End-to-end deep learning model for underground utilities localization using GPR

Abstract

Underground utilities (UUs) are key infrastructures in urban life operations. The localization of UUs is vital to governments and residents in terms of asset management, utility planning, and construction safety. UUs localization has been investigated extensively via the automatic interpretation of ground-penetrating radar B-scan images. However, conventional image processing methods are time consuming and susceptible to noise. Deep learning-based methods cannot optimize parameters globally because of their box-fitting mode, which requires the separation of a task into region detection and hyperbola fitting problems. Thus, the accuracy and robustness of the localization task are reduced. Hence, an end-to-end deep learning model based on a key point–regression mode is proposed and validated in this study. Experimental results show that the proposed method outperforms the current mainstream models in terms of localization accuracy (97.01%), inference speed (125 fps), and robustness on the same platform (NVDIA RTX 3090 GPU).