基于Kinect 3D节点的连续HMM手语识别  

Continuous HMM based on Kinect 3D joint nodes for sign language recognition

作　　者：沈娟[1,2] 王硕[2] 郭丹[2] 

机构地区：[1]安徽新华学院信息工程学院,安徽合肥230088 [2]合肥工业大学计算机与信息学院,安徽合肥230009

出　　处：《合肥工业大学学报（自然科学版）》2017年第5期638-642,共5页Journal of Hefei University of Technology：Natural Science

基　　金：国家自然科学基金资助项目(61305062);安徽新华学院自然科学重点研究资助项目(2016zr007)

摘　　要：文章提出了一种基于Kinect 3D节点的连续隐马尔科夫模型(hidden Markov model,HMM)手语识别方法。首先对Kinect 3D节点三维坐标采用距离换算的方法获取其骨架特征表达,转换后的特征维度减少为原来节点特征的2/3,降低了计算过程中的存储开销;再针对人体体型大小所带来的差异,设置最小-最大归一法及最大值归一化方法;在此基础上,为强化骨架特征变化的特征表达,文章进而提出了用来捕获对骨架动态变化度量的相对归一化法;最后,在最终所获得的骨架特征表达上,构建基于高斯混合的隐马尔科夫模型(hidden Markov model with Gaussian mixture models,GMM-HMM)进行手势识别。实验证明采用相对归一法能够消除原3D节点中坐标漂移、体型各异、动态手势变化导致难以表达的弊端,实现了有效的骨架特征表达,并在识别精度上有了较好的提升。This paper presents a continuous hidden Markov model（HMM） method based on Kinect 3D joint nodes for sign language recognition. Firstly, the skeleton feature of each frame is obtained by distance calculation among five Kineet joint nodes, namely head, left elbow, right elbow, left hand and right hand. The dimension of skeleton feature is reduced to 2/3 of the original coordinate feature, and the storage cost is also reduced. Then the minimum-maximum and the maximum normalization methods are used to eliminate the diversity of different signers＇ shape and habitus. In order to capture dynamics among skeleton gestures in a video, a relative normalization approach on above normalized skeleton feature is proposed. Finally, the hidden Markov model with Gaussian mixture models（GMMHMM） based on above normalization approaches on skeleton features for sign language recognition is constructed. Experimental results show that the relative normalization method can solve shortcomings of the drift of 3D coordinates of joint nodes, diversity of signers＇ shape and habitus and complexity of dynamic gesture changes. The method is effective and improves the recognition accuracy.

关 键 词：Kinect传感器 3D节点 归一化 手语识别 隐马尔科夫模型(HMM) 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]