Content uploaded by Peter Brida
Author content
All content in this area was uploaded by Peter Brida on Mar 21, 2018
Content may be subject to copyright.
SPECIAL ISSUE PAPER
Impact of optimization algorithms on hybrid indoor positioning
based on GSM and Wi-Fi signals
Juraj Machaj*
,†
and Peter Brida
University of Zilina, FEE, Department of Telecommunications and Multimedia, Zilina, Slovakia
SUMMARY
In the recent time indoor positioning becomes an extremely hot topic between researchers worldwide. This
is mainly because of lack of available positioning solutions in the indoor environment and possibility to
offer novel services based on location of users. Such services can also open new markets for service
providers and thus increase their income. In this paper we will focus on application of optimization algo-
rithms on performance of hybrid indoor positioning system, which utilize radio signals from both Global
System for Mobile communication and Wi-Fi networks simultaneously. The system is based on fingerprint-
ing framework, which is widely used for indoor positioning because of its performance in strong multipath
environment. Tests of the optimization algorithms were performed via simulations that were performed
using simulation model created in Matlab environment. Copyright © 2016 John Wiley & Sons, Ltd.
Received 1 March 2016; Revised 20 April 2016; Accepted 13 June 2016
KEY WORDS: indoor positioning; position estimation; received signal strength; simulations
1. INTRODUCTION
Indoor positioning still attracts attention of large number of research teams and service providers. This
is mainly because of fact that commonly used Global Navigation Satellite Systems (GNSS) cannot be
used in the indoor environments because of high signal attenuations and presence of multipath signal
propagation [1]. These two effects are the reason that GNSS signal is not feasible for accurate position
estimation or in some cases is not even present in the area where mobile users need to estimate their
positions.
In recent time large number of alternative positioning systems, which can be used in the indoor
environment, was proposed. These systems can help to provide various Location Based Services
(LBS) to the users, which may produce additional income to service providers [2, 3]. Based on data
utilized to estimate position of mobile user these systems can be divided into systems that use
Inertial Measurement Unit (IMU) sensors [4], sound waves, light emitters, image processing [5] and
changes of magnetic field [6] and systems that utilize signals from radio networks. All of these
systems have both advantages and drawbacks. In our work we focus mainly on positioning using
signals from radio networks.
Most of the positioning systems utilize signals from one of the following radio networks: ZigBee
[7], Bluetooth [8], Ultra-WideBand (UWB) [9], Wi-Fi [10] or cellular networks, i.e. Global System
for Mobile communication (GSM) [11], Universal Mobile Telecommunication System (UMTS) [12]
*Correspondence to: Juraj Machaj, University of Zilina, FEE, Department of Telecommunications and Multimedia,
Zilina, Slovakia.
†
E-mail: juraj.machaj@fel.uniza.sk
Copyright © 2016 John Wiley & Sons, Ltd.
CONCURRENCY AND COMPUTATION: PRACTICE AND EXPERIENCE
Concurrency Computat.: Pract. Exper. (2016)
Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/cpe.3911
and Long Term Evolution (LTE) [13]. However, recently systems that utilize signals from more than
one network begin to emerge. The most common combination seems to be combination of Wi-Fi and
Bluetooth [14, 15]. This is given by the fact that low cost, low power Bluetooth beacons become
available. In contrast we will focus on positioning system that will utilize combination of Wi-Fi and
GSM signals in our work.
The most common data used in positioning based on radio networks in the indoor environment is
Received Signal Strength (RSS). The advantage is that RSS data are easily available at every device.
On contrary, main drawback of RSS data is that RSS is strongly affected by shadowing and
multipath effects, especially in the indoor environment. However, these drawbacks can be overcome
by use of fingerprinting positioning algorithms. These algorithms seem to be immune to multipath
propagation and can utilize fading effects of non-movable obstacles to improve positioning
performance.
In the previous papers we have published paper that proved that combination of Wi-Fi and GSM
signals can further improve positioning performance, when compared to positioning using signals
from only one radio network [16]. Moreover, we have proposed optimization algorithms that can
significantly reduce complexity of modular positioning system by both reducing number of
transmitter used during position estimation process and reducing number of reference points
extracted from database during the position estimation process [17].
In this paper we will analyze feasibility of these algorithms and their impact on positioning using both
GSM and Wi-Fi signals. Because there is a relatively large number of parameters that may have impact
on the performance of positioning algorithms all tests will be performed in Matlab environment using
simulation model. This will help us to alter all relevant variables in each position of mobile device
and thus provide us with results achieved under the exact same conditions for each variation of settings.
The rest of paper is organized as follows: related work in the area of indoor positioning is presented
in Section 2. Section 3 describes the proposed positioning system. Simulation scenarios and achieved
results will be presented in Sections 4 and 5, respectively, and Section 6 will conclude the paper.
2. RELATED WORK
As stated above one of the most common approaches to the positioning in the indoor environment is
use fingerprinting framework. Most of the positioning systems that utilize fingerprinting framework
can be described as mobile assisted positioning systems. This means that positions of mobile
devices are estimated by localization server located in the network based on measurements
performed by mobile device.
