Low-Rank Matrix And Tensor Completion For Video Background Recovery

Most existing video background recovery methods suffer from the deficiency of accuracy or online implementation,which limits their scalability to streaming or big video.Meanwhile,most existing low rank matrix completion algorithms tend to as-sume that the location of missing entries is at random,which cannot handle the case of entire row/column missing.Therefore,this paper proposes an online video background recovery model based on low rank matrix reconstruction.Both a low rank matrix com-pletion model and a low rank tensor completion model are further proposed based on the sparsity of data.The main contents and innovations of this thesis are as follows:1.An online background recovery model from video data is proposed based on ma-trix recovery and motion estimation(OMA-RPCA).Optical flow method is exploited to estimate the motion information of foreground objects.The binary weighting matrix generated from the motion information is integrated into the low rank and sparsity ma-trix decomposition model.The background matrix is decomposed by low rank matrix factorization to avoid the inherent batch-mode implementation of nuclear norm.2.A low-rank matrix completion model against missing rows and columns is pro-posed with separable 2D sparse priors(JPLOSS).The latent matrix is regularized by low rank prior based on the strong correlation between rows/columns.And then,row and column oriented sparse priors are utilized to regularize the matrix according to intra-row and intra-column sparseness.A reweighting scheme is adopted to enhance the low-rankness and sparseness for accuracy boosting.3.A low-rank tensor completion model against structural missing entries is pro-posed based on tensor-train decomposition and nD sparse priors(TCSME).A low rank regularizer is imposed on the matrices unfolded via tensor-train decomposition.The matrices unfolded via Tucker decomposition have sparse representation under over-complete dictionaries based on the sparseness inside the fibers.4.Efficient algorithms are derived for the above proposed models.Experimental results demonstrate the excellent performance of the proposed methods.