A framework for background detection in video

Abstract

This paper presents a framework for background detection in video. First, key frames are extracted to capture background change in video and reduce the magnitude of the data. Then we analyze the content of the key frames to determine whether there is an interesting background in them. A timeconstrained clustering algorithm is exploited for key frame extraction. Background detection in the key frame is done with color and texture cues. The illumination varies much in natural scenes. To deal with the varying illumination, color is modeled with three sub-models: strong light, normal light and weak light. The connectivity of background pixels is used to reduce the computing cost of texture. Experimental results show that background can be detected simply and efficiently under the framework.