Enhanced YOLOv8: improving vehicle detection in hazy conditions with Swin Transformer and AOD-Net

Vehicle detection is a critical task in autonomous driving and intelligent transportation systems. However, adverse weather conditions, particularly haze, significantly impair the performance of traditional object detection models, resulting in a considerable decline in detection accuracy. While existing detection models perform well under normal conditions, they struggle to cope with low contrast and blurred edges in hazy environments. To address this issue, this study proposes an improved YOLOv8 model that integrates the Swin Transformer and AOD-Net. This model leverages the dehazing capabilities of AOD-Net and the multiscale feature extraction capabilities of the Swin Transformer, effectively improving vehicle detection performance in hazy weather conditions. In this study, the detection performance of YOLOv8n, YOLOv5s, and YOLOv8s was compared to evaluate their effectiveness under haze conditions. Experimental results demonstrate that the proposed improved model significantly outperforms YOLOv8n across all metrics, particularly in scenarios with extremely low visibility. Additionally, by testing the model under varying haze intensities, its robustness was further confirmed. This research offers a novel approach to vehicle detection under hazy weather conditions, demonstrating the feasibility and superiority of vehicle detection in adverse weather.