Map Registration of Lidar-based 2D NDT Map and Vision-based 3D NDT Map with Known Initial Robot Poses

Abstract

This paper focuses on the multi-robot mapping, especially the collaboration of vision-and 2D lidar-mounted robots with known initial poses. Since the scale of the map built by a single vision sensor is hardly the same to the real world scale, the scale should be estimated to register the 2D map built by 2D lidar and 3D map built by a vision sensor. In this paper, we propose a method of converting maps into NDT maps to rapidly register the two maps in multiple resolutions based on minimizing the modified objective function of distribution-to-distribution normal distributions transform (NDT-D2D). Instead of scaling the 3D map, the proposed method updates the 2D map within optimization. In the experiment, we showed the registration of 2D NDT map built by gmapping with 2D lidar and 3D NDT map built by direct sparse odometry (DSO) with vision sensor. As a result, the runtime was 0.041 second, and the error of the estimated scale was 18.0%.