Description

This paper combines FastSLAM and EKF-SLAM using the principle of Stefan Williams' constrained local submap filter. A frontier map is generated using FastSLAM, exploiting its ability to deal with nonlinearities and data association. This map is periodically converted to a Gaussian distribution and registered with a Gaussian base map, which is able to preserve long-term correlation information.

Abstract

This paper presents HybridSLAM: an approach to SLAM which combines the strengths and avoids the weaknesses of two popular mapping strategies: FastSLAM and EKF-SLAM. FastSLAM is used as a front-end, producing local maps which are periodically fused into an EKF-SLAM back-end. The use of FastSLAM locally avoids linearisation of the vehicle model and provides a high level of robustness to clutter and ambiguous data association. The use of EKF-SLAM globally allows uncertainty to be remembered over long vehicle trajectories, avoiding Fast- SLAM’s tendency to become over-confident. Extensive trials in randomly-generated simulated environments show that HybridSLAM significantly out-performs either pure approach. The advantages of HybridSLAM are most pronounced in cluttered environments where either pure approach encounters serious difficulty. In addition, the HybridSLAM algorithm is experimentally validated in a real urban environment.

Download

Full paper [pdf] (1.6 Mb, 16 pages)