USING MOBILE DEVICES FOR TOPOLOGICAL INFERENCE

OF INDOOR ENVIRONMENTS

Marco Paiva

, Marcelo Petry

1,2

and Rosaldo J. F. Rossetti

Artificial Intellegence and Computer Science Laboratory, Department of Informatics Engineering,

Faculty of Engineering, Univeristy of Porto, Rua Dr. Roberto Frias, S/N, 4200-465 Porto, Portugal

INESC Porto, Institute for Systems and Computer Engineering of Porto,

Faculty of Engineering, Univeristy of Porto, Rua Dr. Roberto Frias, S/N, 4200-465 Porto, Portugal

Keywords: Indoor Localization, Local Positioning Systems, Topology Mapping.

Abstract: Nowadays location systems are used within a large variety of applications. The application of these systems

within indoor environments is already provided by several solutions. However, the need for high accuracy

within these environments to pursue such purpose implies the use of specific infrastructures designed

towards it. Our project tries to meet the requirements for a simple, low-cost, and scalable location system

through different approaches. The main idea of it is to re-construct topological maps of indoor spaces

through location estimation, i.e. using off-the-shelf technologies. We try to perform location estimations and

then re-create the indoor maps as topological maps as a means of reducing the precision requirements other

systems have, and develop a scalable and highly applicable system using sensors featuring mobile devices.

1 INTRODUCTION

Location-aware services have become an important

area for today’s ubiquitous and pervasive computing

applications. The increasing use of such services

brought the need to implement them in mobile

platforms. Thus, location estimation systems that

can provide those services in indoor environments

have become subject of several studies and

developments.

Today’s solutions provide location-aware

services in applications that are oriented towards

localization in maps that give a metric perception of

environments; thus, maps resemble their real

environment. However, we believe that this metric-

oriented approach is not the optimum approach for a

scalable implementation of indoor location-aware

services used in mobile devices. In this sense,

topological mapping approaches bring an abstraction

to that metric information and are more likely to deal

with low accurate measurements. In addition, they

could be implemented with techniques that use less

a priori information and in systems with limited

computational power. This might be a good

approach to improve scalability in such systems.

These two approaches are well studied in

robotics and their development has been made based

on the SLAM (simultaneous localization and

mapping) methodology, for instance. While metric

maps try to precisely represent the world by

measuring and relating distances, as it is proposed in

(Yiguang et al., 2010), the main idea behind the

topological approach is to provide distinctiveness for

global localization and mapping within the

environment as it is done in (Beeson et al., 2005) or

in (Sogo et al., 2001). More specifically, topological

maps provide a sketch map with labelled places and

abstract paths between them, in contrast to the

metric maps that try to sketch all the features within

the environment with geometric precision. A

topological map approach can be used also in

location-aware services oriented towards human

utilization. For instance, humans would not require

geometric locations to navigate in indoor

environments and could easily do it through

semantic locations and connections between them.

The same idea can be applied to many intelligent

systems which make use of location-aware services.

Our premise is to depart from a system with very

little information about the environment. No map

and topology is known a priori. Thus, its goal must

be to achieve localization in the indoor environment

and through that localization start building a

topological map of the environment along the time

and through a collaborative fashion among users.

427

Paiva M., Petry M. and J. F. Rossetti R..

USING MOBILE DEVICES FOR TOPOLOGICAL INFERENCE OF INDOOR ENVIRONMENTS.

DOI: 10.5220/0003835304270432

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 427-432

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

In order to materialize this idea, we have to

emphasize the available sensor technologies that can

be used to perform localization in indoor

environments. The first technology that comes to

mind when thinking about localization is GPS

(Global Positioning Systems). However, this

technology has a poor performance in indoor

environments due to the lack of satellite coverage.

Some works have been made also with cellular

networks, namely using GSM (Global System for

Mobile Communication), but it suffers from the

same effects as referred to GPS. Nevertheless, better

performances can be achieved through the use of a

priori data about the environment, situation which

we intend to minimize in our research.

One of the most explored technologies to

perform localization in indoor environments is Wi-

Fi Wireless LAN. This is a technology widely

implemented in public infrastructures and that is also

built-in in today’s ordinary mobile devices.

