Algorithm Research On Object Tracking Based On Sparse Representation

With the rapid development of computer technology, video image processing and computational ability have been greatly improved, and computer vision technology has also made rapid development. Visual object tracking technology is a key problem in computer vision field. Although many experts and scholars have studied for a long time and achieved better research results, it is still faced with various challenges to track targets under complex background due to interference factors such as occlusion, illumination changes, rotation, etc. How to design an efficient and robust object tracking algorithm to deal with all sorts of challenges is still an urgent problem to be solved.This paper mainly studies object tracking algorithm based on sparse representation, and does main work as follows: research and improvement based on L1 tracking algorithm: there are two aspects of problems about this algorithm, one is that sparse assumption does not always work; second, computational complexity is too high, leading to bad real-time algorithm. To solve above problems, an improved algorithm was proposed, using greatly disturbed templates instead of small positive and negative templates to construct a sparse dictionary. Experiments show that this algorithm is improved in real-time, but tracking robustness is affected under complicated situations. An improved tracking algorithm based on local structural sparse representation: first, cross pixel sampling similar checkerboard from the gray feature space; then, a new local structural method is adopted to construct a sparse dictionary based on sampled image. Experiments show that the improved algorithm improved real-time while guaranteeing tracking accuracy.Discriminative tracking algorithm based on sparse representation: a hybrid-template method based on global template and local structural template is adopted to construct object appearance model; reconstruction error is achieved via sparse representation of target candidates using global template, and similarity function is achieved via sparse representation of target candidate using local sparse dictionary; then, the optimal discriminative results are achieved as tracking results using logistic discriminative function combining two modeling methods. Experiments on six sequences interfered by different noise are carried out and compared with methods using global template and local structural template alone. Experimental results demonstrate that the proposed algorithm can deal with challenges caused by occlusions, illumination changes and scale in terms of efficiency and robustness.