Smart Wearable Vision-Assistance System for Visually Impaired Using Edge-Based Deep Learning
TL;DR
The findings indicate that implementing effective assistive vision systems on ultra-low-power embedded systems is achievable and can be used in practice as a portable solution for everyday use.
Abstract
Assistive technologies based on computer vision have a huge potential in improving mobility of visually impaired people. However, they are not very widely adopted because they depend on high-power processors, cloud access, or complicated sensor configurations. This paper outlines a low-cost and lightweight wearable vision-assist system that is completely edge-based and provides real-time information in the environment. The framework is constructed using the Raspberry Pi Zero 2W and a lightweight object detection model optimized to run on a monocular camera on the device using the Tensorflow Lite. The design proposed offers object identification, rough distance estimation, and spatial position (right, left, centre) with audio feedback in real-time to enhance MSIA both indoors and outdoors. Experimental analysis demonstrates a mean detection rate of 92%, spatial localization rate of 85% and audio feedback latency of below 2 seconds at a power consumption in the range of less than 5W. These findings indicate that implementing effective assistive vision systems on ultra-low-power embedded systems is achievable and can be used in practice as a portable solution for everyday use.