Appearance Model Based Object Tracking

Object Tracking has been a central problem related with various applications like Surveillance, Control Based on Vision, Human-Computer Interface, Augmented Reality and so on. Among all the approaches devoted to the object tracking problem, we can roughly divide into two categories, the deterministic one and stochastic one. For the deterministic algorithms, the tracking task is considered as a optimization problem to find a region with largest similarity. One popular example is Mean-shift algorithm. Another category is based on probability, like Particle Filter. In this thesis, firstly, we will discuss about the mainstream object tracking algorithms and the research focus currently. Secondly, three major compo-nents will be given to illustrate an object tracking algorithm and a detailed descriptions will also be given. Later, in the following three chapters, three different algorithms will be talked respectively.In Chapter 2, the major problems we want to handle is to overcome the light changes. Our algorithm is based on subspace methods and combine the feature from the orientation of derivative. In the article, we not only give a theoretical proof but also some experiments to show that our algorithm is superior to original one and some widely used methods.Then in Chapter 3, we turn to the problem of occlusion. Since Occlusion may cause a lot of trouble and difficulties to most of object tracking algorithms, we replace PCA with Robust PCA in order to solve that problem. And the experimental results are quite promising with a high precision even high occlusion happens. Besides, Robust PCA can also be employed in the updating stage due to the occlusion handle scheme. The idea behind it is to relate each target object in a frame with a state. Some specific actions will be performed for each state and the transitions between them are determined by the results of Robust PCA. From the scheme, we successfully overcome the occlusion problem.In Chapter 4, rather than employing the subspace methods, we focus on how to improve the proposal distribution of particle filter framework. Due to the introduction of Hough Transform, our object tracking algorithm can be regarded as two steps. The first one is to predict a rough center position of the target objects by Hough Transform. And in the following, Haar features based particle filter algorithm is performed to refine the results. In the updating stage, we also propose a novel scheme which seamlessly integrates the above two parts and successfully handles most of problems.In the last chapter, I conclude this article and present some of the future works.