Object Tracking Based On Multi-Layered Salient Foreground Patches

Visual tracking is one of the hot topics in computer vision, which is the basis of many high-level applications. As a basic computer application research, it has been widely applied and researched in recent decades, and considerable developments are made. However its performance for the real world application is still limited due to the harsh circumstances, such as illumination variation, deformation, occlusion, scale change and noise. A stable, reliable and efficient visual tracking algorithm is urgently requiredThis thesis studies the robust visual tracking method and consists of the following main research contents.First, the status quo of visual tracking research is investigated. Since the focus of the tracking methods in this thesis is mainly constructing the appearance model for the Bayesian tracking, the theories of appearance model and Bayesian based tracking method are introduced in detail in this thesis.Second, multi-layered salient foreground patches model is proposed. In most online tracking methods, the object is usually marked by a rectangular box. The box contains not only all of the foreground information, but also some background information. This affects the accuracy of object template, weakens the distinction between the foreground and background, and reduces the performance of object tracking. To overcome this problem, a new method based on similarity measure from foreground and background is introduced to extract salient foreground for constructing more accurate salient foreground model. Meanwhile, a template update method is also presented to adapt the changes of foreground appearance caused by illumination change, and then a multiple foreground model is constructed. Experimental results demonstrate that the proposed tracking algorithm can effectively suppress the illumination variation, deformation and background clutter.Third, a new method for the computation of confidence map is proposed, and a statistical contour model for object foreground is presented to track object. First, from the perspective of the classification of foreground and background, we analyze and propose the formula expression of the theory of classification between foreground and background. Then, a new method for confidence map is proposed to construct statistical contour model. Statistical contour model is based on the value of statistical contour distribution proportion of pixels. The value of proportion not only reflects the foreground confidence of pixels, but also reflects the distribution situation of foreground. Based on the statistical contour distribution proportion, a mask is formed, which is incorporated for object tracking ultimately. Experimental results demonstrate that the proposed tracking algorithm can make effective tracking results.