基于SMPL模型人体三维重建算法研究  

Research on 3D reconstruction algorithm of human body based on SMPL model

作　　者：王栾栾 严群 姚剑敏[1,2] 林志贤 WANG Luanluan;YAN Qun;YAO Jianmin;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]博感电子科技有限公司,福建晋江362200

出　　处：《传感器与微系统》2022年第10期59-63,共5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资助项目(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：为了实现人体模型的高效快速建模,获得丰富的重建细节,提出了基于SMPL模型的三维人体重建算法。首先,通过图像分割与关键点检测,将2D图像映射到3D人体表面,实现人体关节点精确定位和姿势估计;其次,通过U-Net++和PatchGAN对抗生成网络,计算出人体三维图像的法线贴图和位移贴图;最后,将贴图信息采用线性差值方法添加到SMPL模型表面,实现模型顶点和SMPL模型参数的校正。在Human 3.6 M数据集上将本方法与SMPLify, HMR,NBF等算法进行了对比试验,本文提出的方法可以较完整重建人体的表面褶皱及凹凸细节,在13.7 s内实现完成重建任务。In order to achieve efficient and rapid modeling of human body model and obtain rich reconstruction details, a 3 D human body reconstruction algorithm based on the skinned multi-person linear(SMPL) model is proposed.Firstly, through image segmentation and key point detection, the 2 D image is mapped to the 3 D human body surface to get precise positioning and pose estimation of the human body joint points.Secondly, the normal map and displacement map of the human body 3 D image is calculated through U-Net++ and PatchGAN confrontation generation network.Finally, the mapping information is added to the SMPL model surface using the linear difference method to achieve the correction of the model vertices and the SMPL model parameters.This method is compared with SMPLify, HMR,NBF and other algorithms on the Human 3.6 M dataset.The proposed method can completely reconstruct the surface folds and bumps of the human body, and complete the reconstruction task in 13.7 s.

关 键 词：SMPL模型 深度学习 三维重建 法线贴图 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]