Active markers for outdoor and indoor robot localization

Active markers for outdoor and indoor robot
localization
Riccardo Cassinis, Fabio Tampalini and Roberto Fedrigotti
Department of Electronics for Automation
University of Brescia
Via Branze 38, I-25123, Brescia, Italy
{riccardo.cassinis, fabio.tampalini}@ing.unibs.it, roberto.fedrigotti@libero.it
Abstract
Localization is one of the fundamental problems
in mobile robot navigation. In this paper a
study is presented on a new methodology aimed at
localizing a mobile robot in indoor and outdoor
environments using active markers and commercial
off-the-shelf webcams.
The marker detection system is based on the difference
of working frequencies of the shutter of a
webcam and of a signal from the marker. This
study is part of a more complex project for a patrolling
robot and will eventually become a part of
a message-oriented middleware called DCDT (Device
Communities Development Toolkit) developed
in the past few years at the University of Brescia.
1 Introduction
Localization is a key problem in making truly autonomous
robots. Although many different methodologies
have so far been developed in order to find the location of
a robot relative to the environment in which it operates,
most of them are quite expensive, and many exhibit restrictions
that limit their practical application to some
environments and/or situations.
In this paper a system is presented, that relies on the
recognition of an active marker by a vision system based
on a webcam. Although it requires some structuring of the
environment, this method is appealing because the tooling
is simple and inexpensive, and in many cases already existing
cameras can be used. Since DCDT allows an easy
construction of distributed robots, the traditional placement
of sensors and markers has been reversed, i.e. the
marker is on the robot and the sensor sits in a fixed location.
Computation can take place off-board the robot, relieving
its processor from the burden of image processing
and coordinates computation.
The original part of the method is the use of the beat
created by the shutter of a webcam and a signal from the
marker working at slightly different frequencies.
Moreover, the system is structured to be a part of the
message-oriented middleware DCDT (Device Communities
Development Toolkit) (Cassinis et al., 2001); (Cassinis
et al., 2003), developed at the University of Brescia,
with the purpose of taking advantage of already existing
computing, sensing and communication resources, limiting
the cost of additional equipment.
2 The localization problem
Many different strategies have been applied to the localization
problem.
Tracking or local techniques, like odometry readings
and other dead reckoning methods, keep track of the position
of a robot while it is navigating in the environment,
but since the position estimates are based on earlier positions
and due to the fact that this method accumulates
sensor errors and wheel slippage, the error in the estimates
increases over time, and at least periodical recalibration
procedures that use other localization methods are
required. Other techniques rely on the fact that a robot
operates in a structured or semi-structured environment
and use different types of environment features on different
levels of perceptual abstraction as easily and reliably
recognizable objects (Panzieri et al., 2001). Some of them
are map-based positioning or model matching techniques,
that use geometric features of the environment, like lines
that describe walls in hallways or offices (Davison, 2003);
(Borenstein et al., 1996); (Gutmann et al., 2004).
The matching of maps can come with a significant
amount of uncertainty regardless they are known in advance
or not, so many solutions have been presented including
approaches employing Kalman filtering (Georgiev
and Allen, 2004); (Davison, 2003); (Arras and Tomatis,
1999), grid-based Markov localization, or Monte
Carlo methods (Wolf and Pinz, 2003); (Jensfelt, 2001).
Other methods try to obtain absolute measurements
through the use of beacons (Batalin and Sukhatme, 2003)

or natural or artificial landmarks (Jang et al., 2002),
(Betke and Gurvits, 1997). Many solutions adopt artificial
landmarks (Batalin and Sukhatme, 2003); (Maeda et al.,
2003); (Lamon et al., 2001); (Ma et al., 2002); (Parker et
al., 2004); (Roh et al., 1997); (Fiala, 2004) because they
can be designed for optimal detectability even under adverse
environmental conditions.
The system presented in this paper provides absolute
localization from scratch to a robot in a dynamic environment,
by means of a webcam and the use an optical
active marker.
It is quite obvious that absolute positioning systems
offer great advantages over relative methods, because
their readings can be immediately used. In many cases
high precision is also not an issue, especially when the
system is used for “driving the robot around”, as it happens
for instance in surveillance robots.
Furthermore, absolute systems do not suffer from “robot
stealing” problems, since each reading is either independent
from all the previous ones, or, if previous readings
are used to quicken or to enhance the current one, a
procedure exists that allows not to use them with little
performance degradation.
