Video human segmentation based on multiple-cue integration

Abstract

A multiple-cue integration based video human segmentation model is proposed in this paper. The proposed model firstly describes the local color distribution in an image via training competing 1SVMs and computes shape priors of video human objects by optical flow-based motion estimation. And then the local color model and the shape priors are integrated into the image segmentation framework. Additionally, the weights of different energy terms in the segmentation model are adaptively adjusted according to relevant feedback cues. The experiments have indicated that the proposed model can describe well the foreground (for example, human object) change from video in complex environments. Besides improving the segmentation accuracy for the current frame, the proposed model improves the automatic segmentation accuracy for the subsequent frames because of significantly reducing the adverse impacts on the segmentation for the subsequent frames caused by segmentation error of the current frame.