Moving object detection via low-rank total variation regularization

Abstract

Moving object detection is a challenging task in video surveillance. Recently proposed Robust Principal Component Analysis (RPCA) can recover the outlier patterns from the low-rank data under some mild conditions. However, the e-penalty in RPCA doesn’t work well in moving object detection because the irrepresentable condition is often not satisfied. In this paper, a method based on total variation (TV) regularization scheme is proposed. In our model, image sequences captured with a static camera are highly related, which can be described using a low-rank matrix. Meanwhile, the low-rank matrix can absorb background motion, e.g. periodic and random perturbation. The foreground objects in the sequence are usually sparsely distributed and drifting continuously, and can be treated as group outliers from the highly-related background scenes. Instead of e-penalty, we exploit the total variation of the foreground. By minimizing the total variation energy, the outliers tend to collapse and finally converge to be the exact moving objects. The TV-penalty is superior to the e-penalty especially when the outlier is in the majority for some pixels, and our method can estimate the outlier explicitly with less bias but higher variance. To solve the problem, a joint optimization function is formulated and can be effectively solved through the inexact Augmented Lagrange Multiplier (ALM) method. We evaluate our method along with several state-of-the-art approaches in MATLAB. Both qualitative and quantitative results demonstrate that our proposed method works effectively on a large range of complex scenarios.