Operation of these systems can be divided in two stages [18]. During the first stage called offline or
calibration phase the radio map is created. Measurements of RSS from surrounding Access Points
(APs) are taken in the localization area by the mobile device at points with known ground truth
position. Measured RSS data together with ground truth position are sent to the localization server,
where they are stored in database called radio map. The principle of radio map construction is
shown in the Figure 1.
Figure 1. Principle of radio map.
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
During the second stage called online, or positioning stage, RSS data are measured by the mobile
device at unknown position. RSS measurements are then sent to the localization server, which
utilize implemented algorithms to compare measured RSS data with data stored in radio map
database and estimate position of mobile device.
Algorithms that can be used for estimation of mobile devices position can be divided into two main
groups –deterministic and probabilistic algorithms [19]. In this work we will deal with deterministic
algorithms, because they can achieve the same results as probabilistic algorithms. On the other hand
performance of the probabilistic algorithms can be affected by accuracy of the statistical model.
Deterministic fingerprinting framework is based on two assumptions. The first assumption is that
RSS is not randomly distributed in the area [20]. The second basic assumption is that measured RSS
depends on position of the mobile device. Both of these assumptions are fulfilled because RSS
values are affected by propagation conditions and obstacles in the area, and are relatively stable at
each position. The regulation of transmission power of WLAN APs does not affect measured RSS,
because measurements are performed in passive mode and only RSS from APS beacon signals are
measured. These signals are always transmitted at highest power allowed at each AP.
In the deterministic framework the position of mobile device is estimated based on similarity
between the RSS data measured by mobile device and the fingerprints stored in the radio map
database. The position estimate is computed using following estimator:
b
x¼
X
N
i¼1
ωipi
X
N
i¼1
ωi
;(1)
where ω
i
is a non-negative weighting factor and p
i
represents position of i-th reference point. Weights
can be calculated as the inverted value of the distance between RSS vectors from the current
measurement and RSS data stored in the radio map database. The most commonly used measure is
the Euclidian distance; however, other distance metrics are also possible [21].
The estimator (1) which keeps the Klargest weights and sets the others to zero is called the
Weighted K-Nearest Neighbor (WKNN) method [18]. WKNN with all weights ω
i
= 1 is called the
K-Nearest Neighbor (KNN) method. The simplest method, where K= 1, is called the Nearest
Neighbor (NN) method [20]. In [22] it was found that WKNN and KNN methods perform better
than the NN method, particularly when values of parameter Kare 3 or 4. On the other hand, NN
method can achieve similar results as KNN and WKNN modifications, when radio map density is
high enough.
3. HYBRID POSITIONING SYSTEM
In the previous paper we have proven that combination of Wi-Fi and GSM signals has positive impact
on localization accuracy in the indoor environment. Combination of these signals seems to be quite
promising for use in positioning system, because transmission powers of transmitters in both
networks are stable, in case that infrastructure of the network does not change. Another point is that
RSS measurements of GSM signals are significantly more stable compared to Wi-Fi signals as there
is less interference in GSM frequency band.
From the achieved results in [16] it can be seen that using combination of GSM and Wi-Fi signals
the positioning accuracy is improved significantly, by 50% of Root Mean Square Error (RMSE), when
it is compared to GSM-based localization. Although, when comparing to Wi-Fi-based localization the
difference was not significant as combination of GSM and Wi-Fi signals provided approximately 5%
improvement of RMSE.
It is important to notice that combining signals from both systems can also improve availability of the
positioning. This is because of fact that in some areas there are not enough signals from a single network
to perform positioning, as we assume that minimum requirement is to have RSS measurements from
IMPACT OF OPTIMIZATION ALGORITHMS ON HYBRID INDOOR POSITIONING
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
three transmitters required for position estimation in 2D space. The example of environment for hybrid
positioning system is shown in Figure 2.
The principle of hybrid positioning system is pretty straightforward; RSS measurements are made
on both GSM and Wi-Fi wireless receivers. RSS samples are coupled with IDs that are unique for
each transmitter –MAC address for Wi-Fi APs and CellID for GSM base stations. Measured data
from both networks are transmitted together to the positioning server. The server treats the data in
the same way as it is performed for traditional fingerprinting system which works with single
wireless technology.
As stated above, we have proven that by utilization of signals from both GSM and Wi-Fi networks
the positioning accuracy can be improved. Another issue that has to be focused is complexity of the
localization system. Complexity of the system is especially important when it comes to real world
application, because user needs to have information about estimated position as soon as possible.
Complexity of system is not affected only by number of users and coverage area of positioning
system, but also by number of data that has to be processed by the localization server in order to
estimate position of mobile user. Because number of data is increased by utilization of signals from
both networks, implementation of optimization algorithms is required.
In the previous papers [17,23, 24] we have proposed optimization algorithms which help to decrease
complexity of modular positioning system. The modular positioning system was proposed to estimate
position of mobile device in various environments using optimal localization platform (GNSS, GSM or
Wi-Fi) [24]. These algorithms can also be used to reduce complexity of hybrid positioning system,
which utilize signals from both GSM and Wi-Fi networks.