3 Description of the system
The localization system was designed to aid MORGUL,
MObile Robot for Guarding University Laboratories, in
its task of patrolling an area in the campus of the University
of Brescia (Cassinis et al., 2005a), (Cassinis et al.,
2005b). MORGUL (Figure 1) is a Pioneer 3AT mobile
autonomous robot equipped with an on-board PC, which
is connected to the local area network and to the Internet
via several different channels, even if the most used one
in indoor and short-range outdoor missions is a wi-fi communication
link.
Unlike most absolute localization systems, that use
robot-mounted sensors to gather information from fixed
landmarks, the solution we adopted reverses this philosophy,
using fixed sensors that “look” at the robot. This
solution may seem complicated and non-economical, but
it should be kept in mind that modern networking systems
(specially wireless ones) make it usually very easy to
place internet-connected devices around the environment.
In this approach robots are low-cost mobile devices,
that only take on board actuators and sensors that are
needed for a reliable accomplishment of their task, embedding
the remaining portion of sensors and processing
units into the environment, using whenever possible resources
that are already available for other reasons. In the
patrolling area, for instance, the used cameras can be the
same used for a fixed surveillance system.
This point of view in the recent years has led to the
development of a message-oriented middleware called
DCDT (Device Communities Development Toolkit) that
implements a reliable and stable communication layer as
required by this approach.
Every device outside the robot can autonomously perform
its task, offering its services only when and where
required in this sort of distributed sensing and processing
system, in which the DCDT takes responsibility of delivering
messages throughout the system, managing each
single active and independent software agent, called member.
In our project, the robot’s localization is conceived as
an autonomous task provided by a member of the DCDT
system that interacts with the community of interconnected
devices in a transparent way, regardless of the
physical communication means used.
Figure 1: MORGUL robot.
The localization is performed by a visual-based system
that recognizes an active marker when the robot gets
unsure about its position. The visual system consists of a
commercial off-the-shelf webcam, namely a Philips
ToUCam Pro II PCVC840K with a resolution of 640x480
pixels, connected via USB link to a PC. As far as the active
marker is concerned, several tests have demonstrated
that in outdoor environments the best results can be obtained
with a series of high brightness LEDs, while for
indoor use infrared LEDs can be used with good results,
even if the webcam is not very sensitive in the infrared
region. The LED clusters are controlled by a microcontroller
unit connected via a serial line to the robot PC that
runs a suitable autonomous DCDT member.
Cameras should be placed as high as possible in order
to minimize errors due to geometrical distortions of the
lenses: this translates in placing them close to the ceiling
indoor, and at least at the height of the first floor when
operating outdoor. This, by the way, is a normal placement
for surveillance cameras. The system can be composed
of several cameras, although, in a flat environment

where the height of the marker is constant, a single camera
is sufficient to solve the localization problem. More
cameras can be used to cover wide areas and, if their
fields of view are overlapping, to check the operation of
the system.
This paper mainly deals with the marker detection
problem, and does not go into the details of the localization
algorithm. So far, we used a very simple method
based on a grid of calibration points and on linear interpolation
among these points.
4 Marker recognition algorithm
While designing the marker and the recognition algorithm,
we had to keep in mind several points, among
which the most important ones are that in daylight, and
particularly in full sunshine, the light emitted by any artificial
device cannot compete with the natural light emitted
by the sun, unless enormous amounts of energy are used.
This means that the use of a time-variant marker is mandatory,
and that its characteristics must be carefully chosen
in order to avoid confusion with other direct or indirect
light sources.
A first attempt at using colour-varying markers was
discarded, because this would require processing the image
in at least two different colours. The idea of using a
low frequency flashing light beacon was also abandoned,
because it would make it more complex to adapt the
marker brightness to ambient light. Finally, it was decided
to use a marker that emits flashes at a fixed frequency,
whose period is quite close to the camera frame acquisition
time.
Since the used camera, as almost all CCD webcams,
has an electronic shutter during whose opening time light
is integrated on the sensor, the idea was to have this shutter,
that operate at constant frequency and at a fixed duty
cycle, beat with the light emitted by the marker, that is
also turned on and off at preset times.
