sensors

Article
Smartphone-Based Indoor Localization with
Bluetooth Low Energy Beacons
Yuan Zhuang 1 , Jun Yang 1, *, You Li 2,3 , Longning Qi 1 and Naser El-Sheimy 2
1 National ASIC System Engineering Research Center, Southeast University, 2 Sipailou, Nanjing 210096,
China; zhy.0908@gmail.com (Y.Z.); longn_qi@seu.edu.cn (L.Q.)
2 Department of Geomatics Engineering, The University of Calgary, 2500 University Drive, NW, Calgary,
AB T2N 1N4, Canada; liyou331@gmail.com (Y.L.); elsheimy@ucalgary.ca (N.E.-S.)
3 GNSS Research Center, Wuhan University, 129 Luoyu Road, Wuhan 430079, China
* Correspondence: dragon@seu.edu.cn; Tel.: +86-25-8379-3265

Academic Editors: Mihai Lazarescu and Luciano Lavagno

Received: 9 March 2016; Accepted: 20 April 2016; Published: 26 April 2016

Abstract: Indoor wireless localization using Bluetooth Low Energy (BLE) beacons has attracted
considerable attention after the release of the BLE protocol. In this paper, we propose an algorithm
that uses the combination of channel-separate polynomial regression model (PRM), channel-separate
fingerprinting (FP), outlier detection and extended Kalman filtering (EKF) for smartphone-based
indoor localization with BLE beacons. The proposed algorithm uses FP and PRM to estimate the
target’s location and the distances between the target and BLE beacons respectively. We compare
the performance of distance estimation that uses separate PRM for three advertisement channels
(i.e., the separate strategy) with that use an aggregate PRM generated through the combination of
information from all channels (i.e., the aggregate strategy). The performance of FP-based location
estimation results of the separate strategy and the aggregate strategy are also compared. It was found
that the separate strategy can provide higher accuracy; thus, it is preferred to adopt PRM and FP for
each BLE advertisement channel separately. Furthermore, to enhance the robustness of the algorithm,
a two-level outlier detection mechanism is designed. Distance and location estimates obtained from
PRM and FP are passed to the first outlier detection to generate improved distance estimates for
the EKF. After the EKF process, the second outlier detection algorithm based on statistical testing is
further performed to remove the outliers. The proposed algorithm was evaluated by various field
experiments. Results show that the proposed algorithm achieved the accuracy of <2.56 m at 90% of
the time with dense deployment of BLE beacons (1 beacon per 9 m), which performs 35.82% better
than <3.99 m from the Propagation Model (PM) + EKF algorithm and 15.77% more accurate than
<3.04 m from the FP + EKF algorithm. With sparse deployment (1 beacon per 18 m), the proposed
algorithm achieves the accuracies of <3.88 m at 90% of the time, which performs 49.58% more accurate
than <8.00 m from the PM + EKF algorithm and 21.41% better than <4.94 m from the FP + EKF
algorithm. Therefore, the proposed algorithm is especially useful to improve the localization accuracy
in environments with sparse beacon deployment.

Keywords: indoor localization; polynomial regression model; fingerprinting; extended Kalman

filtering; outlier detection; BLE beacons

1. Introduction
Currently, indoor localization has become a significant research topic as it is the fundamental of
numerous Internet of Things (IoT) applications (e.g., human tracking and precision advertisement) [1].
Several technologies can be used for indoor localization, such as wireless localization [1–4],
sensor-based relative navigation (i.e., dead-reckoning) [5–12], and image-based navigation [13–17].

Sensors 2016, 16, 596; doi:10.3390/s16050596 www.mdpi.com/journal/sensors

Sensors 2016, 16, 596 2 of 20

Wireless localization has been widely used among these technologies. Especially, WiFi localization is
the most common consumer wireless localization technology [1,18,19]. Another important candidate
for wireless localization on consumer smart devices is Bluetooth. The traditional Bluetooth has a
significantly long scan time (~10 s), which limits its value for localization. However, the new Bluetooth
protocol (i.e., Bluetooth Low Energy, BLE), supported by most current smart devices, has overcome the
limitations of long scan time. Moreover, the BLE beacons have the following advantages: small size,
light weight, low cost, power saving and are widely supported by smart devices. Therefore, BLE has
the potential to become a dominant wireless localization technology.
In the BLE protocol definition, 40 channels, each 2 MHz wide, around the 2.4 GHz radio band
are used to transmit messages. The duration for transmitting messages is extremely short to save
battery power. Among these 40 channels, there are three channels (i.e., 37, 38, and 39) for broadcasting
advertisement messages. The received signal strengths (RSS) from these three channels can be used for
estimating the target’s location. The BLE advertising rate can be set up to 50 Hz. The transmission
power for BLE beacons are also set from 0 dBm to ´40 dBm. To reduce power consumption, BLE
advertising rate and transmission power are usually set to less than 10 Hz and ´16 dBm, respectively.
Comparing with WiFi localization, BLE localization has the following advantages:

‚ BLE RSS signals can have a higher sample rate than WiFi RSS signals (0.25 Hz~2 Hz)
‚ BLE consumes less power than WiFi
‚ BLE RSS signals can be obtained from most smart devices, while WiFi RSS signals cannot be
provided by Apple portable devices and
‚ BLE beacons are usually battery powered, which are more flexible and easier deployed than WiFi.

Therefore in this paper, we proposes an algorithm for smartphone-based indoor localization using
RSS signals from BLE beacons.
Currently, there are mainly three RSS-based BLE localization techniques: proximity, trilateration
(or, range-based) and fingerprinting (FP) [20]. The proximity algorithms pre-set an event-triggering
threshold for a coverage area. If the RSS values are stronger than the threshold, the target is indicated
in the area. In the trilateration algorithm, the propagation model (PM) is used to estimate the distances
between the target and BLE beacons. Then, the estimation technique (e.g., least squares or Kalman
filtering) is used to estimate the target’s location from these distances and the locations of BLE beacons.
FP first builds a radio map database where RSSs from available BLE beacons are mapped to absolute
positions. Then, the target’s location is determined by using the average or weighted average of the
locations of several most matched fingerprints in the database.
To further improve accuracy and robustness, we propose an algorithm for indoor localization with
BLE beacons by using the combination of polynomial regression model (PRM), FP, outlier detection and
extended Kalman filtering (EKF). Since indoor environments are complicated by reflection, shadowing
and multipath, the radio PM may be not accurate for distance estimation in the indoor environment.
Therefore, the proposed algorithm uses PRM to estimate the distances between the target and BLE
beacons. Meanwhile, FP is also used to determine the target’s location.
Most of existing research combines RSS from all channels to obtain an aggregate signal (the
aggregate mode). However, we notice that the signals and noises from various advertisement channels
may be different, which is shown in Figure 1. In Figure 1, RSS values from three advertisement channels
of a BLE beacon were collected by keeping an iPhone 4S static and five meters away. As reported
by iOS 8, the mean and standard deviation (STD) of RSS values from various advertisements were
different (i.e., Channel 37: Mean ´71.9 dBm, STD 2.4 dBm; Channel 38: Mean ´68.1 dBm, STD 0.8 dBm;
and Channel 39: Mean ´78.8 dBm, STD 2.7 dBm). This difference might be caused by the different
channel gain and multipath effect. Therefore, one advantage of this research is that it considers that
signals and noises in RSS measurements from various channels may be different and thus uses separate
PRM and FP for three advertisement channels (i.e., the separate mode), that is, the PRM and FP are
used for each advertisement channel separately. Therefore, there are a maximum of three PRM distance
Sensors 2016, 16, 596 3 of 20

Sensors 2016, for

estimates 16, xeach 3 of 20
observed BLE beacon and three FP location estimates. Furthermore, to enhance
the robustness of the algorithm, a two-outlier-detection mechanism is designed. These estimates are
mechanism is designed. These estimates are fed into the first outlier detection to generate improved
fed into the first outlier detection to generate improved distance estimates for observed BLE beacons.
distance estimates for observed BLE beacons. Next, these improved distance estimates are utilized as
Next, these improved distance estimates are utilized as measurements for the EKF. Finally, the second
measurements for the EKF. Finally, the second outlier detection algorithm is further performed to
outlier detection algorithm is further performed to remove the outliers from the measurements by
remove the outliers from the measurements by using the statistical testing method.
using the statistical testing method.

