E va lu ation of P ose On ly S L A M - University of Missourivigir.missouri.edu/~gdesouza/Research/Conference_CDs/IEEE_IROS... · E va lu ation of P ose On ly S L A M G ibson H u,
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Evaluation of Pose Only SLAM
Gibson Hu, Shoudong Huang and Gamini Dissanayake
Abstract— In recent SLAM (simultaneous localization andmapping) literature, Pose Only optimization methods havebecome increasingly popular. This is greatly supported by thefact that these algorithms are computationally more efficient,as they focus more on the robots trajectory rather than dealingwith a complex map. Implementation simplicity allows these tohandle both 2D and 3D environments with ease. This paperpresents a detailed evaluation on the reliability and accuracyof Pose Only SLAM, and aims at providing a definitive answerto whether optimizing poses is more advantages than optimiz-ing features. Focus is centered around TORO, a Tree basednetwork optimization algorithm, which has gained increasedrecognition within the robotics community. We compare thiswith Least Squares, which is often considered one of the bestMaximum Likelihood method available. Results are based onboth simulated and real 2D environments, and presented in away where our conclusions can be substantiated.
I. INTRODUCTION
One of the main focuses on current SLAM research is
the development of solutions to improve SLAM efficiency
without compromising the accuracy. SLAM itself can be
considered as an optimization problem, solved by combining
the information gained from sensor observations and robot
odometry. Researchers have typically resorted to simplifying
data and fitting it around point feature based solutions which
aims to compute optimal locations of both features and robot
poses [1].
Recently, Pose Only SLAM approaches have gained pop-
ularity [2] [3]. A Pose Only implementation typically divides
SLAM into two separate phases, one for the identification of
relative pose constraints and the second for the optimization
of robot poses. During the first phase, the consistency of
information use must be monitored and information reuse
must be avoided. For the second phase, the major focus is
on efficiency and accuracy, that is, how to get a good quality
solution quickly.
A popular Pose Only SLAM algorithms is Tree-based Net-
work Optimizer or TORO. It has been evaluated to be much
faster than most standard maximum likelihood approaches
and stated to work well in many different applications [3].
When it is difficult or impossible to extract features from
the sensor data, Pose Only SLAM seems to be a good
choice for optimizing the robot poses and locating the robot
in an unknown environment. However, if there are good
This work is supported by the ARC Linkage Grant (LP0884112),the ARC Centre of Excellence program (funded by the Australian Re-search Council (ARC) and the NSW State Government), the Pem-pek Systems Pty Ltd and the University of Technology, Sydney, Aus-tralia. Gibson Hu, Shoudong Huang and Gamini Dissanayake arewith Faculty of Engineering and Information Technology, Universityof Technology Sydney, PO Box 123 Broadway NSW 2007, Australia{ghu,sdhuang,gdissa}@eng.uts.edu.au
quality features that can be extracted from the environment,
how much accuracy or consistency is compromised for the
efficiency in TORO or Pose Only SLAM?
In this paper, we assume the point feature based SLAM
set up and ask the above question. We want to know
whether it is necessary to compute the optimal locations of
features at all if accurate feature positions can be gained
from Pose Only SLAM. In other words, how accurate are
the implementations of Pose Only SLAM when compared
with the Full Least Squares solution. Also how does TORO’s
result differ from a Least Squares based optimization where
information use is maximized?
The paper is structured as follows. Section II explains the
three SLAM techniques we used to conduct our experiments.
Section III explains our approach to maintain consistency
when obtaining Pose Only constraints. Section IV describes
our evaluation methods. Section V presents experimental
results and Section VI discusses related work. Lastly, Section
VII draws conclusions on our findings.
II. THREE SLAM ALGORITHMS
A fair comparison to a feature based SLAM solution can
be made by providing a Least Squares benchmark. Unlike
methods such as Extended Kalman Filtering (EKF), the Least
Squares solution keeps all the robot poses and features in its
state vector to avoid any information loss.
The computational efficiency can be improved by means
of map joining or exploiting the sparseness of its information
matrix [6]. The Least Squares result is arguably one of the
most accurate estimations one can achieve.
