Background modeling is one of the key techniques in video surveillance system. When the training images contain more moving objects or its number is not sufficient, the existing methods normally end up with incorrect background estimates. In this paper, we study a type of method on data analysis, i.e., Robust Principle Component Analysis (RPCA), and present its application on the background modeling. Unlike previous approaches based on statistics, the new method uses an advanced convex optimization technique that is theoretically guaranteed to be robust to large errors. Experimental results demonstrate that the proposed solution can robustly estimate the background from relatively few training images, even in the case of sudden change of lighting.