained through corner detection. These comers are clustered by K-means clustering method, allowing us to get the splitting line of the adhesive vehicles from the clustering results. Experimental results show that not only can this method relieve the over-segmentation problem in adhesive vehicle separation effectively, but also can this method reduce separation time needed. Based on the proposed algorithm, contour corner method is also presented, which further reduces the processing time.The specific moving target has been tracked with the detected results. To overcome the shortcomings in the CT algorithm such as features extraction only from the grayscale image and tracking loss when target acceleration, Compressive Tracking based on Multi-Channel Haar-like feature (short as MCCT) is proposed in this thesis, Haar-like features are generated from three channels of RGB, and can effectively alleviate the drawbacks above by the target prediction. Experimental results on large number of public data sets show that the algorithm proposed in this thesis produces less average error of the target center compared with CT algorithm, other improved algorithms and other algorithm; it also performs favorably at the circumstances of the change of illumination, camera shaking and target acceleration.