Figure 1.
Figure RSSvalues
1. RSS valuesreceived
received from
from three
three advertisement
advertisement channels
channels by
by keeping
keeping the
the device
device static.
static.

The contribution
The contribution of this paper is listed as follows:
(1) We
We propose
propose the usage
usage of
of the
the separate
separate PRM
PRM toto improve
improve both
both the
the location
location and
and distance
distance estimation
estimation
for each
for eachadvertisement
advertisement channel
channel of
of BLE
BLEbeacons.
beacons.Moreover,
Moreover, wewe originally
originally generate
generate separate
separate radio
radio
map database
map database for
for each
each BLE
BLE advertisement
advertisement channel
channel for
for the
the FP process.
We originally
(2) We originally propose
proposeananalgorithm
algorithm forfor
BLE-based
BLE-basedindoor localization
indoor by combing
localization separate
by combing PRM,
separate
separate
PRM, FP, EKF
separate FP,and outlier
EKF detection.
and outlier detection.
(3) We
We propose a two-level outlier detection algorithm to improve the robustness of the system. system.
The proposed
The proposed algorithms
algorithms are
are validated
validated inin detail
detail by
by field experiments
experiments with
with both
both sparsely
sparsely and
and
densely distributed
densely distributed BLE beacons. Part
Part of
of the
the outcomes
outcomes are:
(1) Compared
Compared withwith results
results that
that use
use traditional
traditional PM,
PM, the
the distance
distance estimation
estimation accuracy
accuracy is
is improved
improved byby
18.42% using
18.42% using the
the PRM.
PRM.
In the case
(2) In caseofofdense
densedeployment
deployment of BLE beacons,
of BLE the proposed
beacons, algorithm
the proposed achieves
algorithm averageaverage
achieves 35.82%
and 15.77%
35.82% improvement
and 15.77% of the location
improvement accuracyaccuracy
of the location in two trajectories, compared
in two trajectories, with classical
compared with
PM + EKF
classical PMand FP and
+ EKF + EKF,
FP respectively. The improvement
+ EKF, respectively. changes
The improvement to 49.58%
changes and 21.41%
to 49.58% in the
and 21.41%
sparse
in deployment.
the sparse deployment.
This paper
This paper is
is organized
organized as follows. Section
Section 22 reviews
reviews the
the related
related work. Section
Section 33 presents
presents the
the
proposed algorithm, Section 4 evaluates the algorithm with various field experiments and Section 55
proposed algorithm, Section 4 evaluates the algorithm with various field experiments and Section
draws the
draws the conclusions.
conclusions.

2. Related
2. RelatedWork
Work
Current localization
Current localizationalgorithms
algorithmsforfor
BLE beacons
BLE can be
beacons candivided into three
be divided into classes: proximity
three classes: [21–23],
proximity
range-based
[21,22,23], [21,24–26] and
range-based FP [4,21,27,28].
[21,24–26] In the first class,
and FP [4,21,27,28]. In thethe work
first [21]the
class, evaluates
work [21]theevaluates
performance
the
performance of BLE localization by using proximity in a multi-floor building. A triggering threshold
optimization method is proposed for proximity based positioning [22]. The research [23] proposes a
particle filtering for proximity based indoor positioning. In the second class, a PM is used in a multi-
neighbors (k-NN), neural networks, and support vector machine (SVM). The result shows that k-NN
is a good candidate for localization in real-life applications. The research [4] provides a detailed study
of the effects of beacon density, transmit power and transmit frequency for BLE FP. A FP solution
using Weibull probability distribution is proposed in [28] to improve the reliability and accuracy of
the positioning. In addition, the paper [29] presents a comparative analysis of contemporary BLE
Sensors 2016, 16, 596 4 of 20
indoor positioning solutions, taking into account the classification, comparison and various
considerations that are required for designing new indoor positioning approaches. All the above
research
of combines by
BLE localization RSS fromproximity
using all channels to obtain anbuilding.
in a multi-floor aggregate signal, instead
A triggering of using
threshold separate
optimization
model for
method is each separate
proposed forBLE advertisement
proximity channel. [22]. The research [23] proposes a particle
based positioning
filtering for proximity based indoor positioning. In the second class, a PM is used in a multi-floor
3. Algorithm
building Description
for BLE localization [21]. The research [24] presents a modified PM, called iRingLA, for
BLE localization, in
This section first which the ring
presents is used instead
the overview of circle
of the whole for trilateration.
system. Next, PRMAnother approach
is described called
for distance
stigmergic
estimation,iswhich
presented for range-based
is followed by FP for indoor
locationlocalization
estimation.[25].
Then,This approach
“Outlier can reduce the1”
Detection—Level effects
uses
of multipath, fading, and shadowing for BLE positioning. The work
the information from PRM and FP to remove the outliers from the measurements and generate [26] proposes several empirical
PMs for BLE
improved based indoor
distance localization
measurements for in different
EKF. conditions
Finally, the detailssuch as indoor/outdoor
of EKF and line-of-sight
and “Outlier Detection—Level
(LOS)/non-line-of-sight (NLOS). In the third class, the research [21]
2” are described to further remove outliers and achieve an enhanced indoor localization presents a FP for BLE solution.
localization
in a multi-floor building. The work [27] compares three FP algorithms: k-nearest neighbors (k-NN),
neural networks,
3.1. System Overviewand support vector machine (SVM). The result shows that k-NN is a good candidate
for localization in real-life applications. The research [4] provides a detailed study of the effects of
beaconThedensity,
overview of the proposed
transmit power and algorithm
transmitis frequency
shown in Figure
for BLE 2, which
FP. A mainly consists
FP solution of FP,
using PRM,
Weibull
EKF, and a distribution
probability two-level outlier detection
is proposed mechanism.
in [28] to improve Thetheprocessed
reliabilitydata
and for the proposed
accuracy algorithm
of the positioning.
are RSS values from three advertisement channels of observed BLE beacons.
In addition, the paper [29] presents a comparative analysis of contemporary BLE indoor positioning The first process for
these RSS values is passing a smoother. Next, we originally use channel-separate
solutions, taking into account the classification, comparison and various considerations that are
models/databases
to process RSS values in FP and PRM to generate location and distance estimates. Since BLE RSS
required for designing new indoor positioning approaches. All the above research combines RSS from
values from three advertisement channels are processed separately, there could be three different FP-
all channels to obtain an aggregate signal, instead of using separate model for each separate BLE
derived locations and PRM-derived distances. These derived location and distance estimates are then
advertisement channel.
fed into the “Outlier Detection—Level 1” to generate improved FP + PRM distance estimates using
statistical
3. Algorithmmethods. The EKF is then used to process these enhanced FP + PRM distance estimates to
Description
estimate the target’s location. The second outlier detection based on statistical testing is utilized to
This section first presents the overview of the whole system. Next, PRM is described for distance
further remove the outliers. After the EKF computation with outlier detection, EKF outputs the final
estimation, which is followed by FP for location estimation. Then, “Outlier Detection—Level 1”
BLE localization solution. The final BLE localization solution is also fed to the radio map database to
uses the information from PRM and FP to remove the outliers from the measurements and generate
select appropriate fingerprints to improve the FP performance. Overall, the purpose of the proposed
improved distance measurements for EKF. Finally, the details of EKF and “Outlier Detection—Level 2”
algorithm is to achieve a robust and accurate localization solution for the target.
are described to further remove outliers and achieve an enhanced indoor localization solution.

