基于运动特征与序列袋的人体动作识别  

Body motion recognition based on moving feature coupled bag-of-sequence

作　　者：冯小明 冯乃光[2] 汪云云 FENG Xiao-ming;FENG Nai-guang;WANG Yun-yun(Engineering Training Center,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;College of Engineering and Technology,Sichuan Radio and TV University,Chengdu 610073,China)

机构地区：[1]南京邮电大学工程训练中心,江苏南京210003 [2]四川广播电视大学工程技术学院,四川成都610073

出　　处：《计算机工程与设计》2018年第10期3220-3227,共8页Computer Engineering and Design

基　　金：国家自然科学基金青年基金项目(61300165);南京邮电大学引进人才基金项目(NY213033);高等学校博士学科点专项科研基金项目(20133223120009);2016年四川省教育厅重点科研课题基金项目(16ZA0425)

摘　　要：针对当前的动作识别算法难以有效识别复杂动作的问题,提出一种人体行为识别算法。利用改进的密集轨迹,将视频表示为多个基础动作序列(primitive actions,PA),将其编码为PA的特征序列,通过仿射传播,将其变为索引序列;基于序列模式挖掘(sequential pattern mining,SPM),形成不同的序列袋(bag of sequence,BOS)模型;对BOS模型进行学习,计算其序列比对特征、外观匹配特征、序列集特征,构成动作的评分函数;引入线性判别分析(linear discriminat analysis,LDA),对动作的评分值进行分类学习,完成动作识别。在MSR3D与UCF-Sport数据集上进行测试,实验结果表明,面对各种复杂动作,所提算法具有更高的识别精度与稳定性。Aiming at the problem that it is difficult to recognize complex actions effectively in the current action recognition algorithm,a complex motion recognition algorithm based on motion feature coupling bag of pocket was proposed.The video was represented as multiple PA(primitive actions)sequences using an improved dense trajectory,which encoded a sequence of features.By affine propagation,the feature sequence was transformed into PA index sequence.The resulting PA index sequences were combined by sequential pattern mining(SPM)to form a different sequence of BOS.The BOS model was learnt,its sequence alignment features,appearance matching features,sequence set features were calculated.Linear discriminate analysis(LDA)was introduced to classify learning and complete the action recognition.MSR3D and UCF-Sport data were used to test the algorithm,experimental results show that the proposed algorithm has higher recognition accuracy and stability in the face of complex actions.

关 键 词：动作识别 基础动作序列 序列袋 仿射传播 序列模式挖掘 线性判别分析 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]