Objects Localization And Contour Tracking On Video Sequence Based On Online Learning

Research on computer vision is one of the most active and challenging area, and target tracking becomes the core subject in the field of computer vision. The target tracking involves image processing, pattern recognition, mathematical modeling, automatic control and other knowledges.It has practical value and good market propect in the field of traffic safety control,medical diagnosis,military,automotive and other active aspects. The main research is as follows:1. For non-rigid target contour tracking in a complicated environment tracking failure problems, this paper proposes a Snake model and its contour tracking algorithm based on online learning. The algorithm utilize the Tracking-Learning-Detection mechanism to achieve the goal of fast tracking, and update Snake model constraints through the tracking results to improve the accuracy of the target contour tracking. In the phase of initialization, the target to be tracking is divided into several blocks on the basis of Grab Cut algorithm, and the algorithm realize the sub-targets locating and tracking by the use of TLD in the subsequent tracking process, which forming the confident map of target outline. At the same time, the algorithm produce positive and negative samples and update each target tracking model for each target feature extraction. Finally, the constraint of parameterized Snake model is built through confident map and the contour of target is obtained. The experimental results show that the algorithm can adapt to the changing light and dark, and even more complex tracking environment, and obtain precise contour.2. To solve the problem of TLD tracking drifts or fail, a robust objects tracking algorithm based on geometric blur is proposed within the framework of online learning. Under the tracking- detection- learning mechanism, Lucas- Kanade algorithm is used to obtain the rough tracking estimation of the target. Based on the idea of geometric blur matching instead of traditional detection methods, the tracking drifts is efficiently corrected. Then integrator is designed to compare the similarities between the previous frame and the results of the tracker and the detector. Their confidences are obtained by calculating normalized correlative coefficients between positive and negative samples and the detected region. An online learning is then developed to use the current result to update the tracker and the detector. Experimental results show that when applied to the fact moving target tracking under the condition of high background similarity, the proposed method performs well and outperforms other state-of-the-art methods with higher position accuracy.3. To solve the problem of rapid movement and severe deformation, a robust objects contour tracking algorithm based on Regression Learning is proposed. First, the training samples are generated by cyclic shifts of the current tracking area with a cyclic matrix, which is used for kernel correlation based regression training. According to regression model of last frame, correlation map between target and test area in frequency domain is calculated, and returns to spatial space to form a target position maps. The maps are fused with the gray image of the test frame to establish a contour confident map. Target contour is extracted with active contour model by taking contour confident map as auxiliary information. When the contour distortion is detected by a designed distortion evaluation scheme, the contour will be modified in the next frame. The tracking result is then feedback to the kernalized correlation filter, and helps to update the tracking template. Experiments show than in various tracking cases, the proposed method achieves more accurate objects position and contour tracking results than other state-of-the-arts methods with satisfactory speed.