Journal of Global Positioning Systems (2006)
Vol. 5, No. 1-2:110-118
An Intelligent Multi-sensor System for Pedestrian Navigation
G. Retscher
Institute of Geodesy and Geophysics, Vienna University of Technology, Austria
Gusshausstrasse 27-29, A – 1040 Wien, Austria
Abstract. In the research project “Pedestrian Navigation
Systems in Combined Indoor/Outdoor Environements”
(NAVIO) we are working on the development of modern
intelligent systems and services for pedestrian navigation
and guidance. In the project modern and advanced
intelligent mobile multi-sensor systems should be
employed for 3-D position determination of a user. Due
to the fact that satellite positioning with GNSS (Galileo,
GPS, etc.) does not work under any environmental
condition (e.g. in urban “canyons” with no satellite
visibility and indoor) a combination and integration with
other sensors (e.g. dead reckoning sensors, inertial
navigation systems (INS), indoor location techniques,
cellular phone positioning, etc.) is essential. In our
approach a loose coupling of the employed sensors
should be achieved and it is proposed to develop a multi-
sensor fusion model which makes use of knowledge-
based systems. As far as we can see now knowledge-
based systems can be especially useful. Thereby the
decision which sensors should be used to obtain an
optimal estimate of the current user’s position and the
weightings of the observations shall be based on
knowledge-based systems. The new algorithm would be
of great benefit for the integration of different sensors as
the performance of the service would be significantly
improved. In this paper the basic principle of the new
approach will be described. To test and to demonstrate
our approach and results, the project takes a use case
scenario into account, i.e., the guidance of visitors to
departments of the Vienna University of Techology from
nearby public transport stops. The results of first field
tests could confirm that such a service can achieve a high
level of performance for the guidance of a pedestrian in
an urban area and mixed indoor and outdoor
environments. Standard deviations in the range of few
meters can be achieved for 3-D positioning in urban areas
although obstructions cause frequent loss of lock for
satellite positioning. Thereby GPS outages of up to 150 m
can be bridged using dead reckoning observations with
the required positioning accuracy. For indoor areas
satellite positioning can be replaced by indoor positioning
systems (e.g. WiFi, UWB). Due to the development of
advanced sensors it can be expected that such multi-
sensor solutions will be deployed in pedestrians
navigation services. We believe that these services will
play an important role in the field of location-based
services in the near future as a rapid development has
already started which is driven by their possible
applications.
Keywords. Pedestrian Navigation, Integrated
Positioning, Multi-Sensor Fusion, Kalman Filter,
Knowledge-Based Systems.
1 Introduction
In the research project NAVIO (Pedestrian Navigation
Systems in Combined Indoor/Outdoor Environements)
we are working on the development of modern intelligent
navigation systems and services for pedestrian navigation
and guidance. Thereby the research work is performed in
three different work packages, i.e., the first on “Integrated
positioning”, the second on “Pedestrian route modeling”
and the third on “Multimedia route communication”. To
test and to demonstrate our approach and results, the
project takes a use case scenario into account, i.e., the
guidance of visitors to departments of the Vienna
University of Techology (Gartner et al., 2004).
In this paper we will concentrate on the research work
and findings in the first work package. Challenging tasks
that are dealt with here are:
- the capability to track the movements of a pedestrian
in real-time using different suitable location sensors
and to obtain an optimal estimate of the current
user’s position,
- the possibility to locate the user in 3 dimensions with
high precision (that includes to be able to determine
the correct floor of a user in a multi-storey building),
and

Retscher et al.: An Intelligent Multi-sensor System for Pedestrian Navigation 111
- the capability to achieve a seamless transition for
continuous positioning determination between indoor
and outdoor areas.
Thereby a navigation support must be able to provide
location, orientation and movement of the user as well as
related geographic information matching well with the
real world situation experienced by pedestrians. Usable
location sensors have been classified and the most
suitable ones for guidance and navigation services were
selected. For pedestrian navigation systems suitable
location technologies include GPS/GNSS and indoor
location techniques, cellular phone positioning, dead
reckoning sensors (e.g. magnetic compass, gyros and
accelerometers) for measurement of heading and traveled
distance as well as barometric pressure sensors for
altitude determination (Retscher, 2004).
