Development of a Safety System for Drivers Using Recognition of Facial Expressions in Agricultural Vehicles

Driving agricultural vehicles often requires prolonged attention and exposure to intense working environments, which makes drivers more susceptible to fatigue, drowsiness, and inattentiveness. This situation results in many agricultural accidents that lead to injuries, productivity losses, and even fatalities. Since most traditional safety systems are reactive or subject to human bias, there is a growing demand for intelligent, automated inspection systems capable of detecting driver drowsiness or fatigue. This study proposes a real-time safety system that uses deep learning-based facial recognition to detect driver drowsiness. The system is specifically designed for agricultural vehicles. A novel dataset of four driver states, normal, drowsy, sleepy, and yawning, was developed from the recordings of agricultural operators under diverse conditions. For model training, a diverse and comprehensive dataset of 2546 keyframes was provided for detailed and balanced annotation. The YOLOv8 object detection model was trained to recognize key facial signs, and a temporal logic algorithm was introduced to distinguish between natural eye blinks and fatigue-related prolonged eye closure, as well as true yawning events. The system was deployed on an NVIDIA Jetson Nano edge device and integrated with audible buzzer alerts. Laboratory and field tests demonstrated strong detection accuracy, stability under diverse lighting conditions, and reliable real-time performance. The results revealed that the YOLOv8 model achieved a mean Average Precision (mAP50) of 90.1%, with class-wise precision values of 0.98 (normal face), 0.94 (drowsy face), 0.85 (normal eye), 0.79 (sleepy eye), normal mouth (0.91), and 0.902(yawning), demonstrating robust detection across all driver states. During field testing, the system successfully detected facial expressions even under harsh conditions such as vibration, partial occlusion, and high sunlight. Temporal smoothing logic reduced false blink detections, and the alert mechanism responded accurately with rapid beeping for early warning and continuous alerting for critical fatigue states. These results confirmed that the proposed safety system using a facial expression recognition behavior system provides a reliable, low-cost, and real-time solution for improving agricultural machinery safety.