【摘要】 人体姿态估计是计算机视觉领域的重要分支,是人机交互领域的关键问题.现有的三维人体姿态估计算法识别结果大多以三维关节点、线的形式来体现,缺乏人体细节信息且表达形式抽象.引入参数化人体模型(SMPL-X),研究基于形变模型的人体姿态估计与形状重构算法,通过二维图像预测模型参数,实现标准人体模板与真实数据的非刚体配准.首先使用HMR生成对抗网络从彩色图像中提取模型姿态参数,再将模型重投影回二维,利用人体关键点和轮廓的约束构造能量函数并对姿态和体型参数进行优化求解,从而重构出与图像中人物具有相似姿态和形状的三维人体模型.在3DPW和EHF公开数据集及真实数据的实验结果表明,相较基于SMPL模型的方法,重建出包含面部和手部细节的人体表示,姿态与体型估计更贴合人体,提高了重建精确度.更多还原

【Abstract】 Human pose estimation is an important branch in the field of computer vision and a key issue in the field of human-computer interaction.The recognition results of existing 3D human body pose estimation algorithms are mostly embodied in the form of 3D joint points and lines, lacking human body detail information and the expression form is abstract.This paper introduces the parametric human body model(SMPL-X) to study the human body pose estimation and shape reconstruction algorithm based on the deformation model.Through the two-dimensional image prediction model parameters, the non-rigid body registration of the standard human body template and the real data is realized.Firstly, the pose parameters of the model are extracted from the color image by using HMR generated adversation network, and then the model is projected back to two-dimensional, the energy function is constructed by using the constraints of the key points and contour of the human body, and the pose and shape parameters are optimized and solved, so as to reconstruct the three-dimensional human body model with similar pose and shape to the figure in the image.The experimental results on 3DPW and EHF public data sets and real data show that compared with the method based on the SMPL model, the human body representation containing the details of the face and hands is reconstructed.The pose and body shape estimation are more suitable for the human body, and the reconstruction accuracy is improved.更多还原