基于中国人数据集的参数化人体建模

Parametric human modeling based on Chinese database

为解决基于中国人数据集的参数化人体模型重建问题,本文首先采集了152名中国成年女性净体样本,并对其进行头发部位去除、泊松重建、降采样和脚底平面切削。然后采用“粗—精”两步配准策略,先基于BPS对点云进行高效学习,将SMPL模型网格点快速初始配置至扫描网格点附近;再采用基于ICP的非刚性网格配准算法进行精配准,只优化顶点位移D分量,生成精准匹配的SMPLD模型。配准完成后,比较其和LoopReg配准的配准精度和效率。之后,对配准数据进行位姿校正和主成分分析,从而获得能够反映中国人体特征的形体参数。最后进行模型重建,并将其与SMPL平均模板进行比较。由平均模型拟合结果可知,相比SMPL模型,基于重建模型的拟合结果关节误差和顶点误差均有所降低,3个实验对象的关节平均误差分别降低了26.2%、19.4%、21.7%;顶点平均误差分别降低了20.0%、16.1%、12.5%。

The high-precision modeling of the virtual human body is an important prerequisite for various human-related studies.At present,high-precision 3 D human modeling technology mainly relies on complex and expensive multi view camera system or laser scanner.With the proposed SMPL model,the complex 3 D reconstruction problem can be transformed into a series of parameter optimization problems by using the human body prior constraints provided by the model.However,as the model is generated based on Western human body data,there are certain problems of compatibility with Chinese human body.In order to solve this problem,we first collected the net body samples of 152 Chinese adult women,and removed the hair,and carried out Poisson reconstruction,downsampling and plantar plane cutting.Then,a two-step "coarse-fine" registration strategy was adopted:firstly,the point cloud was efficiently learned based on BPS,and the grid points of SMPL model were quickly and initially configured near the scanning grid points;then,the non rigid mesh registration algorithm based on ICP was used for fine registration,and only the D component of vertex displacement was optimized to generate an accurately matched SMPLD model.After registration,the registration accuracy and efficiency of the above-mentioned SMPLD and LoopReg registration were compared.Then,pose correction and principal component analysis were carried out on the registration data,so as to obtain the shape parameters that can reflect the characteristics of the human body of Chinese people.Finally,the model was reconstructed and compared with the SMPL average template.The data show that compared with the advanced registration algorithm LoopReg,the average accuracy of BPS coarse registration is improved by more than two times.Moreover,based on the same platform,the non-rigid matching of the traditional human body after three-dimensional scanning takes about 15 minutes per sample(including posture pre-registration,nonlinear iteration,etc.).The BPS pre-registration based on deep learning used in this paper not only ensures relatively high matching accuracy,but also takes only 0.5 seconds per sample on average.Meanwhile,the fine registration process based on Pytorch 1.17.0 + cu11.0 takes only about 50 seconds per sample on average.The two-step registration method of BPS coarse registration + ICP based non-rigid mesh fine registration not only ensures the registration accuracy of nonlinear models,but also greatly improves the registration efficiency.Compared with the fitting results based on the SMPL model,the joint error and vertex error of the fitting results based on the reconstruction model are reduced.The joint average error and vertex average error of experimental object 1 are reduced by 26.2% and 20.0%,respectively;the average joint error and the average vertex error of subject 2 are reduced by 19.4% and 16.1%,respectively;the average joint error and the average vertex error of subject 3 are reduced by 21.7% and 12.5%,respectively.On the whole,the reconstructed model has smaller fitting error and higher accuracy.