<p>With the rapid development of computer vision (CV) technology, object detection models (e.g., YOLO) have increasingly been applied to bridge traffic identification. However, nighttime traffic detection faces unique challenges due to poor lighting conditions, vehicle light reflections, etc. These issues significantly limit the performance and generalization of established CV-based bridge traffic identification methods in nighttime applications. In this regard, this study proposes a data augmentation framework based on CycleGAN-enabled image translation. By transforming collected daytime traffic data into synthetic nighttime data, the dataset for training object detection models can be efficiently enriched, improving the performance of nighttime bridge traffic detection cost-effectively. We demonstrate the proposed framework by training YOLO model with different datasets and examining the consequent traffic detection accuracy on a prototype bridge. The results show that with the proposed data augmentation method, the nighttime traffic detection accuracy could significantly improve for various YOLO variants (e.g., for YOLO-V5S, mAP50 can improve as much as 15.8% after augmentation). In addition, the demonstrative example reveals that the proportion of synthetic nighttime data to that of real nighttime data could affect the performance of the trained YOLO model, and a ratio of 2:3 for real to synthetic is found to be an optimal choice in this example. Overall, this study provides an effective solution that improves nighttime traffic detection accuracy without sacrificing the robustness of daytime implementation. It helps reduce dependence on labor-intensive annotation, offers a practical annotation strategy for bridge managers, and presents a scalable approach for domain adaptation in bridge traffic identification.</p>