Research On Sparse Representation Based Visual Tracking Methods Under Complicated Conditions

Visual tracking is a hot problem in computer vision,and has huge development potential in such fields as intelligent monitoring,car navigation,advanced human-computer interaction,and so forth.However,affected by the posture and shape change of the target,fast motion,noises,occlusions,cluttered background etc.,it is still a challenging problem to realize the robust target tracking.Recently,sparse representation based visual tracking becomes a more popular tracking method,especially performs satisfactorily in the scenes of noises and partial occlusions.The potential problems of sparse representation based trackers are efficiency and accuracy.In terms of efficiency,this method needs to solve a 1 norm regularized minimum problem for each candidate,therefore brings the computational burden and time consuming.In terms of accuracy,the potential instability of 1 decomposition could affect the accuracy of sparse representation,and consequently affect the accuracy of visual tracking.Aiming at the above mentioned two problems,this paper focuses on the following four aspects to improve performance of sparse representation based visual tracking methods in various complex scenes.First,aiming at the rich multimodal feature information contained in the tracked target of an image sequence,two multimodal feature based visual tracking methods are proposed.Firstly,each feature of each candidate is represented respectively by the sparse representation,and the corresponding reconstruction error is computed.Then the total reconstruction error of each candidate is obtained by adding all the ones of multimodal features of the candidate,and is used to determine the target.Secondly,aiming at the deviation problem caused by the previous method due to ignoring the correlation of multimodal features of the same candidate,a tracking method based on joint sparse representation for the multimodal features of a candidate is proposed.The total reconstruction error of a candidate is computed by adding the reconstruction error of each feature,which is used to compute the observation probability of each candidate and determine the target.Second,the data locality correlation between candidates and target templates is integrated into the sparse representation based tracking framework.First of all,the target templates are weighted according to their similarity with candidates to reflect their locality.Then a weighted sparse representation is used to make sure that the more similar target template to a candidate is apt to be chosen to represent it,which can reduce the deviation and improve the tracking accuracy.Third,a multi-cue weighted sparse representation based visual tracking is proposed.Not only the data locality correlation between candidates and target templates and the correlation among the moltimodal features of candidates exist as mentioned above,but also a similarity correlation among candidates exists due to the fact that all the candidates are sampled in a relatively small region.By considering these points,the proposed tracking method integrates the data locality between candidates and target templates,the correlation of multimodal features of candidates and the similarity among candidates into sparse representation based tracking framework,and effectively improves the tracking accuracy in various complex scenes.Fourth,a kernel sparse representation based tracking method is proposed to deal with the fast motion and blur of tracked object.In the scene of fast motion and blur,the candidates and target templates are mapped nonlinearly into a high dimensional Helbert space by using an appropriate kernel function.Then the candidates are represented sparsely in the high dimensional space.Thus the nonlinear relation between candidates and target templates,which can't be dealt with by the sparse representation based visual tracking,can be effectively coped with,and the tracking accuracy of the whole system is increased in the scene of fast motion and blur.The research fully considers available priori information of the visual tracking problem,effectively improve the performance of sparse representation based visual tracking method in various complex tracking scene,and lays the foundation to some degree for further research and practical application of visual tracking.