Robust Visual Tracking Via Online Low-rank Sparse Representation
Fanqiang KONG,Dandan WANG,Qiu SHEN,Chending BIAN,Xiaole YAN
DOI: https://doi.org/10.15961/j.jsuese.201600511
2017-01-01
Abstract:The gray value features of images were used to describe targets in robust visual tracking via L1 minimization (L1 tracking algorithm).However,the ignorance of structural information between the templates and inaccurate modeling of target appearance changes leaded to a low tracking accuracy.What's more,since the low resolution image block with 12× 15 was adopted to balance the tracking speed and tracking effect,it was difficult to obtain sufficient information to represent the target.To solve the problem of the L1 tracking algorithm,a visual tracking method based on online low-rank sparse representation was proposed.Firstly,in order to take advantage of the ability to represent target appearance changes of principal component analysis (PCA) basis vectors,the PCA basis vector template was used to describe target appearance changes.To take occlusion into account,the occlusion and other abnormal noise were modeled with the trivial template.Consequently,the candidate target was represented as a linear combination of PCA basis templates and trivial templates.Secondly,for the optimization problem of object representation model,a low-rank and L1,1 regularization constraint were added on the PCA basis template coefficients and a L1,1 regularization constraint on the trivial template coefficients.Then the inexact augmented lagrange multiplier(IALM) method was adopted to solve the representation coefficients.Thirdly,visual tracking was achieved by taking the reconstruction error of the unoccluded part and sparse error term as an observation model in particle filter framework.Finally,to avoid model drifting,an occlusion detection updating mechanism was proposed to update templates.In the experiments of 8 video image sequences with 32×32 image block resolution,compared with four state-of-the-art trackers,the proposed algorithm had achieved the highest average overlap rate of 0.78 and the lowest average center error of 4.05.Accordingly,the experimental results demonstrated that the proposed algorithm had better tracking accuracy and robustness.