Target Appearance Modeling And Similarity Measure In Visual Tracking

With the advent of the big data, abundant video data exists in the Internet. How to analysis、identification and understating the semantic information of the video attract much attention in recent. In order to understand the video, we need to analysis the action and pose of the moving target in the video, and one of basic task is to model the motion of the target, which can also be named as visual target tracking. It is the fundamental problem in computer vision. Due to the complex and diverse of the video data, and the uncertainty character of the target, visual tracking is still a challenging task in computer vision. The main issues include:complex appearance variation of the target, e.g., in plane and out-of-plane rotation, scale variation. In addition, the variations of the environment often lead to the change the target’s appearance, e.g., illumination variation, occlusion, cluttered background and motion blur.In this dissertation, the basic formation for visual tracking task is presented at first and then it is decomposed into several components. Hence the task is to address two critical issues:1) How to design a robust target appearance model, to address both the internal and external target appearance variation.2) How to design a robust similarity measure to locate the real target in the forthecoming frame.This dissertation mainly focus on visual tracking, face to these two critical issues, our work unfold the research on the three aspects as follows:(1) From the prespective of human vision, a novel spatio-temporal structural saliency part model is presented. A fully study of the saliency properties of the local image regions is introduced, which not only include the spatial property of local image regions, but also the temporal charachterize, which guarantee that the local image region with high saliency score can supply sufficient tracking information. In addition, the structural constraint is fused into the selection process of the local image region, which is formulated as a mutual exclusion; it ensures that the redundant information can be reduced from the appearance model. Forthemore, the strategy for target locatelization and the strategy for appearance model updating is also introduced, which guarantee that the target appearance model can address the complex variations. (2) A novel target appearance model named low dimensional feature subspace is presented. The feature subspace includes of a set of local image parts, which is selected from an over-complate feature space, it is the low-dimensional approximate representation of the original over-complate high dimensional feature space, and it captures sufficient tracking information. The whole target is represented by the subspace vector, and the subspace matching is employed to locate the real target in the forthcoming frame. A key advantage of the presented appearance model is its effective and efficient, it can be utilized in the real-time applications.(3) We present a unique fused similarity measurement which is suitable for robost visual tracking. It incorporates multiple cues into a uniform similarity and exploits the complementary character among different cues. In addition, it exmplys the context structure of the candidates in the forthcoming frame to further enhance the measurement. The presented novel similarity measure is simple but computational efficient and arbitrary image cues can be utilized as the input of the similarity measure.We present fully study of the target appearance model and the robost similarity measure in this dissertation, and give several critical solutions. From both the theory justification and experimental results, we find that the presented two kinds of target appearance models and the novel similarity measure can efficiently inprove the tracking performance.