Traffic Sign Recognition Using Deep Learning with Interactive Voice Output for Drivers

Traffic safety is a critical issue with rising road accidents. World Health Organization (WHO) reports 1.19 million deaths and 20-50 million injuries annually due to traffic accidents. In Indonesia, $\mathbf{6 1 \%}$ of accidents are due to human factors, mainly traffic violations. Deep learning-based traffic sign recognition offers potential solutions, yet existing systems primarily provide visual information. This research aims to develop an Indonesian-specific traffic sign identification system using deep learning with interactive voice outputs. It integrates the YOLOv8 algorithm, ChatGPT API for information extraction, and Text-to-Speech technology for voice instructions. The model was trained on a combined dataset of Indonesian Traffic Sign Dataset and Road Sign Detection Dataset, employing data augmentation techniques like blurring and highlighting objects. The YOLOv8 model achieved a precision of $\mathbf{9 6. 8 \%}$, recall of $\mathbf{9 5. 3 \%}$, mAP50 of $\mathbf{9 7. 5 \%}$, and mAP50-95 of $\mathbf{8 2. 9 \%}$. Optimal hyperparameter tuning and data augmentation significantly improved performance. The GPT-3.5 Turbo API effectively provided contextual information, and Google’s Text-to-Speech technology delivered clear voice instructions.