基于动态规划融合多模态的足球视频事件分析

Event Analysis in Soccer Video by Dynamic Programming Based Fusion of Multiple Modalities

为了更好地满足用户浏览和检索视频的需要,提出一种融合文本的足球视频事件分析框架.分别从文本和视频中提取事件信息,采用动态规划的算法对2种信息进行全局匹配,对于未匹配的文本事件信息,采用一个全局概率模型估计其在视频中的事件边界.通过寻找文本与视频事件信息的最优全局匹配,有效地避免了局部匹配方法造成的漏检和误检.实验结果表明,文中方法能够快速、准确地检测事件,获得详尽的事件内容信息,性能优于局部匹配算法.

To better satisfy audience＇ demand on browsing and retrieving soccer video, a framework for event analysis in soccer video is proposed that fuses text knowledge and video information. Event information is extracted from text and video, respectively, and then they are globally matched based on dynamic programming （DP） algorithm. For unmatched event information extracted from text, their corresponding event boundaries in video stream are estimated using a global probability model. This approach searches the best global matching for event information extracted from text and video, which avoids missing and mistaking errors causedby other methods using local matching. Experimental results show that the proposed method detects events accurately and efficiently and gets detailed event content information, achieving better performance than other work using local matching methods.