Abstract
Safety-critical driving data is crucial for developing safe and trustworthyself-driving algorithms. Due to the scarcity of safety-critical data innaturalistic datasets, current approaches primarily utilize simulated orartificially generated images. However, there remains a gap in authenticitybetween these generated images and naturalistic ones. We propose a novelframework to augment the safety-critical driving data from the naturalisticdataset to address this issue. In this framework, we first detect vehiclesusing YOLOv5, followed by depth estimation and 3D transformation to simulatevehicle proximity and critical driving scenarios better. This allows fortargeted modification of vehicle dynamics data to reflect potentially hazardoussituations. Compared to the simulated or artificially generated data, ouraugmentation methods can generate safety-critical driving data with minimalcompromise on image authenticity. Experiments using KITTI datasets demonstratethat a downstream self-driving algorithm trained on this augmented datasetperforms superiorly compared to the baselines, which include SMOGN andimportance sampling.