UMEED: Deep Learning based walking and reading assistant for visually impaired people

Abstract

Millions of people in India are influenced by vision loss. Numerous advancements now are done in the field of smart assistance for visually impaired that uses ultrasound sensors and hardware-centric devices. These implementations generally increase the cost of the device making it unattainable for visually impaired. In this paper, a cheap, robust, complete walking solution is introduced, targeted at aiding the blind and low-vision people. This system is standalone and doesn’t require any network access to provide assistance which in turn makes it less complex and economical. It is built by focusing on network coverage issues in the rural regions of the country. The framework is fabricated focusing on the plight of visually challenged people in India and provides specialized functionalities including object detection, face recognition and OCR coupled with audio feedback. The OCR part of the framework is built to recognize Indian regional languages like Gujarati, Hindi, Bengali etc. along with English. The proposed device utilizes low-cost equipment and provides complete assistance to the visually challenged people focusing on the software part of the system. The facial recognition, object detection and object character recognition together help in replacing the imprecise use of traditional methods like white cane and guide dogs. With the enhancement of the applications in the field of machine vision, the scope of providing aid through a camera is endless. This framework provides precise detection resulting in a fully automated and highly accurate guiding assistant.