AI-based estimation of vehicle dwell time at signalized intersections in motorcycle-dominated mixed traffic environment: A case study in Hanoi

Abstract

This study addresses the critical challenge of accurately measuring vehicle dwell times at signalized intersections characterized by motorcycle-dominated traffic flows, a common yet understudied scenario in many Southeast Asian urban centers. We propose a novel computer vision framework that integrates YOLOv8-based vehicle detection with an innovative tracking approach utilizing the H20 reference point to replace conventional centroid-based methods. This strategic reference point selection demonstrates enhanced stability against
common challenges in mixed traffic environments, including partial occlusions, and perspective distortions inherent in surveillance camera setups. Applied to a comprehensive case study at the Nguyen Trai – Nguyen Van
Loc intersection in Hanoi, Vietnam. Our area-based dwell time measurement algorithm successfully captured
stopping durations by combining velocity thresholding with geometric analysis of vehicle trajectories within
a precisely defined monitoring zone. Experimental results demonstrate that the proposed system achieves significantly higher accuracy compared to conventional centroid-based approaches, with the mean absolute error reduced to approximately 2–3 seconds across all vehicle classes. These findings offer transportation authorities in developing countries an automated, scalable solution for intersection performance analysis, enabling data-driven traffic management decisions and supporting the optimization of signal timing plans for heterogeneous traffic conditions characterized by high motorcycle dominance. The method compatibility with existing surveillance infrastructure further enhances its potential for practical implementation in urban traffic monitoring systems.