Object detection is a very critical component of an autonomous driving system for correct decision-making, as knowledge of road conditions—which come in various weather scenarios like dry, wet, and snowy—must be taken into consideration in the control of autonomous vehicles. There exist a lot of annotated datasets, scenes taken in daylight, but little in adverse conditions, especially in fog. The collection of large amounts of annotated data requires substantial human resources and time. Many GAN-based image translation methods have been proposed to handle issues by generating synthetic data. Traditional methods of image translation have a high degree of limitation: generated images often lack consistency with their original counterparts. We would propose a method using transfer learning for characteristic-preserving image translation. Experimental results demonstrate that the method proposed by us can generate foggy images more consistent with the originals as compared to traditional methods. In this paper, we introduce one of the image enhancement techniques using CycleGAN to enhance the different features of input images. This paper trains an unsupervised CycleGAN network for the enhancement of road images in severe weather conditions like fog, and then obtains rich feature information from images. Experimental results show that our CycleGAN-based system can be directly combined with YOLOv7 to support end-to-end training, significantly enhancing the performance of the original network in detecting road features under severe weather conditions.