三维人体模型姿态与形状重构

为了解决单张照片人体重构出现的姿态翻转问题,提高重构模型的准确度,提出相邻帧姿态约束和人体轮廓线匹配的姿态与形状序列同时重构算法.对视频中的每一帧,首先估计出图像中人物的二维关节点、人物脸部特征点及其边缘轮廓线;然后将参数化模型SMPL所表达的三维人体投影到二维平面上,使得投影后的二维信息与对应的视频帧二维信息相匹配;最后通过调整SMPL的姿态与形状参数来最小化匹配能量函数,从而重构出与视频帧中人物具有相似姿态与形状的三维人体模型.此外,为了使重构结果显得更真实,也对图像帧中人体的头部姿态进行了检测和匹配.该算法在MPI-INF-3DHP数据集、Youku视频和自拍视频帧上均进行了实验,实验结果表明,与SMPLify算法等相比,该算法能有效地避免重构结果中出现姿态翻转的现象,且能在保证模型整体姿态相似性的前提下重构出准确的头部姿态和相似的模型形状.

Human Model Pose and Shape Reconstruction

In order to avoid pose inversion in human reconstruction from a single image,this paper presents an approach to automatically reconstruct a 3D human pose sequence from a video based on adjacent frame pose constraint and human contour matching.Given a video frame,it firstly estimates 2D joint positions,and extracts facial feature points as well as 2D contour lines of the character.Then it fits the SMPL,a statistical body shape model,to the aforementioned 2D information by minimizing an energy function.Finally,the approach reconstructs a sequence of 3D mesh models similar to the poses and shapes of the character in the video.The algorithm has been tested on videos from MPI-INF-3DHP datasets and Youku and Selfies.The experimental results show that the proposed method can effectively avoid pose inversion,reconstruct correct head postures and obtain better body shapes compared to the SMPLify approach and other existing methods.