A feature point matching based approach for video objects segmentation

Abstract

In this paper, we propose an approach to segment the multiple objects in video. For a video sequence with stationary background, our approach combines the feature points with the color and contrast information to extract the multiple objects of different sizes. The idea is that the local features of the feature points are more robust than that of the pixels, and more accurate than the global color feature. So we integrate the local cues of the feature points into the basic color model in graph cut. Our method matches the feature points in the known background and the current image, and classifies them in three categories. Then the influences to their neighbor pixels are computed according to the category, and integrated in the pixels' color model. The max-flow algorithm is applied to obtain the last result of the segmentation. Experimental results demonstrate the effectiveness of our approach.