Background subtraction based on tensor nuclear and L_1,1,2 norm

Background subtraction technology is a very important part in the field of video surveillance applications. The common matrix decomposition methods based on robust principal component analysis vectorized video sequences, which destroys the spatial structure and spatio-temporal continuity of videos. Aiming at this problem, a model based on tensor robust principal component analysis was proposed. In the new model, an improved tensor nuclear norm was used to constrain the background, which strengthened the low rank of background and improved the accuracy of foreground background separation. And L_1,1,2 norm was applied to constrain foreground and enhanced the tube sparsity and spatio-temporal continuity of foreground, which improved the accuracy of foreground extraction. Experimental results show that the proposed algorithm outperforms multiple state-of-the-art methods both in qualitative and qualitative aspects, especially, for the videos with the multi-target or in bad weather.