Research On Sparse Representation Method Of Target Tracking Based On Multi-feature Fusion

With the development of computer vision technology,the application of computer vision has spread in everyone's life.Besides,the research of computer vision technology has also attracted the attention of domestic and foreign researchers.As one of the key technologies in the field of computer vision,target tracking has been widely applied in military,industry,business et.al.Although target tracking technology has been widely used at present,the research of robust visual tracking still faces many challenges such as dynamic occlusion and illumination changes et.al.Based on the sparse representation theory,aiming at some key problems in visual tracking,this paper proposes two new sparse representation based visual tracking methods to improve tracking accuracy.The main works are summarized as follows:(1)Since the sparse representation theory is the key point in our research,we first deeply study the sparse representation theory and analyze the limitations of several classical sparse representation models in visual tracking.And then,we establish a sparse representation based particle filter framework for robust visual tracking.(2)Due to the limitation of using single feature to representate particle samples in traditional sparse representation based tracking algorithm,a new multi-kernel fusion based sparse representation tracking method is proposed.This method is based on the particle filter framework,and uses multi-kernel fusion to realize the complementary of the pixel intensity and the target state.The proposed method constrains the similarity between different particles by using a mixing norm,and realizes the multi-task sparse representation by the observation grouping according to the similarity of different observation vectors.(3)To use the target appearance to represent the sampled particles well,a correlation analysis based sparse representation model is proposed.The proposed model first uses two appearance features to represent the sampled particles,and then uses the observational vectors to build sparse representation model.The constructed model can realize the complementary fusion of different features by exploring the correlation between the features in the subspace,and enhance the robustness of the sparse representation model in the complex monitoring environment.