The result is that the pixels pertaining to the marker
appear with a brightness that periodically oscillates following
a quasi-triangular waveform whose frequency is
the beat between the shutter and the marker frequency,
and whose shape is only influenced by the ratio of the
shutter and the marker’s duty cycle.
The recognition algorithm consists of several steps
that will be described separately in the sequel: acquisition,
background elimination, pixel pattern recognition, blobs
identification and finally marker detection.
4.1 Acquisition
During this step images are acquired from the webcam in
VGA resolution and saved in a suitable memory area.
While in indoor environments we could rely on the
marker being one of the brightest objects the camera can
see, in outdoor operation this is unrealistic: an image differencing
technique has to be used, and further processing
takes place on differential images obtained subtracting
each frame from the previous one.
Figure 2 shows the marker, seen from a distance of
about 26 metres, in a slightly cloudy day. It is easy to
imagine that, in bright sunshine, it can be much more difficult
to identify.
Figure 2: Outdoor image - Marker ON (a) and OFF (b).
4.2 Background subtraction
In image differencing, locations of change are indicated
by large values in the difference image obtained subtracting
the corresponding pixels of two consequent frames.
Due to its simplicity, this method is widely applied in
change detection algorithm used in object tracking, intruder
surveillance systems, vehicle surveillance systems,
and inter-frame data compression.
Figure 3: Shutter - marker relationship.
In our recognition algorithm, we make subtraction in
the luminance channel of two consequent frames. A positive
value in a pixel in the image difference represent the
fact that a bright spot has appeared in an image, while a
negative value indicates that it has disappeared.
4.3 Pixel pattern recognition
In order to localize the robot, we must find the pixels
whose behaviour matches the pattern expected from the
marker, i.e. a periodical repetition of n frames where it
appears to be ON and of m frames where it is OFF. In the
difference images collected from the previous phase with
the chosen timing this behaviour is represented by a positive
value when the marker appears in an image, n-1 values
close to 0 while it remains present, a negative value
while it is disappearing and m-1 values close to 0 when it
is in the OFF state. This sequence can be binarised against
a given threshold, and it generates a sequence that, as it
can be seen in Figure 3, is 11100001110000…, with n=3
and m=4. Recognizing this pattern could be simply done
with a mask, but inaccuracies in the timing of the various
components require a slightly more sophisticated approach.

Due to such inaccuracies, the transition between the
two states may shift left or right, resulting in sequences as
for instance 11100011100011000, as it can be seen in
Figure 4.
Figure 4: Actual results from the thresholding algorithm.
Those sequences have different values of n and m for
each ON/OFF period; so, instead of searching for pixels
whose behaviour exactly matches the theoretical sequence,
we assign each pixel a value according to the
weights shown in Figure 5.
The table should be read as follows: if, for a given
pixel, both transitions occur at the correct time, it is assigned
a value of 25. If one transition occurs one frame
earlier or later, the weight of that pixel is 15. If both transitions
are out of timing the weight for that pixel is 5. Any
transition occurring more than one frame earlier or later
than the correct time yields a value of -20. The procedure
is repeated for several cycles, and is similar to integrating
a rectified signal after passing it through a band pass filter.
At the end of this phase we obtain a sort of “values
image” in which the value of every pixel represents how
much that pixel’s pattern matches the behaviour expected
from the marker.
Figure 5: Coefficients used to weigh transitions.
4.4 Blobs detection
The phase just described is performed separately on each
pixel. Pixels with similar behaviour must now be grouped
together, in order to obtain blobs whose centre of gravity
will indicate the geometrical position of the marker. Since
at this stage there normally are several candidates to
choose from, data on the number of pixels, on the maxi-
mum and the medium value of each blob are also collected.
4.5 Marker localization
In this phase, the information collected in the blobs identification
is used to localize the robot.
The centre of gravity of the blob that has the highest
maximum value is chosen as the most probable position
of the marker in the image. If two or more blobs have
similar maximum values, the blob with the highest medium
value is chosen.
5 Experimental results
In order to validate the algorithm two sessions of experiments
were made in outdoor environment. Experiments
were also performed indoor, but they are not reported here
because in that case the task of identifying the marker is
very simple, and no significant experience was gained.