In this paper we focus on implementation of the optimization algorithms for transmitter reduction
and two-phase map reduction into the hybrid positioning system. The hybrid positioning system
utilize signals from GSM and Wi-Fi networks; therefore, there are parameters that have to be set in
order to achieve reduction of system complexity and keeping its performance from the accuracy
point of view. In order to achieve this goal we have to find optimal values for optimization
parameters, especially for thresholds that will be used in transmitter reduction algorithm.
The transmitter reduction algorithm was proposed to reduce complexity of the localization system in
case that mobile device senses large number of transmitters [23]. The basic assumption for the
Figure 2. Example of environment for hybrid positioning system.
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
algorithm is that localization accuracy is mostly affected by transmitters with higher RSS values and
transmitters with low RSS have not significant impact. This assumption was proven by analysis
published in [25].
Another assumption that has to be considered in the transmitter reduction algorithm is that, similarly
to trilateration based positioning, the positioning algorithm needs RSS data from at least N+1
transmitters in order to successfully estimate position in Ndimensional space. Therefore the
minimum number of transmitters has to be set in the algorithm.
The operation of the transmitter reduction algorithm is quite simple, because it was proposed to
reduce complexity of the positioning system and is assumed to be implemented at the mobile
device; the algorithm should reduce amount of data that has to be sent to localization server and
thus reduce response time of the positioning system. The flowchart of the modified algorithm for
use in hybrid positioning is depicted in Figure 3.
In the figure, N
GSM
and N
Wi-Fi
represent number of transmitters after reduction for GSM and Wi-Fi
networks respectively and N
min
represents minimum number of transmitters that has to be used in the
positioning process, in order to satisfy the assumption above.
The second implemented optimization algorithm is the two-phase map reduction algorithm. The
basic idea of this algorithm is to reduce complexity of the positioning system by reduction of number
of reference points used in the estimation process. From the name it is obvious that algorithm works
in two stages. During the first stage all relevant reference points are found based on RSS data
measured by the mobile device during the positioning process. In the second stage reference points
are further reduced based on similarity of detected transmitters. The flowchart of the hybrid
positioning system with implemented two-phase map reduction algorithm is shown in Figure 4.
In the first step the algorithm extracts all reference points, which contain at least one transmitter
detected in fingerprint measured by mobile device in localization phase. In the localization system
this can be easily performed via SQL database. This step helps to reduce the radio map quite
significantly, especially in case that radio map covers large area. The principle of this process is
shown in Figure 5.
For better understanding the process of the two-phase map reduction is explained on the case when
area is covered by three Base Transceiver Stations (BTSs). The mobile device measured the signals
from the surrounding transmitters and signals from the BTS2 and BTS3 were detected.
In the figure, the case in which the mobile device does not detect signals from BTS1 is depicted.
Areas from which reference points are chosen by the algorithm in the first phase are marked with
diagonal pattern. However, there are still some reference points selected from areas where signals
from BTS1 should be detected by the algorithm. On the other hand, the selected area is still quite
large and contains areas where signals from both BTS2 and BTS3 should not be detected.
Figure 3. Block diagram of modified transmitter reduction algorithm.
IMPACT OF OPTIMIZATION ALGORITHMS ON HYBRID INDOOR POSITIONING
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
Therefore, the second stage of the map reduction algorithm selects reference points from the most
appropriate areas. These areas are detected based on similarity between detected transmitters. In this
stage number of matching transmitters in data measured by mobile device in the positioning phase
Figure 5. Phase 1 of map reduction algorithm.
Figure 4. Flowchart of the hybrid positioning system with optimization algorithms.
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
and data stored for individual reference points is computed. Algorithm then extracts from the radio map
all reference points with the highest number of matching transmitters. The output of the second stage of
the map reduction algorithm is shown in Figure 6.
Because the algorithm selects the highest number of matches, it can handle power fluctuations,
which may cause the measured RSS data from the mobile device to contain more (or less)
transmitters when compared to data detected at the reference points in the same area during the
radio map construction.
4. SIMULATION PARAMETERS
Simulations were performed in the simulation model, which was created in Matlab environment. The
layout of the building in the model was proposed to simulate the Department of Telecommunications
and Multimedia at the University of Zilina. The shape of localization area is shown in Figure 7. The
localization area has size 512 m
2
and there are 154 reference points, depicted as dots in the figure.
Reference points are placed in a grid with 2-m spacing between them. The area is covered by
signals from both Wi-Fi and GSM signals provided by nine APs and seven BTSs respectively.
Positions of Wi-Fi APs are depicted in the figure by squares; lines represent walls of the building.
The placement of APs was chosen in a way to provide good Wi-Fi signal coverage in the area.
Positions of BTSs are shown in Figure 8; in the simulation the BTSs are placed in hexagonal shape.
Thus maximum of seven signals can be detected. The distance between BTSs is set to 2 km. In the
Figure 6. Phase 2 of map reduction algorithm.
Figure 7. Localization area.
IMPACT OF OPTIMIZATION ALGORITHMS ON HYBRID INDOOR POSITIONING
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
simulation model we assumed sensitivity threshold for GSM signals 113 dBm. Signals from all BTSs
were detected at most of reference points, because of dependency of building structure (number of
penetrated walls).
