Having landmarks All Over the Scene - a Visual-Assisted Relocalization Strategy for Indoor Robots

It is widely known that certain indoor scenarios, e.g., server rooms with repetitive rows of server cabinets, are challenging for robot positioning strategies using 2D LiDAR (light detection and ranging) sensors. An efficient start-up relocalization function enabling the robot to localize itself at any random spot of its operating area is not yet available in state-of-the-art in-door robot products. In this paper, we present a promising visual-assisted relocalization scheme for indoor robots as a mature engineering solution to the aforementioned critical issues. The main idea is to use the features captured in image frames as a description of the scenes and to establish a link between these features and the robot's pose such that they can be regarded as natural landmarks. Then via an efficient feature matching procedure, initial pose estimates can be obtained and provided to the primary LiDAR-based relocalization scheme to compute a final pose estimate. A thorough experimental analysis with our robot products in various operating scenarios verify that by combining the advantages of visual and LiDAR sensors, our proposed relocalization strategy is able to achieve a satisfactory performance, and its extensions tailored to meet engineering requirements further enhance the robustness and flexibility of this scheme.