Takeover Performance Prediction Based on Function-Based Causal Inference in Level-3 Autonomous Vehicles

Abstract

The need to ensure the safety and effectiveness of level-3 (L3) autonomous vehicles in emergency scenarios requires understanding the factors influencing takeover performance. This study applies functional causal inference to enhance takeover-performance prediction in L3 autonomous vehicles. While existing research has identified numerous influencing factors through statistical analysis, the presence of noncausal variables continues to compromise prediction accuracy. This study proposes an improved linear non-Gaussian acyclic model (LiNGAM) to establish causal relationships between influencing factors and takeover performance, subsequently evaluating prediction models using common machine-learning algorithms. Key findings are that traffic conditions, driver perception, and reaction time emerge as critical determinants of takeover performance. Moreover, braking reaction time is a confounding variable that distorts the perceived relationship between takeover performance and maximum brake-pedal force. Maximum brake-pedal force is noncausal to takeover performance. Removing noncausal variables enhances prediction-model performance. Specifically, eliminating the maximum value of brake force from correlated variables yields an average 0.013 improvement in the area under curve across four algorithms. This study demonstrates that integrating correlation analysis with functional causal inference significantly improves interpretability, accuracy, and simplicity. The findings suggest that eliminating noncausal variables through causal inference provides crucial insights for optimizing driver behavior and advancing L3 autonomous-vehicle safety. These methodological advancements contribute substantially to developing more reliable autonomous driving systems.