Using the C ONDENSATION Algorithm for

Robust, Vision-based Mobile Robot Localization

Frank Dellaert† Wolfram Burgard‡ Dieter Fox† Sebastian Thrun†

†
Computer Science Department, Carnegie Mellon University, Pittsburgh PA 15213
‡
Institute of Computer Science III, University of Bonn, D-53117 Bonn

Abstract viding a mobile robot with autonomous capabilities.”

To navigate reliably in indoor environments, a mobile The current state of the art in localization often uses
robot must know where it is. This includes both the abil- laser range finders or sonar [2], but these sensor modalities
ity of globally localizing the robot from scratch, as well have a number of problems. They tend to be easily con-
as tracking the robot’s position once its location is known. fused in highly dynamic environments, e.g., when the laser
Vision has long been advertised as providing a solution to or sonar beams are blocked by people [12]. Being limited
these problems, but we still lack efficient solutions in un- in range, they cannot be expected to provide accurate lo-
modified environments. Many existing approaches require calization in large open areas, where no landmarks can be
modification of the environment to function properly, and observed [15]. Finally, these systems tend to be expensive.
those that work within unmodified environments seldomly Vision has long been advertised as providing a solution
address the problem of global localization. to these problems, but efficient solutions in unmodified en-
In this paper we present a novel, vision-based local- vironments are still lacking. Many of the approaches re-
ization method based on the C ONDENSATION algorithm viewed in [2] require artificial landmarks such as bar-code
[17, 18], a Bayesian filtering method that uses a sampling- reflectors [10], reflecting tape, or visual patterns that are
based density representation. We show how the C ONDEN - easy to recognize, such as black rectangles with white dots
SATION algorithm can be used in a novel way to track the [1]. But even approaches that use more natural landmarks
position of the camera platform rather than tracking an ob- [26, 20, 23, 15] are used mainly to track the position of a
ject in the scene. In addition, it can also be used to globally robot, and are not able to globally localize the robot.
localize the camera platform, given a visual map of the en-
In this paper, we present a vision-based approach to lo-
vironment.
calization based on the C ONDENSATION algorithm, which
Based on these two observations, we present a vision-
provides both robustness and the capability to globally lo-
based robot localization method that provides a solution to
calize the robot. The idea is to use a visual map of the
a difficult and open problem in the mobile robotics commu- ceiling, obtained by mosaicing, and localize the robot us-
nity. As evidence for the viability of our approach, we show
ing a simple scalar brightness measurement as the sensor
both global localization and tracking results in the context
input. Because the information coming from this sensor
of a state of the art robotics application.
is highly ambiguous, a Kalman filter approach is not ap-
1 Introduction plicable. The C ONDENSATION algorithm was introduced
To operate autonomously, mobile robots must know in the vision community precisely to deal with this type
where they are. Mobile robot localization, that is the pro- of measurement uncertainty [17, 18]. It belongs to the
cess of determining and tracking the location of a mobile general class of Monte Carlo filters, invented in the sev-
robot relative to its environment, has received consider- enties [16], and recently rediscovered independently in the
able attention over the past few years. Accurate local- target-tracking [13], statistical [21], and computer vision
ization is a key prerequisite for successful navigation in literature [17, 18]. These algorithms use a set of random
large-scale environments, particularly when global models samples or particles to represent the propagation of arbi-
are used, such as maps, drawings, topological descriptions, trary probability densities over time. It is this property that
and CAD models [22]. As demonstrated by a recent survey makes them ideally suited to our application.
of localization methods by Borenstein, Everett, and Feng The resulting localization method, which we have called
[2], the number of existing approaches is diverse. Cox [7] the Monte Carlo Localization method, draws on the fol-
noted that “Using sensory information to locate the robot lowing contributions:
in its environment is the most fundamental problem to pro- 1. We show how the C ONDENSATION algorithm can
be used in a novel way to track the position of the camera Prediction Phase In the first phase we use a motion
platform rather than tracking an object in the scene. model to predict the current position of the robot in the
2. We show how in addition to providing robust track- form of a predictive PDF p(xk |Z k−1 ), taking only mo-
