Street View Image-Based Road Marking Inspection System Using Computer Vision and Deep Learning Techniques

Junjie Wu, Wen Liu, Yoshihisa Maruyama

Road markings are vital to the infrastructure of roads, conveying extensive guidance and information to drivers and autonomous vehicles. However, road markings will inevitably wear out over time and impact traffic safety. At the same time, the inspection and maintenance of road markings is an enormous burden on human and economic resources. Considering this, we propose a road marking inspection system using computer vision and deep learning techniques with the aid of street view images captured by a regular digital camera mounted on a vehicle. The damage ratio of road markings was measured according to both the undamaged region and region of road markings using semantic segmentation, inverse perspective mapping, and image thresholding approaches. Furthermore, a road marking damage detector that uses the YOLOv11x model was developed based on the damage ratio of road markings. Finally, the mean average precision achieves 73.5%, showing that the proposed system successfully automates the inspection process for road markings. In addition, we introduce the Road Marking Damage Detection Dataset (RMDDD), which has been made publicly available to facilitate further research in this area.