基于稀疏表示的视觉跟踪新方法研究

视觉跟踪的目标是要确定目标的位置、尺度、姿态、速度等状态信息。但是由于光照、遮挡、旋转、变形、突然运动、复杂背景等因素的干扰，视觉跟踪是一个非常具有挑战性而且非常重要的基础问题。另一方面，稀疏表示的研究近来非常活跃，而且在不断完善的过程中，我们拟针对跟踪这个富有挑战性的问题，围绕稀疏表示理论，设计出能够处理目标旋转、姿态、光照、遮挡、以及复杂背景等鲁棒的跟踪算法。本课题的具体目标是探讨如何将稀疏表示理论与目标跟踪问题的先验知识（时间上的连续性，特征的冗余性，背景和前景的区分性等）相结合，克服稀疏表示理论时间和空间复杂度都较高的难点，拟从有效的视觉特征(Mid-level)描述及选择，基于稀疏表示的判决模型和生成模型的构建与融合，子空间信息的利用与合理的更新机制的设计，2范数约束下的稀疏表示，汇聚算法与局部稀疏表观模型等几个方面深入。

The goal of visual tracking is to determine the states of the target (e.g., position, scale, pose and velocity). Visual tracking is a very important basic problem yet it is also quite challenging due to the interferences of illumination change, occlusion, rotation, deformation, fast movement, complex background, etc.. The research of sparse representation is active recently and of great improvement. We plan to solve the challenging problem of visual tracking with sparse representation and aim to design a robust tracking algorithm that can simultaneously deal with the problem of rotation, pose, illumination change, occlusion, complex background, etc... The goal of our project is to explore the method to combine the sparse representation with the prior knowledge of visual tracking (e.g., the temporal continuity, the redundancy of features, the discrimination between the foreground and the background) so that it can overcome the high temporal and spatial complexity problem of sparse representation. We plan to do deep researches in terms of the following aspects: first, the description and selection of effective visual feature (especially for Mid-level feature); second, the construction and fusion of discriminative and generative models based on sparse representation; third, the exploitation of sub-space information and the design of reasonable update scheme; fourth, the study of sparse representation restrained by l-2 norm; fifth, the design of pooling method and local sparse appearance model.