Moving object detection via RPCA framework using non-convex low-rank approximation and total variational regularization

Moving object detection is one of the significant tasks in computer vision. Robust Principal Component Analysis (RPCA) is a common method for moving object detection. However, this algorithm fails to effectively utilize the low-rank prior information of the background and the spatio-temporal continuity of the foreground, and exhibits degraded performance in the existence of shadows, photometric variations, rapidly moving objects, dynamic backgrounds, and jitters. Therefore, a new model via RPCA framework using non-convex low-rank approximation and total variational regularization is proposed to detect moving objects. Firstly, this paper introduces a non-convex function to deal with the problem that the nuclear norm excessively penalizes large singular values for the sake of constraining the low-rank characteristics of video background availably. Then, the sparsity is strengthened with the l(1)-norm and the spatio-temporal continuity of the foreground is explored by TV regularization. Finally, the augmented Lagrange multiplier algorithm is expanded by the alternating direction multiplier strategy to solve the model. Extensive experiments show that the proposed method outperforms existing methods in terms of the accuracy of moving object detection and foreground extraction effect.