改进高斯核函数的人体姿态分析与识别  

Human postures recognition based on the improved Gauss kernel function

作　　者：林海波[1] 王浩[1] 张毅[1] 

机构地区：[1]重庆邮电大学智能系统及机器人研究所,重庆400065

出　　处：《智能系统学报》2015年第3期436-441,共6页CAAI Transactions on Intelligent Systems

基　　金：科技部国际合作项目(2010DFA12160);重庆市工业振兴专项资金资助项目(渝财金[2013]442号)

摘　　要：为了提高人体动作姿态的识别率,利用Kinect平台构建人体骨骼模型,提出一种基于关节角度的人体姿态表示方法。同时针对传统的高斯核函数中采用欧氏距离计算方法难以完全反映人体关节运动数据样本点与测试点之间位置关系的问题,提出了改进的高斯核函数多类支持向量机(MSVM)人体动作姿态识别方法。在高斯径向基核函数中使用测地线距离代替欧氏距离,建立了基于测地线距离的姿态核函数,采用二叉树方法构建多类支持向量机完成12种上肢姿态的分类。实验结果表明,该算法取得了较好的识别效果,能更加有效识别人体姿态。In this paper,a method based on the joint angles of human postures is proposed in order to improve the human posture recognition rate through building a human skeleton model on the Kinect platform. For the traditional method of human postures recognition,Euclidean distance is used in Gaussian kernel function,but the positional relationship of sample point and test point of human body joint can not be reflected completely. So the method of improved Gaussian kernel function and multi-class support vector machines（ MSVM） is proposed. Using the geodesic distance instead of the Euclidean distance in the Gaussian radial basis kernel function,a posture kernel function based on the geodesic distance is established. Using the binary tree method,a multi-class support vector machine is built to complete classification of 12 kinds of upper limb postures. Experimental results showed that the improved algorithm can identify body postures more effectively than before,achieving a good recognition effect.

关 键 词：人体动作姿态 识别 高斯核函数 KINECT 欧氏距离 测地线距离 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]