基于同步多双目立体视觉的高精度人体建模

Even if a high-speed 3D scanner or RGBD camera is adopted to acquire and reconstruct a human body shape, the temporal and spatial inconsistence of the acquired data will occur because the acquisition time is too long to keep the human motionless. Thus it is still a challenge topic to acquire the non-rigid human body shape synchronously and achieve its high-precision reconstruction. In this project, a hardware system of human body acquisition is constructed to capture a human body in terms of high-resolution images, which is composed of tens of digital single lens reflex cameras. All of cameras can be synchronized in 0.01 seconds. To make the system feasible, the following key problems will be investigated deeply, i.e., efficient and accurate multi-camera calibrations, dense point clouds recovery from the skin images of weak texture; robust and accurate registration of multiple pieces of large scale point clouds, detail-preserving human body shape surface reconstruction, compatible parameterizations and high-quality compression of human body models with various poses, etc. As a result, the human body shape can be reconstructed with high-precision, whose error is less than 0.5mm. The research will develop and enrich the related theories and methods in digital geometry processing and stereo vision. Meanwhile, the developed techniques will be integrated into a prototype system of human body shape acquisition and reconstruction. The system will solve the problems of the temporal and spatial inconsistence of acquired data and the reconstruction accuracy in the traditional methods. The research and system will facilitate the human body shape related applications, such as 3D printing, special effects in film and TV, digital entertainment, somatometry, clothing custom, etc.

在进行三维人体模型获取与真实感重建时，即使采用高速三维扫描仪或RGBD相机，仍然存在采集时间长所导致的数据时空不一致问题。因此如何同步地获取非刚性的人体外形数据并实现高精度重建，是一个具有挑战性的问题。本项目拟采用多台数码单反相机、在百分之一秒瞬间同步地采集人体外形的高分辨率图像数据，并通过研究多相机高效精确标定、面向皮肤弱纹理的稠密点云生成、多片大规模点云的鲁棒和精确配准、保细节人体模型曲面重建、不同姿态人体模型一致参数化和高质量压缩等一系列关键科学问题，重构精度可达0.5毫米的三维人体模型，完善和创新数字几何处理和立体视觉的相关理论和方法。同时，将上述研究成果集成为一个面向人体等非刚性对象的三维采集和建模系统，解决传统方法中存在的采集数据时空不一致及其建模精度不高问题，促进三维打印、影视特效、数字娱乐、人体测量、服装定制等应用发展。

人体模型的采集和重建在三维打印、影视特效、数字娱乐、人体测量、服装定制具有重要的应用价值。当前基于三维扫描仪或者RGB-D相机的采集方式，存在采集时间长、采集数据时空不一致的问题，高精度人体模型的采集与重建是一个具有挑战性的问题。..本项目创新性地提出并实现了一个基于同步多双目立体视觉的高精度人体采集和重建系统，该系统有效地克服了人体模型采集过程中存在的数据时空不一致问题。系统的硬件部分包括64台数字单反相机，组成32个双目立体视觉子系统，可以同步曝光拍摄。系统的软件部分包括图像和视觉计算、数字几何处理两个模块。..围绕高精度人体采集和重建系统，项目组成员在图像处理、立体视觉、数字几何处理、三维形状理解和分析等方面取得了一系列重要进展，提出了立体图像的局部颜色校正、加权相机校正、自适应视差计算、层次多视图点云刚性配准、基于多薄板样条的多视图点云非刚性配准、曲率自适应的有向距离场多尺度曲面重建、骨架引导的高效位移细分曲面生成、面向模型协同分割的体形状上下文描述子、三维模型协同对齐的一般框架、模型序列的插值与渐变、特征敏感的光滑自由变形、非刚性模型重建精度测量等一系列新的理论和算法。..实验结果表明，采用新系统进行人体模型采集重建，其平均误差2.48mm、最大误差5.04mm，重建精度远高于当前商品化软件PhotoScan(平均误差9.91mm、最大误差14.35mm)和开源的多视图重建系统COLMAP (平均误差7.544mm、最大误差11.31mm)。上述理论研究、算法研究和系统开发，为该系统在各领域的应用奠定了理论和方法基础。