BLE RSS Radio BLE Localization Solution

(Channel 37, Map
38, and 39) Database

FP+PRM EKF Innovation

FP and Covariance
Positions Distances
Matrix
FP
Outlier Outlier
BLE RSS
RSS Detection Detection
Initial PRM EKF
– –
Process Distances
Level 1 Level 2
PRM Outlier Detection
Result

Figure 2. Overview
Figure 2. Overview of
of the
the smartphone-based
smartphone-based indoor
indoor localization
localization algorithm
algorithm with
with BLE
BLE beacons.
beacons.

3.1. System Overview

The overview of the proposed algorithm is shown in Figure 2, which mainly consists of FP, PRM,
EKF, and a two-level outlier detection mechanism. The processed data for the proposed algorithm
are RSS values from three advertisement channels of observed BLE beacons. The first process for
these RSS values is passing a smoother. Next, we originally use channel-separate models/databases to
process RSS values in FP and PRM to generate location and distance estimates. Since BLE RSS values
from three advertisement channels are processed separately, there could be three different FP-derived
locations and PRM-derived distances. These derived location and distance estimates are then fed into
the “Outlier Detection—Level 1” to generate improved FP + PRM distance estimates using statistical
methods. The EKF is then used to process these enhanced FP + PRM distance estimates to estimate the
target’s location. The second outlier detection based on statistical testing is utilized to further remove
the outliers. After the EKF computation with outlier detection, EKF outputs the final BLE localization
Sensors 2016, 16, 596 5 of 20

solution. The final BLE localization solution is also fed to the radio map database to select appropriate
fingerprints to improve the FP performance. Overall, the purpose of the proposed algorithm is to
achieve a robust and accurate localization solution for the target.

3.2. Polynomial Regression Model

In indoor localization using BLE beacons, the radio PM is usually used to model the relationship
between measured RSS readings and corresponding distances. The mostly used model is the lognormal
shadowing model as follows [30,31]:
ˆ ˙
d
P pdq “ P pd0 q ´ 10γlog10 ` Xσ (1)
d0

where γ represents the path-loss exponent, P pd0 q represents the RSS at the reference distance, d0 ,
P pdq represents the RSS at the distance between the access point (AP) and the receiver, d, and Xσ
represents a Gaussian random variable, with zero mean, caused by shadow fading [31]. After the
model calibration for propagation parameters, the PM model can be used to convert the RSS values to
distances between wireless APs and the receiver.
PM is derived from the outdoor environment and works well for LOS scenarios; however, PM
may be inaccurate for distance estimation indoors since the indoor environment is complicated by
reflection, shadowing and multipath. Actually, indoor wireless signals consist of both LOS and NLOS
signals. These signals may have different propagation parameters. Furthermore, it is difficult to
differentiate the LOS and NLOS signals in indoor environments. Therefore, the fitted theoretical PM
from all training data may not work well for both LOS and NLOS signals and result in large errors in
distance estimation.
To overcome this problem, this paper proposes the PRM to model the relationship between RSS
and distance for BLE beacons. Different from PM, PRM assumes that the RSS-distance relationship is
a nth-degree polynomial and the polynomial coefficients are estimated from the training data in the
calibration process. The PRM is given by
n
ÿ
dˆPRM “ ci ¨ RSSi (2)
i “0

where ci are the coefficients of the nth-degree polynomial, RSS is the RSS value, and dˆPRM is the
estimated distance. The sum of the model fitting error squares is expressed as
˜ ¸2
M ´
ÿ ¯2 M
ÿ n
ÿ
E pc0 , c1 , ¨ ¨ ¨ , cn q “ dˆPRM,j ´ d j “ dj ´ ci ¨ RSSij (3)
j “1 j “1 i “0

where M is the number of the calibration points, d j represents the true distance between the jth
calibration point and the BLE beacon, j “ 1, 2, . . . , M, and RSS j is the RSS at the jth calibration point.
From the criterion of least squares estimation, E pc0 , c1 , ¨ ¨ ¨ , cn q should be minimized. To achieve the
minimum of E pc0 , c1 , ¨ ¨ ¨ , cn q, we equate its partial derivation to zero with respect to c0 , c1 , ¨ ¨ ¨ , cn ,
and get ˜ ¸
M n
BE ÿ
i
ÿ
i
“ 2 RSS j d j ´ ci ¨ RSS j “ 0 (4)
Bci
j “1 i “0

Equation (4) can be rewritten as follows:

AC “ B (5)
Sensors 2016, 16, 596 6 of 20

where » fi
M M M
RSS0j RSS0j RSS0j RSS1j RSS0j RSSnj ffi
ř ř ř
— ¨¨¨
— j “1 j “1 j“1 ffi
— M M M ffi
RSS1j RSS0j RSS1j RSS1j 1 n
— ř ř ř ffi
— ¨¨¨ RSS j RSS j ffi
A“— j“1 j “1 j “1 (6)
— ffi
ffi
— .. .. .. ffi
—
— . . ¨¨¨ . ffi
ffi
— M M M ffi
RSSnj RSS0j RSSnj RSS1j n n
ř ř ř
RSS j RSS j
– fl
¨¨¨
j “1 j“1 j “1
« ffT
M M M
RSS0j d j RSS1j d j RSSnj d j
ř ř ř
B“ ¨¨¨ (7)
j “1 j “1 j “1

and ” ıT
C“ c0 c1 ¨¨¨ cn (8)

Finally, the polynomial coefficients can be calculated by

C “ A ´1 B (9)

After the calibration, PRM can be used to convert RSS values to distances. In this paper, three
PRMs are used to estimate distances for three advertisement channels of each BLE beacon. Therefore,
there can be three distance estimates for each BLE beacon. When using PRM, the polynomial degree, n,
should be carefully determined. If the polynomial degree is small, the distance estimation error will
be large. In contrast, if the polynomial degree is large, the computation load is large. Therefore, the
polynomial degree is the balance of the distance estimation error and computation load. The field
experiments will further discuss the set of the polynomial degree for PRM.

