![](bg1.png)
BEARING-ONLY SAM USING A MINIMAL INVERSE DEPTH
PARAMETRIZATION
Application to Omnidirectional SLAM
Cyril Joly and Patrick Rives
INRIA Sophia Antipolis M´editerran´ee, 2004 route des Lucioles, Sophia Antipolis Cedex, France
Keywords:
Simultaneous localization and mapping (SLAM), Smoothing and mapping (SAM), Extended Kalman filter
(EKF), Bearing-only, Inverse depth representation.
Abstract:
Safe and efficient navigation in large-scale unknown environments remains a key problem which has to be
solved to improve the autonomy of mobile robots. SLAM methods can bring the map of the world and the
trajectory of the robot. Monucular visual SLAM is a difficult problem. Currently, it is solved with an Extended
Kalman Filter (EKF) using the inverse depth parametrization. However, it is now well known that the EKF-
SLAM become inconsistent when dealing with large scale environments. Moreover, the classical inverse depth
parametrization is over-parametrized, which can also be a cause of inconsistency. In this paper, we propose to
adapt the inverse depth representation to the more robust context of smoothing and mapping (SAM). We show
that our algorithm is not over-parameterized and that it gives very accurate results on real data.
1 INTRODUCTION
Simultaneous localization and mapping (SLAM) is a
fundamental and complex problem in mobile robotics
research which has mobilized many researchers since
its initial formulation by Smith and Cheeseman
((Smith and Cheeseman, 1987)). The robot moves
from an unknown location in an unknown environ-
ment and proceeds to incrementally build up a navi-
gation map of the environment, while simultaneously
using this map to update its estimated position. Tra-
ditional methods are based on the extended Kalman
filter (EKF). It is now well known that EKF based
methods yield inconsistencies ((Julier and Uhlmann,
2001; Bailey et al., 2006a)). This is essentially due
to linearization errors. Furthermore, these algorithms
suffer from computational complexity (O(N
2
) where
N is the number of landmarks in the map).
The FastSLAM method, introduced by Thrun and
Montemerlo ((Montemerlo et al., 2003)), is based on
the particle filter. Each position is approximated by a
set of M random particles and one map is associated to
each particle. It can be shown that MlogN complex-
ity is achievable. However, the algorithm is sensitive
to the number of particles chosen. Furthermore, the
issue of particles diversity can make the FastSLAM
inconsistent ((Bailey et al., 2006b)).
In the above-mentionedmethods, SLAM is solved
as a filtering problem: the state-space contains only
the state of the robot at the current time and the state
of the map. SLAM can also be stated as a smooth-
ing problem: the whole trajectory is taken into ac-
count (Thrun and Montemerlo, 2006; Dellaert and
Kaess, 2006). This approach is referred as simulta-
neous smoothing and mapping (SAM). Although it is
based on the linearization of the equations, the ap-
proach gives better results than EKF because all the
linearization points are adjusted during the optimiza-
tion.
A very interesting problem is the monocular vi-
sual SLAM which is also referred as Bearing-Only
SLAM. Since camera does not measure directly the
distance between the robot and the landmark, there
is an observability issue. In standard approach, land-
marks are initialized with a delay. Recent approaches
try to avoid this and propose undelayed formulation.
The method presented in (Civera et al., 2008) consists
in changing the parametrization of the landmarks.
This parametrization includes the location of the first
position where the landmark was observed and the in-
verse depth between the landmark and this position.
Also the latter algorithm produces good results
in real situations, it has some drawbacks. First,
the parametrization of the landmarks is not minimal.
Then, this parametrization is often associated with the
EKF (although it is inconsistent).
281
Joly C. and Rives P. (2010).
BEARING-ONLY SAM USING A MINIMAL INVERSE DEPTH PARAMETRIZATION - Application to Omnidirectional SLAM.
In Proceedings of the 7th International Conference on Informatics in Control, Automation and Robotics, pages 281-288
DOI: 10.5220/0002949702810288
Copyright
c
SciTePress