摘要
针对现有三维重建研究中对线特征利用不充分,重建精度低的问题,提出了一种联合地物轮廓线的三维重建算法。利用BDCN模型和相应预处理算法,提取出二维影像上的边缘轮廓线;基于对极几何原理的核线引导和三角测量原理,建立线段的初始匹配假设;通过相似性计算和图聚类算法重建出三维轮廓线模型结果;将结果应用于三维网格模型的优化。使用Building-Imagery-P36和DTU MVS数据集对本文算法进行验证,结果表明:本文方法提取出的轮廓线段精简而完整。相比经典的Line3D++算法,MAE和标准差值平均降低了0.321、0.699,重建时间平均降低了58%,精度和效率都有明显提升。附加三维轮廓线约束后,生成的网格模型视觉效果更佳,细节信息更为丰富。
In response to the insufficiency in utilizing line features and the low reconstruction accuracy in existing 3D reconstruction research,this study proposes a 3D reconstruction algorithm joint for building footprint contours.Firstly,the Bi-Directional Cascade Network(BDCN)model and corresponding preprocessing algorithms are employed to extract edge contours from 2D images.Next,initial matching hypotheses for line segments are established based on the epipolar geometry principles and triangulation principles.Subsequently,a combination of similarity calculations and graph clustering algorithms is used to reconstruct the 3D contour line model.Finally,the results are applied to the optimization of the 3D mesh model.Experimental validation of this algorithm is conducted using the Building-Imagery-P36 and DTU MVS datasets.The results demonstrate that the method presented in this study produces concise and complete contour line segments.Compared to the classical Line3D++algorithm,it exhibits significant improvements in both reconstruction accuracy and efficiency.Furthermore,with the additional 3D contour line constraints,the generated mesh models offer enhanced visual quality and richer detail information.
作者
宫志群
许锋
杨世廷
张栋樑
胡润田
袁婷
GONG Zhiqun;XU Feng;YANG Shiting;ZHANG Dongliang;HU Runtian;YUAN Ting(China Construction Infrastructure Co.,Ltd.,Beijing 100029,China;Guangzhou Urban Planning&Design Survey Research Institute,Guangzhou 510060,China;Ditu(Beijing)Technology Co.,Ltd.,Beijing 100089,China)
出处
《测绘科学》
CSCD
北大核心
2024年第1期97-105,共9页
Science of Surveying and Mapping
关键词
三维重建
线特征
网格优化
图论
深度学习
轮廓提取
3D reconstruction
line feature
grid optimization
graph theory
deep learning
contour extraction