3.3. Fingerprinting
The BLE radio map database for FP was constructed by surveying some points’ locations and
taking corresponding RSS values. Gaussian Process Regression (GPR) was used to fill the gaps between
the surveyed points. We constructed a separate database for each advertisement channel (i.e., the
separate mode). Another database was also constructed by dealing all advertisement channels together
(i.e., the aggregate mode). In the field experiments, these databases will be used for the evaluation of
the FP performance.
On the FP positioning phase, the search space for BLE FP can be limited to an ellipse as determined
by the proposed localization solution at the previous epoch, as shown in Figure 3. This localization
solution at the previous epoch is used as the center of the ellipse. The major semi-axis, minor semi-axis,
and the azimuth of the ellipse are calculated by using the location covariance matrix of the previous
epoch from the EKF, that is « ff
σE2 σEN
P“ 2 (10)
σNE σN
2 and σ2 represent the north and east variances; σ
where σN E NE and σEN are the north/east and east/north
covariances. The major semi-axis of the ellipse is
d c
1 2 1 2 2 q2 ` σ 2
a “ s¨ pσ ` σE2 q ` pσ ´ σN (11)
2 N 4 E NE

The minor semi-axis of the ellipse is

Sensors 2016, 16, 596 7 of 20

d c
1 2 1 2 2 q2 ` σ2
Sensors 2016, 16, 596 b “ s¨ pσ ` σE2 q ´ pσ ´ σN (12)
7 of 20
2 N 4 E NE

And the azimuth of the major semi-axis is

1 ´11  22σ NE 
˜ ¸
1 tan 4  2 NE2 
θ “  tan (13)
2 2 4   2
(13)
 σNN
σE2E ´ 

where tan´411p¨q
where tan () represents
represents aa four-quadrant
four-quadrant arctangent operator. ss represents
arctangent operator. representsthe
thescale
scalefactor
factor for
for the
the
4
size change of the ellipse, which is set at 5 in this paper. Readers can refer to [32] for the details of the
confidence ellipse for measurements.
measurements.

Figure
Figure 3.
3. Ellipse-based
Ellipse-based search
search space
space for
for BLE
BLE FP.
FP.

After the search space is determined, the Euclidean distances are calculated by using current
After the search space is determined, the Euclidean distances are calculated by using current
measured fingerprint and the fingerprints in the search space, and the equation is given by
measured fingerprint and the fingerprints in the search space, and the equation is given by


g N `RSS  RSS 2 ˘
2

fN
 ÿ
EDi f
,j
RSSmm,j i,j
´ RSS i,j (14)
EDi “ e
j 1 N (14)
N
j“1
where EDi is the Euclidean distance, RSSm is the measured RSS vector, RSSi is the RSS vector of
whereth EDi is the Euclidean distance, RSSm is the measured RSS vector, RSSi is the RSS vector of the
ithe i fingerprint
th fingerprint in the in the space
search searchofspace of the
the radio map radio map and
database database and
N is the N is the length of the
length of the measured RSS
measured RSS vector. Then, the target’s location is calculated as the weighted
vector. Then, the target’s location is calculated as the weighted average of the locations average ofofthe
thelocations
selected
of the selected fingerprints, and the weight is determined by the inverse of the Euclidean
fingerprints, and the weight is determined by the inverse of the Euclidean distance. The equation for distance.
The calculation
this equation foristhis calculation
given by is given by
KK
 1 1
rFPrFP“  řKK i i rir,i ,ω
ÿ ω
“
i i
(15)
(15)
i“i 11 j“ 1
j 1

ωj j EDED
i i

where rFP is the estimated location of the target, ri is the location of the ith selected fingerprint, ωi is
where rFP is the estimated location of the target, ri is the location of the i th selected fingerprint,
the weight corresponding to the ith selected fingerprint, EDi is the Euclidean distance between the ith
i is the
selected weight corresponding
fingerprint to the
and the measured i th selected
fingerprint fingerprint,
and K EDi is the
is the total number Euclidean
of selected distance
fingerprints.
th
In this paper,
between the FP
i isselected
used forfingerprint
each advertisement channel, and
and the measured can achieve
fingerprint and three
K islocation estimates
the total numberfor
of
each BLEfingerprints.
selected beacon. In this paper, FP is used for each advertisement channel, and can achieve three
location estimates for each BLE beacon.
3.4. Outlier Detection—Level 1
3.4. Outlier
As perDetection—Level 1
the previous discussion, three PRM-derived distances for each observed BLE beacon and
threeAs
FP-derived locations
per the previous can be imported
discussion, to “Outlier Detection—Level
three PRM-derived distances for each1”observed
to generate
BLEanbeacon
enhanced
and
FP + PRM distance estimate for each observed BLE beacon. In this paper, three FP-derived locations
three FP-derived locations can be imported to “Outlier Detection—Level 1” to generate an enhanced are
FP + PRM distance estimate for each observed BLE beacon. In this paper, three FP-derived locations
are represented by rFP 37 , rFP 38 , and rFP 39 . And, three PRM-derived distances are represented by
dPM 37 ,i , dPM 38 , i , and dPM 39 , i for the i th BLE beacon. In the first step of “Outlier Detection—Level 1”,
FP-derived locations in Equation (15) are used to calculate the FP-derived distances for each observed
BLE beacon through the following equation:
Sensors 2016, 16, 596 8 of 20

represented by rFP37 , rFP38 , and rFP39 . And, three PRM-derived distances are represented by d PM37,i ,
d PM38,i , and d PM39,i for the ith BLE beacon. In the first step of “Outlier Detection—Level 1”, FP-derived
locations in Equation (15) are used to calculate the FP-derived distances for each observed BLE beacon
through the following equation:

dˆFP37,i “ norm rFP37 ´ rBLE,i

` ˘

dˆFP38,i “ norm rFP38 ´ rBLE,i

` ˘
(16)
dˆFP39,i “ norm rFP39 ´ rBLE,i
` ˘

where rBLE,i is the coordinate of the ith BLE beacon. d FP37,i , d FP38,i , and d FP39,i are three FP-derived
distances for the ith BLE beacon. norm is the Euclidean norm (2-norm), and is given by
bÿ
Nx 2
2 ´ norm : k x k 2 “ i“1 |xi | (17)

where Nx is the length of the vector x, and xi is the term of the vector x.
The statistical method is used to remove the outliers in these six distance estimates: three
PRM-derived distance estimates and three FP-derived distances. The confidence interval for these six
distance estimates is given by ” ı
dˆi P µdˆ ´ σdˆ µdˆ ` σdˆ
i i i i
(18)

where dˆi are the distance estimates (i.e., FP-derived and PRM-derived distances) for the ith BLE beacon.
µdˆ and σdˆ are the mean and STD of dˆi . If any distance estimate is out of the confidence interval, it is
i i
moved out as an outlier. Finally, the enhanced FP+PRM distance estimate, dˆi,enhanced , for the ith BLE
beacon is given by ´ ¯
dˆi,enhanced “ mean dˆi,trust (19)

where dˆi,trust are FP-derived and PRM-derived distance estimates in the confidence interval.

