Generating realistic traffic scenarios: A deep learning approach using generative adversarial networks (GANs)

Alam, M. S., Martens, M., Bazilinskyy, P.

Submitted for publication.

ABSTRACT Diverse and realistic traffic scenarios are crucial for testing systems and human behaviour in transportation research. Leveraging Generative Adversarial Networks (GANs), this study focuses on video-to-video translation to generate a variety of traffic scenes. By employing GANs for video-to-video translation, the study accurately captures the nuances of urban driving environments, enriching realism and breadth. One advantage of this approach is the ability to model how road users adapt and behave differently across varying conditions depicted in the translated videos. For instance, certain scenarios may exhibit more cautious driver behaviour, while others may involve heavier traffic and faster speeds. Maintaining consistent driving patterns in the translated videos improves their resemblance to real-world scenarios, thereby increasing the reliability of the data for testing and validation purposes. Ultimately, this approach provides researchers and practitioners with a valuable method for evaluating algorithms and systems under challenging conditions, advancing transportation models and automated driving technologies.