Underground infrastructure detection and localization using deep learning enabled radargram inversion and vision based mapping

Abstract

Underground pipeline strikes, a pressing problem due to inaccurate subsurface data, are addressed in this paper with a pipeline detection and localization framework. First, abundant radargrams are generated to relieve radargram data shortage by simulating Ground Penetrating Radar (GPR) scans along the urban roadway and enhancing their realism with Generative Adversarial Network (GAN) technique. Second, a deep learning network is designed to directly reconstruct permittivity maps from radargrams for accurate pipeline detection and characterization, instead of detecting pipeline features within the radargram. Third, Simultaneous Localization and Mapping (SLAM) is employed for GPR position estimation, enabling precise georegistration of pipelines. The proposed method attains an R-squared (R2 15) value of 0.957 in permittivity map reconstruction and 96.2% precision in pipeline detection. Additionally, it provides satisfactory performance with a deviation of 1.71% in depth and 20.44% in diameter for the detected pipelines. Real-world experiments validate the effectiveness of the proposed framework, highlighting its potential to prevent excavation accidents, reduce project delays, and offer significant benefits to utility companies, contractors, and urban planners