3.5. Extended Kalman Filtering

The EKF is used in the proposed algorithm to estimate the target’s current location by fusing
current and historical information. In this paper, the state vector of the EKF is given by
” ıT
x“ re rn ve vn (20)

where re and rn are 2D position components (i.e., east and north) in the horizontal plane. ve and vn are
their corresponding 2D velocity components. In the proposed algorithm, the EKF system model is the
typical kinematic model and is given by

xk`1|k “ Φk,k`1 xk|k ` ωk (21)

where xk`1|k is the predicted state vector, xk|k is the previous state vector at epoch k and Φk,k` 1 is
a 4 ˆ 4 transition matrix: » fi
1 0 ∆t 0
— 0 1 0 ∆t ffi
Φk,k`1 “ — (22)
— ffi
ffi
– 0 0 1 0 fl
0 0 0 1
where ∆t is the time difference between two epochs. ωk is the process noise with the covariance matrix,
Qk “ E ωk ωkT , and is given by
` ˘

” ıT
ωk “ 0 0 ve,k vn,k (23)
Sensors 2016, 16, 596 9 of 20

where ve,k and vn,k are modelled as white noises for velocity in the east and north directions at epoch k.
The measurement model is given by

zk “ Hk xk |k ` υk (24)
” ıT
where zk “ dˆ1,enhanced dˆ2,enhanced ¨ ¨ ¨ dˆm,enhanced uses distance estimates from the “Outlier
Detection—Level 1” as the measurement vector. Hk is the design matrix, and υk is the measurement
noise modelled as a Gaussian white noise, and its covariance matrix is Rk “ E υk υkT . In this paper,
` ˘

we define that re,t and rn,t are 2D position components of the target’s location, which are predicted by
the EKF; re,i and rn,i , i “ 1, 2, . . . , m, are 2D position components of the BLE beacons’ locations. The
design matrix is given by
» re,1 ´re,t rn,1 ´rn,t fi
´b 2 2
´b 2 2
0 0
— pre,1 ´re,t q `prn,1 ´rn,t q pre,1 ´re,t q `prn,1 ´rn,t q ffi
— re,2 ´re,t rn,2 ´rn,t ffi
— ´b ´b 0 0 ffi
— pre,2 ´re,t q2 `prn,2 ´rn,t q2 pre,2 ´re,t q2 `prn,2 ´rn,t q2 ffi
Hk “ — ffi (25)
— .. ffi
.
— ffi
— ffi
– re,m ´re,t rn,m ´rn,t fl
´b ´b 0 0
pre,m ´re,t q2 `prn,m ´rn,t q2 pre,m ´re,t q2 `prn,m ´rn,t q2

There are two phases in the EKF process: prediction and update. In the prediction process, the
state vector and covariance matrix are predicted from the system model:
#
xk`1|k “ Φk,k`1 xk|k
(26)
Pk`1|k “ Φk,k`1 Pk|k Φk,k
T
`1 ` Q k

In the update process, the state vector and covariance matrix are updated from the
measurement model: $ ´ ¯´1
T T
& Kk “ Pk|k´1 Hk Hk P´k|k´1 Hk ` Rk ¯
’
’
xk|k “ xk|k´1 ` Kk zk ´ Hk xk|k´1 (27)
’
’
Pk|k “ pI ´ Kk Hk q Pk|k´1 pI ´ Kk Hk qT ` Rk
%

where Kk is the Kalman gain.

3.6. Outlier Detection—Level 2

In this section, statistical testing on the innovations of the EKF is presented as the second level of
outlier detection. When using EKF, we assume the following two conditions: (1) the measurement
noise is zero-mean, white and Gaussian distributed; and (2) the process noise is zero-mean, white, and
Gaussian distributed. Based on these assumptions, the innovation sequence will be zero-mean, white
and Gaussian distributed. The equation for the innovation sequence can be given as

τk “ zk ´ ẑk|k´1 (28)

where τk is the innovation, zk is the observed measurement, and ẑk|k´1 “ Hk xk|k´ 1 is the predicted
measurement. The innovations have the following covariance matrix [33]:

Cτk “ Hk Pk|k´1 HkT ` Rk (29)

where Cτk is the covariance matrix of the innovation, Hk is the design matrix, Pk|k´1 is the state
covariance matrix and Rk is the measurement covariance matrix. Given the assumptions stated above,
the innovation sequence is distributed as
 
τ k ~ N 0 , Cτk  (30)

where N   ,C  represents the normal distribution with mean of  and covariance of C . The
confidence intervals for the individual measurements are then calculated [33]. If these are violated,
the measurement
Sensors 2016, 16, 596 is considered as an outlier, and removed from the EKF. 10 of 20

4. Field Experiments
` ˘
4.1. Experimental Setup τk „ N 0, Cτk (30)

where ToNevaluate
pµ, Cσ q the performance
represents of the proposed
the normal algorithm,
distribution with meanwe conducted
of µ and experiments
covariance of in C
anσ .office
The
environment (60 m × 40 m). These experiments were conducted with two different
confidence intervals for the individual measurements are then calculated [33]. If these are violated, thedeployments of
BLE beacons as
measurement is shown in Figure
considered as an 4: (a) dense
outlier, and deployment
removed from (total: 20 beacons; average: 1 beacon per 9
the EKF.
m) and (b) sparse deployment (total: 8 beacons; average: 1 beacon per 18 m). All these beacons were
4. Field Experiments
installed on the wall at a height of approximately 1.5 m. To balance the power consumption and
accuracy, each beacon was set to 10 Hz sample rate with −16 dBm transmit power [4]. The locations
4.1. Experimental Setup
of BLE beacons were measured by using a commercial laser rangefinder and the floor plan of the
building. Since the
To evaluate thefloor plan was of
performance verified to be very
the proposed accurate we
algorithm, (which had been
conducted used for engineering
experiments in an office
construction),
environment (60 themlocation
ˆ 40 m).errors ofexperiments
These BLE beaconswere wereconducted
less than 1with
m. Two iPhone 4S with
two different deployments of
iOS
BLE8beacons
were used by different
as shown testers
in Figure 4: in
(a)the experiments.
dense deploymentThe(total:
BLE beacons used in
20 beacons; the experiments
average: 1 beacon perare
based on Texas
9 m) and Instruments
(b) sparse (TI) BLE
deployment chip8CC2540.
(total: beacons;This chip broadcasts
average: 1 beacon perchannel
18 m).information
All these and RSS
beacons
value together on
were installed forthe
each advertisement
wall at a height ofchannel, which can
approximately be To
1.5 m. received
balancebythe
anpower
iOS device when using
consumption and
iOS 7 or above.
accuracy, each beacon was set to 10 Hz sample rate with ´16 dBm transmit power [4]. The locations
of BLEMany trajectories
beacons were testedby
were measured inusing
the experiments
a commercialandlaser
two of them were and
rangefinder selected for performance
the floor plan of the
demonstration
building. Sincein thethis paper.
floor planSeveral reference
was verified to bepoints with surveyed
very accurate (whichlocations
had beenwereusedused to generate
for engineering
the ground truth.
construction), the A stopwatch
location was
errors of used to recordwere
BLE beacons the time when1 the
less than m. testers passed iPhone
Two different these reference
4S with
points and thereafter,
iOS 8 were the locations
used by different testersofinthe
theground truth between
experiments. The BLEthese reference
beacons used points were generated
in the experiments are
through
based oninterpolation.
Texas Instruments The interpolation
(TI) BLE chip method
CC2540.used
This the
chipstopwatch
broadcastsinformation and assumed
channel information the
and RSS
person walked for
value together with a constant
each speed between
advertisement channel,two reference
which can bepoints,
receivedwhich asiOS
by an perdevice
the instructions
when using to
testers
iOS 7 orwhen collecting the data.
above.

(a) ( b)
Figure
Figure 4.
4. Experimental
Experimentalarea
area(red
(redcircles
circles= =
BLE beacons):
BLE (a)(a)
beacons): dense deployment
dense (1 beacon
deployment per per
(1 beacon 9 m); (b)
9 m);
sparse deployment (1 beacon per 18 m).
(b) sparse deployment (1 beacon per 18 m).

