Automatic Real-Time Video Background Segmentation System

Abstract

This paper proposes an automatic background seg- mentation system based on mixture background models and lbp contrast. An adaptive per-pixel background model is developed to set the data cost of an image graph. lbp constrast is adopted to capture the contrast information. We verify our algorithm in different video sequences including public video sequences. Experimental results of composite video demonstrate that our proposed algorithm is efficient. I. INTRODUCTION Foreground segmentation is important and has a lot of ap- plications including background substitution and composition, human-computer interaction, compression, object tracking, etc. In this paper, we focus on how to segment foreground objects from monocular video streams. Recently, some automatic video matting methods for natural scenes are proposed using binocular and monocular video seg- mentation methods. Kolmogorov et al. (1) presented binocular stereo video segmentation algorithm to fuse color, contrast and stereo matching information so as to infer layers accurately and efficiently. However, this method requires synchronized stereo input so as to obtain valuable depth information. In contrast, monocular video segmentation is more popular for its convenience and low cost. The authors of (2)-(4) fused dif- ferent cues to achieve the monocular foreground segmentation based on graph cut algorithm. Although satisfactory results are obtained in their experiments, they pay less attention to the problem of shadows that foreground objects cast on the background. In fact, it is often necessary to separate the foreground object from their shadows in video segmentation (6). In this paper, we present an automatic background segmen- tation system under the circumstance of a static background and fixed web-camera. In the dynamic graph cut framework, an adaptive background model which combines the advantages of both Eigenbackground and pixel-based gaussian models is proposed to set the data cost. And the link cost based on lbp contrast is introduced to improve the segmentation spatial smooth. The automatic segmentation system we proposed can automatically generate high-quality and real-time composite video with shadow removal.