Our proposal and a part of our future research work is
focused on the development of modern and advanced
intelligent mobile multi-sensor systems that can be
employed for any personal navigation application
especially in the field of location-based services. Due to
the fact that satellite positioning with GNSS (Galileo,
GPS, etc.) does not work under any environmental
condition (e.g. in urban ‘canyons’ with no satellite
visibility and indoor) a combination and integration with
other sensors (e.g. dead reckoning sensors, inertial
navigation systems (INS), cellular phone positioning,
etc.) is essential. In our approach a loose coupling of the
employed sensors in the sense of a hybrid multi-sensor
system should be achieved. Therefore it is proposed to
develop a multi-sensor fusion model which makes use of
knowledge-based systems. As far as we can see now
knowledge-based systems can be especially useful.
Thereby the decision which sensors should be used to
obtain an optimal estimate of the current user’s position
and the weightings of the observations shall be based on
knowledge-based systems. The new algorithm would be
of great benefit for the integration of different sensors as
the performance of the service would be significantly
improved. The main development will be focused on the
deduction of a multi-sensor fusion model based on
knowledge-based systems. In this paper the basic
principle of the new approach will be described.
As a practical example, the guidance of a visitor of the
Vienna University of Technolgy from public transport
stops to our department is investigated in the NAVIO
project. First test results of the dead reckoning sensors
will be presented in the paper.
2 Concept of an Intelligent Multi-Sensor Fusion
Model
The integration of different sensors and location methods
shall be based on an intelligent multi-sensor fusion model
in the project NAVIO. Thereby the current position of a
user is estimated using a Kalman filter approach which
makes use of knowledge-based systems. Figure 1 shows a
process flow of the intelligent multi-sensor fusion model.
Firstly the observations of each sensor and location
technique of the multi-sensor system are analyzed in a
knowledge-based preprocessing filter. In this step the
plausibility of the observations is tested as well as gross
errors and outliers are detected and eliminated. The
analyzed and corrected observations are then used in the
following central Kalman filter for the optimal estimation
of the current user’s position and its velocity and
direction of movement. In this processing step all suitable
sensor observations as identified before are employed and
the stochastic filter model is adapted using the knowledge
of the preprocessing step. For example, the weightings of
the GPS observations can be reduced in the case if the
current GPS positioning accuracy is low due to a high
GDOP value (i.e., bad satellite-receiver geometry). Then
the optimal estimate of the user’s position should be more
based on the observations of other sensors (e.g. dead
reckoning observations).
Fig. 1 Process flow of the intelligent multi-sensor fusion model (after
Retscher, 2005)
In the following the principle of the knowledge-based
preprocessing filter will be discussed in more detail.
3 Principle of Knowledge-Based Systems
To provide an automated preprocessing of the sensor
observations, a knowledged-based approach has been
choosen. In the following, the basics underlying
knowledge-based systems are briefly described.
Programs which emulate human expertise in well defined
problem domains are called knowledge-based systems
(see e.g. Stefik, 1998) and they are the results of research
in the area of artificial intelligence. Their main
Sensor observations
Knowledge-based
preprocessing filter
Central Kalman filter
Optimal estimate of
current user’s position

112 Journal of Global Positioning Systems
advantages in comparison with conventional
programming languages (such as Delphi, Fortran and
C++) are (see Reiterer et al., 2003):
- the knowledge about the problem domain is
seperated from general problem-solving knowledge
which makes it easier for the knowledge engineer to
manipulate this knowledge;
- experts knowledge that exists very often in form of
rules can be captured in this form without converting
into forests of data definitions and procedures.
Thereby the knowledge-based system consists of the
following major components: a knowledge base, an
interference engine, an user interface, a knowledge
acquisition tool and an explanantion tool (Stefik, 1998).
Various schemes for knowledge representation can be
employed, e.g. rules, frames, semantic nets and others.