ing, the C ONDENSATION algorithm can be used to globally tion into account. We assume that the current state xk is
localize the robots, given a visual map of the environment. only dependent on the previous state xk−1 (Markov) and
3. Based on these two observations, we present a vision- a known control input uk−1, and that the motion model is
based robot localization method that provides a solution to specified as a conditional density p(xk |xk−1 , uk−1 ). The
a difficult and open problem in the robotics community. predictive density over xk is then obtained by integration:
As evidence for the viability of our approach, we show Z
both global localization and tracking results in the context p(xk |Z k−1
) = p(xk |xk−1, uk−1 ) p(xk−1 |Z k−1) dxk−1
of a state of the art robotics application. In this applica- (1)
tion, the mobile robot Minerva functioned as an interac-
tive robotic tour-guide in the Smithsonian’s National Mu-
seum of American History. During its two weeks of per- Update Phase In the second phase we use a measure-
formance, it interacted with over 50,000 people, travers- ment model to incorporate information from the sensors
ing more than 44 km. Our successful results within this to obtain the posterior PDF p(xk |Z k ). We assume that
unmodified and highly dynamic environment attest to the the measurement zk is conditionally independent of earlier
promise of our approach. measurements Z k−1 given xk , and that the measurement
In the remainder of this paper we first discuss the gen- model is given in terms of a likelihood p(zk |xk ). This term
eral problem of robot localization, in Section 2. This is expresses the likelihood of the state xk given that zk was
made more explicit in Section 3, where we discuss how a observed. The posterior density over xk is obtained using
simple vision-based sensor can give rise to complex like- Bayes’ theorem:
lihood densities. In Section 4, we provide a taxonomy of p(zk |xk )p(xk |Z k−1 )
algorithms based on the way they deal (or not) with these p(xk |Z k ) = (2)
p(zk |Z k−1 )
types of densities. In Section 5, we show how the C ON -
DENSATION algorithm provides an elegant solution to the
problem, and use it as the basis for a novel localization After the update phase, the process is repeated recur-
method. Finally, Section 6 shows global localization and sively. At time t0 the knowledge about the initial state x0
tracking results in our museum testbed application. is assumed to be available in the form of a density p(x0 ).
In the case of global localization, this density might be a
2 Robot Localization uniform density over all allowable positions. In tracking
work, the initial position is often given as the mean and co-
In robot localization, we are interested in estimating
variance of a Gaussian centered around x0 . In our work,
the state of the robot at the current time-step k, given
as in [4], the transition from global localization to track-
knowledge about the initial state and all measurements
ing is automatic and seamless, and the PDF evolves from
Z k = {zk , i = 1..k} up to the current time. Typically, we
spanning the whole state space to a well-localized peak.
work with a three-dimensional state vector x = [x, y, θ]T ,
i.e. the position and orientation of the robot. This estima- 3 Vision-Based Localization
tion problem is an instance of the Bayesian filtering prob- In this section, we explain the basic idea behind our
lem, where we are interested in constructing the posterior vision-based approach to localization, and show that it
density p(xk |Z k ) of the current state conditioned on all gives rise to complex and multi-modal likelihood densities
measurements. In the Bayesian approach, this probability p(z|x). This provides the motivation for using the C ON -
density function (PDF) is taken to represent all the knowl- DENSATION algorithm rather than a Kalman filter based
edge we possess about the state xk , and from it we can approach.
estimate the current position. Often used estimators are the The basic idea of our approach is to generate an image
mode (the maximum a posteriori or MAP estimate) or the mosaic of the ceiling of the environment in which the robot
mean, when the density is unimodal. However, particu- needs to operate, and use that during operation to localize
larly during the global localization phase, this density will the robot. An example of such a ceiling mosaic from our
be multi-modal and calculating a single position estimate testbed-application is shown in Fig. 1. The figure shows
is not appropriate. a large portion (40 m. deep by 60 m. wide) of the first
Summarizing, to localize the robot we need to recur- floor ceiling of the Museum for American History (MAH),
sively compute the density p(xk |Z k ) at each time-step. one of the Smithsonians in Washington, DC. The mosaic
This is done in two phases: was generated from 250 images by globally aligning the
Figure 2: The likelihood densities p(z|x) of being at position x given a brightness measurement z for three values of z.
Left: high z indicates the robot is under a light. Middle: intermediate z. Right: low z: not under a light.

