YOLOv8-Powered Driver Monitoring: A Scalable and Efficient Approach for Real-Time Distraction Detection

Driver distraction remains a significant factor in road accidents, necessitating efficient and real-time monitoring systems. This study proposes an advanced driver distraction detection system using YOLOv8. The dataset comprises 30,000 images, including the AUC Distracted Driver Dataset and self-captured images from dashboard cameras. The dataset covers various distractions such as phone usage, texting, eating, adjusting the radio, and drowsiness. Roboflow and LabelImg tools were used for precise annotation, ensuring high-quality labeled data. A dataset partitioning strategy of 75% training, 15% validation, and 10% testing was employed to maximize learning efficiency and model evaluation. Extensive data preprocessing techniques, including image resizing, normalization, and augmentation methods such as rotation, flipping, brightness adjustment, Mosaic, and MixUp, were applied to enhance model generalization. Histogram Equalization was incorporated to improve contrast in low-light conditions. The YOLOv8 model architecture includes a CSP-Darknet53 backbone, a PANet neck, and an anchor-free detection head. For effective training, the AdamW optimizer was implemented with lx10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup>learning rate and a batch size of 32. The model was trained using the PyTorch-based Ultralytics YOLOv8 framework on an NVIDIA RTX 3090 GPU, with early stopping to prevent overfitting. Model evaluation was conducted using precision, recall, mean Average Precision (mAP@50), inference speed (FPS), and latency. Comparative analysis with Faster R-CNN and EfficientDet demonstrated YOLOv8's superior performance, achieving 95.2% mAP@50, 110 FPS, and a latency of 7ms per frame. The system ensures accurate classification of distracted driving behaviors with a false positive rate below 3%. Deployment in advanced driver assistance systems (ADAS), fleet monitoring, and real-time vehicle surveillance applications can enhance road safety. Future improvements include the integration of Vision Transformers (ViTs) for better contextual understanding and deployment on edge devices such as Jetson Nano for efficient on-device inference.