Road surface state monitoring and pothole localization by opportunistic, in-vehicle monitoring
Résumé
Degradation of roads affects not only rolling noise emissions but also indicates the need for maintenance in order to avoid even stronger damage. Our opportunistic in-vehicle noise and vibration monitoring system is applied to detect this degradation. To this end, noise, vibration and location data for a fleet of vehicles was collected during 2 years, combining map-matching, lateral lane detection and big data. This paper focuses on which road features are captured and how these measurements complement existing, but expensive techniques to monitor road infrastructure. Road waviness, potholes, bridge joints and misaligned concrete slabs are object of this study, focusing on interpretability of the AI system over complexity. A model using sound level spectra peaks is proposed to localize potholes, bridge joints and concrete slab joints. Road features on a location remain constant in short time frames (months), but lane detection and map matching errors spread out the observations. In addition, very localized road features are only sporadically sampled while driving over the same road section (such as potholes). This causes complications in the predictions system. Hence a strategy is proposed to combine multiple noisy road feature observations into single predictions on a location using ordered weighted averages. Thus, pinpointing the location of these potholes on a scale of 20m. Waviness (<2.5m texture) was validated using ARAN measurements by Agentschap Wegen en Verkeer Vlaanderen (AWV). Annotated data from AWV was used to validate the detections of potholes.
Origine : Accord explicite pour ce dépôt