一种基于语义轨迹的事件规则学习算法  

A Rule Learning Algorithm for Event Detection Based on Semantic Trajectory

作　　者：姜广[1,2] 王晓峰[3] 史忠植[1] 

机构地区：[1]中国科学院计算技术研究所智能信息处理重点实验室,北京100190 [2]中国科学院大学,北京100049 [3]中国科学院计算技术研究所无线传感器网络实验室,北京100190

出　　处：《计算机研究与发展》2012年第12期2623-2630,共8页Journal of Computer Research and Development

基　　金：国家自然科学基金项目(61035003;60775035;60933004;60970088;60903141;61072085);国家"九七三"重点基础研究发展计划基金项目(2007CB311004)

摘　　要：视频上的事件探测对于视频检索与语义理解是一个很重要的工作.视频中的轨迹不仅记录了物体的移动信息,也反映了物体移动的动机,并与事件的发生密切相关.主要探讨了如何从轨迹抽取事件.然而,基于内容的视频事件分析中,从视频中抽取的低层特征与高层的语义特征存在一定的鸿沟.因此,利用领域知识标记的兴趣区域,提出一种新的语义轨迹表示方法,从而将视频中得到的原始轨迹转化为语义轨迹.同时,使用物体与兴趣区域关系的正则表达式描述视频中的语义事件.基于归纳学习的事件规则学习算法显示了正则表达式比传统的一阶谓词上的合式公式更易于学习.利用学习得到的事件规则可以很好地用于视频中语义事件的探测.最后,实验表明了事件探测的有效性。Event detection from video is important work for video retrieval and semantic understanding. Trajectories of moving objects in the video not only record the moving information, but also reflect the motivations of the moving objects, and are closely related with the event. However, the raw trajectory is only geographic information of an object without any domain knowledge. Meanwhile, semantic gap exists between the low-level feature extracted and the according high-level concept in the content-based video analysis. Thus, it is critical to combine both the raw trajectory and its semantic information. In this regard, extracting event using the semantic trajectories which are analyzed from the video is studied, domain knowledge is utilized to label interest areas in the video, and a new semantic trajectory representation is proposed which includes interest areas the object stops and passes by. Moreover, the original trajectory of the interest object can be converted into an according semantic trajectory, so video event can be represented as regular expressions of relationship between objects and interest areas. Inspired by FOIL （first order inductive learner）, an inductive-based event rule learning algorithm is proposed, and the regular expression is illustrated to be easier learned than the traditional well-formed formula in the first-order predicate logic. Finally, experiment results indicate the performance.

关 键 词：视频事件 视频语义分析 事件探测 语义轨迹 规则学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]