However, this kind of technology suffers from

significant signal attenuation effects in indoor

environments, such as: multipath; refractions;

reflections; scattering and shadowing. These

problems are normally overcome through

fingerprinting approaches, as in GPS or in GSM.

However, several techniques can be performed

without fingerprinting methods and using Wi-Fi

WLANs. These methods rely on simple lateration or

angulation techniques or even doing an on-line

calibration phase similar to fingerprinting.

Another technology that can be used for

localization purposes and that is widely spread in

mobile devices is Bluetooth. Despite similar to Wi-

Fi in some aspects, Bluetooth measurements are

only able after establishing a connection between

devices. This is usually overcome through learning

algorithms, such as Artificial Neural Networks to

estimate device positions. Another approach relies

on using other sensor technologies that are also

highly available on today’s mobile devices, such as

accelerometers, magnetometers and gyroscopes.

Such an approach usually performs dead-reckoning

of devices when the users are on the move through

acceleration and orientation measures.

We are specifically interested in studying the

potential of available technology to support easy and

practical inference of topological maps for indoor

applications. We want to provide some guidelines to

select mobile-enabled sensors to base topological

map inference methods, and build such maps

through localization of devices in indoor

environments in a collaborative fashion. With the

localization estimates of many devices we expect to

achieve a better basis for performing topological

map inference methods.

Firstly, we analyse the best techniques and

technologies to perform localization with mobile

devices, which can provide us with a better basis for

topological map inference through simple, low-cost

and common infrastructures. Secondly, we show

how such a system is implemented in those devices

and make an analysis of the preliminary

implementation and results using multiple devices.

2 RELATED WORK

Several works have been developed in order to

achieve localization of mobile devices in indoor

environments. Topological map inference has been

made mainly in the robotics field, focusing on the

SLAM methodology, as we mentioned before.

Nevertheless, we must first focus on localization

issues for mobile devices.

2.1 Solutions for Localization

Applications

Several GPS-based systems and approaches have

been proposed for localization purposes. SnapTrack

(Biacs, 1999), uses A-GPS and achieves accuracies

between 5-50m. However, GPS is not the ideal

technology to use in such a purpose due to its poor

performance in this kind of environments. With

GSM based systems, (Otsason et al., 2005)

presented a solution that led to an accuracy as low as

2.5m. The main idea of it was to use wide signal-

strength-fingerprints in a training phase, what

implies several a priori configurations within the

environment. This is a down point regarding the idea

of using as less a priori configurations as possible to

perform localization with a specific system, as we

intend to do.

Lots of Wi-Fi-based solutions have been also

developed and we give emphasis to two of them.

The RADAR solution proposed in (Bahl and

Padmanabhan, 2000), which is one of the most cited

systems in related literature, provides localization

through two main ways. One work is in a basic

fingerprinting approach where several off-line

measurements are made in a prior training phase,

whereas the other is based on a signal propagation

model. The approaches use the k-nearest neighbour

algorithm to achieve positioning estimations. The

number of neighbours, the number of off-line

measured points, user orientations, and the number

of samples achieved in the real-time measurement

ICAART 2012 - International Conference on Agents and Artificial Intelligence

428

phase affect the accuracy of the system. They are

reported to achieve accuracy around 2-3m. Another

work (Ekahau, 2010) is one of the most accurate

systems that use the Wi-Fi sensor technology. It uses

a Bayesian approach combined with stochastic

complexity and a competitive online learning

algorithm. Although it achieves accuracy as high as

1m, it is achieved through a training phase that

implies a strict calibration that has to be performed

in a very complex and precise way.

We also found Bluetooth solutions in our

literature review. The one presented in (Tadlys,

2010) reported an accuracy of about 2m, with 95%

reliability with a positioning delay between 15-

30sec. The solution relies on wireless tags and

wireless access points. The use of inertial sensors as

referred in the Introduction finds also some

developed solutions in literature. They are mainly

done through a pedestrian tracking approach using a

proper algorithm to estimate the device’s orientation

and in a next phase they take samples of the

accelerometer sensor to acquire the mobility of the

device. In (Yiguang et al., 2010), estimation is made

according to a walking pattern that provides a

realistic representation of movements of a human

carrying the device. Reported results have shown a

better trend with an increase of devices being

tracked and given errors between 2-3m on the floor

map estimation, not considering angle errors.

