Drone and Computer Vision-Based Analysis of Service Levels and Traffic Density at Signalized Intersections

Abstract

Manual traffic surveys at signalized intersections are labor-intensive, hazardous, and prone to counting error under heterogeneous, high-density traffic typical of Indonesian cities. Purpose: This study proposes and evaluates a drone (unmanned aerial vehicle, UAV) and computer vision-based framework for measuring traffic density and determining the level of service (LOS) of a four-arm signalized intersection in Padang City, West Sumatra. Methods: Aerial video was recorded over three consecutive weekdays across three peak periods using a UAV hovering at 100–120 m. A YOLOv8 detector, fine-tuned on a locally annotated dataset of four vehicle classes, was combined with a DeepSORT tracker to count and classify vehicles on each approach. Counts were validated against synchronized manual ground-truth counts and converted into passenger car units (PCU) following the 2023 Indonesian Road Capacity Guidelines (Pedoman Kapasitas Jalan Indonesia/PKJI 2023) to compute capacity, degree of saturation (DS), delay, and LOS. Results: The model achieved an overall mean average precision (mAP@0.5) of .918 and a mean absolute percentage error (MAPE) below 6% relative to manual counts. DS ranged from 0.58 to 1.12; the East approach exceeded the 0.85 serviceability threshold in the morning and evening peaks and reached LOS F in the evening peak. Implications: UAV-based computer vision can reliably replace or substantially augment manual counting for intersection performance evaluation, offering wider coverage, repeatability, and safer field surveys. Conclusion: The framework is accurate and efficient; future studies should extend it to real-time signal control and larger training datasets.