Automated Road Surface Condition Monitoring–Evaluation of Pilot Study

This paper evaluates the detection quality of an automated road surface condition monitoring system tested in a fully automated pilot study. The system integrates multiple image-processing modules, including a road damage detection model based on YOLOv8 and a road surface segmentation model using DeeplabV3. The system was deployed in an automated pilot study setup, from March 2024 to Sept 2024. To assess the system’s performance, a subjective evaluation of the model’s performance during pilot study was conducted with the help of road maintenance professionals using a five-point qualitative scale with categories "very poor", “poor”, “satisfactory”, “good” and "excellent". The results showed that 65% of the analysed images were rated as satisfactory or better for detection quality. In contrast, an F1-score of 0.45 was achieved by the YOLOv8 model during finetuning. The study also identifies key errors and limitations in the system’s automated detection, with a primary detection error identified as confusion between visually similar objects.