Data-Driven Traffic Management: Implementing YOLOv8 for Vehicle Detection and Classification in Mixed Traffic Conditions

抄録

This study validates a YOLOv8-based framework for vehicle detection and turning movement classification in the complex mixed traffic of emerging economies. We address the need for real-time traffic data by fine-tuning a model on a large, custom-annotated Indonesian dataset, utilizing a class-consolidation strategy to manage data imbalance. The system achieves high accuracy (mAP@0.5 of 90.0%) and real-time capability (~41 FPS). Crucially, automated vehicle counts show excellent agreement with manual ground truth, with Geoffrey E. Havers (GEH) statistic values consistently below the 5.0 threshold. This confirms the system's validity for practical traffic engineering. The research provides a validated, low-cost, and scalable tool for gathering granular turning movement data, essential for developing advanced Intelligent Transportation Systems (ITS) and Adaptive Traffic Control Systems (ATCS).