Pedestrian flow counting at urban intersections using a retrained YOLOv8 model

Abstract

This study addresses the limited application of computer vision techniques for counting pedestrians at urban intersections by developing an integrated deep learning-based framework. Three models are proposed: a detection model trained on a large, diverse pedestrian dataset; a tracking model incorporating a novel reference point on bounding boxes with an enhanced identity-switch handling algorithm; and a counting model tailored to pedestrian crossing behaviors with movement-specific algorithms. The framework was applied to a case study in Vietnam, where pedestrian flow is often complex due to mixed traffic environments. A comprehensive dataset comprising 22 video footages under both daytime and nighttime conditions yielded over 120,000 labeled pedestrian instances. The tracking model effectively captures pedestrian trajectories across crosswalks, while the counting model introduces a multi-line crossing technique to enhance accuracy at signalized intersections. Evaluation results show the counting model achieves over 98% accuracy compared to manual annotations across various time frames and pedestrian densities. These findings offer valuable tools for urban transport planners and policymakers in Vietnam and similar countries, enabling automated pedestrian data collection, improving intersection safety assessment, and supporting infrastructure design.