The basic principle behind Least Squares is the minimiza-
tion of the error function
(Z − F (X))T P−1(Z − F (X)) (1)
where X is the state vector, Z is the measurement infor-
mation and P is the covariance of the measurement. The
problem itself is not always linear therefore the state vector
should be solved iteratively by
(JT P−1J) ·Xk+1 = JT·P−1(Z −F (Xk) + J ·Xk)) (2)
where J is the Jacobian.
Often the algorithm can be made more robust by using
a Levenberg Marquardt implementation where a damping
factor is introduced to improve the convergence.
In this paper we use Least Squares in two ways.
The 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems October 18-22, 2010, Taipei, Taiwan
testing using larger data sets, the tradeoff between efficiency
and accuracy involved in TORO is still undetermined.
In the future, we are planning to use more large-scale data
sets to compare the three algorithms in both 2D and 3D
scenarios. Some performance comparison with other SLAM
algorithms based on local map joining will also be very
interesting.
REFERENCES
[1] F. Dellaert and M. Kaess, “Square root SAM: Simultaneous localizationand mapping via square root information smoothing”. International
Journal of Robotics Research, vol. 25, no. 12, December 2006, pp. 1181-1203.
[2] F. Lu and E. Milios. Globally consistent range scan alignment forenvironment mapping. Autonomous Robots, 4:333-349, 1997.
[3] Giorgio Grisetti, Cyrill Stachniss, Slawomir Grzonka, and WolframBurgard; ”A Tree Parameterization for Efficiently Computing Maximum
Likelihood Maps using Gradient Decent.”, Robotics: Science and Systems
(RSS), 2007[4] G. Grisetti, D. L. Rizzini, C. Stachniss, E. Olson and W. Burgard, “On-
line constraint network optimization for efficient maximum likelihoodmapping”. In Proceedings of 2008 IEEE International Conference on
Robotics and Automation (ICRA), Pasadena, California, on May 19-23,2008.
[5] W. Burgard, C. Stachniss, G. Grisetti, B. Steder, R. Kuemmerle, C.Dornhege, M. Ruhnke, A. Kleiner, J. D. Tardos, “A comparison of SLAMalgorithms based on a graph of relations.” IROS, 2009
[6] S. Huang, Z. Wang, G. Dissanayake, and U. Frese, “Iterated D-SLAMmap joining: evaluating its performance in terms of consistency, accuracyand efficiency.” Autonomous Robots, (2009) 27: 409-429
[7] J. L. Blanco, F. A. Moreno, J. Gonzalez, A collection of outdoor roboticdatasets with centimeter-accuracy ground truth. Autonomous Robots,(2009) 27: 327-351.
[8] S. Ceriani, G. Fontana, A. Giusti, D. Marzorati, M. Matteucci, D.Migliore, D. Rizzi, D. G. Sorrenti, P. Taddei, Rawseeds ground truthcollection systems for indoor self-localization and mapping. Autonomous
Robots, (2009) 27: 353-371.[9] Eustice, R. M., Singh, H., and Leonard, J. 2005. Exactly sparse delayed-
state filters. In Proc. IEEE International Conference on Robotics and
Automation, pp. 2428-2435.[10] Giorgio Grisetti, Cyrill Stachniss, and Wolfram Burgard: Non-linear
constraint network optimization for efficient map learning, IEEE Trans-
[11] Grisetti Giorgio, Slawomir Grzonka, Cyrill Stachniss, Patrick Pfaff,and Wolfram Burgard: Efficient Estimation of Accurate Maximum Like-lihood Maps in 3D., IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), 2007[12] R. Kummerle, B. Steder, C. Dornhege, M. Ruhnke, G. Grisetti, C.
Stachniss, A. Kleiner, On measuring the accuracy of SLAM algorithms,Autonomous Robots, (2009) 27: 387-407.
[13] E. Olson, J. Leonard, and S. Teller. Fast iterative optimization of posegraphs with poor initial estimates, In Proc. of the IEEE Int Conf on
Robotics and Automation (ICRA), pages 2262 -2269, 2006[14] Chen, L., Arambel, P. O., and Mehra, R. K. 2002. Estimation under un-
known correlation: covariance intersection revisited. IEEE Transactions
on Automatic Control 47(11):1879-1882.[15] Julier, S. J., and Uhlmann, J. K. 2001. Simultaneous localization and
map building using split covariance intersection. In Proc. IEEE/RSJ
International Conference on Intelligent Robots and Systems, pp. 1257-1262.