基于局部特征词袋模型人体动作识别关键帧选取方法  

Key frame selection method for action recognition based on local characteristics with bag-of-words model

作　　者：柳似霖 王颖[2] 吴峰 LIU Silin;WANG Ying;WU Feng(Beijing University of Chemical Technology,Beijing 100029,China;Institute of Microelectr onic,Chinese Academy of Sciences&University of Chinese Academy of Sciences,Beijing 100094,China;Patent Examination Cooperation Tianjin Center,Patent Office of Naitonal Intellectual Property Adminstration,PRC,Tianjin 300304,China)

机构地区：[1]北京化工大学,北京100029 [2]中国科学院微电子研究所&中国科学院大学,北京100094 [3]国家知识产权局专利局专利审查协作天津中心,天津300304

出　　处：《应用光学》2019年第2期265-270,共6页Journal of Applied Optics

基　　金：国家自然科学基金(61340056);北京化工大学中日友好医院联合基金(PYBZ1804)

摘　　要：针对原始动作视频帧数多、信息冗余、计算量大的问题,提出了基于离散粒子群算法的人体动作识别关键帧选取方法。提取视频图像中的时空兴趣点建立视觉词典,统计视频图像中视觉词汇概率分布,采用离散粒子群算法进行关键帧选取,引入原始视频和选取的关键帧的视觉词汇概率分布向量的夹角余弦值,作为最优适应度函数评价关键帧选取前后的动作特征相似性。采用离散粒子群关键帧选取方法对KTH和Weizmann数据库进行实验验证。实验结果显示,该文提出的关键帧选取算法可有效去除动作视频中的冗余帧,提高了动作识别效率,动作识别准确率保持89.17%和98.89%不变。A method of key frames selection based on discrete particle swarm optimization(DPSO)was proposed to avoid large number of frames,redundant information and large amount of computation of original motion video.First,the temporal and spatial interest points of the video images were extracted to create a visual dictionary.Secondly,the probability distribution of the visual words were calculated.Then the key frames were selected based on discrete particle swarm optimization.The angle cosine value of the probability distribution vectors was used to evaluate the similarity between the original video and the selected key frames.The KTH and Weizmann databases were experimentally verified by the discrete particle swarm key frame selection method.Experimental results show that the redundant frame can be removed,the recognition efficiency is improved and the accuracy keeps 89.17%and 98.89%invariable by the proposed method.

关 键 词：时空兴趣点 关键帧 离散粒子群 视觉词汇 动作识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]