基于时空图像分割和交互区域检测的人体动作识别方法  

Human action recognition method based on spatio-temporal image segmentation and interactive area detection

作　　者：张杰[1] 吴剑章[2] 汤嘉立[1] 范洪辉[1] 

机构地区：[1]江苏理工学院计算机工程学院,江苏常州213001 [2]东南大学计算机科学与工程学院,南京210096

出　　处：《计算机应用研究》2017年第1期302-305,320,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61302124);江苏省自然科学基金资助项目(BK20130235);江苏省产学研联合创新基金资助项目(BY2014038-05);江苏省高校自然科学研究面上项目(16KJB520010)

摘　　要：针对现有人体动作识别方法没有考虑到非人体目标的作用,提出一种基于时空图像分割和目标交互区域检测的人体动作识别方法。在视频流中检测出人体轮廓,并将其进行时空图像分段形成关键段区域;然后,扩展分段使其包含与人体交互的非人体目标,通过时空梯度方向直方图(HOG)和光流场方向直方图(HOF)描述符来表示关键段的静态和动态特征,并通过K-均值算法构建成码书,同时采用局部约束线性编码(LLC)技术来优化码书;最后采用非线性支持向量机(SVM)对特征进行学习并进行动作识别。实验结果表明,与现有基于兴趣点的方法相比,该方案获得了较高的动作识别率。For the issues that the existing human action recognition method has not considered the role of non human targets, this paper proposed a human action recognition scheme base on spatio-temporal image segmentation and interactive area detection. First, this scheme detected the human body contour in the video stream, and formed the key section by spatio-temporal image segmentation. Then, it extended segment to make it contain a non human object interaction with human body. After that, it used spatio-temporal histogram of oriented gradient（HOG） and histograms of optical flow（HOF） descriptors to represent the static and dynamic characteristics of the key segments, and used K-means algorithm to build a code book, while it used local linear encoding（LLC） technology to optimize the code book. Finally, it used the nonlinear support vector machine （SVM） to study the characteristics of the feature and perform the action recognition. Experimental results show that the proposed scheme has a higher recognition rate compared with the existing methods.

关 键 词：人体动作识别 时空图像分割 交互区域 局部约束线性编码 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]