Application of Machine Learning to Study the Behavior of Roundabout Users

Abstract

Roundabouts have gained popularity as an infrastructure solution to enhance traffic safety and reduce congestion in urban areas. Their design facilitates smoother vehicle flow, mitigating severe accidents caused by decision-making at traditional intersections and traffic signals. However, the effectiveness of roundabouts largely depends on driver behavior, which poses significant safety risks and contributes to accident occurrences within roundabouts. This study applies computer vision techniques to detect driver behavior using video data collected from drones at real-world roundabout locations. The selected roundabouts include pedestrian crossings and clear traffic lines to capture diverse interaction scenarios. Machine learning is employed to analyze roundabout user behavior. The study found that traffic rule violations occurred in approximately 47% of observed cases. Supervised learning is used to identify and classify risk-related driving behaviors within the roundabout. These behaviors include overspeed limit 79%, failure to yield the right-of-way to vehicles on the right 21%, sudden lane changes, and failure to stop for pedestrians. Furthermore, the study measured the minimum accepted gap distance for merging behavior, which was found to be approximately 0.73 meters.