Research Of Video Object Tracking Method

The technology of video based object detection and tracking is one of the hotspots in the field of computer vision,and provide an important data source for visual analysis and understanding.In this dissertation,the research is focused on the three critical problems of visual target tracking - object detection,shadow detection and object tracking.The main contributions of this dissertation are summarized as follows:1.An object contour tracking algorithm based on particle filter and graph cut has been proposed.We improve the accuracy of the foreground extraction by embedding texture information and discriminative features selection method into the graph cut algorithm.An adaptive tracker based on the integration of particle filter and mean shift has been proposed,which is conciser and effective compared with the existing methods of a kind.The object model utilizing probabilistic principal component analysis gives some guidance to the handling of the occlusion. Experiments on variety real-world video data demonstrate the proposed schem not only locates the target accurately but extracts the objcect contour quite well.2.A video object detection algorithm and a moving shadow detection algorithm based on texture which is represented by the histogram of the uniform pattern of the local binary pattern(LBP) have been proposed.Video object detection algorithm can extract the moving target quite well and deal with the issue of moving shadow to some extent because of the characteristic of LBR The moving shadow detection algorithm requires only a small number of parameters and experiments on variety real-world video data demonstrate the favorable performance and robustness of the proposed scheme.3.A tracking algorithm based on the adaptive appearance model has been proposed,in which the adaptive model is utilizing the probabilistic principal component analysis.The object model based on the probabilistic principal component analysis is robuster and more accurate compared with the object model in most existing object tracking method and we can deal with the issue of "object drifting" based on the model,furthermore we can locate the target more accurately and handle the issue of occlusion to some extent by some prediction and local search.We demonstrate the favorable performance and robustness of the proposed scheme by experiments of the integration of this adaptive object model and mean shift method.4.An enperimental video object tracking system has been designed and verified by experiment results.Our main work is focused on object detection and shadow detection under a fixed camera scene.When it comes to object tracking,we mainly research on single object tracking.The following work will be focused on the research of object detection and shadow detection under a moving camera scene,and multi-object tracking.