resolution.
The likelihood densities p(z|x) associated with this
simple sensor model are in general complex and multi-
modal, which explains the need for more powerful prob-
abilistic reasoning. To illustrate this, three different likeli-
hood densities are shown in Fig. 2, respectively when the
robot is under a light, at the border of a light, and not under
a light. It is clear that these densities do not resemble any-
thing remotely like a Gaussian density, and that more pow-
erful density representations are required than provided by
the Kalman filter.

4 Existing Approaches:
A Tale of Density Representations
The solution to the robot localization problem is ob-
Figure 1: Example of a large-scale ceiling mosaic. The tained by recursively solving the two (1) and (2). Depend-
area shown is 40 m. deep by 60 m. wide. ing on how one chooses to represent the density p(xk |Z k ),
one obtains algorithms with vastly different properties:
The Kalman filter If both the motion and the measure-
images, and estimating the texture of the ceiling from the ment model can be described using a Gaussian density, and
aligned images. The alignment was done using a variant of the initial state is also specified as a Gaussian, then the den-
an algorithm by Lu and Milios [25] for map-building from sity p(xk |Z k ) will remain Gaussian at all times. In this
laser range finder scans. The details are beyond the scope case, (1) and (2) can be evaluated in closed form, yield-
of the current paper, and are described elsewhere [8]. ing the classical Kalman filter [27]. Kalman-filter based
At run time, we use a simple scalar measurement to lo- techniques [24, 31, 14] have proven to be robust and accu-
calize the robot. Rather than working with entire images, rate for keeping track of the robot’s position. Because of
the idea is to use a very simple sensor that can be inte- its concise representation (the mean and covariance matrix
grated at high frequencies while requiring very little com- suffice to describe the entire density) it is also a particu-
putation and memory. This is an important property in en- larly efficient algorithm. However, it is clear that the basic
abling small, low cost robots. In particular, the brightness assumption of Gaussian densities is violated in our applica-
of a small ceiling patch directly above the robot is mea- tion, where the likelihood densities are typically complex
sured. In our testbed application, the measurement is done and multi-modal, as discussed above.
using a camera pointed at the ceiling, by extracting a small Topological Markov Localization To overcome these
window of 25 by 25 pixels from the center of the image. disadvantages, different approaches have used increasingly
A Gaussian filter with a standard deviation of 6 pixels is richer schemes to represent uncertainty, moving away from
then applied, to create a single scalar measurement. The the restricted Gaussian density assumption inherent in the
blurring is essential to filter out frequency components that Kalman filter. These different methods can be roughly dis-
cannot be represented within the mosaic, which is of finite tinguished by the type of discretization used for the rep-
resentation of the state space. In [28, 32, 19, 34], Markov In analogy with the formal filtering problem outlined in
localization is used for landmark-based corridor navigation Section 2, the algorithm proceeds in two phases:
and the state space is organized according to the topolog-
ical structure of the environment. However, in many ap- Prediction Phase In the first phase we start from the set
plications one is interested in a more fine-grained position of particles Sk−1 computed in the previous iteration, and
estimate, e.g., in environments with a simple topology but apply the motion model to each particle sik−1 by sampling
large open spaces, where accurate placement of the robot from the density p(xk |sik−1 , uk−1 ):
is needed. (i) for each particle sik−1 :
draw one sample s0 k from p(xk |sik−1 , uk−1 )
Grid-based Markov Localization To deal with multi- i

