RoadVision AI: An Intelligent Vision-Based Road Infrastructure Monitoring System Using YOLOv8 and Depth Estimation

Abstract

Road defects such as potholes, damaged drain holes, and missing or broken sewer covers pose a persistent challenge to road safety, ride comfort, and vehicle maintenance costs. Field-based inspection, which remains the conventional method of assessing pavement condition, depends on manual observation by maintenance personnel and is therefore slow, subjective, and difficult to repeat at scale across large road networks. This paper presents RoadVision AI, a vision-based road infrastructure monitoring system that automates the detection and severity assessment of common road defects directly from images. The system performs detection in a structured pipeline. First, a trained YOLOv8 object detection model localizes potholes, drain holes, and sewer covers within an uploaded road image. Once a defect region has been identified, OpenCV is used to extract visual descriptors such as contour geometry, edge density, texture, and darkness from the cropped region, while the MiDaS monocular depth estimation model generates a relative depth map to estimate pothole depth without requiring dedicated depth sensors. The extracted visual descriptors and depth values are then fused to compute a Visual Severity Score and a Road Damage Index, from which the system determines the risk category, evaluates overall road health, and generates an appropriate maintenance recommendation. A Streamlit-based web application allows users to upload images, review annotated detections, and monitor historical trends through an interactive Plotly dashboard. The YOLOv8 model was trained on 3,813 annotated road images collected from Roboflow and achieved a precision of 91.20%, recall of 87.79%, mAP@50 of 91.25%, and mAP@50–95 of 56.70%. This framework contributes an integrated road-monitoring pipeline that combines object detection, depth-aware severity estimation, and dashboard-based analytics within a single deployed system.