Indoor navigation for visually impaired

Ng, Wei Yu (2025) Indoor navigation for visually impaired. Final Year Project, UTAR.

PDF Download (3619Kb)

Abstract

The increasing prevalence of visual impairments globally has highlighted the importance of effective assistive technologies to enable independent navigation for the visually impaired in indoor environments. Traditional indoor navigation systems often suffer from high costs, signal interference, and reliance on pre-installed infrastructure, making them inaccessible to many. This project proposes a cost-effective and infrastructure-free indoor navigation system that leverages a smartphone's camera and advanced computer vision. The system's core is a novel localization approach that uses a DINOv2 vision transformer to generate robust semantic embeddings from real-time images. For superior accuracy, initial embedding comparisons are verified by a Vision-Language Model (VLM), which provides contextual understanding of the scene. To handle dynamic environments, the system integrates YOLOv8-seg and Stable Diffusion 2.0 to detect and remove people from the camera feed, ensuring reliable performance. Implemented as a Flutter mobile application with a high-performance FastAPI and Supabase backend, the system provides users with step-by-step guidance along pre-recorded visual routes, delivering clear instructions via audio feedback and a hands-free voice assistant. By integrating state-of-the-art deep learning models into an accessible platform, this project addresses the limitations of existing solutions and contributes meaningfully to advancing assistive technology for the visually impaired community.

Actions (login required)