Global Localization from a Single Image in Known Indoor Environments

We address the problem of estimating the position and orientation of robots within a pre-constructed global map, which is called global localization. We construct a system based on the client-server model where a client has a camera. The global map is represented as a 3D wireframe model and stored in a server. The basic idea is image matching between the line segment image obtained from the image sent by the client and the database images generated by projecting the pre-constructed wireframe model to various viewpoints. If the number of database images is large, highly accurate localization can be expected, but the search cost increases. In this paper, we propose a method to localize the client from a single image in a practical time. Our method reduces the number of database images by distance transform. Distance transform provides the robustness to the shift of viewpoint while retaining the characteristics of original line segments. We also propose a new line matching method to handle distance transformed image. We experimentally showed that our method worked successfully in an actual indoor environment.