Synthetic Collision-Prone Trajectory Data Generation Using CTGAN for Connected Autonomous Vehicles

The safety and reliability of connected autonomous vehicles (CAVs) hinge on their ability to navigate complex and unpredictable driving environments. Traditional testing methods, relying on real-world data and simulations, often fall short of providing the diverse and comprehensive set of collision scenarios needed for rigorous evaluation. On the other hand, mobility trajectory data acquisition is challenging due to privacy concerns, commercial considerations, missing values, and expensive deployment costs. This paper presents a novel approach to generating synthetic collision-prone trajectories using the Conditional Tabular Generative Adversarial Network (CTGAN). By leveraging CTGAN’s ability to handle imbalanced and mixed data types, we create a rich synthetic dataset that enhances the validation and testing processes for CAVs. Our methodology includes the preparation of real-world trajectory data, the configuration and training of the CTGAN model, and the generation of a diverse set of collision-prone scenarios. We test our approach using two large-scale real-world trajectory datasets and comparative analyses with other trajectory generation methods, including TimeGAN and SocialGAN, to underscore the superiority of our approach in generating realistic and varied scenarios. This research contributes to the field of autonomous vehicle testing by offering a robust and scalable solution for scenario generation. The synthetic dataset not only improves safety and reliability metrics but also provides a valuable resource for future research and development. Our findings suggest that integrating CTGAN-generated data into CAV testing frameworks, not only in terms of statistical and quality metrics, is better than the other approaches but also can significantly enhance the preparedness of autonomous vehicles, ultimately contributing to safer and more efficient transportation systems.

I. Introduction

Connected Autonomous Vehicles (CAVs) represent a significant advancement in the automotive industry, promising to enhance road safety, reduce traffic congestion, and provide more efficient transportation solutions. [1]. These vehicles rely on complex algorithms and vast amounts of data to navigate and make decisions in real time. To ensure the safety and reliability of CAVs, extensive testing is essential. Traditionally, this testing involves real-world driving scenarios, simulations, and the use of datasets comprising various driving conditions and incidents.