Towards Seamless Sidewalk Navigation: On-Device Machine Learning for Real-Time Obstacle Detection in Visually Impaired Assistance

Undeniably, visual impairment severely affects the quality of life and impacts many daily activities of visually impaired individuals. Visually impaired individuals have difficulty navigating on side-walks. There are many assistive tools available for navigational assistance for visually impaired individuals. The majority of assistive technologies for sidewalk navigation in visually impaired individuals rely on server-based models, introducing challenges of latency, data costs, and privacy. This research investigates on-device machine learning as an alternative, emphasizing real-time feedback and user experience. We assessed algorithms, including EfficientDet-Lite, SSD, and YOLOv4, optimizing them for mobile deployment. The resultant Android application, embedding the top-performing model, demonstrated the potential for immediate, server-independent feedback. This research not only bridges a notable gap in the literature but also paves the way for research on more accessible, immediate, and discreet navigation tools for the visually impaired community.