A Real-Time Driver Drowsiness Detection Method Using a Hybrid of Deep Learning and Fuzzy Logic

Driver drowsiness contributes to approximately 21% of traffic accidents. Three deep neural networks were trained, with ResNet50 achieving the best performance: 99.74% training accuracy, 99.62% validation accuracy, an average test F1-score of 0.99, no overfitting, and real-time inference at 0.023 seconds per frame on an RTX 3060 laptop GPU with 6GB RAM. Following drowsiness detection, a fuzzy inference system integrating Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR) is applied to determine the drowsiness severity. The system was tested with an infrared camera on 453 volunteers from various Middle Eastern ethnic groups under different lighting conditions. A local database was compiled from these tests, recording 31 misclassifications and achieving an overall accuracy of 93.16%. The findings demonstrate that the system performs reliably under both day and night conditions and across diverse ethnicities, supporting its suitability for integration into advanced driver-assistance systems (ADAS).