The Impact of Weather, Lighting, and Camera Viewpoint on the Traffic Monitoring Performance of YOLO-SORT

Abstract

This study investigates the condition sensitivity of a camera-based traffic monitoring pipeline that combines YOLOv8s detection with SORT tracking for real-time vehicle counting. Video data were collected from a pedestrian bridge above a two-lane highway in Győr, Hungary, under multiple weather conditions (clear, rainy, foggy), lighting (day, night) conditions, and viewpoint (central, moderate shift, maximum available shift) configurations. Pixels-per-meter calibration, normalized trajectories, and bounding-box dimensions across angles, and a trajectory-based count line produced event-level ground truth. Performance was evaluated using minute-level precision, recall, and F1-score, with condition-wise and angle-wise differences assessed via Wilcoxon signed-rank tests with Holm's correction and associated effect sizes. Across all angles, the pipeline maintained high F1-scores, with nighttime conditions yielding the highest and most stable performance, supported by consistent street lighting and lower traffic flow. Fog produced the most persistent degradation, while rain mainly caused brief precision dips during onset. Aggregated across conditions, no statistically significant differences were observed between viewpoints, indicating that within the studied range of angles, environmental factors dominate over camera orientation. Overall, the results suggest that YOLOv8s and SORT can support accurate real-time vehicle counting in CCTV-like deployments while remaining sensitive to reduced visibility and peak-traffic occlusions.