There are several propagation models implemented in the simulation model. However, based on
comparison with real world data, the most reliable results were achieved by Multi Wall and Floor (MWF)
propagation model for Wi-Fi signals [26] and AMATA propagation model for GSM signals [27].
The MWF model is based on assumption that there is a nonlinear relationship between the
cumulative penetration loss and the number of penetrated floors and walls [26]. Total propagation
loss L
WMF
in distance dbetween transmitter and receiver can be computed from equation:
LMWF ¼L0þ10nlog dðÞþ
X
I
i¼1
X
Kwi
k¼1
Lwik þX
J
j¼1
X
Kfj
k¼1
Lfjk;(2)
where L
0
is path loss in distance of 1 m in [dB], nrepresents path loss exponent, dis distance between
transceiver and receiver in [m], Istands for number of walls types, K
wi
is number of penetrated walls of
the i-th category, L
wik
represents attenuation because of i-th wall type and k-th traversed wall in [dB].
Similarly, Jstands for number of floor types, K
fj
is number of traversed floors of j-th category and L
fjk
represents attenuation because of j-th floor type and k-th traversed floor in [dB].
In the simulations, the L
0
was set to 20 dB, based on calculation using Friis equation and the path
loss exponent was set to n= 2. Single floor building was used in simulation, therefore attenuation
caused by floors can be neglected. On the other hand, three types of wall were assumed:
•interior wall –attenuations L
w1
= 8; 6; 4; 2; 1 [dB],
•external wall –attenuation L
w2
= 9; 7 [dB]
•elevator shaft or pillars L
w3
= 10 [dB].
The RSS was calculated from L
MWF
based on assumption that all APs transmit signal with power of
10 dBm. All the propagation parameters were set according to results published in [26].
AMATA propagation model was created as modification of ITU propagation model in order to allow
calculating signal propagation losses in the indoor environment [27]. Propagation loss L
GSM
is given by:
LGSM ¼20log10 fþ10n0log10 dþLout þXaþLf;(3)
where fis frequency of the GSM signal in [MHz], n
0
represents path loss exponent after isolation of the
internal wall effect, dstands for the distance from the transmitter in [m], L
out
represents attenuation loss
factor of outer wall and L
f
represents (multi)floor attenuation. X
a
is the internal multi wall attenuation
loss factor, which can be estimated by:
Figure 8. Positions of BTSs in the simulation model.
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
Xa¼0:0075H40:18H3þ1:1H2þ2:9H;(4)
where Hrepresents the number of penetrated walls of standard type. The equation above is applicable
for up to 11 penetrated indoor walls between transmitter and receiver. It is also important to notice
that, based on real world data, signal attenuations become relatively low to measure after penetration
of eight walls.
In the simulation the parameters for AMATA propagation model were set as follows: frequency of
GSM signal was set to f= 900 MHz, path loss exponent n
0
= 1.5, attenuation of external wall was set to
L
out
= 8 dB and attenuation by floors was neglected, because we assumed single floor building. The
transmit power of BTSs stations was assumed to be P
T
= 30 dBm and was used to compute the RSS
at the receiver position.
In the simulation model random variable is added to the computed propagation loss in order to
simulate signal fluctuations because of fading and multipath propagation. Random variables were
different for GSM and Wi-Fi attenuations, because these signals have different fluctuations, because
of different frequencies and different sources of interference in respective frequency bands. For the
Wi-Fi signals we have used random variable which was achieved as a product of two random
variables with lognormal and uniform distribution respectively. This helped us to simulate
distribution of RSS samples similar to distribution achieved from the real world measurement [21].
On the other hand signal fluctuations for GSM signals were modeled using zero mean random
variable with normal distribution and standard deviation of 5 dB.
Because there is a need to reduce impact of signal fluctuations on localization performance the
average received signal strength RSS
――― is used by the fingerprinting positioning systems. The average
received signal strength is calculated as:
RSS
―――
¼1
NsX
NS
i¼1
RSSi;(5)
where N
s
is number of RSS samples measured at the given position by the mobile device and RSS
i
stands for i-th sample. In the simulations we used N
s
= 5, because based on previous results [17] we
can conclude that further increase of RSS samples does not have significant impact on improvement
of localization accuracy.
In the simulations all three deterministic algorithms described in the Section 2 –NN, KNN and
WKNN –are used to estimate position of the mobile node. As the metric for difference between
RSS data measured in the online phase and RSS data in the radio map database the Euclidean
distance was used. Number of reference points used for position estimation in KNN and WKNN
was chosen based on results from the preliminary simulations shown in Figure 9.
From the results it can be seen, that with increasing number of reference points used for position
estimation –K, the localization error is decreased. It can also be seen that the most significant
impact of K is between values 1 and 2 and then between 3 and 4. When K is further increased it has
only small impact on the localization accuracy. Therefore we have decided to set K= 4 for both
algorithms.
Figure 9. Impact of number of used reference points Kon localization error of WKNN and KNN algorithms.
