Alignment of visual maps in multirobot FastSLAM

摘要

This paper focusses on the study of the Map Alignment problem in a multirobot SLAM context. The map fusion problem can be tackled in two stages: map alignment and map merging. The alignment stage consists in obtaining the tranformation between the reference systems of the robots. Then, in the map merging stage, the maps built by different robots can be fused into a single one. In this paper, we concentrate on the alignment stage. Particularly, we have a team of robots, each one building its own local map independently. At some point, the fusion of the maps may be required. In this case, these maps must be aligned. We therefore evaluate a set of aligning methods. These methods establish correspondences between each pair of maps and compute an initial estimate of the alignment. Finally, we apply the least squares minimization to obtain a more accurate solution. Each robot extracts distinctive 3D points from the environment with a stereo camera. Then, the robots use these observations as landmarks to build their maps while simultaneously localize themselves. The SLAM problem is solved using the FastSLAM algorithm. The movements of the robots occur only in 2D plane, so although the maps are 3D landmark-based, the alignment takes place only in 2D.