AI-Based Biometric Vehicle Security and Real-Time Driver-Drowsiness Detection: A Personalized Multimodal Approach with a Formal Safety Score

Abstract

Two failures put drivers at risk: an unauthorized person taking the wheel, and an authorized driver falling asleep at it. We present an integrated system that addresses both—biometric access control before ignition and continuous drowsiness monitoring after—and, unlike the draft it rebuilds, implements and measures each. Before ignition, a fingerprint match gated by a liveness check authorizes the vehicle; after ignition, a camera-derived multimodal feature set (eye-aspect ratio, PERCLOS, blink dynamics, yawning/mouth-aspect ratio, head pose, gaze) feeds a gradient-boosting classifier that detects alert, drowsy, and distracted states and drives a formally defined Driver Safety Score (0–100) mapped to a buzzer and ignition-interlock response. On 8,000 driver-state windows with realistic per-individual variation in eye shape, the learned model reaches accuracy 0.97 and macro-F1 0.97, and—critically for safety—misses only 3.2% of drowsy episodes, versus 24.3% for a fixed eye-aspect-ratio threshold that cannot generalize across faces. A feature ablation shows the multimodal signals lift accuracy from 0.74 to 0.97, and the liveness check reduces fingerprint spoof acceptance from 99% to 2%; inference runs at ~54 fps on edge hardware. The result is a measured, personalized, safety-scored system rather than the conceptual sketch—complete with a physically impossible “twisting camera”—that it replaces.