Reading minds on the road: decoding perceived risk in automated vehicles through 140K+ ratings

2025-08-26

AI Summary

Perceived risk assessment for autonomous driving passengers remains fundamentally challenged by the latent and dynamic nature of psychological states. To address this, we propose the first large-scale, temporally continuous perceived risk decoding framework. Our approach leverages 236 hours of high-fidelity driving videos annotated with 141,628 human safety ratings to construct the largest publicly available continuous-risk dataset to date. We introduce a novel methodology that jointly optimizes discrete rating reconstruction and deep neural modeling, enabling millisecond-level risk prediction solely from vehicle kinematic signals. Furthermore, we integrate interpretable AI techniques to identify salient behavioral and environmental risk drivers, thereby establishing a new paradigm for dynamic, psychologically grounded risk measurement in autonomous vehicles. Experimental results demonstrate an average relative prediction error below 3%, significantly advancing both real-time accuracy and model interpretability in passenger risk perception estimation.

Technology Category

• Reasoning under Uncertainty: Sequential Decision Making
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Intelligent Robots: State Estimation

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Perceived risk in automated vehicles (AVs) can create the very danger that automation is meant to prevent: a frightened rider may hesitate when seconds matter, misjudge hazards, or disengage. However, measuring how perceived risk evolves in real time during driving remains challenging, leaving a gap in decoding such hidden psychological states. Here, we present a novel method to time-continuously measure and decode perceived risk. We conducted a controlled experiment where 2,164 participants viewed high-fidelity videos of common highway driving scenes and provided 141,628 discrete safety ratings. Through continuous-signal reconstruction of the discrete ratings, we obtained 236 hours of time-continuous perceived risk data - the largest perceived risk dataset to date. Leveraging this dataset, we trained deep neural networks that predict moment-by-moment perceived risk from vehicle kinematics with a mean relative error below $3%$. Explainable AI analysis uncovers which factors determine perceived risk in real time. Our findings demonstrate a new paradigm for quantifying dynamic passenger experience and psychological constructs in real time. These findings can guide the design of AVs and other machines that operate in close proximity to people, adjusting behaviour before trust erodes, and help realise automation's benefits in transport, healthcare, and service robotics.

Problem

• Decoding real-time perceived risk in automated vehicles
• Measuring hidden psychological states during driving scenarios
• Predicting moment-by-moment risk perception from vehicle kinematics

Innovation

• Continuous-signal reconstruction of discrete safety ratings
• Deep neural networks predicting risk from vehicle kinematics
• Explainable AI analysis uncovering real-time risk factors