Implementasi YOLO V8 untuk Deteksi Kantuk pada Pengemudi Kendaraan Roda Empat Secara Real-Time

Abstract

Sleepy driving is one of the main causes of traffic accidents. It is estimated that in recent years, 20-30% of road accidents are caused by sleepy driving. Sleep-deprived driving (<6 hours) is associated with a 4-5 times increased risk of accidents. According to the Central Bureau of Statistics (BPS), traffic accidents in Indonesia increased from 100,028 in 2020 to 103,645 in 2021. Therefore, a drowsiness detection system for four-wheeled vehicle drivers was built using the You Only Look Once (YOLO) algorithm version 8 by considering head, eye, and mouth movement activities. The training data in the research model uses four-wheeled vehicle driver image data taken from roboflow and kaggle while the test data is collected independently in real-time. The best model was obtained using the k-cross validation technique with k=5 and a combination of hyperparameter fold 3 epoch 250 and batch size 16. The model training results were implemented into an android-based mobile application with an accuracy rate of 95% using the confusion matrix calculation method. The application system can display the bounding box with a delay time of 226 ms to 366 ms or 0.2 to 0.3 seconds on devices with 12 MP front camera specifications. Testing produces maximum results using devices with front camera specifications of 12MP and above compared to devices with front camera specifications of only 5 MP.