The Algorithm Of Complex Target Description And Robust Tracking

Visual target tracking is a hot research field in computer vision. It has important value and significance on both academic research and practical application. Target speed, position and rotation information can be acquired in real time by processing image sequences in visual target tracking. The uncertainty of tracking object, non-regularity of target motion trace and the complexity of environment bring many challenges to robust tracking. The appearance modeling and motion estimation are two most important modules in a typical visual object tracking system. This paper focus on how to model the target appearance and estimate target motion. The main research work are as follows:1. The appearance modeling methods, especially the sparse representation method are studied. A two-step sparse representation method is proposed. The global target images and partial target images are used to construct sparse representation dictionary. The global sparse representation exploits the global appearance information of the target and the local sparse representation exploits both partial information and spatial information of the target. The coefficients of global sparse representation and local sparse representation are alignment-pooled to compute the importance parameter of global description and the feature vector of local description.2. Template update approaches based on online learning are studied in this paper. The SKL algorithm combining with sparse representation updating strategy is mainly studied. Some improvements are put forward. A cumulative probability sequence based on quadratic function is used to represent the probability of the updating template set. This sequence not only ensures the informations of template are real-time, but also maintain the diversity of templates. The occlusion problem is handled by the spatial information of target characteristic vector.3. Dynamic state estimation algorithms are studied. The importance parameter of global description is used to decide which particle will be discarded. This approach can improve the system efficiency. Both qualitative and quantitative evaluations on challenging benchmark image sequences demonstrate that the proposed tracking algorithm performs better than several state-of-the-art methods.