一种基于区域选择的视频对象分割算法

视频对象分割对于行动识别和视频检索等领域具有重要作用。针对现有方案的不足,提出一种新的视频对象分割算法。首先,基于物质度构建出视频主要对象的区域图,然后,视频对象的区域选择被建模为区域图中最大权重派系的寻找问题,为了避免不合理的区域选择所导致的无法求解问题,对同一派系的区域引入两种互斥约束:帧内约束和帧间约束,最后提出了一种新的最大权重派系(MWC)算法来计算满足约束条件的最大加权派系,从而实现视频对象的精确分割。将本文算法用于SegTrack数据库中的多个高难度基准视频进行测试,实验结果表明,本文算法能够实现每帧视频主要对象检测和分割自动化,且分割误差也要明显小于已有的算法。

An Video Object Segmentation Algorithm Based on the Region Selection

video object segmentation is important for many potential applications, such as activity recognition and video retrieval. Aiming at the deficiencies of the existing schemes, new video object segmentation is proposed. Firstly, the region graph of the primary video object is build with the material degree, secondly, the region selection problem of video object as the problem of finding constrained Maximum Weight Cliques (MWC), in order to avoid can not solve the problem caused by the unreasonable area selection, two types of mutex constraints on regions in the same clique are introduced: intra-frame and inter-frame constraints. Finally, in order to achieve the accurate segmentation of video objects, a new algorithm is proposed to calculate the MWC satisfy the constraints of the Maximum Weight Cliques. Our algorithm is test with the challenging benchmark videos on the SegTrack dataset, the results show that our algorithm can automatically identify the primary object, and segment that object out in every frame, in addition, the Segmentation error of our algorithm is less than the traditional algorithms.