针对现有动态背景下目标分割算法存在的局限性,提出了一种融合运动线索和颜色信息的视频序列目标分割算法。首先,设计了一种新的运动轨迹分类方法,利用背景运动的低秩特性,结合累积确认的策略,可以获得准确的运动轨迹分类结果;然后,通过过分割算法获取视频序列的超像素集合,并计算超像素之间颜色信息的相似度;最后,以超像素为节点建立马尔可夫随机场模型,将运动轨迹分类信息以及超像素之间颜色信息统一建模在马尔可夫随机场的能量函数中,并通过能量函数最小化获得每个超像素的最优分类。在多组公开发布的视频序列中进行测试与对比,结果表明,本文方法可以准确分割出动态背景下的运动目标,并且较传统方法具有更高的分割准确率。
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 dynarnic 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 o ver-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 dem- onstrate that the proposed method can accurately segment the moving objects from the dynamic back- ground,and it has better segmentation accuracy compared with classical methods.