Robustness Evaluation of Lightweight YOLO Detectors for Traffic Monitoring Under Visual Corruptions and Adversarial Attacks

Abstract

Reliable traffic-object detection remains challenging when visual conditions degrade or when the detector is exposed to adversarial perturbations. This study investigates these robustness challenges through a systematic evaluation of lightweight YOLO detectors for intelligent traffic monitoring. Several nano-scale variants are first compared, and YOLO26n is selected for its favorable trade-off between detection accuracy and computational efficiency, achieving an mAP50 of 79.48% and an mAP50:95 of 59.65% on clean data. The selected detector is then evaluated under ten common image corruptions at five severity levels and under white-box, black-box, localized, and task-oriented adversarial attacks. ISO noise and motion blur cause the largest corruption-induced degradation, while PGD produces the strongest adversarial effect, reducing mAP50:95 to 6.18% at 16/255. Class-wise analysis also shows that small and visually ambiguous traffic categories are generally more vulnerable. These findings reveal that strong clean-data performance does not ensure reliable behavior under degraded or hostile conditions and provide practical insights for the development of more robust traffic-monitoring systems.