Gao, Dingcheng (Tsinghua University), Qin, Yanjun (Tsinghua University), Tao, Xiaoming (Tsinghua University), Lu, Jianhua (Tsinghua University)

Diversifying Latent Flows for Safety-Critical Scenarios Generation

Scheduled for presentation during the Regular Session "Generating driving scenarios I" (ThBT11), Thursday, September 26, 2024, 15:30−15:50, Salon 19/20

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 8, 2024

Abstract

Accidents have become a widespread societal problem on a global scale as the automotive industry has progressed. In contrast to the long-term safe driving environment, the likelihood of encountering safety-critical scenarios leading to traffic accidents while driving is minimal. However, these rare events are crucial for enhancing human or autonomous driving skills. Therefore, evaluating and refining the decision-making processes of human or autonomous vehicles requires scalable generation of long-tail traffic scenarios. These scenarios should be both realistic and challenging, but still partially solvable. In this paper, we propose an automated method for generating challenging scenarios. Our primary objective is to use a safety-critical scenario generation model based on a Conditional Variational Autoencoder (VAE) to increase the variety of scenarios by diversifying latent flows on the pre-trained trajectory representation model. The results show that our method can produce plausible scenarios, surpassing the baseline by over 10% in collision metric for scenario generation.