Vision-Based Road-Following Using Results of Semantic Segmentation for Autonomous Navigation

Recent research into autonomous navigation of a robot have used accurate and dense three-dimensional sensors such as 3DLiDAR and RADAR for map building and localization. However, humans move from a given position to their destination without accurate metric maps in urban scenes: a topological map including only landmarks and their connections enables navigation. Our study endeavors to develop a visual navigation scheme based on a topological map, similar to the scheme used by humans, and this paper proposes a unique road-following scheme using a combination of image processing schemes with the results of semantic segmentation. The proposed scheme identifies a target point toward which the robot moves. To confirm the feasibility of the proposed scheme, a moving experiment on a 500-meter-long course in our university campus, where several people were moving back and forth, was conducted with a robot named 'Emu' using only three webcams as external sensors. The experimental results demonstrated that the robot controlled by the proposed scheme could navigate the course adequately.