Scene/Object Recognition Based Navigation System for the Blind (Device Application - Raspberry Pi 3)

Abstract

The life of any people with visual impairment hasn’t been easy, helping them to ease their inconvenience with the aid of technology has always been an active research field. This paper presents a scene/object recognition-based navigation system for the blind and visually impaired implemented on a Raspberry Pi 3 device, which would notify the user the environment he/she is in via audio output. It is found that the method and the scene classification algorithm used by the predecessor is not strong enough to give a good performance, rather than training a Convolutional Neural Network from the scratch, it’s better, both in terms of quality and efficiency, to use off-the-shelf pre-trained network and readjust it to fit this project’s requirement, such approach has been proven successful throughout this project as the image classification accuracy has vastly increased after. However, compared to the original algorithm, the recognition time have dropped from 2 seconds to 5-6 seconds, this is not an optimized performance and should be studied in future. Furthermore, a more user-friendly user interface is designed.