Background Subtraction Algorithms Based On Tensor

Background subtraction technology is one of the important technologies in the field of video surveillance,which subtracts the background from videos to extract the moving foreground.We further study background subtraction on the basis of Robust Principal Component Analysis(RPCA),which has been well developed in recent years.At the same time,background subtraction extended to high-dimensional RPCA is also further studied.The research results are as follows:(1)A background subtraction model based on logarithm rank function and structured sparsity is proposed.Segmentation and index tree are used to dynamically process the image foreground,which enhances the appearance similarity and spatial continuity between pixels.And the C(2,1)norm is applied to constrain the sparsity of the image blocks,which strengthens the structured sparsity of foreground.Then,the model utilizes the logarithmic rank function to constrain the low rank of background,which considers the influence of different singular values for the rank function.Experimental results show that the proposed algorithm can effectively deal with dynamic background,slow movement and camera jitter.(2)A new model based on tensor nuclear norm and 3D Total Variation(3D-TV)is proposed.On the basis of tensor RPCA,improved tensor nuclear norm is exploited to constrain the background,which strengthens the low rank of background and preserves the spatial information of videos.Then the moving objects are regularized by 3D-TV.It considers the spatio-temporal continuity of moving foreground and effectively suppresses the interference of dynamic background and target movement on foreground extraction.Experiments show that the proposed model effectively improves the accuracy of foreground and background separation,and suppresses the interference of complex weather and target movement on foreground extraction.(3)In order to further improve the performance of the above model(2),the L_1,1,2norm instead of 3D-TV is applied to constrain foreground,which enhances the tube sparsity and spatio-temporal continuity of foreground and improves the accuracy of foreground extraction.The experiments demonstrate that our algorithm can effectively separate the foreground and background of videos.And from the qualitative and quantitative aspects,our new approach has superior performance in the multi-target and bad weather videos.