AI-BASED AUTOMATIC POTHOLE DETECTION AND MAPPING VIA COLLABORATIVE MULTI-MODAL SIGNAL PROCESSING

Road potholes pose a significant safety hazard to drivers and cyclists while accelerating vehicle wear and incurring substantial maintenance costs for municipalities, which often rely on slow, manual inspection processes to identify them. Traditional reporting methods are reactive, inconsistent, and unable to keep pace with the rate of road degradation in urban and rural networks alike. This work proposes an AI-based distributed system integrated with GPS to automatically detect, classify, and geolocate potholes in real time using multi-modal signal processing (vision and inertial sensors). The solution enables proactive, large-scale road monitoring, allowing authorities to prioritize repairs efficiently and improve overall road safety. The first goal of the thesis project is to implement and test a vision-based pothole detection pipeline based on classic AI models like YOLO, to then build a first working prototype to be tested in a real-world scenario. The prototype must also integrate GPS localization of pothole detections, a distributed remote database, and a user visualization interface. Upon building a working prototype, the research part of the project will focus on enhancing the detection capabilities to previously unseen scenarios (e.g. countryside roads, dirty roads, partial occlusion) by means of Vision Foundation models with text-to-image-alignment. The second research goal will be to investigate pothole detection pipelines based on inertial sensors embedded in common devices like smartphones.