SafeAug: Safety-Critical Driving Data Augmentation from Naturalistic Datasets

TL;DR

SafeAug introduces a depth- and geometry-driven data augmentation framework that enriches safety-critical driving scenarios in naturalistic KITTI data. By detecting vehicles with YOLOv5, estimating depth with Depth-Anything, creating 3D models, and perturbing front-vehicle distances, it generates hazard-rich images while maintaining realism. The augmented dataset improves downstream car-following models, especially on safety-critical cases, and outperforms SMOGN and importance sampling baselines. This approach narrows the realism gap between synthetic hazard data and real-world driving, offering a practical path to safer autonomous systems. The technique is formalized through a training objective that combines original and augmented data via $\theta^* = \arg\min_{\theta} D(\theta; I + I_{\text{aug}})$ and validated on KITTI with a targeted hazard-focused evaluation.

Abstract

Safety-critical driving data is crucial for developing safe and trustworthy self-driving algorithms. Due to the scarcity of safety-critical data in naturalistic datasets, current approaches primarily utilize simulated or artificially generated images. However, there remains a gap in authenticity between these generated images and naturalistic ones. We propose a novel framework to augment the safety-critical driving data from the naturalistic dataset to address this issue. In this framework, we first detect vehicles using YOLOv5, followed by depth estimation and 3D transformation to simulate vehicle proximity and critical driving scenarios better. This allows for targeted modification of vehicle dynamics data to reflect potentially hazardous situations. Compared to the simulated or artificially generated data, our augmentation methods can generate safety-critical driving data with minimal compromise on image authenticity. Experiments using KITTI datasets demonstrate that a downstream self-driving algorithm trained on this augmented dataset performs superiorly compared to the baselines, which include SMOGN and importance sampling.

Figures (5)

• Figure 1: Autonomous Driving Data Augmentation Pipeline
• Figure 2: Street View to Depth
• Figure 3: YOLO-v5 Result
• Figure 4: Comparison of Vehicle Distance Adjustments
• Figure 5: Distribution of Acceleration