四维时空兴趣点提取结合多流形判别分析的人体动作识别  

Human Action Recognition Based on Extracting Four-Dimensional Space-Time Interest Points and Multiple Manifolds Discriminant Analysis

作　　者：王向前[1] 张月芬[2] 

机构地区：[1]平顶山学院师范教育学院,河南平顶山467000 [2]云南大学网络与信息中心,云南昆明650091

出　　处：《湘潭大学自然科学学报》2016年第1期111-114,共4页Natural Science Journal of Xiangtan University

基　　金：河南省教师教育课程改革研究项目(2015-JSJYYB-139)

摘　　要：针对人体动作识别方法较差的稳定性和视角选择的局限性问题,提出了四维时空兴趣点提取结合多流形判别分析(MMDA)的人体动作识别方法.首先,将每个动作通过三维空间体和四维时空兴趣点投影到任意视角;然后,构建运动历史图像和非运动历史图像,并使用类增强主成分分析进行降维;最后,将降维后的矩阵构建为多流形,计算测试图像流形与各个训练图像流形之间的距离,利用最近邻分类器完成识别.在IXMAS数据集上的实验结果表明,相比其他几种动作识别方法,提出的方法取得了更高的识别率,且对任意视角都具有较好的鲁棒性.As the poor stability of action recognition method and the limitation of choosing viewpoints,a new view invariant action recognition based on four-dimensional space-time interest points and multiple manifolds discriminant analysis is proposed.Firstly,each action is projected to arbitrary viewpoints with 3D space volumes and 4D-STIP.Then,motion history images and non-motion history images are constructed,and the feature dimension is reduced by class augmented principal component analysis.Finally,matrixes with reduced dimension are constructed into multiple manifolds,the manifold distances between testing image and training mages are calculated,and nearest neighbor classifier is used to finish recognition.The experimental results on the IXMAS data set show that the proposed method has better recognition accuracy than several other outstanding algorithms,and it is robust to arbitrary universal perspective.

关 键 词：视角不变 人体动作识别 多流形判别分析(MMDA) 四维时空感兴趣点 最近邻分类器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]