Enhancing road safety with DL vision: do driver distraction alerts hold the key?

Driver distraction is a significant contributor to road accidents and traffic safety challenges. This study evaluates the performance of various YOLO models trained on a combination of custom and publicly available datasets, with their performance assessed using metrics such as mAP@50, precision, recall, F1-score, inference time, FPS, and model parameters. The aim is to determine the best trade-off between accuracy, speed, and computational efficiency, providing insights into the practical applicability of these models for distracted driving detection. Among the models, YOLOv8l achieves the highest accuracy with a mAP@50 of 0.921, precision of 0.911, recall of 0.895, and an F1-score of 0.899, making it suitable for scenarios prioritizing detection precision. YOLOv5m balances accuracy and efficiency, achieving a mAP@50 of 0.917 and an F1-score of 0.896, while maintaining moderate inference times. For real-time applications, YOLOv5n and YOLOv8n offer the fastest speeds, with inference times of 0.9 ms and FPS of 1111, at the cost of reduced accuracy. YOLOv8m provides a balance between performance and speed, achieving a mAP@50 of 0.915 and F1-score of 0.898, with an inference time of 3.4 ms and FPS of 294. A post-simulation survey reveals that participants find the visual warning system effective in mitigating driver distraction and enhancing perceived road safety. By integrating simulation data and user feedback, this study highlights the potential of advanced deep learning models like YOLO in improving driver assistance systems and promoting safer driving practices.