The Research Of Single-target Visual Tracking Algorithms In Complex Scenes

Object tracking is an important part of intelligent video surveillance system, which is a hot research topic in computer vision field. Object tracking has the good market application prospect and has been applied for city security surveillance, missile-guidance, city transportation, bank surveillance and entertainment. Although visual tracking has been studied for over thirty years, the difficulties such as occlusion, scale change, and illumination change that exist in complex scenes make it very challenging to design a robust, fast, and accurate tracking method. In this dissertation, I analyze the research status of object tracking and propose two visual based single-target tracking algorithms that can run well in complex scenes.(1) Recent years, sparse representation (SR) based tracking method has attracted much attention. Many SR based tracking methods have been proposed, but they all have a serious problem——low timeliness, caused by solving l1 minimization in every frame, which seriously restricts the market application of SR based tracking method. I develop a real time object tracking algorithm based on l2 norm minimization in section Ⅲ. I orthogonalize the object templates to learn an orthogonal subspace for modeling the appearance of the target, which removes the redundancy between templates, solve the object representation model with l2 norm minimization, and introduce a threshold parameter to decompose the residual into two parts to establish a more robust observation model. The proposed method is fast and effective.(2) Current object tracking methods can be classified into two categories: generative method and discriminative method, which both have their own dis-advantages. Generative tracking method can effectively model the appearance of the target, but without using the background information; discriminative tracking method combines the foreground information and background information to learn a discriminative classifier, but it has a shortcoming in modeling the appearance. To solve this problem, in section IV, I propose a complementary tracking method combining generative model and discriminative model.I evaluate the two proposed tracking methods on a mass of challenging sequences, compared with several state-of-art tracking methods. The quantitative and qualitative evaluation results demonstrated that our methods can track the target effectively and accurately, and can deal well with occlusion, illumination change, scale change, and no-rigid appearance change in complex surveillance scenes.