Drowsy driving is a major factor in traffic accidents, necessitating more accurate and rapid detection systems. This paper compares two advanced visual recognition frameworks, YOLOv5 and YOLOv8, applied to real-time driver drowsiness detection. Algorithms underwent training and evaluation procedures using a collection of visual samples. Performance assessment utilized multiple evaluation criteria: accuracy indicators (precision, recall), combined metrics (F1 score), and threshold-dependent evaluations (mAP calculated at 0.5 and 0.5:0.95 IoU levels) Results showed both models performed well across all measures, with precision and recall consistently exceeding 0.98. YOLOv8 demonstrated slight advantages in F1-score and inference speed, indicating its suitability for real-time applications, particularly in embedded environments. YOLOv5, while somewhat slower, exhibited robust performance at higher IoU thresholds, reflecting its reliability in precise localization tasks. This comparative analysis confirms the effectiveness of both YOLOv5 and YOLOv8 in detecting driver drowsiness with high reliability. The findings highlight practical trade-offs between computational efficiency and detection accuracy, providing guidance for selecting appropriate models based on deployment constraints and performance priorities in intelligent transportation systems.