Driving Safety Area Classification for Automated Vehicles Based on Data Augmentation Using Generative Models

Abstract

The integration of automated vehicles (AVs) into existing road networks for mobility services presents unique challenges, particularly in discerning the driving safety areas associated with the automation mode of AVs. The assessment of AV’s capability to safely operate in a specific road section is contingent upon the occurrence of disengagement events within that section, which are evaluated against a predefined operational design domain (ODD). However, the process of collecting comprehensive data for all roadway areas is constrained by limited resources. Moreover, challenges are posed in accurately classifying whether a new roadway section can be safely operated by AVs when relying on restricted datasets. This research proposes a novel framework aimed at enhancing the discriminative capability of given classifiers in identifying safe driving areas for AVs, leveraging cutting-edge data augmentation algorithms using generative models, including generative adversarial networks (GANs) and diffusion-based models. The proposed framework is validated using a field test dataset containing disengagement events from expressways in South Korea. Performance evaluations are conducted across various metrics to demonstrate the effectiveness of the data augmentation models. The evaluation study concludes that the proposed framework significantly enhances the discriminative performance of the classifiers, contributing valuable insights into safer AV deployment in diverse road conditions.