Adaptive vehicle detection in foggy conditions based on improved YOLOv5

In order to reduce traffic accidents and improve driving safety in foggy conditions, a foggy conditions adaptive vehicle detection model based on improved YOLOv5 is proposed. First, the detector is enhanced through optimized anchor assignment, a refined loss function, and an embedded attention mechanism to improve the recognition of small and occluded vehicles. Second, Based on the atmospheric scattering model, synthetic images are generated by applying varying concentrations of fog effects to clear-weather vehicle images, thereby constructing a multi-source foggy vehicle dataset that integrates real foggy scene images with synthesized images. Third, a hybrid processing pipeline is designed to seamlessly integrate the GCA-Net defogging algorithm with the improved YOLOv5 model. Extensive experiments demonstrate that our method runs at 33.52 FPS and achieves an mAP of 97.7%, outperforming the baseline by 4.2%. The model performs reliably in light and medium fog without preprocessing, while dense fog necessitates prior defogging. Accordingly, a threshold-based adaptive strategy is proposed to switch between direct detection and the defog-then-detect pathway. This work presents a novel adaptive framework that integrates GCA-Net with an attention-augmented YOLOv5, enabling real-time vehicle detection across a wide range of fog densities.