The used marker is a black panel with a circle of high
brightness LEDs as shown in Figure 6. Obviously, such
marker is highly directional, and therefore unsuitable for
being used on a mobile robot. We are currently working
in two directions to solve this problem: a) building the
marker on a hemi-spherical surface to make it omnidirectional,
and b) mounting it horizontally, over a revolving
diffusing mirror, with a device that keeps the device approximately
pointed against the webcam. The latter solution
is more complex, but has the advantage of requiring
less energy and, to some extent, of being less intrusive
because it does not emit unnecessary light.
Figure 6: The marker used in the experiments.
The webcam was set to acquire 15fps. Shutter opening
time was 17ms for all tests. The marker was operated at
13.33Hz, and its “on” time was 33ms. These settings
were chosen to have a beat period between the two signals
of 9 camera frames. A study of the behaviour of the system
led to the conclusion that, with these settings, the
most probable detected sequence is 4 frames where the
marker appears to be ON, and of 5 where it appears to be
OFF. This sequence was considered the most probable
taking into account any possible phase shift between the
marker and the camera, and several other factors, as daylight
effect and noise in captured images. Changes of
these parameters could yield slightly different sequences
of frames but, as it has already been said, the algorithm
has been designed to accommodate such situations.

The two sessions of experiments took place around
noon, the first one with a cloudy sky and the second one
under a very bright sunshine. Each session included 5
experiments with the marker placed in different locations,
as shown in Figure 7. Distance of each location from the
webcam is shown in Table 1.
Each experiment consisted in 10 acquisitions of 75
frames taken from a first floor window facing north. During
every of these 10 tests, students were freely moving in
the campus, cars passed in a road in the background, and
moving tree leaves added some noise to the scene.
Figure 7: Placements of the marker during the experiments.
Exp. number Distance
1 26m
2 52m
3 64m
4 73m
5 100m
Table 1: Distance of experiment points from the webcam.
For each experiment, no matter if the marker identification
had been correctly done or not in a test, we analyzed
the mean value of the maximum and of the minimum
value, and the apparent area of the blob chosen by
the system as the most probable candidate.
For all tests that yielded a correct result we also statistically
analysed the computed position of the centre of
gravity of the detected marker with the actual one in the
image space.
We also considered data concerning blobs that were
discarded, to have an idea about the selectivity of the algorithm
in real world images.
Finally, even if the algorithm has not been optimized
yet, the timing of the various phases of the algorithm was
measured.
5.1 Session performed on May 5, 2005
This session, that took place on a day with light, uniform
cloud coverage gave very satisfactory results: all tests
succeeded, except two that took place in a location where
a constant flux of walking students kept interfering with
the marker light.
Mean errors in the position estimate of the marker average
one pixel. The only significant error occurred on the
closest location, and is due to the fact that the marker, at
this distance, appears as several, separate blobs. The introduction
of a region growing technique would help reducing
this kind of error. Something similar also happened
in experiment 3.
In experiment 2 the identified blob appears very small.
This is due to camera saturation phenomena. Nevertheless,
the marker was correctly identified in all tests.
Experiment 4 was very demanding, because the
marker was far from the camera, and especially because
the marker was covered several times by people passing
by. This explains two wrong localizations, and the fact
that in general confidence values are lower than in the
other tests.
Experiment 5 was performed at the maximum allowable
distance (100m). The marker was always correctly
identified, and position error is also very small. It should
be noted that the dot that identifies the marker (Figure 8
and 9) is much larger than the actual marker blob, for the
purpose of visibility.
Figure 8: May 5 th session, Exp.5 - Marker at 100m.
Analyzing the false detections of experiment 4, we can
see from Table 2 that blobs are composed of a very small
number of pixels, and that the maximum value of such
pixels is much less than in the correct identification.
These data are close to those pertaining to discarded
blobs, which average an area of 1.26 pixels and have a
value of 14.39.
Exp. Max Val. Min Val. Average Area
4 30 25 27.5 1.5
Table 2: May 5 th session: Exp.4 – Average data of incorrectly
identified blobs.
As far as processing time is concerned, in all tests it
took about 1.5s, i.e. roughly 20ms per frame, excluding
acquisition time.
5.2 Session performed on May 6, 2005
In this session the experiments were done following the
same protocol of the previous day. In this case results

were also very good, and errors only occurred with the
marker at the maximum distance. There was, however, a
general decrease in precision, due to the very bright sunshine
that reduced contrast between the marker and the
background. In most tests, the background (white marble
in experiment 2, for instance) was significantly brighter
than the marker. Also, the marker generally appears
smaller than the previous day. However, all identifications
were successful, except those done at the maximum
range, where we had 4 correct identifications out of 10.
