基于单张图像的人体准确姿势3D重建研究  

3D reconstruction research of accurate human posture based on single image

作　　者：裘志超 姚剑敏[1,2] 严群 林志贤 QIU Zhichao;YAO Jianmin;YAN Qun;LIN Zhixian(College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China;Jinjiang RichSense Electronic Technology Co Ltd,Jinjiang 362200,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350108 [2]晋江市博感电子科技有限公司,福建晋江362200

出　　处：《传感器与微系统》2023年第5期61-64,共4页Transducer and Microsystem Technologies

基　　金：国家重点研发计划资助项目(2016YFB0401503);广东省科技重大专项项目(2016B090906001);福建省科技重大专项项目(2014HZ0003-1);广东省光信息材料与技术重点实验室开放基金资助项目(2017B030301007)。

摘　　要：为了解决单张图像人体重建时存在部分姿势不准确,模型精确度不足的问题,提出了基于单张图像进行人体准确姿势3D重建的方法。首先,提取图像特征,并将特征与SMPL基础模板信息相融合,通过图卷积网络将2D图像映射为3D人体表面;其次,将参数化SMPL模型所表达的3D人体投影到2D平面,通过投影后的3D信息与输入图像的2D信息相匹配,实现体型与姿势的估计;然后,使用生成对抗网络判断模型数据是否来自未配对数据的真实人体网格数据;最后,使用多种损失函数共同训练,生成最终生成器模型。使用Human 3.6M测试时与SMPLify, NBF,Rogez等算法进行了对比,姿势估计准确度有所提升。In order to solve the problem of inaccuracy of some postures and inaccuracy of model in single image human body posture reconstruction,a method of 3D human body reconstruction based on single image is proposed.Firstly,image features are extracted and fused with SMPL basic template information.The 2D image is mapped to 3D human body surface by graph convolution network.Secondly,the 3D human body expressed by parametric SMPL model is projected onto 2D plane.The body shape and posture are estimated by matching the projected 2D information with the 2D information of the input image.Then,the generated confrontation network is used to judge whether the model data is from the real human grid data of unpaired data.Finally,the final generator model is generated by training with a variety of loss functions.When testing with Human 3.6M,it is compared with SMPLify,NBF,Rogez,the accuracy of pose estimation is improved.

关 键 词：3D人体重建 SMPL模型 图卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]