Background Subtraction Based on Online Tensor Decomposition

Background subtraction is the most popular technique for foreground detection to automatically separate foreground objects from background in video sequences. Low- rank subspace learning methods have attracted great attentions in recent years. Compared with the methods based on matrix decomposition which reshape the video data into a matrix, the methods based on tensor decomposition can take good advantage of the spatial information within frames. However, most existing methods based on tensor decomposition are batch-based approaches which generally require heavy memory cost and are unable to adapt to dynamic scenes. In this paper, we propose a background subtraction method based on adaptive online tensor decomposition. The proposed method deals with samples sequentially and also adapts to dynamic environments, which can help reduce the computational burden and memory cost. Besides, adaptive weighting parameter is employed to adjust the degree of attention to the sparsity of estimated sparse component, which can considerably improve the performance of foreground detection. Experimental results demonstrate that, the proposed method can effectively detect the foreground objects with less computational cost.