Indoor Topological Localization Based on a Novel Deep Learning Technique

Abstract

Millions of people in the world suffer from vision impairment or vision loss. Traditionally, they rely on guide sticks or dogs to move around and avoid potential obstacles. However, both guide sticks and dogs are passive. They are unable to provide conceptual knowledge or semantic contents of an environment. To address this issue, this paper presents a vision-based cognitive system to support the independence of visually impaired people. More specifically, a 3D indoor semantic map is firstly constructed with a hand-held RGB-D sensor. The constructed map is then deployed for indoor topological localization. Convolutional neural networks are used for both semantic information extraction and location inference. Semantic information is used to further verify localization results and eliminate errors. The topological localization performance can be effectively improved despite significant appearance changes within an environment. Experiments have been conducted to demonstrate that the proposed method can increase both localization accuracy and recall rates. The proposed system can be potentially deployed by visually impaired people to move around safely and have independent life.