Illumination Uncertainty-Aware YOLOv12s for Real-Time Low-Light Traffic Object Detection

Abstract

Low-light traffic scenes reduce image contrast and weaken object boundaries, making accurate and real-time detection difficult. To address illumination-dependent localization errors, we propose an illumination uncertainty-aware YOLOv12s detector. The method combines five-level low-light augmentation with real-dark training data and introduces brightness-aware dynamic lambda into positive sample assignment. Localization uncertainty is further incorporated into non-maximum suppression to reduce the ranking priority of unreliable detections. Model selection was performed on a 400-image real-dark validation set, while final evaluation used an independent 800-image test set with no data overlap. Compared with the daytime-only YOLOv12s baseline, the proposed model increased mAP50 from 0.3651 to 0.4675 and mAP50-95 from 0.1950 to 0.2536 while achieving 120.19 FPS. Ablation results show that real-dark training data provides the main performance gain, whereas dynamic assignment and uncertainty-aware NMS offer complementary improvements. These results demonstrate a favorable balance between detection accuracy and inference efficiency in low-light traffic scenes.