An Approach to Crowd Segmentation at Macroscopic Level

摘要

In this paper, an approach to crowd segmentation at macroscopic level is proposed. The method is based on salient points within the region of dense optical flow that are tracked through N frames. Salient points were tracked by using the pyramidal Lukas-Kanade tracker. Around each salient point, an area is formed by using combination of a dense optical flow, a morphological operation dilation and Voronoi diagram. To each subarea, i.e. Voronoi cell of an optical flow (VcOF), the direction and magnitude of displacement is assigned based on the tracklet of the salient point. The VcOFs are grouped based on their directions and distances by hierarchical clustering. The method considers the temporal and spatial dynamics of the crowd motion. It has been tested on the video clips of real-world video sequences and has produced the results that are close to human perception and segmentation of a crowd.