Reliable risk assessment remains a central challenge for autonomous vehicles. Although automation levels keep rising, passenger cognition offers a non‑intrusive auxiliary signal that can boost both objective safety and perceived safety without requiring active intervention. This paper introduces an electroencephalogram (EEG)‑based brain‑computer interface (BCI) that decodes passenger neural responses for Risk Prediction (RP) and Danger Identification (DI), explicitly modeling passengers as real‑world occupants. To this end we propose the Passenger Cognitive Model (PCM), Risk‑aware Sequential Labeling (RSL) and the Passenger EEG Decoding Strategy (PEDS), with a 3D Convolutional Recurrent Neural Network (3D‑CRNN) as the core decoder that jointly learns spatio‑temporal EEG features. Experiments show that 3D‑CRNN achieves a Balanced Accuracy (BA) of $95.3\% \pm 2.7\%$ on RP and, with RSL, improves single‑subject DI from $80.9\% \pm 3.9\%$ to $85.0\% \pm 3.2\%$. Event‑wise analysis indicates that 3D‑CRNN consistently outperforms other models across different event types for both RP and DI. In cross‑session DI tests the model reaches $77.0\% \pm 5.3\%$ BA, while cross‑subject evaluation yields $77.4\% \pm 1.1\%$ BA on seen subjects and $64.9\% \pm 8.5\%$ BA on unseen subjects, demonstrating strong intra‑ and inter‑subject generalization. These results validate the feasibility of an EEG decoding framework for AV passenger hazard perception and suggest that passenger cognitive signals can provide auxiliary supervision for future AV decision‑making and Safety of the Intended Functionality (SOTIF).
Review: The study convincingly bridges neuro‑signal decoding with autonomous driving safety, opening a promising avenue for human‑centric risk assessment.