2.2 Localization Techniques

The main techniques used on systems that perform

localization estimation in indoor environments

nowadays are identified in the next sub-section.

2.2.1 Lateration and Angulation Methods

Lateration is a technique that can provide a position

estimate of a mobile device through at least three

measures for the two dimensional case. Essentially it

is the conjunction of three distances between

transmitters and one receiver that can provide the

position through lateration. Given those distances, it

is possible to know that the receiver is within three

circumferences, on which their radius are the

distances computed between transmitters and

receiver. The localization of the receiver can be then

achieved through the Euclidean distance. The best

case scenario of this technique is when the three

circumferences intercept themselves at one point. In

this case the estimated receiver position would be

that point. However, it is possible that the

circumferences do not intercept at a single point,

being the estimation through this technique not

precise or even impossible to define. This is

basically due to error estimations on the distance

between reference points and the device, which

exists due to the use of propagation models that

cannot optimally describe the environment

characteristics. To overcome this problem least-

squares algorithm is usually applied to achieve a

better estimation of the receiver position.

On the other hand, angulation takes advantage

of the “angles” of the received signal on

transmitters. Basically, angulation only requires to

get two reference-point transmitters and measuring

their AoA (angle of arrival). With those values it is

possible to sketch a triangle and then obtain the

position of the device (Hui et al., 2007).

2.2.2 Scene Analysis

Scene analysis makes use of a predefined data set

that maps observed features into object locations.

One example of static scene analysis is the RADAR

system, presented in (Bahl and Padmanabhan, 2000).

This technique is considered one of the best

alternatives to simple triangulation or lateration

methods made without a priori fingerprinting. In

theory, the static scene analysis can be performed

with all the measurements or features available. But

for indoor localization purposes, the most common

metric used is RSSI (received signal strength

indicator). Other metrics can also be achieved and

used, for example AoA (angle of arrival) or ToA

(time of Arrival), or even the use of image analysis

for navigation, which can be done also through

mobile platforms with built-in cameras.

Scene analysis is usually performed in two

phases: one training phase that is usually made off-

line, and where normally a database is built

containing pattern variations of one or more features

in known locations; and, another phase when the

location estimation is done through a relationship

analysis with data originated in the first phase.

Deterministic and Probabilistic models are normally

used to perform the matching in the second phase.

The main advantages of scene analysis concern

the fact that location of “objects” can be inferred

using passive observations. However, this kind of

method requires previous analysis of scenes, so that

a comparison can be performed to achieve location.

2.2.3 Using Neural Networks

Artificial Neural Networks (ANN) is another

approach to yield indoor localization, as it is done in

(Ahmad et al., 2006). The basic concept is to use

USING MOBILE DEVICES FOR TOPOLOGICAL INFERENCE OF INDOOR ENVIRONMENTS

429

ANN to produce a non-linear input-output mapping.

It is used in several applications, such as

classification, approximation or regularization. This

kind of approach is similar to scene analysis but,

instead of using probabilistic methods, it uses ANN

to compute the matching between the fingerprints

and the real time observations. The key advantage of

using ANN is that it has a lower cost in terms of

time and space necessary to process measurement

acquisitions. However, as this kind of approach

mostly relies on learning algorithms, it can induce to

some erroneous estimation. ANN is also well suited

for different complex and dynamic indoor

environment estimations.

3 A PRELIMINARY

ASSESSMENT

To implement a system that could be applied on

scalable basis, which would not require as much a

priori setups, one that could be provided through

ordinary mobile devices and within common

infrastructures, we opted to develop a solution using

three main sensor technologies: Wi-Fi sensors

within WLAN, performing a lateration technique

with a minimum squares optimization procedure,

accelerometers and magnetometers to perform dead

reckoning with the mobile devices through a

pedometer and a compass emulation.

The Wi-Fi sensors and the performed lateration

technique was chosen due to the high availability

that this kind of sensors have in existing

infrastructures and also due to the low complexity

characteristic of the lateration technique that does

not require much setup data to be performed.