Figure 9: May 6 th session, Exp.2 - The marker over white marble
under bright sunshine.
It has been noted that, as ambient light increases, the
apparent size of the blob identified as the marker becomes
smaller. This is also due to a greater fragmentation of the
marker into several blobs, as it was already shown in the
preceding paragraph.
As far as experiment 2 is concerned, the chosen camera
parameters played an important role, because they
caused a strong saturation of the sensor, which in turn
reduced its dynamic range. Better results could have been
obtained using a variable exposure scheme, whose development
will be included in the sequel of the research.
Wrong detections only occurred in experiment 5. As
can be seen in Table 3, however, blobs are very small due
to the distance and their value is very low.
Exp. Max Val. Min Val. Average Area
5 40 25 36.67 1
Table 3: May 6 th session, Exp.5 - Average data of incorrectly
identified blobs.
Conclusions and future research
In this paper we presented the heart of a simple and lowcost
localization system based on the beat created by the
shutter of a webcam and the signal from an active marker
working at different frequencies.
This system behaves well even in full daylight, when
the marker appears dimmer than parts of the environment.
The system is still at an experimental stage, and several
improvements are being implemented. Among them,
a better generalization of the video input (in the current
implementation the system only works with some kinds
of cameras), improvements in the marker detection algorithm,
and a global optimization of the developed software.
Moreover, the system can be improved using more
than one camera and integrating it in the DCDT middleware.
DCDT would allow separating the tasks involved in
the marker recognition, with a webcam member only
transmitting images, one identifying the marker, another
one performing geometric calculations to locate the robot,
etc.
Max Blob Value Blob Mean Error Max Error
Exp. Dist. (m) CI Min Max Average Average Area X Y X Y
1 26 10 165 190 185.00 99.00 23.10 0.8 0.7 1 3
2 52 10 60 200 157.50 110.10 3.80 0.5 0.6 1 1
3 64 10 145 190 173.50 86.30 15.10 0.6 1.2 3 2
4 73 8 80 175 124.25 87.21 3.60 0.8 1.2 2 2
5 100 10 60 175 152.31 108.69 8.77 1 0.1 1 1
Table 4: May 5 th session - Experimental results: correct identifications (CI), statistical analysis of the maximum pixel value in the
marker blobs, marker blob’s average value and apparent area, errors (in pixels) along the image axes.
Exp. Blobs Average Area Total processing time (ms) Image processing time (ms)
1 17.30 16.91 1.28 1504.30 20.09
2 14.90 17.12 1.33 1506.75 20.06
3 23.30 15.15 1.32 1510.59 20.14
4 37.90 12.44 1.22 1489.90 19.86
5 22.30 10.34 1.17 1492.64 19.90
Average 23.14 14.39 1.26 1502.64 20.04
Table 5: May 5 th session – Average number of discarded blobs, blob’s average value and area, processing time.

References
Max Blob Value Blob Mean Error Max Error
Exp. Dist. (m) CR Min Max Average Average Area X Y X Y
1 26 10 155 200 164.58 151.17 1.25 3 0.6 6 3
2 52 10 25 200 125.50 96.40 2.20 0.5 0.6 1 2
3 64 10 90 200 170.00 114.60 7.90 0.1 0.3 1 1
4 73 10 110 200 160.00 133.10 2.30 2.3 1.9 3 2
5 100 4 55 95 103.75 73.00 2.75 0.1 0.2 1 1
Table 6: May 6 th session.
Exp. Blobs Average Area Total processing time (ms) Image processing time (ms)
1 22.08 18.33 1.42 1497.77 19.97
2 21.10 13.43 1.24 1523.28 20.31
3 51.50 10.07 1.42 1517.40 20.23
4 21.00 20.49 1.38 1516.53 20.22
5 21.00 19.80 1.25 1517.36 20.23
Average 27.33 16.42 1.68 1514.46 20.19
Arras, K. O. and Tomatis, N. (1999). Improving robustness
and precision in mobile robot localization by using
laser range finding and monocular vision. 3rd
European Workshop on Advanced Mobile Robots.
