A New Formation Maintenance Technique for Particle Diversity in RBPF-SLAM

Abstract

Rao-Blackwellized particle filter (RBPF) has been to the fore as one of methods to solve simultaneous localization and mapping (SLAM) problem, i.e. RBPF-SLAM. The RBPF-SLAM, however, has been suffering from the particle depletion problem and the convergence problem caused by the improper posterior density and brutal rejection and replication of particles during resampling. We present a new technique to overcome those problems in RBPF-SLAM by keeping the particle diversity using particle formation maintenance (PFM). The triangular mesh structure is adaptively generated as a one form of PFM and it completely replaces the resampling part in RBPF-SLAM. Its considerable improvements regarding robot pose and features were shown in simulation by comparing conventional methods, i.e. FastSLAM 2.0 and PSO based FastSLAM.

Relational FastSLAM: an improved Rao-Blackwellized particle filtering framework using particle swarm characteristics

“…In our previous works, [17][18][19] the weights of particles were compensated for by heuristic rules. In addition, the particles underwent formation by means of the generation of an adaptive triangular mesh 20,21 instead of a resampling step.…”