PILOT-DSM: Risk perception-aligned takeover prompting framework via coverage-gap assessment in conditionally automated driving.

Takeover performance in conditionally automated driving can degrade when drivers fail to perceive critical hazards after a takeover request (TOR), while existing prompting strategies are often environment-driven and rarely assess what drivers have visually covered. In dynamic multi-risk scenarios, this limitation may lead to biased prompting by highlighting already perceived or less relevant hazards while leaving unattended critical risks unaddressed. To address this gap, we propose Programmatic Imitation Learning for Ocular Tracking and Dynamic ScanMatch (PILOT-DSM), a closed-loop takeover prompting framework that combines expert visual scanning logic with assessment of drivers' object-level hazard coverage within a short post-TOR window. PILOT learns interpretable expert gaze-transition policies, and DSM compares a driver's gaze sequence with expert reference strategies to detect coverage gaps and select unattended hazards for intervention through an augmented reality head-up display (AR-HUD). In a high-fidelity simulation with four prompting conditions, linear mixed-effects model analyses showed that PILOT-DSM-HUD significantly reduced gaze latency and increased takeover success rate relative to No-HUD, and outperformed Static-HUD in planned contrasts. Descriptively, it achieved the lowest gaze latency (M = 3.74 s) and highest takeover success rate (M = 71.46%), whereas differences from PILOT-Static-HUD were not significant after Holm correction. Pupil-diameter results indicated that PILOT-DSM-HUD was comparable to No-HUD and lower than Static-HUD, while the difference from PILOT-Static-HUD was not significant after correction. TTC analysis indicated target-dependent safety benefits, mainly for secondary and rear-approaching risks. These findings suggest that coverage-gap-driven prompting can enhance takeover safety while avoiding unnecessary cognitive burden.