Study on A Navigation System for Visually Impaired Persons based on Egocentric Vision Using Deep Learning

There are 310 thousand visually impaired persons in Japan. They use usually a white cane or a guide dog while walking. However, the number of guide dogs is less than the required number, and the white cane is difficult to get surrounding information. So, we are developing a navigation system based on egocentric vision instead of walking support tools such as a white cane. The aim of the research is to develop the navigation system with to recognize objects and estimate distances using deep learning. Therefore, to investigate to refer to landmarks object for non-handicapped persons while walking and dangerous objects for the visually impaired person. As a result, the error of the distance system for people was less than 10%, and the error of the distance system of 4-5m was less than 1m. In other words, we think that obstacle detection and distance can be presented to visually impaired people using the method proposed in this study.