Video foreground and background separation via Gaussian scale mixture and generalized nuclear norm based robust principal component analysis

Abstract

With the rapid development of information technology, the video foreground and background separation has attracted much interests in the computer version. Recently, robust principal component analysis (RPCA) has been widely used in the task of video foreground and background separation. However, the current RPCA methods have more challenge since they are difficult to find the optimal surrogate functions which can describe the sparse and low-rank components of videos very well. This paper proposes a novel RPCA method based on Gaussian scale mixture and general nuclear norm (GSMGNN) to facilitate video foreground and background separation efficiently. In this novel method, we use the Gaussian mixture scale and the nonconvex generalized nuclear norm to describe the sparse foreground and the low-rank background, respectively. Moreover, considering that noise is always existed in the practical scene, we extend the proposed GSMGNN method to the robust Gaussian scale mixture and general nuclear norm (RGSMGNN) method by inducing the noise item. And then, we adopt the efficient alternating direction method of multipliers (ADMM) method to solve these two new proposed methods. At last, we apply these two proposed methods to the video foreground and background separation experiments, and observe their significant encouraging results compared to the existing RPCA-based video foreground and background separation methods.