Robust Object Tracking Via Correction Dictionary And Sparse Concept Coding

Object tracking in video sequences is a major research topic in computer version. It has a wide range of application in real world, which including intelligent transportation, military security, monitoring systems, medical research facilities and human-computer interaction. Object tracking still has many complex issues to handing at current stage, which including: how to build robust appearance model, which not only keeps the basic apparent features of target, but also adapt to changes in the scene that the object is deformed; how to extract features have more discriminating power, and these features can enhance identification and detection the tracker. The current object tracking algorithms have significant limitations in the scene that object deformation and partial occlusion. This dissertation mainly devotes to the research work on the theory of sparse representation and object tracker based on it, the sparse concept coding and its applications. The main contents and contributions as follows:1. The theory of particle filter witch based on sequential importance sampling and its applications in object tracking are studied. The framework of object tracking which based on sparse representation is discussed, the limitations and difficulties of current object tracking algorithm are propose.2. Some typical object tracking algorithm is studied. Based on sparse discriminant classifier, the robust object tracking via dynamic weighted correction dictionary is first proposed. The correction dictionary is efficiently and accurately which is composed by the last tracking result. Extensive experiments show that the dynamic weighted correction dictionary has the role of error correction and this make the tracker more robust in the scene that the object is deformed.3. The theory of sparse concept coding and its application in visual analysis are discussed. Inspired by the matrix factorization based on sparse concept coding, the robust object tracking via sparse concept coding is first proposed. This tracker combined the advantage of structure information of patch and the power of sparse concept coding, which improved the accuracy of the algorithm. Extensive experiments show that this tracker is robust in the scene partial occlusion.