Yang, Yibing (Xi'an Jiaotong University), Zhang, Chi (Xi'an Jiaotong University), Xu, Linhai (Xi’an Jiaotong University), Ma, Shuangxun (Chang'an University), Li, Li (Tsinghua University)

Safety-Critical Scenario Generation by Causal Influence Detection

Scheduled for presentation during the Regular Session "Generating driving scenarios II" (FrBT10), Friday, September 27, 2024, 14:30−14:50, Salon 18

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 7, 2024

Abstract

Appropriate behavior intervention for a specific background vehicle at the right moment is a crucial step in identifying safety-critical scenarios. Existing methods generate scenarios by estimating the data distribution, but when noise interventions are present, the emergence of safety-critical scenarios becomes collinear with the noise interventions, potentially resulting in unnecessary interventions and unrealistic vehicle behaviors. To address this, we propose to integrate the causal relationship between the background vehicle behavior and the safety status of ego-vehicle as a prior into scenario generation. In this notion, we present a reinforcement learning framework with a causal influence detection module (emph{CausalID}). Specifically, we employ conditional mutual information to quantify the causal influence between vehicle sequential behaviors, and intervene on the background vehicle's behavior when the causal influence exceeds a certain threshold. Additionally, to prevent unrealistic collision-oriented driving, we sample the final intervention action from a set of candidate actions according to their probabilities of occurring in natural driving scenarios. Our experiments in 3-lane highway scenarios with multiple vehicles validate the effectiveness of the proposed framework.