Semantic Mapping and Navigation with Visual Planar Landmarks

Abstract

We propose a semantic map representation and human-like navigation strategies for the mobile robot with a monocular camera. First, we develop a method to automatically detect landmarks, which make up a perceived planar region. Next, we build a vision-based map with the detected visual planar landmarks. To build a map with a single camera, we use the concept of bearing-SLAM. The landmark bearings are measured by a camera from the detected planar regions. By measuring two bearings between three feature points in the detected planar regions, we estimate the distance from the robot to the landmark for an observation model. After building a vision-based map, we extract semantic information. The proposed semantic map represents the topology of the environment with nodes (area and landmarks) and their spatial relationships. Next, we attempt to apply human navigation strategies for the robot navigation with semantic map. We apply strategies (path integration, view-dependent place recognition, reorientation, and active searching for additional landmarks) to mobile robots and demonstrate a human-like navigation system based on a semantic map.