利用单目图像重建人体三维模型  

Reconstruction of a Three-dimensional Human Body Model Using Monocular Images

作　　者：钱融 王勇[1] 王瑛[1] QIAN Rong;WANG Yong;WANG Ying(School of Computer,Guangdong University of Technology,Guangzhou Guangdong 510006,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《信息与电脑》2021年第5期60-64,共5页Information & Computer

基　　金：广东省科技项目“消防应急救援车辆关键技术研发与产业化”(项目编号:2015B090923001)。

摘　　要：人体三维模型在科幻电影、网上购物的模拟试衣等方面有广泛的应用场景,但是在单目图像重建中存在三维信息缺失、重建模型不具有贴合的三维表面等问题。为了解决上述的问题,笔者提出基于SMPL模型的人体三维模型重建算法。该算法先预估人物的二维关节点,使用SMPL模型关节与预估的二维关节相匹配,最后利用人体三维模型数据库的姿势信息对重建的人体模型进行姿势先验,使得重建模型具有合理的姿态与形状。实验结果表明,该算法能有效预估人体关节的三维位置,且能重建与图像人物姿势、形态相似的人体三维模型。The human body 3D model are widely used in science fiction movies,online shopping simulation fittings,etc,but there is a lack of 3D information in monocular image reconstruction,and the reconstructed model does not have problems such as a fit 3D surface.In order to solve the above mentioned problems,a human body 3D model reconstruction algorithm based on SMPL model is proposed.The algorithm first estimates the two-dimensional joint points of the character,and uses the SMPL model joints to match the estimated two-dimensional joints;finally,the posture information of the three-dimensional human body model database is used to perform posture prior to the reconstructed human body model,making the reconstructed model reasonable Posture and shape.The algorithm was tested on the PI-INF-3DHP data set.The experimental results show that the algorithm can effectively predict the 3D position of human joints,and can reconstruct a 3D model of the human body similar to the pose and shape of the image.

关 键 词：人体姿势估计 三维人体重建 单目图像重建 人体形状姿势 SMPL模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]