A Multimodal Driver Drowsiness and Distraction Monitoring System: Design and Preliminary Evaluation

Abstract

Driver drowsiness and distraction have long beenrecognised as major contributors to road-traffic crashes, and lowcostdriver-monitoring prototypes have frequently been restrictedto a single visual cue such as eye closure. In this work, amultimodal, webcam-based driver-monitoring system has beendesigned and preliminarily evaluated. Five complementary visualsignals—Eye Aspect Ratio (EAR), Percentage of Eye Closure(PERCLOS), Mouth Aspect Ratio (MAR), head pose, and upperbodyposture—have been extracted in real time, and facialand body landmarks have been obtained using MediaPipe.The features have been normalised and combined through atransparent weighted-linear fusion model, from which a CompositeDrowsiness Score has been derived and mapped to fourgraduated alert levels. A driver-specific calibration stage has beenincorporated so that personalised baselines are established beforemonitoring begins. Three alert-classification configurations—arule-based fusion system, a MobileNetV2 classifier, and a stackedensemble—have been compared on a subject-independent splitof the Driver Monitoring Dataset over 3,605 held-out frames.Overall accuracies of about 46%, 57%, and 54% have beenobtained, macro ROC-AUC values of up to 0.69 have beenrecorded, and an ablation study of the five cues has been reported.The most severe alert category has been found difficult to classifybecause of class imbalance and a face-only camera view. Ithas therefore been concluded that the interpretable multimodalpipeline is technically feasible, while systematic evaluation onpublic benchmarks has been identified as the principal remainingwork.