A Novel Tensor Robust PCA Approach for Background Subtraction from Compressive Measurements

Background subtraction so far has been a fundamental and widely studied task in the field of video analysis, with a wide range of applications to video surveillance, teleconferencing and 3D modeling. Recently, due to the emergence of massive video data, background subtraction from compressive measurements (BSCM) provides a promising way for reducing the burden of storing and transmitting large video data. In this paper, we propose a novel tensor robust PCA approach for the task of BSCM. We utilize tensor low-rank approximation to eliminate the spatial-temporal redundancy in the video backgrounds, and design a multi-scale 3D total variation regularizer to enforce the smoothness of video foreground. Furthermore, we introduce the non-local self-similarity of video cuboid into our model to further reduce the video redundancy. The alternating direction method of multipliers (ADMM) is employed to solve the proposed optimization models. Extensive experiments on simulated and real-world videos demonstrate the superiority of the proposed models over the state-of-the-art approaches.