Nevertheless, we need to remark the low accuracy in

localization estimation that this technique has.

However, our goal was not to achieve the most

accurate estimation but to be able to construct

topological maps which do not require precise

metric information about the environment.

Figure 1: Floor plan where the experiments were made.

We chose the accelerometers and magnetometers

to perform dead-reckoning also due to their high

availability in today’s mobile devices. To perform

the dead-reckoning we make use of the pedometer

emulator to count steps, and compute the orientation

of each step through the fusion of an accelerometer

and a magnetometer from where we are able to

emulate a compass.

In order to implement the system we developed

an application under the Android OS to be used on

mobile phones, and which had the task to collect

data from all those sensors. Then we performed the

localization estimations outside the application and

through that collected data. We must remark that we

perform estimations outside the mobile devices

because our idea is to provide a collaborative

solution among users in order to accomplish a more

reliable map reconstruction.

3.1 Wi-Fi Lateration Results

3.1.1 Path Tracking Estimations of Moving

Devices

In this experiment we performed several estimations

of five different devices within a corridor path in the

floor presented in Figure 1. The results of all merged

estimations achieved accuracies as low as 1.4m to

the real corridor path. However, these results

presented a maximum deviation of 7.37m in some of

the estimations.

Figure 2: Localization estimations made along a walking

path with Wi-Fi, after applying a Kalman filter.

Figure 2 presents merged estimations along that

path; blue dots are estimations achieved after a

Kalman Filter procedure applied over raw

estimations, and the line in black represents the real

path, where the mobile devices moved.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

430

3.1.2 Dead-Reckoning Estimation Made

within a Path

In this experiment we carried out dead-reckoning

within the same path illustrated in Figure 1. As the

orientation estimations made had too much magnetic

influence from the environment, we filtered the

measurements through a median filter and clustered

the orientations to a minimum orientation change of

90º, as all the corridors within the environment can

be modelled like that.

The results presented an improvement of the

accuracy in comparison to the estimations made with

the Wi-Fi lateration, as low as 1.05m with a

maximum error deviation of 3.0m.

In Figure 3, we illustrate one of the results

achieved with one mobile device where the line in

black is the walking path and the blue dots are the

estimations made with dead-reckoning.

Figure 3: Estimation of the walking path through dead-

reckoning.

3.1.2 Inference of a Corridor Topology and a

Room Space

This was the inference experiment that we made

using the previously described results. Its main goal

was to infer a single room as a starting point and the

corridor that connects to that room. To do it we used

a histogram of the estimations made with the Wi-Fi

sensor and which were gathered in the I121 room,

which is illustrated in Figure 4.

Figure 4: Wi-Fi lateration estimations histogram,

performed in room I121.

With that histogram we computed a single

position which corresponds to the centre of mass

between the 3 positions with more estimations along

the time (sc1, sc2, sc3) as it is depicted in Figure 4.

Thus, we inferred it as the working place within the

room; we used the dead reckoning estimations,

starting at that single position and performed along

the corridor topology.

Figure 5: Inference of a room connected to the corridor

topology.

With those estimations we inferred the corridor

topology and gathered the working place within

room I121. When fusing the two inferences, we

were able to achieve the inference of the room

connected to the corridor as it is shown in Figure 5,

where blue dots are the dead-reckoning estimations

and the red circle is the working place estimated

with the histogram. However, the sketched topology

is deviated from the real one due to the error of the

estimation made with the Wi-Fi at the position in the

room. Despite this deviation, the topology is similar

to the real one and we can conclude that it is a good

inference made within the topology map purpose we

pursue in our research.

This method can be applied to several semantic

locations and extended into different environments,

as they can be inferred based on the time users spend

in different spots.