IMPACT OF OPTIMIZATION ALGORITHMS ON HYBRID INDOOR POSITIONING
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
The simulations were performed in order to find appropriate RSS thresholds for both Wi-Fi and
GSM signals in order to be able to reduce complexity of the system, by use of transmitter reduction
algorithm. The algorithm is utilized to reduce data that have to be sent to localization server and
processed and also test impact of the two-phase reduction algorithm on the achieved localization
accuracy.
The thresholds were gradually increased by 1 dB from the sensitivity threshold up to 70 dBm. The
sensitivity threshold is 113 dBm for GSM according to [28]; therefore, the lowest RSS threshold was
set to 110 dBm. However, there is no standard that defines sensitivity threshold for Wi-Fi and
therefore the sensitivity threshold is affected by differences in chipset and antenna design. However,
based on our experiences the most common value is 100 dBm, which was also chosen as lowest
detectable RSS value in simulations.
The other factor that may have impact on localization accuracy, especially when higher thresholds
are defined, is minimum number of transmitters assumed for each network. Because the transmitter
reduction algorithm was proposed to work with single wireless network the minimum number of
transmitters was set to 3, according to the fact that we need signals from at least three transmitters to
define point in 2D space. In the simulation the minimum number of transmitters was set to 3 for
both networks, resulting in minimum of six transmitters in total for position estimation process.
To conclude this section we summarize simulations that will be performed:
•Impact of different RSS thresholds in transmitter reduction algorithm on accuracy of positioning
algorithms.
•Impact of two-phase map reduction algorithm on accuracy of positioning algorithms.
•Analysis of localization errors achieved before and after implementation of optimization
algorithms.
•Comparison of accuracy of optimized hybrid positioning system with hybrid positioning system
and both systems based on single network.
5. ACHIEVED RESULTS
In the first simulation the impact or RSS threshold in the optimization algorithm for transmitter
reduction was tested for each of the implemented algorithms. Simulations were performed for
10 000 independent position estimations; for each trial position of mobile device was randomly
chosen from the localization area with uniform distribution. Results achieved for the NN, KNN and
WKNN localization algorithms are shown in Figures 10, 11 and 12, respectively.
From the results achieved for NN algorithm it is clear that RSS threshold for each technology has
impact on the accuracy of the positioning algorithm. It can be seen that with increasing threshold
value, the localization accuracy decreases. However, it can also be seen that the localization error is
stable when threshold of RSS is set between 110 dBm and 95 dBm for GSM signals along with
RSS threshold between 100 dBm and 90 dBm for Wi-Fi signals. These results are with
Figure 10. Impact of RSS threshold in transmitter reduction algorithm on NN algorithm.
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
correlation of the thresholds for individual technologies as was previously published in [17]. It is
interesting to see that for NN algorithm the accuracy gets significantly lower, when only threshold
for GSM signals is increased and all available Wi-Fi signals are used, i.e. RSS threshold for Wi-Fi
set to 100 dBm.
Similarly to NN algorithm also performance of KNN algorithm is affected by RSS threshold for
both wireless networks. The area where the localization accuracy is stable is exactly the same, i.e.
with RSS threshold 110 dBm and 95 dBm for GSM signals along with RSS threshold between
100 dBm and 90 dBm for Wi-Fi signals. The main difference is that localization error is
significantly decreased with the higher RSS thresholds for GSM signals, when compared to results
achieved by NN algorithms. The errors achieved for higher RSS thresholds together with lowest
RSS thresholds for Wi-Fi signals for GSM signals are approximately 50% lower compared to those
achieved by NN algorithm.
From the figure above it can be seen that there is not a big difference between performance achieved by
KNN algorithm and WKNN algorithm. The behavior of the WKNN and KNN algorithms is the same.
From the results shown it is obvious that performance of fingerprinting algorithms is significantly
affected by RSS threshold from both wireless technologies that are assumed to be used for position
estimation. This is caused by the fact that signals from both networks provide some additional
information about position. Moreover, both signals have different fading characteristics; therefore,
hybrid system can utilize advantages of relative stable GSM signals and Wi-Fi signals that are more
affected by the environment.
It is also important to note that with RSS thresholds set approximately 10 dB above the sensitivity of
receivers for each technology, the achieved accuracy is approximately the same as in the system
without RSS thresholds. It can also be assumed that in real world scenario the impact of RSS
thresholds set 10 dB above sensitivity of the receiver will have positive impact on the accuracy of
the system. This can be assumed based on the fact that there can be reflected signals from the APs
Figure 11. Impact of RSS threshold in transmitter reduction algorithm on KNN algorithm.
Figure 12. Impact of RSS threshold in transmitter reduction algorithm on WKNN algorithm.
IMPACT OF OPTIMIZATION ALGORITHMS ON HYBRID INDOOR POSITIONING
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
and BTSs with low RSS and high signal fluctuations. Such signals might have negative impact on the
accuracy of the system especially in case that high number of transmitters is detected in the area.
In the forthcoming simulation, the impact of the two-phase map reduction algorithm on the
performance of positioning algorithms was tested. In this stage only information about transmitters
which were above the threshold were used for the map reduction. However it is important to notice
that minimum of three strongest transmitters were utilized from both technologies in a case that
there were not enough transmitters above the threshold. The results achieved for NN, KNN and
WKNN algorithms are depicted in Figures 13, 14 and 15, respectively.