Many trajectories were tested in the experiments and two of them were selected for performance
demonstration in this paper. Several reference points with surveyed locations were used to generate
the ground truth. A stopwatch was used to record the time when the testers passed these reference
points and thereafter, the locations of the ground truth between these reference points were generated
through interpolation. The interpolation method used the stopwatch information and assumed the
person walked with a constant speed between two reference points, which as per the instructions to
testers when collecting the data.
Sensors 2016, 16, 596 11 of 20

Sensors 2016, 16, x 11 of 20

4.2. Performance of Polynomial Regression Model for Distance Estimation
The performance of PRM-based distance estimation is discussed in this section. First, we need
The performance
to determine the polynomialof PRM-based
degree, distance estimation
n , for the PRM. Asisperdiscussed in this
the previous section. First,
discussion, we need to
the selection of
determine the polynomial degree, n, for the PRM. As per the previous discussion,
the polynomial degree is a trade-off between distance accuracy and computation load. We collected the selection of the
polynomial degree
RSS values from BLE is beacons
a trade-off betweenpoints
at several distance
withaccuracy and computation
known locations load.
in the dense We collected
deployment RSS
shown
values from BLE beacons at several points with known locations in the dense deployment
in Figure 4. This data was used to estimate the PM and PRM with different polynomial degrees for shown in
Figure 4. This data was used to estimate the PM and PRM with different polynomial
each advertisement channel of the BLE beacons. In the estimation, the selected range for polynomial degrees for each
advertisement
degree is set tochannel
n  1,2,of the ,5 ,BLE
and beacons. In the estimation,
the corresponding the selected
PRM models range
are called for polynomial
POLY1, degree
POLY2, POLY3,
is set to n “ 1, 2, . . . , 5, and the corresponding PRM models are called POLY1, POLY2, POLY3, POLY4,
POLY4, and POLY5. In the dense deployment, there are total 60 advertisement channels (i.e., 3
and POLY5. In the dense deployment, there are total 60 advertisement channels (i.e., 3 channels/beacon
channels/beacon × 20 beacons = 60 channels). Therefore, there are 60 channel models in total for each
ˆ 20 beacons = 60 channels). Therefore, there are 60 channel models in total for each method (i.e., PM,
method (i.e., PM, POLY1, POLY2, POLY3, POLY4 and POLY5). Each channel model has an average
POLY1, POLY2, POLY3, POLY4 and POLY5). Each channel model has an average distance estimation
distance estimation error, therefore there are a total of 60 average distance estimation errors for each
error, therefore there are a total of 60 average distance estimation errors for each method (i.e., PM,
method (i.e., PM, POLY1, POLY2, POLY3, POLY4, and POLY5). The cumulative distribution function
POLY1, POLY2, POLY3, POLY4, and POLY5). The cumulative distribution function (CDF) of these 60
(CDF) of these 60 average distance estimation errors for each method, are plotted in Figure 5. As
average distance estimation errors for each method, are plotted in Figure 5. As shown in Figure 5,
shown in Figure 5, POLY2, POLY3, POLY4 and POLY5 have similar performance and are better than
POLY2, POLY3, POLY4 and POLY5 have similar performance and are better than POLY1 and PM.
POLY1 and PM. On the other hand, POLY2 has the smallest computation load among these four
On the other hand, POLY2 has the smallest computation load among these four methods. Therefore,
methods. Therefore, POLY2 is selected for the PRM and PRM represents POLY2 for the rest of the
POLY2 is selected for the PRM and PRM represents POLY2 for the rest of the paper. From Figure 5, we
paper. From Figure 5, we can find that the 90% distance estimation error using PRM is 3.1 m, which
can find that the 90% distance estimation error using PRM is 3.1 m, which is reduced by 18.42% over
is reduced by 18.42% over the PM (i.e., 3.8 m).
the PM (i.e., 3.8 m).

Figure 5. CDFs
Figure 5. CDFs of
of 60
60 average
average distance
distance estimation errors corresponding
estimation errors to 60
corresponding to 60 channel
channel models
models for
for PM,
PM,
POLY1, POLY2, POLY3, POLY4, and POLY5.
POLY1, POLY2, POLY3, POLY4, and POLY5.

For
For the
the demonstration
demonstrationpurpose,
purpose,one
oneBLE
BLEchannel is used
channel as an
is used example
as an to compare
example the PRM
to compare and
the PRM
PM. The PRM
and PM. and PM
The PRM andfunctions are given
PM functions by by
are given

d
d“ ˆRSS
0.01582
0.01582 RSS22´ 0.03178
0.03178  RSS`0.06341
ˆRSS 0.06341 (31)
(31)
and
and
RSS
RSS “ ´79.5891  10
-79.5891´ 10ˆ1.2954
1.2954 ˆlog
log  d pdq
10 10 (32)
(32)
Figure
Figure 66 shows
shows the
the measured
measured data data and
and distance
distance estimation
estimation results
results using
using PM PM and
and PRM,
PRM, inin which
which
PRM
PRM works better than PM in indoor environments. It seems that the relationship between BLE RSS
works better than PM in indoor environments. It seems that the relationship between BLE RSS
and
and distance
distance does
does not
not obey
obey the
the log
log path-loss
path-loss model
model inin complex
complex indoor
indoor environments
environments due due to
to the
the effect
effect
of
of fading,
fading,reflection,
reflection,multipath,
multipath, etc.etc.
Figure 7 depicts
Figure the CDF
7 depicts functions
the CDF of theofdistance
functions estimation
the distance errors
estimation
errors in the selected channel using PM and PRM. The 90% error of distance estimation using PRM
is 2.6 m, which is only 47.24% of that uses PM (5.5 m).
Sensors 2016, 16, 596 12 of 20

in the selected channel using PM and PRM. The 90% error of distance estimation using PRM is 2.6 m,
Sensors
Sensors 2016, 16,
16, xx47.24% 12
12 of
of 20
which is only
2016, of that uses PM (5.5 m). 20

Figure
Figure 6.
6. Measured
Measured data
data and
and distance
distance estimation
estimation results.
results.
results.

Figure
Figure 7.
Figure 7. CDFs
CDFs of
CDFs of distance
of distance estimation
distance estimation errors
estimation errors for
errors for the
the selected
selected channel
channel by
by using
using PM
PM and
and PRM.
PRM.

Next,
Next, we
we compare
compare thethe performance
the performance of
performance of the
the distance
distance estimation
distance estimation using
using the
the separate
separate PRM
PRM forfor each
each
advertisement channel with that
advertisement channel with that uses uses the
uses the aggregate
the aggregate
aggregate PRM PRM for
PRM for all
for all advertisement
all advertisement channels.
channels. For
advertisement channels. For the
the
illustration purpose,
illustration purpose, one
purpose, one BLE
oneBLE beacon
BLEbeacon is used
beaconisisused
used as
asasan example
ananexample
example for this
forfor comparison.
thisthis
comparison. Figure
Figure
comparison. 8 depicts
8 depicts
Figure CDF
CDF
8 depicts
functions
functions
CDF of
of the
functionstheofdistance
distance estimation
estimation
the distance errors
estimationerrors for
errors the
forfor selected
thethe
selected
selectedBLE
BLEbeacon
BLE beaconusing
beacon usingseparate
using separatePRM
separate PRM andand
aggregate
aggregate PRM.
PRM. TheThe 90%
90% distance
distance estimation
estimation errors
errors of
of three
three separate
separate PRMs
PRMs are
are
The 90% distance estimation errors of three separate PRMs are 5.1 m, 5.2 m and 5.1
5.1 m,
m, 5.2
5.2 m
m and
and 5.0
5.0
m;
5.0 which
m; which are
m; which decreased
areare
decreased by
decreased 19.05%,
byby
19.05%,
19.05%,17.46%
17.46%
17.46% and
and 20.63%
and20.63%
20.63% over
over the
overthe aggregate
theaggregate PRM
aggregatePRM (6.3
PRM(6.3 m)
(6.3m) respectively.
m)respectively.
respectively.
Therefore,
Therefore,our
Therefore, our tests
ourtests indicate
testsindicate that
indicatethat separate
thatseparate
separate PRM
PRMPRM can
can
can provide
provide
provide more
more
more accurate
accurate
accurate distance
distance
distance estimation
thanthan
estimation
estimation than
the
the
the aggregate
aggregate PRM.
PRM. Thus,
Thus, separate
separate PRM
PRM is
is adopted
adopted
aggregate PRM. Thus, separate PRM is adopted in this paper. in
in this
this paper.
paper.
Sensors 2016, 16, 596 13 of 20
Sensors 2016, 16, x 13 of 20
Sensors 2016, 16, x 13 of 20

