Research On Object Tracking Technology Based On Sparse Representation And Particle Filter

Object tracking belongs to the field of computer vision and it has important applications in many fields,such as human life,social production,service and security.Because of its wide and complex application environment,object tracking faces a series of challenges such as occlusion,illumination change,deformation,complex background and so on.In order to solve these problems,key technologies in the object tracking system based on sparse representation and particle filter is researched in this thesis.The algorithm of this thesis is composed of motion model and appearance model.For the motion model,the traditional motion model based on particle filter theory is improved in this thesis.The particle weights are computed by reverse sparse representation to remove particles that do not contribute to tracking.The experimental results show that the proposed algorithm not only reduces the accuracy of the algorithm,but also reduces the computation and improves the tracking speed.The appearance model in this thesis is combined by the discriminative model and the generative model.The discriminative model is constructed based on sparse representation.The confidence value of the discriminative model combines the foreground reconstruction error and the background reconstruction error so that it can judge the foreground and background better.The weight of the local module of the candidate target is obtained by combining the target reconstruction error and the center point distance in the generative model based on the structured sparse representation.Structured sparse representation indicates the integration of global information,local information and spatial information,so there is a certain spatial relationship between sparse solutions.The location relation of sparse representation coefficients of local image blocks is used to weigh,align and assemble the coefficients,then we combine the weights of the local module to construct the generative model of this thesis.We construct a more accurate and robust joint appearance model based on discriminative model and generative model.The dictionary that is fixed in the tracking process cannot cope with changes in the shape and environment of the target.In the update process of the generative model dictionary template,we give the weight of each foreground template in the dictionary and update the foreground dictionary with the similarity between the local image block and the local dictionary template,the occlusion problem is well dealt with.The proposed algorithm is compared with many mainstream algorithm in a series of test videos.We compare the tracking performance of some mainstream algorithms with different tracking problems through qualitative comparison and quantitative comparison.Experiments show that this algorithm has higher accuracy and robustness than other algorithms.