A novel robust statistical method for background initialization and visual surveillance

Abstract

In many visual tracking and surveillance systems, it is important to initialize a background model using a training video sequence which may include foreground objects. In such a case, robust statistical methods are required to handle random occurrences of foreground objects (i.e., outliers), as well as general image noise. The robust statistical method Median has been employed for initializing the background model. However, the Median can tolerate up to only 50% outliers, which cannot satisfy the requirements of some complicated environments. In this paper, we propose a novel robust method for the background initialization. The proposed method can tolerate more than 50% of foreground pixels and noise. We give quantitative evaluations on a number of video sequences and compare our proposed method with five other methods. Experiments show that our method can achieve very promising results in background initialization: including applications in video segmentation, visual tracking and surveillance.