Slawomir Grzonka
Untersuchungen zur Genauigkeit von SLAM-Verfahren mit
Partikel-Filtern
Abstract:
The ability to construct an accurate map of the environment is one of the vital properties of a
truly autonomous agent. In the context of simultaneous localization and mapping (SLAM), the
Rao-Blackwellized Particle Filter (RBPF) has emerged as an efficient approach to this problem.
The convergence of this technique can be guaranteed for an infinite number of map hypotheses.
In practice, however, only a limited number of hypotheses can be maintained, which leads to
an upper limit on the level of noise the system can deal with. Thus, the integration of laser
scans at false or unfavourable map positions cannot be avoided completely, which may lead
to inconsistent or even diverged maps. In this work, three extensions to the standard RBPF
algorithm are introduced that yield accurate maps even in low-structured environments and with
higher noise levels. The three extensions work complementarily at the three main parts of the
mapping algorithm respectively: the point of time before, during, and after the integration of the
current observation. Experiments with real robots in highly structured indoor environments as
well as in low-structured outdoor environments demonstrate that the developed techniques lead
to substantially improved maps. It is furthermore shown that the proposed system is nearly as
accurate as the scanmatching-based FastSLAM algorithm in the highly structured environment
and clearly outperforms it in the low-structured case.
Bibtex:
@MastersThesis {grzonka06thesis,
author = {Grzonka, S.},
title = {Untersuchungen zur Genauigkeit von SLAM-Verfahren mit Partikel-Filtern},
school = {University of Freiburg, Department of Computer Science},
year = {2006},
note = {In German}
}
Document (pdf file):
master thesis (15.2 MB)