People tracking based on predictions and graph-cuts segmentation

Abstract

This paper presents a new approach to segment and track multiple persons in a video sequence via graph-cuts optimization technique. In fact, first, we extract the initial silhouettes that will be modeled by ellipses. Then, a prediction step based on optical flow vectors allows us to detect if an occlusion will handle in the next frame. Hence, we identify the occluding persons by the use of the chi-squared similarity metric based on the intensity histogram and we update the objects models of the interacting persons. Finally, a segmentation based on graph-cuts optimization is performed based on the predicted models. The experimental results show the efficiency of our algorithm to track multiple persons even under occlusion. © 2013 Springer-Verlag.