modal and non-Gaussian densities at a fine resolution, one

can discretize the interesting part of the state space, and In doing so a new set S 0 k is obtained that approximates
use it as the basis for an approximation of the density a random sample from the predictive density p(xk |Z k−1).
p(xk |Z k ), e.g. by a piece-wise constant function [3]. This The prime in S 0 k indicates that we have not yet incorpo-
idea forms the basis of the grid-based Markov localization rated any sensor measurement at time k.
approach [5, 11]. Methods that use this type of represen- Update Phase In the second phase we take into account
tation are powerful, but suffer from the disadvantages of the measurement zk , and weight each of the samples in
computational overhead and a priori commitment to the S 0 k by the weight mik = p(zk |s0 k ), i.e. the likelihood of
i
size of the state space. In addition, the resolution and
s0 k given zk . We then obtain Sk by resampling from this
i
thereby also the precision at which they can represent the
weighted set:
state has to be fixed beforehand. The computational re-
quirements have an effect on accuracy as well, as not all (ii) for j=1..N:
draw one Sk sample sjk from {s0 k , mik }
i
measurements can be processed in real-time, and valuable
information about the state is thereby discarded.
The resampling selects with higher probability samples s0 k
i
Sampling-based Methods Finally, one can represent
the density by a set of samples that are randomly drawn that have a high likelihood associated with them, and in do-
from it. This is the representation we will use, and it forms ing so a new set Sk is obtained that approximates a random
the topic of the next section. sample from p(xk |Z k ). An algorithm to perform this re-
sampling process efficiently in O(N) time is given in [6].
5 Monte Carlo Localization
After the update phase, the steps (i) and (ii) are repeated
In this section, we present a novel localization method
recursively. To initialize the filter, we start at time k = 0
based on a sampling-based representation, using the C ON -
with a random sample S0 = {si0 } from the prior p(x0 ).
DENSATION algorithm [17, 18] to propagate the posterior
density over time. In the localization context, C ONDEN - A Graphical Example
SATION is used to estimate the position of the camera plat- A B C D
form rather than an object in the scene. In addition, the p(xk−1 |Z k−1 ) p(xk |Z k−1 ) p(zk |xk ) p(xk |Z k )
ability to represent arbitrary densities enables us to per-
form global localization without modification.
In sampling-based methods one represents the density
p(xk |Z k ) by a set of N random samples or particles Sk =
{sik ; i = 1..N } drawn from it. We are able to do this be-
cause of the essential duality between the samples and the
density from which they are generated [33]. From the sam-
ples we can always approximately reconstruct the density,
e.g. using a histogram or a kernel based density estimation
technique.
The goal is then to recursively compute at each time-
step k the set of samples Sk that is drawn from p(xk |Z k ). Fig. 3: The probability densities and particle sets for one
A particularly elegant algorithm to accomplish this is the iteration of the algorithm. See text for detail.
C ONDENSATION algorithm [17, 18], known alternatively
as the bootstrap filter [13] or the Monte-Carlo filter [21]. One iteration of the algorithm is illustrated in Figure 3.
These methods are generically known as particle filters, In the figure each panel in the top row shows the exact den-
and an overview and discussion of their properties can be sity, whereas the panel below shows the particle-based rep-
found in [9]. resentation of that density. In panel A, we start out with a
 Fig. 5: Initial uncertainty Fig. 6: Representing ambiguity Fig. 7: Global localization
at iteration 15. at iteration 38. at iteration 126.

resampling to sample from the posterior is called Sam-

pling/Importance Resampling (SIR) [30], and an accessi-
ble introduction to it can be found in [33].

