This study addresses the significant performance limitations of the YOLOv8 algorithm in autonomous driving (AD) under adverse weather conditions, particularly heavy fog and low-light environments. Existing YOLOv8 shows considerable accuracy degradation in such conditions, necessitating targeted enhancements to ensure reliable and safe autonomous navigation. The study optimized the YOLOv8 algorithm through several innovative methods: integrating PEnet for low-light enhancement, FAANet for foggy conditions, the SPD-conv block for improved detection of small objects, and the GAM attention model for better adaptability in complex scenarios. Additionally, EfficientViT, FastCNN, and VanillaNet backbone networks were compared to identify the most effective model. Datasets from ExDARK and fog-specific conditions were utilized, split into training and validation subsets. Post-optimization, YOLOv8 showed marked performance improvement: fog detection recall increased by 8.7%, precision by 3.9%, mAP by 15%, and mAP50-95 by 6.17%. Under low-light conditions, recall and precision improved by approximately 15%. Among tested backbones, VanillaNet exhibited superior performance, showing a 6.42% increase in precision and a 14.42% increase in mAP, demonstrating its efficiency and suitability for real-world deployment. The enhanced YOLOv8 algorithm significantly improves AD systems’ safety and reliability in adverse weather by reducing misidentification risks and improving detection accuracy. These enhancements can substantially decrease AD-related accidents, enhancing road safety. Future directions involve expanding datasets through community-driven campaigns and extensive real-world deployment testing to validate the model’s practical effectiveness.