AWD-YOLO enhancing autonomous driving perception reliability in adverse weather

The perception systems of autonomous vehicles are crucial for their safety. However, their reliability is significantly undermined by adverse weather conditions like rain, sandstorms, and haze, which obscure critical details and lead to object detection failures. To address these issues, a YOLOv8n-based Adverse Weather Dual-backbone YOLO (AWD-YOLO) is proposed. The core contributions are the hierarchical feature fusion strategy and the Faster Local-Region Self-Attention (FLRSA) module. The hierarchical feature fusion strategy employs a dual-backbone network that extracts complementary features from both raw and preprocessed images. These features are integrated using shallow concatenation, a Dual Feature Scale Enhancement (DFSE) module, and a Dual Coordinate and Channel Attention (DCA) module to preserve details and enhance semantics. The FLRSA module is specifically designed to efficiently capture critical local features. Furthermore, a dedicated small-object detection head and the Wise-IoU v3 loss function are utilized to improve localization accuracy. Extensive experiments demonstrate that AWD-YOLO attained an mAP@0.5 of 67.7% and an mAP@0.5:0.9 of 40.8% on the DAWN dataset, marking enhancements of 15.4% and 8.1% over YOLOv8n, respectively. Additionally, AWD-YOLO exhibits robustness on the RTTS and ACDC datasets, outperforming YOLOv8n by 6.9% and 12.5% in terms of mAP@0.5. These results highlight AWD-YOLO's potential to enhance the reliability of autonomous driving perception systems in adverse weather.