This research delves into the development of a fatigue detection system based on modern object detection algorithms, particularly YOLO (You Only Look Once) models, including YOLOv5, YOLOv6, YOLOv7, and YOLOv8. By comparing the performance of these models, we evaluate their effectiveness in real-time detection of fatigue-related behavior in drivers. The study addresses challenges like environmental variability and detection accuracy and suggests a roadmap for enhancing real-time detection. Experimental results demonstrate that YOLOv8 offers superior performance, balancing accuracy with speed. Data augmentation techniques and model optimization have been key in enhancing system adaptability to various driving conditions.