A video object segmentation algorithm for dynamic background combining motion cue with color information

Zhi-gao CUI,Ai-hua LI,Guo-yan FENG

DOI: https://doi.org/10.16136/j.joel.2014.08.022

2014-01-01

Abstract:Video object segmentation is important for object tracking,video surveillance and semantic classification.In order to overcome the limitation of existing video object segmentation methods under dynamic background,a video segmentation algorithm based on motion cue and color information is proposed in this paper.At first,a new motion trajectory classification method is designed.The proposed method can accurately divide the motion trajectory set into background and moving object ones by combining the low rank constraint and cumulative acknowledgment strategy.Then,the superpixels are acquired by over-segmenting method.And the color similarity of adjacent superpixels is computed according to their common boundary.At last,taking the superpixels as node,an energy function of Markov random field model is designed,which has combined motion trajectory classification information and color similarity of superpixels.The classification of each superpixel can be obtained by finding the minimum of the energy function.The proposed algorithm is tested on several publicly available videos.Experimental results demonstrate that the proposed method can accurately segment the moving objects from the dynamic background,and it has better segmentation accuracy compared with classical methods.