基于CNN的三维人体姿态估计方法  

Three-Dimensional Human Pose Estimation Based on Convolution Neural Network

作　　者：肖澳文 刘军[1,2] 张苏沛[1,2] 杜壮 孙思琪 XIAO Aowen;LIU Jun;ZHANG Supei;DU Zhuang;SUN Siqi(Hubei Key Laboratory of Intelligent Robot(Wuhan Institute of Technology),Wuhan 430205,China;School of Computer Science & Engineering,Wuhan Institute of Technology,Wuhan 430205,China)

机构地区：[1]智能机器人湖北省重点实验室(武汉工程大学),湖北武汉430205 [2]武汉工程大学计算机科学与工程学院,湖北武汉430205

出　　处：《武汉工程大学学报》2019年第2期168-172,共5页Journal of Wuhan Institute of Technology

基　　金：国家自然科学基金(61172150;61803286);智能机器人湖北省重点实验室开放基金(HBIR 201802);武汉工程大学第十届研究生教育创新基金(CX2018197;CX2018200;CX2018212)

摘　　要：针对传统三维人体姿态估计受遮挡限制的问题,提出一种基于卷积神经网络(CNN)的三维人体姿态估计方法。首先,实验模型系统采用了几段单目视频为输入源进行人体姿态识别。相对于传统的人体姿态估计方法,改进了一种顺序化的卷积神经网络用于提取人体空间信息和纹理信息。并通过对视频中人体的二维姿态估计,找出了人体头部和四肢关节点的精确位置。最后,通过投影关节点到三维空间,估计出每个人的三维姿态。实验结果表明,本文方法相比传统的姿态估计算法在人体行为上的测试平均误差从98.53 mm降低至92.88 mm,对于视频中的人体三维姿态估计有更优的精度。To solve the problem that the traditional three-dimensional human pose estimation performance was limited by occlusion,this paper presents a three-dimensional human pose estimation method based on convolution neural network.Firstly,some monocular videos were used as the inputs to recognize the human body postures in the experiment model.Secondly,a sequential convolution neural network was constructed to extract the spatial and texture information of human body.Thirdly,the exact position of the joint points of the head and body was found through two-dimensional human pose estimation in the video.Finally,the three-dimensional pose of each person was estimated by projecting the correlation node to the three-dimensional space.The experimental results show that the mean error reduces from 98.53 mm to 92.88 mm compared with the traditional human pose estimation algorithm,and our method has higher precision in the three-dimensional human pose estimation in the testing video.

关 键 词：三维人体姿态估计 卷积神经网络 关节点 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]