Figure CDFs
8. 8.
Figure ofofdistance
CDFs distanceestimation
estimationerrors
errors for the selected
for the selectedBLE
BLEbeacon
beaconby
byusing
usingseparate
separate PRM
PRM and
and
Figure 8. CDFs of distance estimation errors for selected BLE beacon by using separate PRM and
aggregate
aggregatePRM.
PRM.
aggregate PRM.

4.3.
4.3.4.3. Performance
Performance
Performance ofofof
ofof ofFingerprinting
Fingerprintingfor
Fingerprinting forLocation
for Location Estimation
Location Estimation
Estimation
This
This
This section
section
section mainlycompares
mainly
mainly comparesthe
compares theFP-based
the FP-based location
FP-based location estimation
location estimation using
estimationusing separate
usingseparate
separate databases
databases
databases forfor
for each
each
each
advertisement
advertisement
advertisementchannelchannel
channelwith with that
withthat uses
thatuses the
uses the aggregate
the aggregate database. In the
database.InInthe
aggregate database. comparison,
thecomparison,
comparison, RSS
RSS
RSS values
values
values were
werewere
collected
collected
collected at
atat
150150points
150 pointsinin
points inthe
theexperimental
the experimental area.
experimental area. These
area. These RSS
TheseRSS values
RSSvalues were
valueswere used
wereused for
usedfor FPFP
for
FP calculation
calculation
calculation by by
by
using
using two
two types of FP databases:
typesofofFPFPdatabases:
databases:thethe separate
theseparate databases
separate databases
databases and and aggregate database. Furthermore, the
using two types and aggregate
aggregatedatabase.
database.Furthermore,
Furthermore, thethe
parameter “k”
parameter “k” of k-NN
k-NN is is temporarily
temporarily set set to
to 1 here toto avoid
avoid the
the effect
effect of
of k-NN for for the comparison
comparison of of
parameter “k” ofofk-NN is temporarily set to11here
here to avoid the effect k-NN
of k-NN the for the comparison
these two kinds of databases. To achieve a better localization performance, k is set to 3 in other parts
of these
thesetwotwokinds
kindsofof databases.
databases. To To
achieve a better
achieve localization
a better performance,
localization k is setktois3 set
performance, in other
to 3 inparts
other
of this
of this paper.
paper. Figure
Figure 99 depicts
depicts CDF
CDF functions
functions of of the
the location
location estimation
estimation errors
errors using
using separate
separate FP FP
parts of this paper. Figure 9 depicts CDF functions of the location estimation errors using separate
databases and aggregate FP database. The FPs using separate databases have better performance than
FPdatabases
databasesand andaggregate
aggregate FPFPdatabase.
database.TheTheFPs FPs
using separate
using databases
separate have better
databases have performance than
better performance
that
that uses
uses aggregate
aggregate database.
database. The 90%
TheThe
90%90% errors
errors of FP using
of FPofusing separate
separate databases
databases are 6.9
are 6.9arem, 7.0
m,6.9
7.0m,m,
m,7.0and 7.0
andm,7.0
than
m; that uses
which areaggregate
reduced database.
by 4.17%, 2.78%, errors
and 2.78% FP using
over thatseparate
uses the databases
aggregate database (7.2 m), and
7.0m;m; which are reduced by 4.17%, 2.78%, and 2.78% over that uses the aggregate database (7.2 m),
which are
respectively. These reduced
These tests by 4.17%,
tests indicate
indicate that2.78%,
that the and
the separate 2.78%
separate strategyover that
strategy can uses
can provide the
provide more aggregate
more accurate database
accurate FP FP result (7.2
thanm),
result than
respectively.
respectively.
the aggregate These
aggregate strategy.tests indicate
strategy. Thus,
Thus, FP that
FP using the separate
using separate strategy
separate databases
databases are can provide
are used
used inin thismore
this paper.
paper.accurate FP result than
the
the aggregate strategy. Thus, FP using separate databases are used in this paper.

Figure
Figure 9.
9. 9.
Figure CDFsofof
CDFs
CDFs oflocation
locationestimation
location estimation errors
estimation errors by
errorsby using
byusing separate
usingseparate FPFP
separate
FP databases andand
databases
databases and aggregate FP database.
aggregate
aggregate FP database.
FP database.

4.4. Performance
Performance Evaluation
Evaluation for
for the
the Proposed
Proposed Algorithm
Algorithm
4.4.4.4.
Performance Evaluation for the Proposed Algorithm
The performance
The performance of the
the proposed
proposed algorithm
algorithm was
was evaluated
evaluated in indoor environments
environments with
with densely
densely
The performance ofofthe proposed algorithm was evaluatedininindoor
indoor environments with densely
and
and sparsely
sparsely distributed BLE
distributedBLE beacons
BLEbeacons shown
beaconsshown
shown inin Figure
in Figure 4. The proposed algorithm was also compared
and sparsely distributed Figure 4.
4. The
Theproposed
proposedalgorithm
algorithmwas
wasalso compared
also compared
with classical PM + EKF and FP + EKF algorithms. Figure 10 shows the two trajectories which were
Sensors 2016,
Sensors 16,16,596
2016, 596 1420
14 of of 20
Sensors 2016, 16, x 14 of 20

with classical PM + EKF and FP + EKF algorithms. Figure 10 shows the two trajectories which were
with
used forclassical PM + EKF
performance and FP +Figure
evaluation. EKF algorithms.
11 depicts Figure 10 shows
the numbers the two trajectories
of observed whichchannels
advertisement were
used for performance evaluation. Figure 11 depicts the numbers of observed advertisement channels
used
from BLEfor performance evaluation. Figure 11 depicts the numbers of observed advertisement channels
from BLEbeacons
beaconsininthese
thesetwo
twotrajectories
trajectories with dense and
with dense andsparse
sparsedeployment
deploymentofofBLEBLE beacons.
beacons. In In
from
Figure BLE
11,11, beacons
there in these two trajectories with dense and sparse deployment of BLE beacons. In
Figure thereare
are5~19
5~19observed
observed advertisement channelsinintwo
advertisement channels twotrajectories
trajectories with
with dense
dense deployment
deployment
Figure 11, there are 5~19 observed advertisement channels in two trajectories with dense deployment
while
while the
thenumber
numberchanges
changestoto0~9
0~9 with
with sparse
sparse deployment.
deployment. The Thedifferent
differentnumber
number ofof measurements
measurements
while the number changes to 0~9 with sparse deployment. The different number of measurements
leads to various localization performance.
leads to various localization performance.
leads to various localization performance.

