A regularized tensor decomposition method with adaptive rank adjustment for Compressed-Sensed-Domain background subtraction

In intelligent video surveillance, Background Subtraction is the foundation and key to the task of Moving Object Detection (MOD). Recently, the development of Compressive Sensing (CS) theory has made the Compressed-Sensed-Domain Background Subtraction (CSDBS) an interesting task and several related models have been proposed. The latest tensor-based method provides the best performance on both reconstructing video and detecting moving objects at present. Unfortunately, it just simply sets the ranks of the tensor video background fixed along all modes, which makes it impossible to obtain accurate background component when the scene is at different time or places. In this paper, we propose a Regularized Tensor Decomposition Method with Adaptive Rank Adjustment (RTDARA) for CSDBS, which can accommodate backgrounds with differently low-rank property in more scenes to a certain extent. In addition, for the model, we use a non-convex surrogate of the rank instead of the convex nuclear norm. Finally, we develop a fast implementation using the alternative direction multiplier method (ADMM) to solve the proposed model. A large number of experimental results have shown that, on the Compressed-Sensed-Domain video in different scenes, our proposed method is superior over the existing state of the art techniques.