Enhanced Footpath Segmentation Using Attention-Integrated UNET For Assistive Urban Navigation

This study presents an advanced deep-learning ap- proach to enhance navigation in visually impaired individuals in urban environments. We employed a modified UNET architecture integrated with spatial and channel attention mechanisms for semantic segmentation to distinguish footpaths from other urban features effectively. The model was trained from a dataset of 3000 high-quality images, and it had an IoU of 89.82% on the test set, which exceeded the baseline model by 4.5%. We demonstrated the robustness of the model to adverse settings, including Gaussian noise and blurring, by maintaining an IoU greater than 85.32 %. When implemented in a prototype assistive navigation system, the model provided real-time feedback with 50 ms per image inference time, a far more rapid inference time that significantly increased user autonomy. In addition to advancing semantic segmentation techniques, this study also features a practical application for visually impaired navigation.