Image-based roadway hazard detection and risk rating using artificial intelligence

Abstract

Road safety is affected not only by road geometry, but also by the visibility and condition of roadside safety features such as warning signs, pavement markings, and guardrails. This paper presents an image-based framework for roadway hazard detection and risk rating using artificial intelligence. Three YOLOv8-based models were trained to detect right-turn and left-turn warning signs, double road lines, and guardrails from road scene images. In parallel, 148 horizontal curves along the Colombo-Hatton Road (A7) were identified in QGIS and described using radius, direction, and curve class. Roadside drop-off was assessed using a digital elevation model. The detection outputs, curve geometry, roadside drop values, and guardrail condition were then combined through a weighted likelihood–impact matrix. Likelihood was calculated using curve radius, warning sign condition, and road marking condition, while impact was calculated using roadside drop and guardrail condition. The weighted values were converted into matrix levels and multiplied to classify the selected curves as low, medium, or high risk. The results show strong performance for warning sign and double line detection, while guardrail detection is more difficult because of occlusion, roadside clutter, and background similarity. High-risk curves were mainly associated with sharp curvature, significant roadside drop-off, and inadequate roadside protection. The framework therefore provides a practical screening method for identifying hazardous curves and optimum locations for guardrail installation.