Robot vision for the visually impaired

Abstract

We present a head-mounted, stereo-vision based navigational assistance device for the visually impaired. The head-mounted design enables our subjects to stand and scan the scene for integrating wide-field information, compared to shoulder or waist-mounted designs in literature which require body rotations. In order to extract and maintain orientation information for creating a sense of egocentricity in blind users, we incorporate visual odometry and feature based metric-topological SLAM into our system. Using camera pose estimates with dense 3D data obtained from stereo triangulation, we build a vicinity map of the user's environment. On this map, we perform 3D traversability analysis to steer subjects away from obstacles in the path. A tactile interface consisting of microvibration motors provides cues for taking evasive action, as determined by our vision processing algorithms. We report experimental results of our system (running at 10 Hz) and conduct mobility tests with blindfolded subjects to demonstrate the usefulness of our approach over conventional navigational aids like the white cane.