Generating Realistic Traffic Scenarios: A Deep Learning Approach Using Generative Adversarial Networks (GANs)

Shadab Alam, Md; Martens, Marieke; Bazilinskyy, Pavlo · 2025 · Crossref

DOI: 10.54941/ahfe1005927

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the scarcity of nighttime traffic data in transportation research, noting that major datasets like KITTI and Cityscapes are heavily biased toward daylight conditions. To bridge this gap, the authors investigate the efficacy of Unsupervised Recycle Generative Adversarial Networks (Recycle-GANs) for transforming daytime traffic scenes into nighttime environments and vice versa. The primary objective is to generate realistic traffic videos that preserve visual fidelity and structural accuracy, thereby enhancing the robustness of deep learning algorithms for real-world applications. The researchers utilized a dataset from a live street camera in Seoul, South Korea, comprising one hour of daytime footage and one hour of nighttime footage recorded on April 5, 2024. This data was split into 80% for training and 20% for validation. The model was tested on new footage recorded on April 13, 2024. The Unsupervised Recycle-GAN architecture was chosen for its ability to incorporate tonal constraints, ensuring the preservation of vehicle shapes, colors, and movement patterns while maintaining temporal and spatial consistency. Evaluation involved two distinct methods: automated assessment using GPT-4V and subjective human evaluation. GPT-4V analyzed 12 generated frames for signs of artificial manipulation, such as lighting uniformity, shadow behavior, and edge artifacts. The AI found no definitive indications of artificial creation, noting consistent lighting and natural textures. Additionally, 15 participants rated the realism of the generated transition videos on a scale of 1 to 10. The videos achieved a mean realism score of 7.21. Most participants did not detect the artificial nature of the videos until informed; only two participants identified them as deep-fakes, though they could not pinpoint specific flaws. Common feedback focused on the chaotic nature of the traffic rather than visual artifacts, with only one participant noting halos around pedestrians. The findings demonstrate that Unsupervised Recycle-GANs can effectively generate high-fidelity traffic scenarios with minimal perceptible flaws. By successfully bridging the data gap between day and night conditions, this approach offers a robust method for expanding training datasets for autonomous driving and traffic analysis models. The study concludes that such generative techniques can significantly enhance the applicability of deep learning in transportation research, suggesting future work should explore diverse weather conditions, cross-cultural traffic patterns, and the integration of audio to further improve simulation realism.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• simulator validity fidelity