Semi-Automatic Town-Scale 3D Mapping Using Building Information From Publicly Available Maps

The 3D maps are used for self-positioning estimation and path planning for the autonomous navigation of robots in urban areas. This paper presents a framework that generates globally consistent 3D maps from the pose graph of existing simultaneous localization and mapping (SLAM) methods. Our approach corrects a pose graph by performing a 3D alignment with building information on publicly available maps. The framework automatically finds an appropriate anchor pose for the alignment and optimizes the pose graph according to the constraints obtained by the alignment. However, there are situations where it is difficult to automate 3D alignment because of measurement errors as well as errors in publicly available maps. To minimize operational costs, the proposed framework incorporates a user interface (UI) that allows users to check the results of 3D map alignment and make simple corrections. The framework was evaluated by conducting 3D mapping experiments in an urban area in Japan, and 3D mapping was performed over a distance of approximately 15 kilometers. The experimental results showed that the framework could automatically select anchor poses with high probability and generate 3D city maps with an average of approximately five manual operations per km by the user. The accuracy of the 3D mapping was evaluated by comparing it with a manually corrected reference trajectory based on an accurate 3D map from a commercial mobile mapping system (MMS). The 3D maps had an average absolute position error of 5.5, which is the lowest error compared to the maps generated by other open source software (OSS) SLAM methods.