关键词: 稀疏表示/
正则化/
残差矩阵/
目标跟踪
English Abstract
Visual tracking based on the estimation of representation residual matrix
Chen Dian-Bing1,2,Zhu Ming1,
Gao Wen1,
Wang Hui-Li1,2,
Yang Hang1
1.Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, China;
2.University of Chinese Academy of Science, Beijing 100039, China
Fund Project:Project supported by the National Natural Science Foundation of China (Grant No. 61401425), and the Science and Technology Development Plan for Youth Foundation of Jilin Province, China (Grant No. 20150520057JH).Received Date:05 April 2016
Accepted Date:14 July 2016
Published Online:05 October 2016
Abstract:In recent years,sparse representation theory has acquired considerable progress and has extensively been used in visual tracking.Most trackers used the sparse coefficients to merely calculate the position of the target according to the reconstruction error relative to sparse coefficients,and often neglected the information contained by representation residual matrix in representing step.Consequently,we present a novel sparse representation based tracker which takes representation residual matrix into consideration.First of all,at initialization of a new frame,we reconstruct the frame by singular value decomposition (SVD) to eliminate noise and useless information,which contributes a friendly frame for the following representation step.To obtain the compact representation of the target,we build L2-norm regularization according to the distance between the candidates wrapped in particle framework and the reconstruction calculated by dictionary templates and residual matrix.Additionally,we use the L1-norm constraint to restrict the sparse coefficients and the residual matrix of each candidate.Secondly,as the built optimization problem does not have closed-form solution,we design a method to compute the coefficients and the residual matrix iteratively.During each iteration,the coefficients are obtained by solving classical least absolute shrinkage and selectionator operator (LASSO) model,and the residual matrix is achieved by shrinkage operation.After solving the optimization problem,we compute the score of each candidate for evaluating the truth target with considering coefficients and residual matrix.The score is formulated as weighted reconstruction error which consists of dictionary templates,candidates,coefficients and residual matrix. The weight is the exponential value of absolute value of elements in residual matrix.Finally,for capturing the varying appearance of target in series,we update the dictionary template with assembled template,which is composed of residual matrix,selected candidate and dictionary template.In this paper,the template to be replaced is determined according to the score which is inversely proportional to the distance between the selected candidate and each dictionary template. Then we update the dictionary frame by frame during tracking process.Contributions of this work are threefold:1) the representation model captures holistic and local features of target and makes the tracker robust to varying illumination, shape transformation,and background clutter,profiting from preprocessing of SVD reconstruction,the model exhibits a more compact representation of target without disturbance of noisy variance;2) we employ a weight matrix to adjust reconstruction error in candidate evaluation step,as described above,the weight matrix strengthens the effect of error in residual matrix for evaluating candidates from which target is selected,it is noted that weights are all greater than one,which leads to reconstruction error expanding according to the error value of residual matrix,and keeps pixels where there is small error value believable for evaluation;and 3) we adopt an assembled template to update dictionary template and reconstruction of coefficients of selected candidate,which alleviates dictionary degradation and tracking drift problems and provides an accurate description of new appearance of target.In order to illustrate the performance of the proposed tracker,we enforce the algorithm on several challenging sequences and compare the proposed algorithm with five state-of-art methods,whose codes are all supplied by the authors.For complete illustration,both qualitative evaluation and quantitative evaluation are presented in experiment section.Through the experimental results,we could conclude that the proposed algorithm has a more favorable and robust performance than other state-of-art algorithms when dealing with kinds of situations during tracking.
Keywords: sparse representation/
regularization/
representation residual matrix/
visual tracking
