Deep Learning Approach for Sleepiness Detection with YOLOv12

Drowsiness while driving is a significant contributor to traffic accidents. To mitigate such occurrences, a precise and real-time drowsiness detection system is essential. This research aims to create a computer vision-based drowsiness detection system utilizing the YOLOv12 algorithm. The dataset was sourced from Kaggle and manually annotated with the help of Roboflow. It was categorized into two groups: drowsy and non-drowsy, with the original 5,000 images augmented to a total of 6,976 images. The model training utilized the AdamW optimizer (learning rate=0.001667, momentum=0.9) over 100 epochs and a batch size of 4. Performance assessment indicates that the model attained an mAP@50 of 0.732 and an mAP@50-95 of 0.62, alongside a precision of 0.648 and a recall of 0.928. These findings illustrate that YOLOv12 can successfully identify drowsiness in real-time. Nevertheless, the performance of the model is significantly influenced by the quality and balance of the dataset. Consequently, enhancing the structure and distribution of the dataset is vital for improving detection accuracy.