The 3D localization of subsurface pipes from Ground Penetrating Radar images using edge detection and point cloud segmentation
Tsukasa Mizutani, Jingzi Chen, Shuto Yotsumoto
Abstract
This research introduces an innovative algorithm using Ground Penetrating Radar (GPR) for inspecting underground infrastructure non-destructively, surpassing R-CNN methods by combining signal processing with computer vision techniques including hyperbola correlation, point cloud segmentation and noise reduction based on linearity, density and echo patterns. Utilizing on-vehicle multi-channel GPR and without training, it maps pipes in 3D with high precision on both experimental fields and real roads. The method demonstrated remarkable accuracy, with an average error of 0.1175 meters in locating pipes, outperforming deep learning methods requiring numerous training data. The ablation study and sensitivity analysis on hyperparameters further prove the improvement from the proposed method and its robustness. These outcomes offer notable improvements in infrastructure monitoring, highlighting the potential for automated underground utility detection.