Research On Moving Object Detection And Tracking Based On Sparse Coding Shape Classification And Random Walking Model

While tracking target under complex backgound, the main reason of tracking failure is the interference of the background surrounding the target and target appearance changes. Furthermore, when camera moves, the initial detection of moving targets will become more complex. Taken together, these factors make robust visual tracking in videos become very difficult. Therefore, this paper integrates the shape feature information and takes a deep research on the moving object detection and tracking based on the sparse representation and random walk model from the entry point on background information suppression. The main contents are as follows:The first chapter expounds the research background, significance and purpose of visual object tracking. The research status of related theory in visual target tracking are summarized. We analyze some existing problems in visual tracking and give the main research contents of this paper aiming at these problems.In chapter 2, moving target detection method under moving background is studied. A motion detection method based on the abnormal edge analysis of spatiotemporal image is proposed. This method provides a new idea for motion detection under moving camera and achieves good experimental results. We obtain the spatiotemporal image (STI) by extracting the corresponding rows from each frame and displaying them one above another after vertical motion estimation by median-flow tracker. We cluster the edges in the spatiotemporal image according to their slopes, and then analyze the number of edges in each class and the dispersion of the edge distribution. The area where the edge distribution is clustered is determined as the area where the real moving object is located. The algorithm is evaluated in a series of video sequences collected in diverse real scenarios. Experimental results demonstrate that our algorithm is more robust than other classical methods for occlusion, background interference, different camera rotation speed.In chapter 3, a more efficient shape feature descriptor is proposed based on the latest shape feature classification method. The new shape descriptor based on the sum of distance between the contour fragment gravity and each contour point, which is invariant to rotation, translation, and scaling. In addition, sparse coding method that can effectively simulate the simple cellular response mechanism of the mammalian visual system V1 region is applied to encode the contour fragment feature. In order to make the codes contain spatial layout information, we utilize spatial pyramid matching method to perform max-pooling on these codes. Finally, SVM classifier is used to realize shape classification. Experimental results on several well-known shape benchmarks show that our algorithm can guarantee high classification accuracy and has a high efficiency.In chapter 4, aiming at the background supression problem of visual tracking under complex background, a robust visual tracking algorithm based on structural local sparse representation model is proposed. This method can accurately analyze and identify the region surrounding the target, and find out the real background region.Several particle swarm regions are generated around the tracking target. The weight of these particles is calculated by the structural local sparse representation theory. The regions most likely to belong to the background are selected according to the weight of each region. The background information in these areas is suppressed to enhance the anti-jamming capability of target model. Furthermore, the new target information is added into the model according to the reliability of the tracking result in each frame.In addition, a weighted search method in the back project image is proposed. To a certain extent, this method solves the local optimization problem in visual tracking.The proposed tracking algorithm is evaluated in the international standard dataset and video sequences captured in real scenarios. Experimental results show that the proposed tracking scheme can adapt to target appearance changes, and robust to background interference.In chapter 5,aiming at the small human target tracking problem in surveilance video, the random walk image segmentation method with weak boundary property is introduced into tracking. The proposed method can discriminate the human target region according to the distribution of the segmentation results, and improve tracking robustness. Four layers seed points are uniformly selected in the target and its surrounding area. Each layer contains six seed points and each seed is assigned a label respectively. Random walk image segmentation algorithm is used to image segmentation. In the segmentation results, the area similar to human body is selected as the target area. Furthermore, the principal component analysis method is applied to distinguish human target from the background. The probability of prominent features in the background is reduced and the unique feature of the target is highlighted. The proposed algorithm is evaluated in some standard datasets and video sequences captured in the parking lot. Experimental results show that the proposed algorithm achieves good tracking performance.In chapter 6, we integrate the proposed shape feature representation method into the tracking framework to form a tracking strategy. During tracking, the tracking results are classified according to the shape feature of edge curves. Some classification results are collected to form the set of classification results. The classification result in the current frame is compared with the results in the set to determine whether it is consistent with the previous target category. The target model is updated based on the classification result and the reliability of tracking result. The proposed tracking method is tested in several very challenging aerial video sequences.Results show that the algorithm presents good robustness.In chapter 7, the major work of this thesis is summarized. The main contents,conclusion and innovation of our research are briefly discussed. Furthermore, the next research work is prospected.