From Figure 13 it can be seen that map reduction algorithm has just a small impact on localization
accuracy; however, it is clear that in comparison to results in Figure 10 the localization error is reduced
when higher RSS threshold for Wi-Fi signals is used in the localization process. The localization error
was reduced by more than 2 m in case that RSS threshold for Wi-Fi signals was higher than 85 dBm
and RSS threshold for GSM was below 95 dBm.
The results for KNN algorithm show that two-phase map reduction algorithm does not have
significant impact on the localization accuracy of the hybrid positioning system; however, the same
as for NN algorithm can be concluded. The localization accuracy was increased in case that RSS
threshold was below 95 dBm for GSM above 85 dBm for Wi-Fi.
Results achieved for the last tested localization algorithm show the similar results as for NN and
KNN algorithms. Therefore, it can be concluded that two-phase map reduction algorithm does not
have significant impact on localization accuracy. However, main goal of this algorithm is to reduce
complexity of localization system by reducing amount of data that needs to be processed in order to
estimate position of the mobile device.
In order to better understand differences of the positioning accuracy achieved by implementation of
both optimization algorithms cumulative distribution functions (CDF) of localization errors are
Figure 13. Impact of map reduction on results achieved for NN algorithm.
Figure 14. Impact of map reduction on results achieved for KNN algorithm.
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
compared for each algorithm. In this comparison localization errors achieved by optimization
algorithms for optimal thresholds were used. Based on the results shown above the optimal
thresholds are 100 dBm and 90 dBm for GSM and Wi-Fi networks, respectively. In Figure 16
CDF of localization errors achieved by modifications of NN algorithm are depicted.
From the figure above it can be seen that localization error has very similar characteristics for all
three modifications of the NN algorithm. It can be seen that by implementation of transmitter
reduction optimization algorithm the localization error was slightly increased. However, use of two-
phase map reduction optimization algorithm helped to increase accuracy of the NN algorithm. The
results achieved by KNN algorithm are shown in Figure 17.
From the results shown in Figure 17 it can be seen that again transmitter reduction algorithm slightly
decreased localization accuracy on KNN algorithm. Moreover it can be seen that in this case the
localization error was not improved by the two-phase map reduction algorithm. Implementation of
this algorithm, however, did not have significant impact on the accuracy of the system.
In Figure 18, the CDF of modified WKNN algorithm are shown. From these results it can be seen
that the impact of the optimization algorithms is the same as in case of KNN algorithm. The differences
in the results achieved by both optimization algorithms are not significant. Even though, the
optimization algorithms show slight decrease of localization accuracy in the simulation results. It
can be assumed that in real world conditions, where signal reflections with low RSS are randomly
detected over the area, improvement of the localization accuracy can be achieved. Statistical
parameters of the achieved localization errors are shown in Table I.
For comparison localization errors achieved by the system, which utilizes signals from a single
wireless network, are shown in Table II.
Based on the comparison of the results shown in the tables, the combination of both signals in the
hybrid positioning system allows to improve localization accuracy. The improvement is more
significant when compared to GSM-based localization, because GSM signals are less attenuated by
walls and distance compared to Wi-Fi signals. Another improvement is the fact that by use of
Figure 15. Impact of map reduction on results achieved for WKNN algorithm.
Figure 16. CDF of localization errors achieved by NN algorithm.
IMPACT OF OPTIMIZATION ALGORITHMS ON HYBRID INDOOR POSITIONING
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
Figure 17. CDF of localization errors achieved by KNN algorithm.
Table I. Statistical parameters of achieved localization errors for hybrid positioning system.
NN KNN WKNN
Basic
Map
reduction
Transmitter
reduction Basic
Map
reduction
Transmitter
reduction Basic
Map
reduction
Transmitter
reduction
Localization
error [m]
Mean 3.56 3.51 3.61 2.77 2.88 2.9 2.79 2.88 2.92
Median 3.12 3.12 3.17 2.55 2.59 2.63 2.57 2.59 2.65
std 2.31 2.28 2.31 1.65 1.82 1.81 1.67 1.81 1.8
95% 8.55 8.12 8.59 5.82 6.38 6.25 5.89 6.34 6.32
Figure 18. CDF of localization errors achieved by WKNN algorithm.
Table II. Statistical parameters of achieved localization errors for GSM and Wi-Fi based systems.
NN KNN WKNN
Wi-Fi GSM Wi-Fi GSM Wi-Fi GSM
Localization error [m] Mean 3.64 6.83 2.99 5.54 3.01 5.53
Median 3.27 4.65 2.72 4.02 2.74 4.02
std 2.24 5.86 1.78 4.37 1.78 4.38
95% 8.43 19.14 6.77 14.37 6.85 14.4
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
hybrid positioning system the density of radio map and coverage area can be increased, because at
some points in the area the mobile device is not able to detect at least three transmitters from a
single network.
It is also important to notice that the localization accuracy of optimized hybrid system is slightly
better compared to localization systems based on utilization of signals from a single network. This
was achieved even though implemented optimization algorithms slightly decreased localization
accuracy of the hybrid positioning system. Moreover as already stated the implemented algorithms
are assumed to have positive impact on the localization accuracy, because these will help to reduce
reflected signals that cause errors in position estimation process.
