YOLOv8-EVdet: An emergency vehicle detection method for autonomous vehicles

This work presents a deep-learning based approach for emergency vehicle detection and classification for various driving conditions. Conventional vision-based methods often struggle with poor lighting, motion blur, and occlusions caused by surrounding vehicles, leading to inaccurate detections that may eventually slow the emergency vehicles' response. To address these challenges, we propose a modified object detection algorithm, tailored to identify emergency vehicles reliably in varied traffic and environmental conditions. Since existing datasets provide limited coverage of emergency vehicle classes, we additionally provide a custom dataset that includes diverse emergency vehicle classes in challenging visual and traffic scenarios. The training process involved extensive data augmentation and fine-tuning, to ensure robustness against occlusion, low lighting, and diverse backgrounds. The proposed method was evaluated on the custom EVdet dataset achieving a mAP@50 of 0.782 ± 0.006, across three independent training runs, confirming training stability. Cross-dataset evaluation on nuImages and EMS3D-KITTI demonstrated generalization to unseen domains. Additionally, a zero-shot comparison against four COCO-pretrained baselines (YOLOv5m, YOLOv8m, YOLO11m, RT-DETR-L) revealed that general-purpose detectors score 0.0 mAP on all emergency vehicle classes, establishing the necessity of task-specific fine-tuning; paired bootstrap and Wilcoxon signed-rank tests confirm that the improvements over all baselines are statistically significant (p < 0.01).