电影中吸烟活动识别  

Smoking recognition in movies

作　　者：叶果[1] 程洪[1] 赵洋[1] 

机构地区：[1]电子科技大学自动化工程学院,四川成都611731

出　　处：《智能系统学报》2011年第5期440-444,共5页CAAI Transactions on Intelligent Systems

基　　金：国家"973"计划资助项目(2011CB707000);国家自然科学基金资助项目(61075045);中央高校基本科研业务费专项基金资助项目(ZYGX2009X013);新世纪优秀人才支持计划资助项目(Y02020023901067)

摘　　要：电影中的活动识别是计算机视觉领域的一个难点问题.传统识别算法受到电影中镜头视角变化、场景变化和光照变化等因素的影响,使得其对于真实场景活动识别的效果较差.针对上述问题,提出一种新颖的基于互信息的组合识别方法.该方法以纯贝叶斯互信息最大化构造初始框架,针对"吸烟"这类极具代表性的动作,将活动的SIFT信息和STIP信息融合得到最优的组合分类器.该方法在电影《咖啡和烟》中进行了测试,实验结果表明,该方法具有很好的鲁棒性,并且很大程度上提高了抽烟活动的识别率.Action recognition in movies is a difficult problem in the computer vision domain. Traditional approaches have a bad recognition effect because they are subjected to viewpoint changes, scene changes, and illuminationchanges in real scenes. This paper presented a novel combined recognition approach, using mutual information to solve the problems mentioned above. This method builds the initial skeleton using naive-Bayesian mutual informa-tion maximization （NBMIM） and combines the shape information with the motion information to recognize smoking, which is a typical activity in movies. The proposed smoking recognition approach was evaluated in the film Coffeeand Cigarettes. The results indicate that the proposed method is robust, and it significantly improves the recognition rate.

关 键 词：电影 吸烟活动识别 纯贝叶斯互信息最大化 计算机视觉 模式识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]