Sub-Map Dividing and Realignment FastSLAM by Blocking Gibbs MCEM for Large-Scale 3-D Grid Mapping

Abstract

In this article, we propose a novel Simultaneous Localization and Mapping (SLAM) method by using a sampling-based approach. FastSLAM is well-known approach as a sampling-based SLAM method. FastSLAM utilizes a theorem that map errors are decidable under a sample of trajectories. From this theorem, FastSLAM samples many trajectories and maps to find a minimum error map. However, in case of constructing a large-scale grid map, FastSLAM becomes unuseful since the method requires huge memory to generate many grid maps. Our proposed method requires only one map that is deformable corresponding to a trajectory. In order to find a minimum error map, our method only generates trajectories. Our method enables construction of a minimum error map by little memory in comparison with original FastSLAM. Experimental results demonstrate our method is able to construct a large-scale 3-D grid map by low memory usage in comparison with original FastSLAM.