Batalin, M. A. and Sukhatme, G. S. (2003). Coverage,
exploration and deployment by a mobile robot and
communication network. In Proceedings of the International
Workshop on Information Processing in Sensor
Networks.
Betke, M. and Gurvits, L. (1997). Mobile robot localization
using landmarks. In IEEE International Conference
on Robotics and Automation, 13(2):251-263.
Borenstein, J., Everett, H. R., and Feng, L. (1996).
Where am I? Systems and Methods for Mobile Robot
Positioning. A. K. Peters, Ltd., Wellesley, MA.
Cassinis, R., Meriggi, P., Bonarini, A., and Matteucci,
M. (2001). Device communities development toolkit:
an introduction. In Eurobot ‘01, pages 155-161, Lund,
Sweden.
Cassinis, R., Meriggi, P., and Panteghini, M. (2003). A
very low cost distribuited localization and navigation
system for mobile robot. In 1st European Conference
on Mobile Robots (ECMR’03).
Cassinis, R., Tampalini, F., and Bartolini, P. (2005a).
Wireless Network Issues for a Roaming Robot. Technical
report, Università degli Studi di Brescia.
http://www.ing.unibs.it/~cassinis/docs/papers
/05_010.pdf.
Cassinis, R., Tampalini, F., Bartolini, P., and Fedrigotti,
R. (2005b). Docking and Charging System for
Autonomous Mobile Robots. Technical report, Università
degli Studi di Brescia. http://www.ing. unibs.it/~cassinis/docs/papers/05_008.pdf.
Table 7: May 6 th session.
Davison, A. J. (2003). Real-time simultaneous localization
and mapping with a single camera. In IEEE International
Conference on Computer Vision, pages 1403-
1410.
Fiala, M. (2004). Pseudo-random linear image marker
(plim) self-identifying marker system. National Research
Coucil Canada.
Georgiev, A. and Allen, P. K. (2004). Localization
methods for a mobile robot in urban environment. In
IEEE Trans. on Robotics and Automation, 20(5):851-
864.
Gutmann, J.-S., Fukuchi, M., and Sabe, K. (2004) Environment
identification by comparing maps of landmarks.
In IEEE International Conference on Robotics
and Automation, pages 262-267.
Jang, G., Kim, S., Lee, W., and Kweon, I. (2002). Color
landmark based self-localization for indoor mobile robots,
In IEEE International Conference on Robotics
and Automation, pages 1037-1042.
Jensfelt, P. (2001). Approaches to mobile robot localization
in indoor environments, Ph.D. thesis, Royal Institute
of Technology, SE-100 44 Stockholm, Sweden.
Lamon, P., Nourbakhsh, I., Jensen, B., and Siegwart, R.
(2001). Deriving and matching image fingerprint sequences
for mobile robot localization. In Proceedings
of ICRA 2001.
Ma, L., Berkemeier, M., Chen, Y., Davidson, M., Bahl,
V., and Moore, K. L. (2002). Wireless visual servoing
for odis an under car inspection mobile robot. In Proceedings
of IFAC World Congress, Barcelona, Spain.
Maeda, M., Habara, T., Machida, T., Ogawa, T., Kiyokawa,
K., and Takemura, H. (2003). Indoor localization
methods for wearable mixed reality. In Proceedings
of the 2nd CREST Workshop on Advanced Com-

puting and Communication Tehniques for Wearable
Information Playing.
Panzieri, S., Pasucci, F., Setola, R., and Ulivi, G. (2001).
A low cost vision based localization system for mobile
robots. In 9th Mediterranean Conf. on Control and
Automation, Dubronvnik, Croatia.
Parker, L., Kannan, B., Tang, F., and Bailey, M. (2004).
Tightly-coupled navigation assistance in heterogeneous
multi-robot teams. In Proceedings of IEEE International
Conference on Intelligent Robots and Systems
(IROS).
Roh, K. S., Lee, W. H., and Kweon, I. S. (1997). Obstacle
detection and self-localization without camera
calibration using projective invariants. In Proceedings
of the IEEE International Conference on Intelligent
Robots and Systems (IROS).
Wolf, J. and Pinz, A. (2003). Particle filter for self localization
using panoramic vision. In 26th Workshop
of the Austrian Association for Pattern Recognition,
pages 157-164, Laxenburg, Austria.