Automatic 3D Reconstruction for As-built Underground Utilities

Abstract

This thesis addresses three challenges in the 3D reconstruction of underground utilities. Firstly, it introduces a high-precision ground penetrating radar-based deep learning model to enhance localization accuracy. Secondly, it presents an unsupervised deep learning model for effective 3D reconstruction under low-light conditions. Finally, it proposes a graph convolutional network-based model to accurately complete missing topological data of utility networks. Experimental results demonstrate significant improvements in accuracy, speed, and data integrity.