Designing Reliability-Aware Mobile Driver Alerts From Cross-Model Perception Conflict

Abstract

Mobile and in-vehicle interfaces increasingly surface automated-driving perception outputs to drivers, yet detector confidence scores alone provide a poor basis for deciding when a system should interrupt driver attention. This study proposes a reliability-aware alert model converting spatial disagreement among heterogeneous detectors into four interface states—normal, caution, request attention, and sensor degraded—using a Gaussian-process reliability field, reliability-gated attention, and conflict-aware suppression. Benchmark results indicate improved detection and calibration under corruption. An offline, synthetic stress-test shows a hysteresis mapping with explicit thresholds reduces anomalous state oscillation by 66.5% versus a naive design, at a tunable precision-recall operating point. This evidence assesses technical plausibility only, not driver comprehension, workload, trust, or behaviour. The paper contributes design requirements, a quantified evidence-to-alert mapping, and a preregistration-style evaluation protocol. A controlled user study remains necessary before usability or safety claims can be made.