A Target Tracking Algorithm Based On Subregion Sparse Representation

Object tracking is one of the most important research fields in computer vision.It is widely used in fields of guidance,traffic and so on.In recent years,the tracking algorithm based on sparse coding has become a hot research topic in the field of tracking,and its important steps include constructing sparse dictionary,sparse representation and so on.The sparse coefficient and its distribution can reflect the important information of the object image,which can reduce the influence of the surrounding complex environment on characteristics of the target by means of the high dimensional space.Through the analysis of existing tracking algorithms based on sparse dictionary,object representation methods based sparsing are divided into two categories: one is that the object model is constructed by off-line learning.Although the constructed over complete sparse dictionary contains rich information of the target,it can not achieve real-time when it is applied to the target tracking algorithm.Another method is obtaining templates of different characteristics by a variety of affine transformation of the target image obtained in the first frame.Global features of the target are obtained from the space of these templates.This representation method can achieve real-time,but it can not be said that the local hidden information of the target,because global features are used to represent the target which contains more useless information.It can not effectively express the target when environment is complicated.In order to express the target quickly and effectively by using the limited target information,and distinguish the foreground and the background according to structural characteristics of the target itself which is that the edge region of target is mainly the edge feature,while the central region of it mainly includes scale invariance,rotation invariance,brightness invariance and other features.HOG feature vectors are extracted from the edge region of the target,and the central region is represented by SIFT,which constructs a sparse block dictionary with two blocks of local features.At the same time,in order to distinguish the background and the foreground effectively,the vector basis of the background is added when the sparse dictionary is constructed.L1 and L2 mixed norm are used to express the sparsity of the sparse dictionary proposed in this paper,and the APG algorithm is used to solve the sparsity of the target.In order to update the template library and recapture the lost target,on the basis of a method of block structure with multi feature partition domain sparse dictionary presented in this paper,it is determined whether the target is occluded according to the difference of the reconstruction error among candidate blocks,so as to determine whether to update the template library.In order to prevent the error accumulation,a new method of recapturing the lost target is proposed which can prevent the problem of drift and loss in the process of target tracking.Through experiments and algorithm evaluation,method based on the domain partition block multi feature sparse dictionary object representation presented in this paper can effectively use known information of the target to express object in the complex environment and when the target changes fast.The template update principle and target loss recapture method based on the target representation method proposed in this paper can effectively update the target template library,capture the lost target and solve drift,loss and other issues in the process of target tracking.Compared with other target tracking algorithms based on sparse encoding,the proposed target tracking algorithm is less time-consuming.