DiffCrash: leveraging denoising diffusion probabilistic models to expand high-risk testing scenarios using in-depth crash data

doi:10.1016/j.eswa.2025.128140

DiffCrash: leveraging denoising diffusion probabilistic models to expand high-risk testing scenarios using in-depth crash data

2025-05-22

0

PRE

AI

DOI：10.1016/j.eswa.2025.128140

Original

Original request for help

Save

Abstract

En 中文

Scenario-based safety testing for autonomous vehicles (AVs) is a critical component in their development, production and deployment. To fulfill its role effectively, such testing must ensure that the test scenarios are realistic, high-risk, and diverse. To address the issue of low risk and limited testing value in generated test scenarios caused by the rarity of crash data in naturalistic driving data (NDD), this study expands real-world crash data recorded in the China In-depth Mobility Safety Study-Traffic Accident (CIMSS-TA) dataset to achieve better coverage of potential high-risk scenarios and ensure the realism of the generated scenarios. Inspired by Denoising Diffusion Probabilistic Models (DDPMs), this paper introduces a novel approach, DiffCrash, designed to generate high-risk test scenarios for evaluating a well-known autonomous driving system, Baidu Apollo. The experimental results demonstrate that DiffCrash not only generates test scenarios with higher usability and greater similarity to real-world accident scenarios compared to two other baseline methods, but also produces scenarios with risk levels comparable to those in real crash data, while exhibiting improved diversity.

Keywords:

Traffic safety

Autonomous vehicle

Safety testing

Testing scenarios

000

1111

DiffCrash: leveraging denoising diffusion probabilistic models to expand high-risk testing scenarios using in-depth crash data

Abstract

Journal

Organization