基于主题相似度的视频分段  

Video segmentation based on topic similarity

作　　者：周敏[1] 陈昌红[1] 干宗良[1] 刘峰[1] 

机构地区：[1]南京邮电大学江苏省图像处理与图像通信重点实验室,江苏南京210003

出　　处：《南京邮电大学学报（自然科学版）》2016年第6期15-23,共9页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61172118;61471201);江苏省自然科学基金(BK20130867);江苏省高校自然科学研究重大项目(13KJA510004);南京大学计算机软件新技术国家重点实验室开放课题基金(KFKT2014B10)资助项目

摘　　要：视频中的人体行为分析是计算机视觉领域的研究热点,视频分段是人体行为分析的基础。文中提出了基于主题相似度的视频分段方法,首先提取视频的时空特征,并利用词袋模型将视频中的人体行为表示成行为文档,对行为文档进行主题建模得到蕴含在文档中的主题特征,然后计算两个相邻滑动窗口内主题分布的相似度,绘制出相似度曲线,最后根据相似度曲线进行分段。实验结果表明,文中算法能够有效地将视频中的不同行为分割开来,提高后续行为识别工作的效率。Human behavior analysis in video is the hottest field in computer vision. The video segmentation is the basis of human behavior analysis. The video segmentation based on topic similarity is presented. Firstly, the human behavior is expressed as a behavior document using space-time interest points and bag of words model. After representing the human behavior as document, a topic model is established to capture the topic features occurring in it. Then, the similarity of topic distribution between two side-by-side windows is calculated. With these topic similarities, the similarity curve is obtained. Finally, the human behavior in video is segmented according to the curve. Experimental results show that the video segmentation based on topic similarity can effectively separate different behaviors effectively, improving the efficiency of follow-up behavior recognition.

关 键 词：视频分段 主题相似度 词袋模型 LDA 

分 类 号：TN919.8[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]