講演情報

[AI-03]Network-Level Evaluation of Pothole Risk Using a Simple In-Vehicle Camera and U-Net–Based Crack-Shape Analysis

*Rau Tomidokoro1, Takumi Asada1 (1. Muroran Institute of Technology (Japan))

キーワード:

Pothole risk prediction、Crack-shape analysis、U-Net semantic segmentation、Preventive maintenance、Real-field validation

In snowy and cold regions, potholes frequently form under harsh weather conditions, creating an urgent maintenance challenge. Because potholes generate localized and significant surface discontinuities, they degrade ride comfort and traffic safety and may worsen roadside environments through increased noise and vibration. Current practice relies on periodic patrols to detect damage and conduct emergency repairs; however, this reactive approach can delay responses, increase repair costs, and impose additional social burdens on road users and residents. Predictive, preventive maintenance is therefore needed, yet municipalities often face severe budget and staffing constraints that make sufficiently frequent patrols impractical.
To support network-level pothole risk management with minimal field effort, we develop a framework using a simple in-vehicle camera and U-Net–based crack-shape analysis. First, in-vehicle footage captured by a commercially available camera is processed with a U-Net semantic segmentation model to detect cracks at the pixel level. From the resulting crack masks, we compute quantitative crack-shape descriptors—such as crack fineness (texture density), width-related measures, and other morphology indices—that can reflect early-stage surface deterioration. We then link these descriptors to subsequent pothole occurrences to train and validate a pothole occurrence prediction model.
Finally, using two rounds of comprehensive pavement inspection data collected in Muroran City, Japan, we demonstrate a network-level evaluation of pothole risk across the entire urban road system. The proposed approach enables early identification of locations likely to develop potholes and supports prioritization of preventive interventions at the network scale. The results highlight the feasibility and usefulness of integrating vision-based crack-shape information with municipal inspection records to advance data-driven, preventive pothole risk management in snowy and cold regions.