Object Tracking Based On Improved Sparse Representation

Object tracking is one of the most important research fields of computer vision, which is thefoundation of solving many computer vision problems. Tracking is to find the precise location andobtain other useful information of the target of interest in the sequences of successive video images.In recent years, object tracking is widely used in various fields, such as intelligent videosurveillance, intelligent human-computer interaction, military navigation, and so on. As a result, it isworthy to hammer at the object tracking.In this paper, according to the sparsity of the solution of the sparse model, an object trackingalgorithm based on improved sparse representation is proposed. This method is to represent thetarget by sparse representation, and then to track the target in the particle filter framework. Themain research contents of this paper are as follows:(1) L1/2regularization is discussed, and then an iterative algorithm is put forward to solve thisproblem, which transforms the l1/2regularization problem into a series of weighted l1regularizationproblems. And experiments show that the solution of l1/2regularization is sparser than l1regularization.(2) An object representation based on local sparse coding is introduced. By solving the l1/2regularization problem, the target is represented by the sparse coefficients, so it is easier todistinguish the target from the background. Then an effective tracking algorithm based on theparticle filter framework is proposed. During the tracking period, the information of the constantlyupdated initial frame and the last frame is used for accurate tracking. The experiments are carriedout on five challenging video sequences. These sequences cover many complex situations such asocclusion, illumination changes, fast moving and so on. Then the performance of the proposedalgorithm is compared with some state-of-the-art tracking algorithms. The proposed algorithmresults in good tracking performance.(3) At last, an object tracking algorithm via sparse secondary choice is proposed. The secondarychoice of target candidates can screen the candidates effectively, as a result of reducing thedimensionality of image processing in the particle framework. The experiments demonstrate thatthe proposed algorithm can keep the balance between accuracy and real-time capability.