Evaluating the robustness of facial landmark-based indicators for driver drowsiness detection under realistic conditions

Driver Drowsiness is a significant safety concern which creates a need for reliable systems that can detect signs of drowsiness under varying realistic conditions. This study evaluated the robustness of facial landmark-based indicators, including Eye Aspect Ratio(EAR), Mouth Aspect Ratio (MAR), and head pose(pitch, yaw, and roll values) for drowsiness detection under changes in resolution, lighting, and facial occlusions. Images from a driver drowsiness dataset were evaluated using threshold-based models and a logistic regression classifier. Performance was determined through accuracy, precision, recall, and F1 score. EAR, MAR, and head pose thresholds were determined through finding the best training F1 score. Under baseline conditions, logistic regression achieved an F1 score of 0.714, overall performing better than threshold-based models. Environmental conditions affected measurements of facial features and model performance, with facial occlusions and lighting affecting results more than image resolution. Ultimately, this study’s findings suggest that combining multiple landmark-based indicators using machine learning classifiers may be a more effective approach to driver drowsiness detection than fixed threshold based approaches.