6 Experimental Results
As a test-bed for our approach we used the robotic tour-
guide Minerva, a prototype RWI B18 robot shown in Fig.
4. In the summer of 1998, Minerva functioned for two
weeks as an interactive robotic tour-guide in the Smith-
sonian’s National Museum of American History. During
this time, it interacted with thousands of people, travers-
ing more than 44 km. The datasets we work with consist
of logs of odometry and vision-based measurements col-
lected while Minerva operated within the MAH environ-
ment. The upward pointing camera, used both for building
the ceiling-mosaic as performing the brightness measure-
ment, can be seen on the right-hand side of Fig. 4. Using
Figure 4: Our testbed, the robotic museum tour-guide Min- the time stamps in the logs, all tests have been conducted
erva. The ceiling camera can be seen on the right. in real-time.
6.1 Global Localization
cloud of particles Sk−1 representing our uncertainty about One of the key advantages of using a sampling-based
the robot position. In the example, the robot is fairly local- representation over Kalman-filter based approaches is its
ized, but its orientation is unknown. Panel B shows what ability to represent multi-modal probability distributions.
happens to our belief state when we are told the robot has This ability is a precondition for localizing a mobile robot
moved exactly one meter since the last time-step: we now from scratch, i.e., without knowledge of its starting loca-
know the robot to be somewhere on a circle of 1 meter tion. This global localization capability is illustrated in Fig.
radius around the previous location. Panel C shows what 5 through Fig. 7. In the first iteration, the algorithm is ini-
happens when we observe a landmark, half a meter away, tialized by drawing 40,000 samples from a uniform prob-
somewhere in the top-right corner: the top panel shows the ability density save where there are known to be (static)
likelihood p(zk |xk ), and the bottom panel illustrates how obstacles. After 15 iterations of the algorithm, the samples
each sample s0 k is weighted according to this likelihood.
i
are still scattered over the area, but tend to be concentrated
Finally, panel D shows the effect of resampling from this in areas with high brightness. This is shown in Fig. 5.
weighted set, and this forms the starting point for the next After 38 iterations, all but a few possibilities have been
iteration. eliminated, as shown in Fig. 6. This is possible because all
Good explanations of the mechanism underlying the samples that disagree with the actual temporal sequence of
C ONDENSATION algorithm are given in [6, 29]. The en- brightness measurements become very unlikely. It should
tire procedure of sampling, reweighting and subsequently be noted that in this early stage of localization, the ability
 Fig. 8: Estimated path Fig. 9: Corrected path
using odometry only using vision

