Multi-cue-Based Crowd Segmentation in Stereo Visio

摘要

People counting and human detection have always been important objectives in visual surveillance. With the decrease in the cost of stereo cameras, they can potentially be used to develop new algorithms and achieve better accuracy. This paper introduces a multi-cue-based method for individual person segmentation in stereo vision. Shape cues inside the crowd are explored with a block-based Implicit Shape Model. Depth cues are obtained from the disparity values of some foreground blobs, which are calculated concurrently during crowd segmentation. Crowd segmentation is therefore achieved with evidences from both shape and depth cues. The methods were evaluated on two video sequences. The results show that the segmentation performance has been improved when depth cues are considered.