YOLOv8s-WAMNet: enhancing robust vehicle detection under adverse weather via hybrid attention and multi-scale fusion in real time

Vehicle detection in adverse weather is crucial for autonomous driving; however, fog, rain, snow, and low illumination significantly degrade feature quality and detection reliability. This paper presents YOLOv8s-WAMNet (weather-adaptive multi-scale network), a lightweight hybrid attention framework designed to maintain robustness under visibility degradation. The model employs an efficient hybrid vision transformer backbone that combines convolutional and transformer-based feature extraction for resilient representation learning. Cross-dimensional multi-scale attention and a contextual multi-attention fusion neck enhance multi-scale feature refinement and stabilize spatial-contextual reasoning in adverse scenes. A multi-head dynamic attention detection head with a hybrid SIoU-MPDIoU loss further improves localization accuracy and convergence stability. Extensive experiments on the WEather images by DALL-E GEneration dataset demonstrate that YOLOv8s-WAMNet achieves [Formula: see text] mAP@50 and [Formula: see text] mAP@50-95, outperforming YOLOv8s by [Formula: see text] mAP@50 while reducing computational cost by approximately [Formula: see text] (13.24 vs. 28.7 GFLOPs). Additional evaluations on the real-world detection in adverse weather nature dataset confirm the robustness and cross-dataset generalization capability of the proposed model.