to represent ambiguous probability distributions is vital for research is to adapt the number of samples appropriately.
successful position estimation. Finally, after 126 iterations Also, the noisy nature of the estimated position occurs of-
only one localized peak is left, as shown in Fig. 7, which ten when two or more probability peaks arise. In this case,
indicates that the robot has been able to uniquely determine rather than plotting one mean position, it might be more ap-
its position. Note that the samples are shown after the re- propriate to display several alternative paths, correspond-
sampling step, where they are unweighted, and that most ing to the modes of the posterior probability density.
of the off-peak samples have in fact low probability.
6.2 Position Tracking 7 Conclusion and Future Work
Using the same approach, we are also able to track the In this paper, we have shown that C ONDENSATION type
position of a robot over time, even when the robot is mov- algorithms can be used in a novel way to perform global
ing at high speeds. In this experiment, we used recorded and local localization for mobile robots. The ability of
data from Minerva, as it was moving with speeds up to 1.6 these Monte Carlo methods to deal with arbitrary likeli-
m/sec through the museum. At the time of this run there hood densities is a crucial property in order to deal with
were no visitors in the museum, and the robot was remotely highly ambiguous sensors, such as the vision-based sen-
controlled by users connected through the world wide web. sor used here. In addition, the ability to represent multi-
To illustrate this, Fig. 8 shows an occupancy grid map modal posterior densities allows them to globally localize
of the museum along with the estimated trajectory using the robot, condensing an initially uniform prior distribution
odometry alone. In the tracking experiment, the robot po- into one globally localized peak over time.
sition is initialized with the true position at the top of the Based on these properties, the resulting Monte Carlo
map. Due to the accumulated odometry error, the robot is Localization approach has been shown to address one of
hopelessly lost after traversing 200 meters. In contrast, the the fundamental and open problems in the mobile robot
path estimated using the vision-sensor keeps track of the community. In order to support this claim, we have shown
global position at all times. This can be appreciated from results within the context of a challenging robot applica-
Fig. 9, where the vision-corrected path is shown. This path tion. Our method was able to both globally localize and
was generated by running the C ONDENSATION algorithm locally track the robot within a highly dynamic and un-
with 1000 samples, and plotting the mean of the samples modified environment, and this with the robot traveling at
over time. The resulting estimated path is noisier than the high speeds.
one obtained by odometry, but it ends up within 10 cm. of The results in this paper were obtained using a single
the correct end-position, indicated with a black dot at the brightness measurement of the ceiling directly above the
right of the figure. robot’s position. In part, this is done to investigate how lit-
In general, far fewer samples are needed for position tle sensing one can get away with. In future work, this can
tracking than for global localization, and an issue for future be further developed in two opposite directions: one can
envision building robots with even less sensory input, e.g., [16] J.E. Handschin. Monte Carlo techniques for prediction and filtering
by omitting wheel encoders that provide odometry mea- of non-linear stochastic processes. Automatica, 6:555–563, 1970.
surements. This shifts the burden towards the development [17] M. Isard and A. Blake. Contour tracking by stochastic propagation
of even more powerful paradigms for probabilistic reason- of conditional density. In Eur. Conf. on Computer Vision (ECCV),
pages 343–356, 1996.
ing. In the other direction, it is clear that integrating more
[18] M. Isard and A. Blake. Condensation – conditional density propaga-
complex measurements will provide more information at tion for visual tracking. International Journal of Computer Vision,
each iteration of the algorithm, resulting in faster global 29(1):5–28, 1998.
localization and more accurate tracking than shown here. [19] L.P. Kaelbling, A.R. Cassandra, and J.A. Kurien. Acting under un-
In future work, we hope to investigate both these possibil- certainty: Discrete Bayesian models for mobile-robot navigation.
ities in more depth. In Proceedings of the IEEE/RSJ International Conference on Intel-
ligent Robots and Systems, 1996.
References [20] S. King and C. Weiman. Helpmate autonomous mobile robot nav-
[1] J. Borenstein. The nursing robot system. PhD thesis, Technion,