Each has its peculiar strengths and weaknesses. The
structure of a rule-based approach is very similar to the
way how people solve problems. Thereby human experts
find it convenient to express their knowledge in form of
rules (i.e., situation – action pairs). Furthermore rules are
a way to represent knowledge without complex
programming constructs (Reiterer et al., 2003).
For the implementation of a knowledge-based system in
practice different approaches can be selected, e.g.
procedural methods, object oriented methods, logical
based methods, etc. In practice also combinations of the
different methods are employed. For the implementation
of the knowledge-based preprocessing filter a rule-based
object oriented approach was selected (Retscher, 2005).
The rule-based system consists of two components, i.e., a
working memory (WM) and a set of rules or the so-called
rule memory (Brownston et al., 1985). The WM is a
collection of working memory elements which itself are
instantiations of a working memory type (WMT).
WMT’s can be considered as record declarations in
PASCAL or struct declarations in C. The second
component of a rule-based system are the rules. The rule
base is divided into three groups of rules, i.e.,
- rules for the choice of suitable algorithms,
- rules for the predefinition of necessary parameters,
and
- rules to define the order of the algorithms.
A rule is divided into two parts, namely the lefthand side
(LHS) and the righthand side (RHS). In the LHS the
preconditions of the rule are formulated, whereas in the
RHS the actions are formulated. A rule can be applied (or
‘fired’) if all its preconditions on the LHS are satisfied.
Then the actions specified in the RHS are executed. Rules
can be seen as so-called IF-THEN statements, e.g.
IF (condition 1 AND condition 2) THEN (action). (1)
There are algorithms for the so-called matching phase,
i.e., the phase where all rules are checked against all
working memory elements which are efficient in practice.
The result of the matching phase is the ‘conflict set’
which includes all rule instances ‘ready to be fired’. A
conflict resolution strategy selects one rule instance
which is actually fired (Reiterer et al., 2003).
The coding of the rule is performed in the chosen
programming language. For the knowledge-based
preprocessing filter an implementation based on CLIPS
(2005) or wxCLIPS (2005) will be performed.
The processing of the rules is performed as described
above. In the case under consideration this process is
performed in a forward-reasoning following the
recognize-act-cycle. In forward-reasoning a specific rule
is selceted from an existing database which fulfills the
preconditions of the database and then its action part is
applied (or fired) where the action changes the existing
database. This proces is repeated as long as no rule can be
applied anymore (Puppe, 1991). The recognize-act-cycle
consists of the following three steps, i.e.,
- the examination as a first step where all rules are
tested about their feasibility,
- the selection of the rule as a second step where a
specific rule from the preselection is selected, and
- the action as a last step where the selected rule or its
action part is applied.
This cycle is run as long as no rule can be executed
anymore or if a stop signal is given.
4 Central Kalman Filter
After the preprocessing filter, an optimal estimation of
the current user’s position and its velocity and direction
of movement is performed in a central Kalman filter
using all suitable observations from the sensors and
location techniques. Using this recursive approach the
state of the movement of the pedestrian can be estimated
based on the use of theoretical assumptions about the
user’s movement behavior and current observations.
Thereby the user’s movement behavior is formulated in
the system equations and the observations are introduced
in the measurement equation of the filter. The Kalman
filter provides then an exact solution to the linear
Gaussian filtering problem and the problem is
characterized completely by its state vector and
covariance matrix. The filtering process is reduced to the
prediction and updating of these two statistical
parameters (see e.g. in Gelb, 1986; Schrick, 1997).
For the system equations of the filter a 3-D kinematic
motion model is employed which enables the prediction
of the state of the movement of the pedestrian (e.g. the

Retscher et al.: An Intelligent Multi-sensor System for Pedestrian Navigation 113
current position, velocity and heading) from one epoch to
the next. Depending on the type of the model different
parameters can be included in the state vector )(kx
r
. The
following parameters can be used to describe the state of
the system (see Retscher and Mok, 2004; Retscher,
2004):
- 3-D coordinates of the current position y, x, z of the
user,
- 3-D velocities vy, vx, vz,
- 3-D accelerations ay, ax, az,
- direction of motion (heading) φ in the ground plane
xy,
- velocity v in the ground plane xy,
- radial acceleration arad in the ground plane xy.
