An Attempt to Segment Foreground in Dynamic Scenes

In general, human behavior analysis relies on a sequence of human segments, e.g. gait recognition aims to address human identification based on people's manners of walking, and thus relies on the segmented silhouettes. Background subtraction is the most widely used approach to segment foreground, while dynamic scenes make it difficult to work. In this paper, we propose to combine Mean-Shift-based tracking with adaptive scale and Graphcuts-based segmentation with label propagation. The average precision on a number of sequences is 0.82, and the average recall is 0.72. Besides, our method only requires weak user interaction and is computationally efficient. We compare our method with its variant without label propagation, as well as GrabCut. For the tracking module only, we compare Mean Shift with several state-of-the-art methods (i.e. OnlineBoost, SemiBoost, MILTrack, FragTrack).