A Deep Learning Based Framework for Blind Road Occupancy Detection

Abstract

This research paper addresses the pressing concern of road safety for the visually impaired in densely populated regions, especially in China. While tactile paving exists to guide blind individuals, unexpected obstacles pose serious hazards. The study proposes an Artificial Intelligence (AI) system utilizing advanced machine learning techniques to identify obstacles on blind roads, providing real-time feedback for navigation. To overcome data limitations, a virtual environment using Pybullet is created for data generation, combining synthetic and real-world images for training. This study introduces a Convolutional Neural Network (CNN)-based model and integrates a Vision Transformer (ViT) model, comparing their efficacies. The progressive training approach yields a highly effective CNN model, outperforming ViT models. The practical application of the CNN model in real-world scenarios has proven its efficacy in detecting obstacles, underscoring its reliability and significant contribution to improving road safety for visually impaired individuals. This research extends beyond providing a safety enhancement measure; it also sheds light on the wider applicability of such methods in addressing issues like the scarcity of real-world data.