If the state vector )(kx
r
includes only 6 parameters, i.e.,
the 3-D coordinates of the current position y, x, z and the
velocities vy, vx, vz, the kinematic model describes a
constant linear movement. A constant accelerated
movement is described with 9 parameters in the state
vector where in addition to the previous model also the 3-
D accelerations ay, ax, az are included in the kinematic
model. A constant radial movement can be described by
different parameters in the state vector, i.e., the 3-D
coordinates of the current position y, x, z, the heading φ,
the velocity v and the radial acceleration arad in the
ground plane xy and the velocity vz in z-direction. Using
these models the filter predicts slightly different the
movement of the user where in the first model compared
to the second no accelerations are used and in the third
model a radial movement without tangential accelerations
atan is employed. Simulations have shown that the third
model gives a good approximation to describe the
movement behavior of a pedestrian (Retscher and Mok,
2004).
Fig. 2 Architecture of the central Kalman filter (after Retscher, 2005)
Figure 2 shows the architecture of the central Kalman
filter. It consists of four different modules which describe
either the current environment of the pedestrian (outdoor
or indoor area) or the movement of the pedestrian
(pedestrian moves or does not move) or takes into
account a possible failure of the filter. Thereby of great
importance is the detection of bad GPS quality in outdoor
environements due to e.g. bad satellite-receiver geometry
(high GDOP value) or multipath. From the results of the
knowledge-based preprocessing filter an additional
statistical evaluation of the deviations between the
kinematic motion model and the GPS observations (e.g.
using tests of the innovation, i.e., the difference between
the real observations and the predicted measurements, in
Initialization of the filter
Sensor observations
Pedestrian
does not move
Outdoor
GPS
Indoor
no GPS
Failure of
the filter
GPS
deviation ?
Compass
deviation ?
Estimation of
new state
Keep
previous state
•

114 Journal of Global Positioning Systems
the Kalman filter) and an adequate weighting of the GPS
observations in the stochastic filter model is performed.
In the indoor environment the filter estimate is mainly
based on the oberservations of the dead reckoning
sensors. This is the case if no other indoor location
system is employed that provides absolute coordinates of
the user (e.g. WiFi fingerprinting; see Retscher, 2004).
The dead reckoning observations depend thereby mainly
on the output of the heading sensor (i.e., digital compass).
Similar to the analysis of the GPS observations in the
outdoor environment, the observations of the digital
compass are analyzed for gross errors or outliers and their
weight for the Kalman filter is derived. In the case if the
pedestrain does not move the observations are not used to
determine a new position estimate but the previous
determined state is kept. If a failure of the filter occurs a
reinitialization is required (Retscher, 2005).
5 Sensors for Pedestrian Navigation
The integration of different location technologies and
sensors is essential for the performance of modern
advanced navigation systems. Thereby common
navigation systems rely mainly on satellite positioning
(GNSS) for absolute position determination. Losses of
lock of satellite signals are usually bridged using dead
reckoning (DR) observations. Due to the main limitations
of the sensors (i.e., satellite availability in the case of
GNSS and large drift rates in the case of DR) other
positioning technologies should be integrated into the
system design of a personal navigation system to
augment GNSS and DR positioning.
Other radio positioning systems and wireless geolocation
technologies have been developed and can be employed
in personal navigation systems. Following Pahlavan et al.
(2002) two basic approaches can be distinguished in the
development of wireless geolocation techniques, i.e., one
approach where the system is solely designed for
positioning using certain radio signals and the second
where already established wireless infrastructure (e.g.
WiFi or UWB) is employed for location determination.
Thereby the second approach has the advantage that
usually no additional and costly hardware installations are
required. Some of these systems have been especially
developed for indoor applications, but they can also be
employed in indoor-to-outdoor and urban environments
(Retscher and Kealy, 2005). One approach is the use of
WiFi signals for position determination. The basic
principle of this approach has been analyzed and can be
found in Retscher (2004). In a study the performance of a
WiFi fingerprint method has been recently tested and it
can be summarized that positioning accuracies in the
range of 1 to 3 m can be achieved.