6. CONCLUSION
In the paper hybrid positioning system was introduced together with optimization algorithms that are
focused on the reduction of the localization system complexity. The optimization algorithm for
transmitter reduction was tested in the first part of simulations. Impact of different RSS threshold
combinations on the performance of the algorithm and its impact on the localization error was tested.
Based on the achieved results it can be concluded that RSS threshold set 10 dB over sensitivity of the
receiver can provide reduction of RSS data; however, its impact on the accuracy of the localization
system is negligible. Moreover, the performance of the system in real world conditions is assumed to
be improved because algorithm will remove reflected signals from transmitters that are far from the
area of localization. Based on the results it can be stated that transmitter reduction algorithm can
reduce amount of data that has to be sent from mobile device to localization server and therefore
reduce both complexity of the localization process and requirements on data transfer.
The impact of the two-phase radio map reduction algorithm was tested at the same dataset with all
the different RSS thresholds. The impact of radio map reduction on the performance of the localization
system was significant only in case that system worked with the higher RSS threshold for Wi-Fi
signals. In such case the algorithm helped to reduce localization error by 2 m compared to results
achieved only by transmitter reduction algorithm. However, the main goal of this optimization
algorithm was to decrease computational complexity of the system at the server side. This was
achieved by significant reduction of data that has to be compared using Euclidean distance.
From the comparison of localization errors achieved by hybrid positioning system with and without
optimization to the errors achieved by localization using data from only one network it can be seen that
hybrid positioning system achieved better results. It is also clear that by introducing optimization
algorithms the accuracy of the system slightly decreased. However, achieved accuracy is still better
compared to the system-based RSS measurements from single network. Another feature of the hybrid
positioning system is that it allows estimate position of mobile device in areas where systems based on
measurements from single network do not have enough information to estimate position of mobile device.
ACKNOWLEDGEMENTS
This work was partially supported by the Slovak Research and Development Agency under the contract no. LPP-
0126-09, by the Slovak VEGA grant agency, project no. 1/0263/16 ‘Research of integrated positioning system
based on wireless systems and sensors implemented in intelligent mobile devices’,EUREKAprojectno.E!
6752 –DETECTGAME: R&D for Integrated Artificial Intelligent System for Detecting the Wildlife Migration
and ‘Broker Center of Air Transport for Transfer of Technology and Knowledge into Transport and Transport In-
frastructure ITMS 26220220156’. We support research activities in Slovakia. Project is co-financed by EU.
REFERENCES
1. Andreotti M, Aquino M, Woolfson M, Walker J, Moore T. Signal propagation analysis and signature extraction for
GNSS indoor positioning. 2006 IEEE/ION Position,Location,And Navigation Symposium, pp. 913–919, doi:
10.1109/PLANS.2006.1650691
2. Wen YH, Chang HS, Kao HW, Ju GH. Location-aware services based on Wi-Fi network. 16th Asia-Pacific Network
Operations and Management Symposium (APNOMS), 2014, pp. 1–4., doi: 10.1109/APNOMS.2014.6996533
3. Horalek J, Sobeslav V, Krejcar O, Balik L. Communications and security aspects of smart grid networks design.
Communications in Computer and Information Science 2014; 465:35–46. doi:10.1007/978-3-319-11958-8_4.
IMPACT OF OPTIMIZATION ALGORITHMS ON HYBRID INDOOR POSITIONING
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
4. Lategahn J, Muller M, Rohrig C. Robust pedestrian localization in indoor environments with an IMU aided TDoA
system, 2014 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pp. 465–472, doi:
10.1109/IPIN.2014.7275518
5. Mautz R, Tilch S. Survey of optical indoor positioning systems, 2011 International Conference on Indoor Position-
ing and Indoor Navigation (IPIN), pp. 1–7, doi: 10.1109/IPIN.2011.6071925
6. Shen HM, Hu L, Qin LH, Fu X. Real-time orientation-invariant magnetic localization and sensor calibration based on
closed-form models. IEEE Magnetics Letters 2015; 6:1–4. doi:10.1109/LMAG.2015.2460211.
7. Fang SH, Wang CH, Huang TY, Yang CH, Chen YS. An enhanced ZigBee indoor positioning system with an ensemble
approach. IEEE Communications Letters April 2012; 16(4):564–567. doi:10.1109/LCOMM.2012.022112.120131.
8. Jianyong Z, Haiyong L, Zili Ch, Zhaohui L. RSSI based Bluetooth low energy indoor positioning. 2014 International
Conference on Indoor Positioning and Indoor Navigation (IPIN), 2014, pp. 526–533, doi: 10.1109/
IPIN.2014.7275525
9. Muller P, Wymeersch H, Piche R. UWB positioning with generalized Gaussian mixture filters. IEEE Transactions on
Mobile Computing Oct. 2014; 13(10):2406–2414. doi:10.1109/TMC.2014.2307301.
