Theodoraki E. M., Koliou P., Yannis G., “Enhancing Risky Driving Behavior Classification Using Conditional GANs (cGANs): A Data Augmentation Approach”, Proceedings of the the 17th World Conference on Transport Research (WCTR), Toulouse, France, 6-10 July 2026.

Accurately classifying risky driving behaviors is essential for road safety but is hindered by class imbalance, with dangerous behaviors underrepresented in most datasets. This study explores the use of Conditional Generative Adversarial Networks (cGANs) to generate synthetic data for underrepresented risk categories and improve classification performance. Unlike traditional GANs, cGANs enable controlled data generation based on predefined risk levels. Driving behaviors were initially categorized as Normal, Dangerous, or Avoidable Accident. Following augmentation, a binary classification scheme (Normal vs. Avoidable Accident) was adopted. A cGAN was trained on Normal data to generate synthetic high-risk scenarios, which were used to augment the dataset. Classifiers, GAN-based, XGBoost-RF, and RNN-AdaBoost, were then evaluated on both original and augmented datasets.
Results showed that cGAN-generated data significantly improved GAN model accuracy (from 76% to 90% in Belgium; 79% to 91% in the UK). However, hybrid models achieved near-perfect accuracy on augmented data, indicating overfitting and limited generalizability. This study demonstrates the benefits and challenges of using cGANs for synthetic data augmentation in driving behavior classification and underscores the need for careful validation to ensure real-world applicability.