基于神经辐射场算法的混合现实三维重建技术  

3D reconstruction technology of mixed reality based on neural radiance fields algorithm

作　　者：杜旋[1] 黄勇 董惠良[1] 周宇豪 宫正 姚雨龙 曾晰[3] DU Xuan;HUANG Yong;DONG Huiliang;ZHOU Yuhao;GONG Zheng;YAO Yulong;ZENG Xi(China Tobacco Zhejiang Industrial Co.,Ltd.,Hangzhou 310008,China;Zhejiang Machinery and Electrical Group Co.,Ltd.,Hangzhou 310002,China;College of Mechanical Engineering,Zhejiang University of Technology,Hangzhou 310014,China)

机构地区：[1]浙江中烟工业有限责任公司,浙江杭州310008 [2]浙江省机电集团有限公司,浙江杭州310002 [3]浙江工业大学机械工程学院,浙江杭州310014

出　　处：《机电工程》2024年第10期1759-1767,共9页Journal of Mechanical & Electrical Engineering

基　　金：浙江中烟工业有限责任公司科技项目(ZJZY2022E008)。

摘　　要：针对目前在混合现实(MR)环境中高效率建立高质量三维(3D)模型的需求,基于神经辐射场算法(NeRF)的三维重建技术,提出了一种基于Laplacian算子的数据集优化算法。首先,围绕某线切割设备录制了一段1 min 51 s的视频,并采取等距提取视频帧的方式,获取了训练数据集;然后,使用Laplacian算子对数据集进行了优化,同时保留了原始数据集作为对比,使用了基于NeRF算法的重建方式与传统的基于COLMAP的稠密点云重建方式,分别对两组数据集进行了三维重建;最后,在重建精度与重建速度方面,对不同重建方式、不同重建数据集的重建结果进行了比较。研究结果表明:COLMAP稠密点云重建耗时是基于NeRF重建耗时的9.98倍,而相较于COLMAP稠密点云重建,使用NeRF重建方式的模型表面缺陷较少;此外,使用Laplacian算子优化的数据集的NeRF重建在峰值信噪比(PSNR)和结构相似性(SSIM)指标上分别提升了2.43%、0.72%,有利于提升重建模型的质量。研究结果支持混合现实技术在制造业数字化转型中的应用,可为其提供有益的参考。Aiming at the heightened demand for efficiently establishing high-quality three-dimensional(3D)models within mixed reality(MR)environments,a thorough investigation was carried out into the neural radiance fields(NeRF)algorithm-based three-dimensional reconstruction technique.In addition,a dataset optimization algorithm rooted in the Laplacian operator was proposed to augment the overall reconstruction process.Firstly,a 1 min 51 s video around a wire cutting equipment was recorded.Equidistant frame extraction was employed to create a comprehensive training dataset.Then,the Laplacian operator was applied to optimize the dataset,with the original dataset retained for comparative analysis.The NeRF algorithm-based reconstruction method and the traditional dense point cloud reconstruction based on COLMAP method were applied separately to the 3D reconstruction of the two datasets.The reconstruction results,with a focus on precision and speed,were systematically compared across various reconstruction methodologies and datasets.The study results indicate that the reconstruction time of COLMAP dense point cloud is 9.98 times than that of NeRF based reconstruction,and comparing with COLMAP dense point cloud reconstruction,the model surface defects are fewer when using the NeRF reconstruction method.Additionally,the NeRF reconstruction of the dataset optimized by the Laplacian operator has respectively improved the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)metrics by 2.43%and 0.72%,which is beneficial for enhancing the quality of the reconstructed model.This research presents findings that endorse the application of mixed reality technology in the digital transformation of the manufacturing sector,offering valuable insights for pertinent fields.

关 键 词：高质量三维模型 神经辐射场算法 混合现实 重建速度 重建精度 LAPLACIAN算子 数据集优化算法 

分 类 号：TH166[机械工程—机械制造及自动化] TP391.41[自动化与计算机技术—计算机应用技术]