10. Grzenda M. Reduction of signal strength data for fingerprinting-based indoor positioning. Intelligent Data Engineer-
ing and Automated Learning–IDEAL 2015, Springer International Publishing, 2015. p. 387–394, DOI 10.1007/978-
3-319-24834-9_45
11. Ibrahim M, Youssef M. CellSense: an accurate energy-efficient GSM positioning system. IEEE Transactions on Ve-
hicular Technology Jan. 2012; 61(1):286–296. doi:10.1109/TVT.2011.2173771.
12. Veletic M, ŠunjevarićM. Optimal positioning in UMTS using least mean squares algorithm on circular lateration.
20th Telecommunications Forum (TELFOR), 2012, Belgrade, 2012, pp. 334–337, doi: 10.1109/
TELFOR.2012.6419215
13. Kangas A, Wigren T. Angle of arrival localization in LTE using MIMO pre-coder index feedback. IEEE Communi-
cations Letters August 2013; 17(8):1584–1587. doi:10.1109/LCOMM.2013.062113.130870.
14. Baniukevic A, Jensen CS, Lu H. Hybrid indoor positioning with Wi-Fi and Bluetooth: architecture and performance.
2013 IEEE 14th International Conference on Mobile Data Management (MDM), Milan, 2013, pp. 207–216, doi:
10.1109/MDM.2013.30
15. Galván-Tejada CE, Carrasco-Jiménez JC, Brena RF. Bluetooth-WiFi based combined positioning algorithm, imple-
mentation and experimental evaluation. Procedia Technology 7:2013, 37-45, ISSN 2212-0173, doi:10.1016/j.
protcy.2013.04.005
16. Machaj J, Brida P. Using GSM and Wi-Fi signals for indoor positioning based on fingerprinting algorithm. Advances
in Electrical and Electronics Engineering 2015; 13(3):242–248. doi:10.15598/aeeev13i31241.
17. Machaj J, Brida P, Benikovsky J. Scalability optimization of seamless positioning service. Mobile information sys-
tems (in review)
18. Honkavirta V, Perälä T, Ali-Löytty S, Piché R. A comparative survey of WLAN location fingerprinting methods, 6th
Workshop on Positioning,Navigation and Communication, WPNC 2009, pp: 243–251, 2009, doi: 10.1109/
WPNC.2009.4907834
19. Saha S, Chauhuri K, Sanghi D, Bhagwat P. Location determination of a mobile device using IEEE 802.11b access
point signals. 2003 IEEE Wireless Communications and Networking,WCNC 2003, vol. 3, pp. 1987–1992, 2003,
doi: 10.1109/WCNC.2003.1200692
20. Bahl P, Padmanabhan VN. RADAR: an in-building RF-based user location and tracking system. Nineteenth Annual
Joint Conference of the IEEE Computer and Communications Societies,INFOCOM 2000, ISBN: 0-7803-5880-5,
2000, doi: 0.1109/INFCOM.2000.832252
21. Machaj J, Brida P. Performance comparison of similarity measurements for database correlation localization method.
3rd Asian Conference on Intelligent Information and Database Systems,ACIIDS 2011, Lecture Notes in Computer
Science, Springer Berlin/Heidelberg, 2011, vol. 6592, ch. 46, pp. 452–461, ISBN 978-3-642-20041-0, 2011, doi:
10.1007/978-3-642-20042-7_46
22. Tsung-Nan L, Po-Chiang L. Performance comparison of indoor positioning techniques based on location fingerprint-
ing in wireless networks. 2005 International Conference on Wireless Networks,Communications and Mobile Com-
puting, Vol. 2, 2005, doi: 10.1109/WIRLES.2005.1549647
23. Machaj J, Brida P, Benikovsky J. Impact of APs removal on accuracy of fingerprinting localization algorithms. 2015
38th International Conference on Telecommunications and Signal Processing (TSP), 2015, pp. 1–5, doi: 10.1109/
TSP.2015.7296384
24. Brida P, Machaj J, Benikovsky J. A modular localization system as a positioning service for road transport. Sensors
2014; 14(11):20274–20296. doi:10.3390/s141120274.
25. Machaj J, Brida P. Performance investigation of the RBF localization algorithm. Advances in Electrical and Elec-
tronic Engineering, Vol. 11, No. 2, pp. 100–107, ISSN 1336-1376, 2013, doi: 10.15598/aeee.v11i2.761
26. Lott M, Forkel I. A multi-wall-and-floor model for indoor radio propagation. 53rd IEEE VTS Vehicular Technology
Conference,2001.VTC 2001 Spring, vol.1, pp.464–468, 2001 doi: 10.1109/VETECS.2001.944886
27. Ata OW, Shahateet AM, Jawadeh MI, Amro AI. An indoor propagation model based on a novel multi wall attenu-
ation loss formula at frequencies 900 MHz and 2.4 GHz. Wireless Personal Communications, Vol. 69, Issue: 1, pp.
23-36, ISSN: 0929-6212, 2013, doi: 10.1007/s11277-012-0558-x
28. ETSI/TC: digital cellular telecommunications system (phase 2); AT command set for GSM Mobile Equipment (ME),
ETS 300 642, GSM 07.07 version 4.4.1, 1999
J. MACHAJ AND P. BRIDA
Copyright © 2016 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. (2016)
DOI: 10.1002/cpe
