Crowd segmentation based on fusion of appearance and motion features

Crowd segmentation is an important topic in a visual surveillance system. In this paper, crowd segmentation is formulated as a problem to cluster the feature points inside the foreground region with a set of rectangles. Coherent motion of feature points in an individual are fused with appearance cues around the feature points for crowd segmentation, which has improved the segmentation performance. Furthermore, three descriptors are proposed to extract the points with a non-articulated movement. Some results on the CAVIAR dataset have been shown. The results show that coherent motion cue can be used more reliably by considering the points with rigid motion only.