igation system. In Proceedings of the SPIE Conference on Mobile
Haifa, Israel, 1987.
Robots, volume 2352, 1990.
[2] J. Borenstein, B. Everett, and L. Feng. Navigating Mobile Robots:
[21] G. Kitagawa. Monte carlo filter and smoother for non-gaussian non-
Systems and Techniques. A. K. Peters, Ltd., Wellesley, MA, 1996.
linear state space models. Journal of Computational and Graphical
[3] R. S. Bucy. Bayes theorem and digital realisation for nonlinear fil- Statistics, 5(1):1–25, 1996.
ters. Journal of Astronautical Science, 17:80–94, 1969.
[22] D. Kortenkamp, R.P. Bonasso, and R. Murphy, editors. AI-based
[4] W. Burgard, A. Derr, D. Fox, and A.B. Cremers. Integrating global Mobile Robots: Case studies of successful robot systems, Cam-
position estimation and position tracking for mobile robots: the Dy- bridge, MA, 1998. MIT Press.
namic Markov Localization approach. In IEEE/RSJ Intl. Conf. on
Intelligent Robots and Systems (IROS), 1998. [23] D. Kortenkamp and T. Weymouth. Topological mapping for mobile
robots using a combination of sonar and vision sensing. In Proceed-
[5] W. Burgard, D. Fox, D. Hennig, and T. Schmidt. Estimating the
ings of the Twelfth National Conference on Artificial Intelligence,
absolute position of a mobile robot using position probability grids.
pages 979–984, 1994.
In Proc. of the Fourteenth National Conference on Artificial Intelli-
gence (AAAI-96), pages 896–901, 1996. [24] J.J. Leonard and H.F. Durrant-Whyte. Directed Sonar Sensing for
[6] J. Carpenter, P. Clifford, and P. Fernhead. An improved particle filter Mobile Robot Navigation. Kluwer Academic, Boston, 1992.
for non-linear problems. Technical report, Department of Statistics, [25] F. Lu and E Milios. Globally consistent range scan alignment for
University of Oxford, 1997. environment mapping. Technical report, Department of Computer
[7] I.J. Cox. Blanche—an experiment in guidance and navigation of Science, York University, April 1997.
an autonomous robot vehicle. IEEE Transactions on Robotics and [26] M. J. Matarić. A distributed model for mobile robot environment-
Automation, 7(2):193–204, 1991. learning and navigation. Master’s thesis, MIT, Artificial Intelligence
[8] Frank Dellaert, Sebastian Thrun, and Chuck Thorpe. Mosaicing a Laboratory, Cambridge, 1990.
large number of widely dispersed, noisy, and distorted images: A [27] P. Maybeck. Stochastic Models, Estimation and Control, volume 1.
Bayesian approach. Technical report, Computer Science Depart- Academic Press, New York, 1979.
ment, Carnegie Mellon University, 1999.
[28] I. Nourbakhsh, R. Powers, and S. Birchfield. DERVISH an office-
[9] A. Doucet. On sequential simulation-based methods for Bayesian navigating robot. AI Magazine, 16(2):53–60, Summer 1995.
filtering. Technical Report CUED/F-INFENG/TR.310, Department
of Engineering, University of Cambridge, 1998. [29] M. K Pitt and N. Shephard. Filtering via simulation: auxiliary par-
ticle filters. Technical report, Department of Mathematics, Imperial
[10] H. Everett, D. Gage, G. Gilbreth, R. Laird, and R. Smurlo. Real-
College, London, October 1997.
world issues in warehouse navigation. In Proceedings of the SPIE
Conference on Mobile Robots IX, volume 2352, 1994. [30] D.B. Rubin. Using the SIR algorithm to simulate posterior distri-
[11] D. Fox, W. Burgard, and S. Thrun. Active Markov localization for butions. In Bayesian Statistics, volume 3, pages 395–402. Oxford
mobile robots. Robotics and Autonomous Systems, 25(3-4):195– University Press, 1988.
207, 1998. [31] B. Schiele and J. L. Crowley. A comparison of position estimation
[12] D. Fox, W. Burgard, S. Thrun, and A.B. Cremers. Position estima- techniques using occupancy grids. In IEEE Conf. on Robotics and
tion for mobile robots in dynamic environments. In Proceedings of Automation (ICRA), volume 2, pages 1628–1634, 1994.
the AAAI Fifteenth National Conference on Artificial Intelligence, [32] R. Simmons and S. Koenig. Probabilistic robot navigation in par-
1998. tially observable environments. In Proc. International Joint Confer-
[13] N.J. Gordon, D.J. Salmond, and A.F.M. Smith. Novel approach to ence on Artificial Intelligence, 1995.
nonlinear/non-Gaussian Bayesian state estimation. IEE Procedings [33] A.F.M. Smith and A.E. Gelfand. Bayesian statistics without tears: A
F, 140(2):107–113, 1993. sampling-resampling perspective. American Statistician, 46(2):84–
[14] J.-S. Gutmann and C. Schlegel. Amos: Comparison of scan match- 88, 1992.
ing approaches for self-localization in indoor environments. In [34] S. Thrun. Bayesian landmark learning for mobile robot localization.
Proceedings of the 1st Euromicro Workshop on Advanced Mobile Machine Learning, 33(1), 1998.
Robots. IEEE Computer Society Press, 1996.
[15] G. Hager, D. Kriegman, E. Yeh, and C. Rasmussen. Image-based
prediction of landmark features for mobile robot navigation. In
IEEE Conf. on Robotics and Automation (ICRA), pages 1040–1046,
1997.