Real-Time Driver Fatigue Detection and Classification Using YOLOv8-FD: A Precision-Driven Approach for Enhanced Road Safety

Objective: The present research proposes to develop a dependable, real-time driver fatigue detection system by integrating facial feature evaluation (eye-blinking, head nodding, pupil size, and yawning) with motion displacement analysis via the Improved Lucas-Kanade Optical Flow. The technology is designed to perform accurately in a variety of real-world scenarios, such as those with different lighting and eyeglasses. To enhance detection reliability, a multi-phase classification strategy that combines deep learning, conventional machine learning, and real-time detection models is employed. Methods: The proposed solution is evaluated using the NTHU dataset and real-time driving footage captured under varying lighting and eyewear conditions. Key features such as pupil size variation, eye-blinking, yawning, and head nodding are extracted using euclidean distances between facial landmarks. Modifications include preprocessing optimization and motion displacement analysis via enhanced Lucas-Kanade optical flow. Performance is assessed using metrics like F1 score, mAP, accuracy, precision, and Euclidean distance, and compared with established algorithms like CNN-ResNet50V2. Findings: The results obtained indicate in terms of accuracy, speed, and resilience, the YOLOv8-FD-based method performs significantly better than conventional models like CNN-ResNet50v2. Even when wearing eyewear, the system consistently recognizes fatigue indications, achieving an accuracy rate of 98.06% on the dataset obtained. Its exceptional performance results from the combination of facial attribute extraction and landmark-based euclidean distance analysis, which guarantees constant accuracy in a variety of situations. Additionally, under difficult circumstances like changing lighting and several kinds of eyewear, the system acquired an F1 score of 97%, mAP of 94%, and precision of 98.5%. Novelty: The novelty of this research lies in combining landmark-based Euclidean distance analysis with multiple facial indicators and employing YOLOv8-FD for real-time fatigue detection. This represents one of the earliest implementations of YOLOv8-FD in this domain. The system's exceptional performance under various eyewear conditions, coupled with its 98.06% accuracy, marks a significant advancement over previous CNN-based approaches, offering a more scalable and reliable solution for driver fatigue detection. Keywords: Fatigue, YOLOV8-FD, Euclidean, CNN, Nodding, Pupil