Place Recognition for Simultaneous Localization and Mapping
The ability to recognize known places is an essential competence of any intelligent system that operates autonomously over longer periods of time. Approaches that rely on the visual appearance of distinct scenes have recently been developed and applied to large scale SLAM scenarios.
Modern SLAM systems are typically based on the efficient
optimization of probabilistic constraint or factor graphs.
These systems are generally divided into a back-end and
front-end. The back-end contains the optimizer that builds
and maintains a map by ﬁnding an optimal solution to
the robot’s trajectory and the landmark positions given the
constraints constructed by the front-end.
This front-end is
responsible for data association in general and, in the context
of pose-only SLAM, place recognition in particular.
Reliable place recognition is a hard problem, especially
in large-scale environments.
Repetitive structure and sensory
ambiguity constitute severe challenges for any place recogni-
tion system. As optimization based back-ends for SLAM like
iSAM, Sparse Pose Adjustment, iSAM2, or g2o are not robust against outliers, even a single wrong loop
closure will result in a catastrophic failure of the mapping
Recent developments in appearance-based place recognition therefore aimed at reaching a high recall rate at
100% precision, i.e. they concentrated on preventing false
positives. This of course leads to computationally involved,
very complex systems.
In parallel work
, we developed a robust formulation to
pose graph SLAM that allows the optimizer in the back-
end to identify and reject wrong loop closures. This can be
understood as enabling the back-end to take back any data
association decision of the front-end.
Given this robust back-end, the need of reaching a precision of 100% during the data
association (i.e. place recognition) process is eliminated. The
place recognition system in the front-end can therefore be
kept simple and focused on a high recall rate, as a reasonable
number of false positive loop closures is acceptable.
We propose BRIEF-Gist, an appearance-based
place recognition system that builds upon the BRIEF descriptor by Calonder et al. We evaluated BRIEF-Gist
and concluded that our approach is suitable to perform place
recognition in large scale scenarios, despite its simplicity
regarding implementation and computational demands.
Please see the paper below for further information and details.
Please contact Niko Sünderhauf
for any additional information.