In addition, for the pedestrian navigation service in our
research project NAVIO the following dead reckoning
(DR) sensors are employed:
- dead reckoning module DRM III from
PointResearch,
- Honeywell digital compass module HMR 3000,
- Crossbow accelerometer CXTD02, and
- Vaisala pressure sensor PTB220A.
The dead reckoning module DRM III from PointResearch
(2005) is a self contained navigation unit where GPS is
not required for operation. It provides independent
position information based on the user’s stride and pace
count, magnetic north and barometric altitude. The
module is designed to self-calibrate when used in
conjunction with an appropriate GPS receiver, and can
produce reliable position data during GPS outages. The
system consists of an integrated 12 channel GPS receiver,
antenna, digital compass, pedometer and altimeter. The
module is clipped onto the user’s belt in the middle of the
back and the GPS antenna may be attached to a hat.
Firmware converts the sensor signals to appropriate
discrete parameters, calculates compass azimuth, detects
footsteps, calculates altitude and performs dead reckoning
position calculation. An internal Kalman filter algorithm
is used to combine dead reckoning position with GPS
position to obtain an optimum estimate for the current
user’s position and track. With the dead reckoning
module and GPS integrated together, a clear view of the
sky is only required for obtaining the initial position fix.
The fix must produce an estimated position error of 100
m or less to begin initialization. Subsequent fixes use
both dead reckoning and GPS data, so obstructed
satellites are not as critical as in a GPS only
configuration. The Kalman filter continuously updates
calbration factors for stride length and compass mounting
offset. The GPS position error must be less than 30 m
before GPS data will be used by the Kalman filter, and
the first such fix will also initialize the module’s latitude
and longitude. Subsequently, the filter will use any GPS
position fix with an estimated position error of 100 m or
less, adjusting stride, body offset, northing, easting,
latitude and longitude continually.
The Honeywell digital compass module HMR 3000 is
employed in the project NAVIO for precise heading
determination of the pedestrian. The HMR 3000 consists
of a magnetic sensor and a two-axis tilt sensor
(Honeywell, 2005). The low power, small device is
housed in a non-magnetic metallic enclosure that can be
easily installed on any platform. A sophisticated auto
compass calibration routine will correct for the magnetic
effects of the platform. Wide dynamic range of the
magnetometer allows the HMR 3000 to be useful in
applications with large local magnetic fields. The
influence of magnetic disturbances on the sensor has been

Retscher et al.: An Intelligent Multi-sensor System for Pedestrian Navigation 115
tested and is presented in Retscher and Thienelt (2005). It
could be seen that deviations of only 2 to 3 degrees
occurred if the source of disturbance (e.g. a notebook
computer or a metallic lighter) is put in a distance of
about 30 cm from the sensor. Higher deviations occur,
however, at shorter distances to the sensor. As a
consequence the sensor should be kept away from mobile
phones, coins, metallic lighters and keys.
For measurement of the accelerations of the pedestrian
the Crossbow accelerometer CXTD02 should be
employed. The CXTD02 is a tilt and acceleration sensor
and measures tilt and acceleration using a triaxial MEMS
accelerometer (Crossbow, 2005). It provides high
performance in more demanding measurement
applications where high accuracy must be maintained
over a wide temperature range. The low noise floor and
true DC response guarantees a long-term stability. It
should be analyzed in detail how the sensor can employed
for the determination of the traveled distance, pitch and
roll of the sensor platform.
In addition, the Vaisala pressure sensor PTB220A is
employed in the project for determination of height
differences from changes of the air pressure. The
PTB220A is designed for measurements in a wide
environmental pressure and temperature range with an
extremly high accuracy (Vaisala, 2005). Starting from a
given height the pressure changes can be converted in
changes in height using the following equation:
() ()
21
12
18464 10,0037lglg
m
HH H
tBB
∆=− =
=⋅+ ⋅⋅−
(2)
where ∆H is the height difference between two stations 1
and 2, B1 and B2 are the pressure observations at station 1
and 2 and tm is the mean value of the temperature of both
stations.
