基于全局优化的实时高精度模型重建  

Real-time high-precision model reconstruction based on global optimization

作　　者：许新傲 李艺璇 钱佳铭 冯世杰 左超[1,2,3] XU Xin-ao;LI Yi-xuan;QIAN Jia-ming;FENG Shi-jie;ZUO Chao(Smart Computational Imaging Laboratory(SCILab),School of Electronic and Optical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;Smart Computational Imaging Research Institute(SCIRI)of Nanjing University of Science and Technology,Nanjing 210019,China;Jiangsu Key Laboratory of Spectral Imaging&Intelligent Sense,Nanjing 210094,China)

机构地区：[1]南京理工大学电子工程与光电技术学院智能计算成像实验室(SCILab),江苏南京210094 [2]南京理工大学智能计算成像研究院(SCIRI),江苏南京210019 [3]江苏省光谱成像与智能感知重点实验室,江苏南京210094

出　　处：《液晶与显示》2023年第6期748-758,I0002,共12页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.62075096,No.62005121,No.U21B2033);国家重点研发计划(No.2022YFB2804600,No.2022YFB2804603);江苏省基础研究计划前沿引领技术(No.BK20192003);江苏省“333工程”科研项目资助计划(No.BRA2016407);江苏省“一带一路”创新合作项目(No.BZ2020007);中央高校科研专项(No.30921011208,No.30919011222,No.30920032101);江苏省光谱成像与智能感知重点实验室开放基金(No.JSGP202105,No.JSGP202201);江苏省研究生科研与实践创新计划(No.KYCX22_0412)。

摘　　要：三维形貌测量在先进制造、航空航天、生物医学等领域发挥着重要的应用。凭借高精度、全视场、非接触等优点,条纹投影轮廓术是目前使用最广泛的一种光学三维测量手段。为了获得物体全局三维信息,通常需要将待测物置于转台之上,通过不断地扫描和拼接来获得物体的全局信息。然而,传统的扫描和拼接是以离线的方式进行的,导致整个三维模型的重建速度缓慢。现有的实时点云配准方法虽然能够有效提高点云扫描与拼接的速度,但实时点云拼接的精度依然受待测物的运动状态影响。本文针对上述问题进行优化改进,提出一种基于全局优化的实时高精度模型重建方法。首先,介绍了一种由粗配准到精配准的快速点云配准算法并提出了基于点云法向量约束的点云初始化算法,能够提升粗配准过程中点云初始位姿计算的稳定性与精度。其次,在精配准阶段引入了图优化算法以获得全局点云位姿的最优解,进一步提升了全局点云配准的精度。实验结果表明,所提方法相比于现有实时模型重建方法,能够实现更高精度且稳定的全局点云配准。特别地,针对动态场景中由于抖动等因素引起的被测物体速度突变等情况,本方法依然能够鲁棒地完成三维模型重建,全方位模型重建的精度达84μm。Three-dimensional(3D)shape measurement plays an important role in advanced manufacturing,aerospace,biomedicine and other fields.With the advantages of high precision,full field of view,and noncontact,fringe projection profilometry is currently the most widely used optical three-dimensional measurement method.In order to obtain 360°global three-dimensional information,it is usually necessary to place the object to be measured on the turntable,and obtain the global information of the object by continuous scanning and stitching.However,the traditional scanning and stitching are performed offline,resulting in slow reconstruction of the entire 3D model.Although the existing real-time point cloud registration methods can effectively improve the speed of point cloud scanning and stitching,the accuracy of real-time point cloud stitching is still limited by the motion state of the object under test.This paper optimizes and improves the above problems,and proposes a real-time high-precision model reconstruction method based on global optimization.Firstly,a fast point cloud registration algorithm from coarse to fine registration is introduced,and a point cloud initialization algorithm based on point cloud normal vector constraints is proposed on this basis,which can improve the stability and accuracy of the point cloud initial pose calculated during the rough registration process.Secondly,a graph optimization algorithm is introduced in the fine registration stage to obtain the optimal solution of the global point cloud pose,which further improves the accuracy of global point cloud registration.The experimental results show that the proposed method can achieve higher precision and stable global point cloud registration than the current real-time point cloud registration method.In particular,this method can still robustly complete the reconstruction of the 3D model,and the accuracy of the omni-directional model reconstruction reaches 84μm,especially for situations such as sudden changes in the speed of the measured object c

关 键 词：条纹投影轮廓术 图优化 实时 三维重建 点云配准 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]