4 CONCLUSIONS

In order to accomplish the goals of this work we

developed a system using sensor technologies

featured on the most common mobile devices: Wi-Fi

sensors, accelerometers and magnetometers. With

the application we developed we were able to gather

important data to perform localization within indoor

environments. We should first remark the low

accuracy that we achieved with the Wi-Fi lateration

USING MOBILE DEVICES FOR TOPOLOGICAL INFERENCE OF INDOOR ENVIRONMENTS

431

estimations, which at a first glance can lead us to

thinking in the unfeasibility of such a purpose using

this technique. However, when looking into the

overall results gathered with the Wi-Fi lateration and

the dead-reckoning made with the accelerometer and

the magnetometer, the achieved results led us to

conclude the feasibility of creating an indoor

environment topological map with those localization

estimations over the time they are gathered. In this

sense we were able to make a simple inference

where we achieved a topological map of a room

connected to a corridor which was very realistic and

which reinforces it is possible to accomplish our

purpose.

As next steps or further developments that can be

taken after this work, we might suggest the

improvement that can be made in order to achieve

even more localization estimations, which can be

made for instance by adding more sensor

technologies to the system, such as Bluetooth.

Another kind of improvement can be made in the

scope of optimizing the inference methods. This can

be achieved, for example, by using regression

models or clustering techniques to estimate the

topology of the indoor environment using all the

estimations gathered with all the sensors at the same

time.

As future applications of our work we should

state that it leads to the suggestion of several future

studies within the area of local positioning systems.

Navigation applications, such as the ones used

nowadays within outdoor environments but applied

to indoor spaces are one of the ways of exploring the

developments of this project. Integration with

autonomous driving systems or other applications in

the robotics field of research is another possibility.

Integration with applications used for indoor

commercial purposes, as a way to provide better

efficiency and easier functionalities when consumers

search for products within indoor areas is also

considered. Using such a system in rescue situations,

as a way to know in advance the topology of the

rescue environment represents a very promising

field of application as well. Location-based social

games are another area of great potential for this

work. Generally all applications that make use of

indoor topology maps can benefit from a system that

provides, through inference, indoor topology maps.

REFERENCES

Ahmad, U., Gavrilov, A., Nasir, U., Iqbal, M., Seong Jin,

C. & Sungyoung, L. (2006). In-building Localization

using Neural Networks. In: IEEE International

Conference on Engineering of Intelligent Systems,

2006. p.1-6.

Bahl, P. & Padmanabhan, V. N. (2000). RADAR: an in-

building RF-based user location and tracking system.

In: Proceedings of the 19th Annual Joint Conference

of the IEEE Computer and Communications Societies

(INFOCOM). vol.2, p.775-784.

Beeson, P., Jong, N. K. & Kuipers, B. (2005). Towards

Autonomous Topological Place Detection Using the

Extended Voronoi Graph. In: Proceedings of the 2005

IEEE International Conference on Robotics and

Automation (ICRA). p.4373-4379.

Ekahau (2010). Ekahau, Wi-Fi Based Real-time Tracking

and Site Survey Solutions? [Online]. Available: www.

ekahau.com [Accessed on 26/Feb./2011].

Hui, L., Darabi, H., Banerjee, P. & Jing, L. (2007). Survey

of Wireless Indoor Positioning Techniques and

Systems. IEEE Transactions on Systems, Man, and

Cybernetics, Part C: Applications and Reviews. (37):

1067-1080.

Otsason, V., Varshavsky, A., Lamarca, A. & De Lara, E.

(2005). Accurate GSM Indoor Localization. In: Beigl,

M., Intille, S., Rekimoto, J. & Tokuda, H. (eds.)

UbiComp 2005: Ubiquitous Computing. Springer.

Sogo, T., Ishiguro, H. & Ishida, T. (2001). Acquisition and

propagation of spatial constraints based on qualitative

information. IEEE Transactions on Pattern Analysis

and Machine Intelligence. (23):268-278.

Tadlys (2010). Topaz local positioning solution [Online].

Available: www.tadlys.com [Accessed on 25/Feb./

2011].

Yiguang, X., Sengupta, R. & Fallah, Y. (2010). Making

indoor maps with portable accelerometer and

magnetometer. In: Ubiquitous Positioning Indoor

Navigation and Location Based Service (UPINLBS).

p.1-7.

Biacs, Z., Marshall, G., Moeglein, M., Riley. W. (1999).

The Qualcomm/ SnapTrack Wireless-Assisted GPS

Hybrid Positioning System and Results from Initial

Commercial Deployments. In: Proceedings of the 55th

ION Annual Meeting, Navigational Technology for the

21st Century.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

432