It must be notetd that this equation is an approximation
formula that is valid for central Europe only (Kahmen,
1997).
Recently performed tests have shown that we are able to
determine the correct floor of a user in a multi-storey
building using this sensor (see also Figure 5).
6 Sensor Tests
Practical tests in the NAVIO project are carried out for
the guidance of visitors of the Vienna University of
Technology to certain offices in different buildings or to
certain persons. Thereby we assume that the visitor
employs a pedestrian navigation system using different
sensors that perform an integrated positioning. Start
points are nearby public transport stops, e.g. underground
station Karlsplatz in the center of Vienna. Tests with two
different GPS receivers have been carried out in this area
and are presented in Retscher and Thienelt (2004).
Because of obstructions caused by the surrounding four
to five storey buildings it frequently happens that GPS
signals are lost so that large parts of the route of the
pedestrian must be bridged by dead reckoning. Only in a
park at the exit of the underground station and on isolated
road crossings it is possible to receive GPS signals with
sufficient quality. This area is therefore suitable for
testing the combination of absolute and relative DR
location sensors. Further sensor tests are schedulded to be
performed in the next months in this area.
First test measurements with the dead reckoning module
DRM III from PointResearch (2005) have been carried
out in another test area in the park of Schönbrunn Palace
in Vienna shown in Figure 3. This test site has been
chosen as it provides free satellite visibility. Figure 4
shows the dead reckoning observations as well as the
GPS measurements along a 475 m long track in the park
of Schönbrunn Palace. In the dead reckoning module,
measurements of accelerometers are employed to count
the steps of the walking pedestrian and the traveled
distance is obtained using a predefined value for the
stride length. Using GPS observations the stride length
can be calibrated. Furthermore a compass and a gyro are
employed for measurement of the heading or direction of
motion. The dead reckoning observations shown in
Figure 4 have been obtained without using the GPS
calibration. They reach deviations in the range of 7 m
over a distance of 150 m and 20 m over 200 m from the
given track. The GPS measurements have a maximum
deviation of 7 m. Figure 4 shows also the resulting
trajectory from the internal Kalman filter of the DRM III
module calculated from a combination of GPS and DR
observations. It can be seen that the large drift rate of the
DR observations can be reduced. Using the DR
observations, GPS outages (i.e., when GPS is
unavailable) of up to 150 to 200 m can be bridged with a
reasonable positioning accuracy. For longer GPS outages,
however, other location technologies have to be
employed providing an absolute position estimate to
correct for the DR drift.
Figure 5 shows test observations with the Vaisala
pressure sensor PTB220A in our office building of the
Vienna University of Technology. This building has 5
storeys and our department is located on the 3rd floor. It
can be clearly seen in Figure 5 that the sensor is able to
determine the correct floor of the user with a high
precision. The standard deviation of the pressure
observation is in the range of ± 0.2 hPa and the maximum
deviation of the determined height is less than ± 1 m for
90 % of the observations.

Journal of Global Positioning Systems (2006)
Vol. 5, No. 1-2:110-118
- 1650 - 1600 - 1550 - 1500 - 1450 - 1400
5338250
5338300
5338350
5338400
5338450
Y-axis [m]
X-axis [m]
Theedgesof the
lawnswere
meas ured
Fig. 3 Field test site in the park of Schönbrunn Palace in the city of Vienna
- 1580 - 1560 - 1540 - 1520 - 1500 - 1480 - 1460 - 1440
5.338.320
5.338.340
5.338.360
5.338.380
5.338.400
5.338.420
5.338.440
East [m]
North
[m]
DRM III Measurement (Park Schönbrunn)
GPS Measurement
Dead Reckoning Measurement
CalculatedTrajectory
Trajectory
Fig. 4 Test measurements with the dead reckoning module DRM III in the park of Schönbrunn Palace in Vienna
The results of the different tests could confirm that a
pedestrian navigation service can achieve a high level of
performance for the guidance of a user in an urban area
and mixed indoor and outdoor environments. Standard
deviations in the range of few meters can be achieved for
3-D positioning in urban areas although obstructions
cause frequent loss of lock for satellite positioning.
