Modeling pedestrian-vehicle interaction risk based on risk perception of experienced drivers in autonomous vehicles.

OBJECTIVE: Reliably determining collision risk in pedestrian-vehicle interactions remains a critical challenge for autonomous vehicles (AVs). Experienced human drivers are able to quickly evaluate the overall trend of a scene and predict potential risks based on their experience. In contrast, existing risk assessment models often lack experience-based judgment. This study develops and validates an expert-guided framework to quantify pedestrian-vehicle interaction risk by learning the takeover behaviors of safety drivers in AVs. METHODS: We extracted 113 real-world, high-risk pedestrian-vehicle interaction cases from an autonomous driving database based on extensive road-testing data. We recorded instances of takeover events during high-risk situations. To thoroughly analyze the limited yet factual data from these events, we developed an XGBoost model to infer drivers' judgments leading to takeovers. Key risk factors were then identified from the perspective of safety drivers. Finally, we validated the model's performance by assessing its consistency with actual human takeover behaviors. RESULTS: Through an analysis of high-risk takeover events in real emergency situations, the developed model was able to predict the takeover decisions of safety drivers with an accuracy of 92.2% and an F1 score of 91.3% using 765 time-sliced samples extracted from 113 high-risk cases. Interpretability analysis revealed that Time-to-Collision (TTC) and lateral distance are the main factors influencing the safety driver's risk perception and subsequent takeover behaviors. Furthermore, a comparison between the feature distributions at model-identified high-risk moments and those at actual driver takeovers revealed strong consistency in key indicators such as TTC and lateral distance. CONCLUSIONS: Based on real-world data from high-risk pedestrian-vehicle interactions, this study models interaction risk based on experienced drivers' risk perception and highlights key factors influencing human risk assessment. These findings provide insights for developing more human-aligned and interpretable risk assessment frameworks in autonomous driving.