Research And Application Of Visual Tracking

Image sequence based object tracking is a fundamental problem for computer vision research and has been widely studied. The main goal of visual tracking is to imitate the motion sensibility of physical visual system, empower the machine with the ability of perceiving the object motion and their relations in the scene and provide an important way for image sequence understanding. Some typical examples of application of visual tracking are video surveillance, video analysis, video indexing, video based motion analysis and synthesis, motion-based human identification. As a topic with many applications, object tracking has drawn much attention of researchers. Many institutes have done a lot of work on it and got achievements. However, there are still many problems to be solved in order to build a robust and practical tracking system.Particle filter implement recursive Bayesian filter via Monte Carlo simulation. The method is suitable for any non-linear system that could be represented with state model. It is more practical than conventional Kalman filter and its precision could approach optimal estimation. Particle filter is flexible and easy to be implemented. Furthermore it is also has parallel structure.Under the framework of sequential Monte Carlo filtering algorithm, this thesis aimed to improve the precision and robustness of the tracking algorithm and try to get insights on some key issues in visual tracking. Some important technologies and solutions are studied which are necessary for robust and practical tracking systems. This thesis mainly consists of the following parts:(1) In order to improve the quality of the state-space exploration and the accuracy of visual tracking, in this paper a particle filter algorithm based on MCD and partial linear Gaussian models is presented. MCD avoids the manner that each pair of pixels in the image contribute to the matching result equally.The proposed method uses neighborhood between pixels as the matching similarity. The obtained correlation curve by the proposed method is much sharper. So the image matching method has high matching precision. A direct consequence of using partial linear Gaussian models is that the optimal importance function is adopted. The combination of them will be the optimal particle filter. The stability of the algorithm has been improved due to the robustness of MCD. Two simulated experiments are finally conducted to confirm the validity of the improved algorithm.(2) A novel template statistical feature matching similarity criterion with a tracking framework using particle filter is proposed. Weighted factor introduced can effectively reduce the influence of boundary noise and background feature, it also emphasizes the importance of target feature. Similarity bias may be got in the two completely different templates because of the intersection of statistical features. The proposed method can correct the bias and improve the peak modality of matching function by fusing similarity based on HSV color system, so it can obtain the global optimal solution and robust tracking. Experimental results show that template matching has an excellent peak distribution. The proposed tracking algorithm exhibits good precision and robustness in the presence of noise, deformation and occlusion.(3)Image matching similarity criterion is the critical factor of visual tracking. Template matching in visual tracking often can not obtain the global optimal solution because of the defect of pixels r,g,b color's many to one in computation and the influence of background feature. This paper presents a concept of fuzzy membership and a novel similarity measure formula. The proposed method can overcome the defect of color computation and make similar color pixels clustered, consequently it can improve the ability of target recognition. The new criterion can effectively reduce the influence of background feature, and emphasizes the importance of target feature. It can improve the peak modality of matching function, so the global optimal solution and robust tracking can be easily obtained. Experimental results show that template matching has an excellent peak-like distribution, and tracking algorithm is precise and robust in the presence of noise, deformation and occlusion.