Thereby GPS outages of up to 150 m can be bridged
using dead reckoning observations in combination with
cellular positioning with the required positioning
accuracy. For indoor areas satellite positioning can be
replaced by indoor positioning systems (e.g. WiFi
fingerprinting; see Retscher, 2004) and the altitude of the
user can be observed using a barometric pressure sensor.
7 Conclusions
From the presented sensor tests can be seen that a high
precision and relability for the position determination of a
pedestrian can be achieved if different location
techniques and dead reckoning sensors are employed and
combined. For the integration of all observations a new
multi-sensor fusion model based on an extended Kalman
filter which makes use of a knowledge-based

Retscher et al.: An Intelligent Multi-sensor System for Pedestrian Navigation 117
preprocessing of the sensor observations can be applied.
The principle of this new approach is presented in the
paper. The knowledge-based preprocessing filter
represents an extension of common multi-sensor fusion
models in a way that the data based system analysis and
modeling is supplemented by a knowledge-based
component and therefore not directly quantifiable
information is implemented through formulation and
application of rules. This rules are tested in the
preprocessing step and if they are fulfilled certain actions
are executed. Due to the knowledge-based analysis of the
sensor observations gross errors and outliers can be
detected and eliminated. In addition, the preprocessing
filter supplies input values for the stochastic model of the
central Kalman filter. Therefore the weightings of the
sensor observations can be adjusted in the Kalman filter
depending on the availability and quality of the current
observations. This approach will be implemented and
further sensor tests will be carried out. Due to the
development of advanced sensors it can be expected that
such multi-sensor solutions will be deployed in
pedestrians navigation services in the near future. We
believe that these services will play an important role in
the field of location-based services.
Fig. 5 Test measurements with the Vaisala pressure sensor PTB220A in
our office building of the Vienna University of Technology
Acknowledgments
The research work presented in this paper is supported by
the FWF Project NAVIO of the Austrian Science Fund
(Fonds zur Förderung wissenschaftlicher Forschung)
(Project No. P16277-N04).
The author would like to thank Mr. Michael Thienelt and
Mr. Michael Kistenich for the performance of test
measurements with the dead reckoning sensors and the
preparation of some of the Figures in this paper.
References
Brownston L., Farrell R., Kant E., Martin N. (1985)
Programming Expert Systems in OPS5: An Introduction
to Rule-based Programming. Addison-Wesley Longman
Publishing Co.
CLIPS (2005) CLIPS – A Tool for Building Expert Systems.
http://www.ghg.net/clips/CLIPS.html (last date accessed:
July 2005).
Crossbow (2005) CXTD Digital Tilt and Acceleration Sensor,
Product Information, Crossbow, USA,
http://www.xbow.com/Products/Product_pdf_files/Tilt_pdf
/6020-0012-01_B_CXTD.pdf (last date accessed: July
2005).
Gartner G., Frank A., Retscher G. (2004b) Pedestrian
Navigation System in Mixed Indoor/Outdoor
Environment – The NAVIO Project. in: Schrenk M. (ed.):
CORP 2004 and Geomultimedia04. Proceedings of the
CORP 2004 and Geomultimedia04 Symposium, February
24-27, 2004, Vienna, Austria, pp. 165-171,
http://corp.mmp.kosnet.com/CORP_CD_2004/archiv/papers
/CORP2004_GARTNER_FRANK_RETSCHER.PDF, (last
date accessed: July 2005).
Gelb A. (ed.) (1986) Applied Optimal Estimation. The MIT
Press, Cambridge, Massachusetts and London, England, 9th
print, 374 pgs.
Honeywell (2005) HMR 3000 Digital Compass Module User’s
Guide, Honeywell International Inc., USA,
http://www.ssec.honeywell. com/magne tic/datashe ets/hmr3
000_manual.pdf (last date accessed: July 2005).
