Research Of Online Visual Object Tracking Algorithms Based On Sparse Representation

Object tracking is a challenging basic research topic in computer vision,which has wide application prospects in video surveillance,intelligent transportation,human-computer interaction,visual navigation,medical analysis and other fields.Object tracking is to track the appearance and state changes of the target in each frame of the video image sequence.There are many technical problems in the tracking task,such as object occlusion,illumination change,complex background,motion blur,in-plane rotation and out-of-plane rotation,which makes designing a robust tracking algorithm become much more challenging.In this paper,the application fields of object tracking are summarized,and the classical object tracking algorithms and theories at home and abroad are analyzed,and then the problems faced in tracking are explained and summarized.On this basis,the existing object tracking algorithms based on sparse representation are improved,and the robust object tracking is achieved by designing an effective appearance model.The main research results of this paper are summarized as follows:(1)Aiming at the problem of low tracking accuracy in sparse prototypes tracking without considering the density of the orthogonal template coefficients,a robust object tracking algorithm based on L1-L2 norm simultaneous constraint is proposed.First of all,an object representation model via L1-L2 norm simultaneous constraint is build.The L2 norm regularization constraint on PCA basis template coefficients ensures the density of the PCA basis template coefficients,and the L1 norm regularization constraint on trivial template coefficients guarantees the sparsity of error term in the optimization model.Then,ridge regression and soft threshold shrinkage methods are used to solve the PCA basis template coefficients and the trivial template coefficients.Finally,in particle filter framework,the object tracking is realized by the observation model established with the reconstruction error of the unoccluded part and sparse noise term,combined with the online template updating mechanism.Experimental results show that the proposed algorithm has better tracking performance compared with other classical algorithms.(2)In view of the fact that the gray feature of image block is used as object template which leads to modeling object appearance inaccurately in the low-rank sparse tracking algorithm,a robust visual tracking based on online low-rank sparse representation is proposed.Firstly,in order to make full use of the PCA basis vector to describe the object appearance change as well as take object occlusion into account,the observation value of target is represented with PCA basis vector templates and trivial templates linearly.Secondly,a low-rank sparse representation optimization model is proposed,which adds a low-rank and L1,1 norm regularization constraint on the coefficients related with the PCA basis templates and a L1,1 norm regularization on the coefficients related with the trivial templates,thus the Inexact Augmented Lagrange Multiplier method is adopted to solve the representation coefficients.Thirdly,object tracking is achieved by taking the reconstruction error of the unoccluded part and the sparse error term as an observation model in particle filter framework.Finally,to avoid model drifting,an occlusion detection updating mechanism is used to update PCA basis templates.Through the comprehensive analysis,the tracking results on several image sequences demonstrate that the proposed method has better tracking accuracy and robustness than those existing state-of-art trackers.