WaveSenseNet: A Self-Supervised Learning Framework for Robust Ground Penetrating Radar Based Localization in Autonomous Systems

Ground Penetrating Radar (GPR) is widely utilized for underground structure detection and surveying, but its adoption in autonomous driving remains limited. Current autonomous positioning technologies, such as GPS, LiDAR and camera-based systems, often struggle in dynamic environments, areas with poor GPS reception, or conditions requiring the use of external features such as lane markings, guardrails, or high definition maps. GPR-based localization offers a robust alternative, particularly in GPS-denied environments such as shipping ports, mining sites, and defense operations, where satellite signals are obstructed or disrupted by ionospheric scintillations. Although prior studies have demonstrated the high accuracy and high availability of GPR-based localization, challenges persist, which include dependency on uniqueness in underground features, low overlap with mapped areas, clock drift, and varying speeds. These issues degrade localization accuracy, with non-unique features hindering extraction of key information, low overlap leading to incorrect matching, clock drift causing temporal misalignment, and varying speeds introducing errors due to inconsistent data acquisition rates. We propose WaveSenseNet, a novel self-supervised learning framework for feature extraction, designed to address these limitations and improve GPR localization accuracy. Being a self-supervised network, there are no data labeling requirements making this solution easy to scale. Trained exclusively on real-world data across various surface types like concrete and asphalt, WaveSenseNet significantly improves the performance of GPR-based localization, enabling highly accurate positioning for autonomous systems. We have successfully demonstrated accuracies of 5 cm for indoor and 10 cm for outdoor tests respectively. Our framework does not rely on any specific underground features for localization and uses the existing subsurface features to find the best match in the map. The performance of this GPR based localization proves it to be a viable solution for localization for autonomy applications. This novel algorithm serves as a foundation for using deep learning based techniques for GPR based localization.