Kahmen H. (1997) Vermessungskunde. 19. Auflage, Walter de
Gruyter, Belin, New York, pp. 475-482.
Pahlavan K., Li X., Mäkelä J.P. (2002) Indoor Geolocation
Science and Technolgy. IEEE Communications Magazine,
February 2002, pp. 112-118.
PointResearch (2005) DRM-III Dead Reckoning Module -
Engineering Development Tools, PointResearch
Corporataion, USA,
http://www.pointresearch.com/drm_eval.htm (last date
accessed: July 2005).
Puppe F. (1991) Einführung in Expertensysteme. 2nd edition,
Springer Verlag, Berlin, Heidelberg, New York (German).
Reiterer A., Kahmen H., Egly U., Eiter T. (2003) Knowledge-
based Image Preprocessing for a Theodolite
Measurement System. in: Gruen, A., Kahmen H. (eds.).
Optical 3-D Measurement Techniques VI, September 22-
25, Zurich, Vol. II, pp. 183-190.
Retscher G. (2004) Multi-sensor Systems for Pedestrian
Navigation. Proceedings of the ION GNSS Meeting, 21-24
September, Long Beach, California, U.S.A. (Institute of
Navigation, Fairfax, Virginia), unpaginated CD-ROM.
Retscher G. (2005) A Knowledge-based Kalman Filter
Approach for an Intelligent Pedestrian Navigation
System. Proceedings of the ION GNSS 2005 Conference,
September 13-16, 2005, Long Beach, California, U.S.A.
height difference [m]
time [s]
ground floor
1st floo
r
2nd floor
3rd floor
roof

118 Journal of Global Positioning Systems
(Institute of Navigation, Fairfax, Virginia), unpaginated
CD-ROM.
Retscher G., Kealy A. (2005) Ubiquitous Positioning
Technologies for Intelligent Navigation Systems. in:
Papers presented at the 2nd Workshop on Positioning,
Navigation and Communication, University of Hannover,
Germany, March 17-18, 2005, Hannoversche Beiträge zur
Nachrichtentechnik, Band 0.2, Shaker Verlag, pp. 99-108.
Retscher G., Mok E. (2004) Sensor Fusion and Integration
using an Adapted Kalman Filter Approach for Modern
Navigation Systems. Survey Review, Vol. 37, No. 292,
April, pp. 439-447.
Retscher G., Thienelt M. (2004) NAVIO – A Navigation and
Guidance Service for Pedestrians. Journal of Global
Positioning Systems, CPGPS, Vol. 3, No. 1-2, pp. 208-
217, see
http://www.gmat.unsw.edu.au/wang/jgps/v3n12/v3n12p26.
pdf (last date accessed: July 2005).
Retscher G., Thienelt M. (2005) Evaluation and Test of a
Digital Compass Module Employed for Heading
Determination in a Pedestrian Navigation System. in
Grün A., Kahmen, H. (eds.): Optical 3-D Measurement
Techniques VII, Papers presented at the 7th Conference on
Optical 3-D Measurement Techniques, October 3-5, 2005,
Vienna, Austria, pp. 389-394.
Schrick K.W. (1977) Anwendungen der Kalman-Filter-
Technik. Methoden der Regelungstechnik, R. Oldenbourg
Verlag, München, Wien (German).
Stefik M. (1998) Introduction to knowledge systems. 2nd
edition, Kaufmann, San Francisco.
Thienelt M., Eichhorn A., Reiterer A. (2005) Konzept eines
wissensbasierten Kalman-Filters für die
Fußgängerortung (WiKaF). Paper submitted to VGI,
Österreichische Zeitschrift für Vermessung und
Geoinformation (German).
Vaisala (2005) PTB220 Digital Barometer, User’s guide,
Vaisala, Finland,
http://www.vaisala.com/businessareas/instruments/
servicesupport/userguides/barometricpressure/PTB220%20
User%20Guide%20in%20English.pdf (last date accessed:
July 2005).
WxCLIPS (2005) wxCLIPS.
http://www.anthemion.co.uk/wxclips/ (last date accessed:
July 2005).