(a) ( b)
(a) (b)
Figure 10. Two trajectories for performance evaluation: (a) Trajectory I; and (b) Trajectory II.
Figure 10.Two
Figure10. Twotrajectories
trajectoriesfor
for performance evaluation:(a)
performance evaluation: (a)Trajectory
TrajectoryI; I;and
and(b)(b) Trajectory
Trajectory II. II.

(a) ( b)
(a) (b)

(c) ( d)
(c) (d)
Figure 11. Numbers of observed advertisement channels of BLE beacons in two trajectories with dense
Figure
Figure 11.11.Numbers
Numbersofofobserved
observedadvertisement
advertisement channels
channels of
ofBLE
BLE beacons
beaconsinBLE
intwo trajectories
two with
trajectories dense
with dense
and sparse deployment of BLE beacons: (a) Trajectory I with 20 deployed beacons; (b) Trajectory
and sparse deployment of BLE beacons: (a) Trajectory I with 20 deployed BLE beacons; (b) Trajectory
and
II with 20 deployed BLE beacons; (c) Trajectory I with 8 deployed BLE beacons; and (d) Trajectory II II
sparse deployment of BLE beacons: (a) Trajectory I with 20 deployed BLE beacons; (b) Trajectory
II with 20 deployed BLE beacons; (c) Trajectory I with 8 deployed BLE beacons; and (d) Trajectory II
with
with208deployed
deployedBLE
BLE beacons;
beacons. (c) Trajectory I with 8 deployed BLE beacons; and (d) Trajectory II with
with 8 deployed BLE beacons.
8 deployed BLE beacons.
Sensors 2016, 16, 596
596 15 of 20
Sensors 2016, 16, x 15 of 20

The localization performance with dense deployment of BLE beacons is discussed first. Figure
The localization performance
performancewith withdense
dense deploymentofofBLE BLE beacons is discussed first. Figure
12 showslocalization
The the estimated trajectories using thedeployment
proposed algorithm, beacons
PM +is EKFdiscussed
and FPfirst. Figure
+ EKF. 12
The
12 shows
shows the the estimated
estimated trajectories
trajectories using using
the the
proposedproposed algorithm,
algorithm, PM + PM
EKF +
and EKF
FP +and
EKF. FP
The+ EKF.
proposedThe
proposed algorithm performs slightly better than FP + EKF and PM + EKF. Figure 13 shows the
proposed performs
algorithm algorithmslightly
performs betterslightly
than FP better
+ EKFthan FP
PM++EKF
andproposed EKF. and
FigurePM13+showsEKF. Figure 13 shows the
localization errors of the two trajectories using the method, PM + EKFthe localization
and FP + EKF errors
in the
localization
of the two errors of
trajectories the
usingtwo trajectories
the proposed using
method,the proposed
PM + EKF method,
and FP +PMEKF + EKF
in theand FP + EKF inwith
environment the
environment with densely distributed BLE beacons. As shown in Figure 13, the proposed algorithm
environment
densely with densely distributed BLE beacons. As shown in Figure 13, the proposed algorithm
has less distributed
localizationBLE errors beacons.
than PM As+shown
EKF and in Figure
FP + EKF13, at
themost
proposed
of thealgorithm
time. Table has less localization
1 summarizes the
has less
errors localization
than PM + EKFerrors
and than
FP + PM
EKF + EKF
at mostand ofFP
the+ EKF
time. at most
Table 1 of the
summarizestime. Table
the 1 summarizes
localization the
errors
localization errors of the two trajectories using the proposed algorithm, PM + EKF and FP + EKF.
localization
of the two errors of
trajectories the two trajectories
usingfunctions
the proposed using the proposed
algorithm,errors
PM +of EKF algorithm, PM + EKF and FP + EKF.
Figure 14 shows the CDF of localization theand
twoFP + EKF. Figure
trajectories using14these
shows the
three
Figure
CDF 14 shows
functions the CDF functions of localization errors of the two trajectories using these three
approaches. Asofshown
localization
in Figureerrors of the
14 and two1,trajectories
Table using theseerror
the 90% localization threeofapproaches.
the proposed Asalgorithm
shown in
approaches.
Figure 14 andAsTable
shown 1, in Figure
the 90% 14 and Table
localization 1, the
error of 90%
the localization
proposed error of the
algorithm is proposed
2.57 m in algorithmI,
trajectory
is 2.57 m in trajectory I, which is reduced by 36.70% over PM + EKF (4.06 m) and 14.33% over FP +
is 2.57 is
which m reduced
in trajectory I, which is reduced by (4.06
36.70%m)over PM + EKF (4.06 m)EKFand(3.00
14.33%m).over FP +
EKF (3.00 m). Theby 90% 36.70% over
localization PM + EKF
error of the proposedand algorithm
14.33% over FP +
is 2.55 m in trajectory The
II, 90%
which
EKF (3.00 m). Theof90%
localization the localization error of the proposed algorithm II,is 2.55 m in trajectory II, which
reduces by error34.95% over proposed
PM + EKF algorithm
(3.92 m) is 2.55
and m in trajectory
17.21% over FP +which
EKF reduces
(3.08 m).byThese
34.95% over
results
reduces
PM + EKF by 34.95%
(3.92 m) over
and PM
17.21% + EKF
over (3.92
FP + m)
EKF and
(3.08 17.21%
m). These over FP
results + EKF
demonstrate(3.08 m).
that These
the results
proposed
demonstrate that the proposed algorithm achieves around 2.5 m 90% localization error in the two
demonstrate
algorithm that the
achieves proposed
around malgorithm
2.5better90% achieves
localization around
error in the2.5 twom trajectories,
90% localization error in the two
trajectories, which performs than both the traditional PM + EKF and FPwhich + EKFperforms
when there betteris
trajectories,
than both thewhich performs
traditional PM better
+ EKF than
and both
FP + the traditional
EKF when there PM
is + EKFdeployment
dense and FP + EKF of when
BLE there is
beacons.
dense deployment of BLE beacons.
dense deployment of BLE beacons.

(a) ( b)
(a) (b)
Figure 12.
12. Estimatedtrajectories
trajectories with dense deployment
of BLEof BLE beacons
using theusing the algorithm,
proposed
Figure 12.Estimated
Figure Estimated trajectorieswith dense
with deployment
dense deployment ofbeacons
BLE beacons proposed
using the proposed
algorithm,
PM PM
+ EKF, PM + EKF,
and +FP and
+ EKF: FP + EKF: (a) Trajectory I; and (b) Trajectory II.
algorithm, EKF, and(a)
FPTrajectory
+ EKF: (a) I;Trajectory
and (b) Trajectory
I; and (b)II.
Trajectory II.

(a) ( b)
(a) (b)
Figure 13. Localization errors of two indoor trajectories with densely distributed BLE beacons using
Figure 13.
13. Localization
Localization errors
errors of
of two
two indoor
indoor trajectories
trajectories with
with densely distributed
distributed BLE beacons
beacons using
using
Figure
the proposed method, PM + EKF, and FP + EKF: (a) Trajectorydensely
I; and (b) TrajectoryBLE
II.
the proposed method, PM + EKF, and FP + EKF: (a) Trajectory I; and (b) Trajectory
the proposed method, PM + EKF, and FP + EKF: (a) Trajectory I; and (b) Trajectory II. II.
Sensors 2016, 16, 596 16 of 20
Sensors 2016, 16, x 16 of 20

(a) (b)
Figure 14. CDFs
Figure 14. CDFs of
of localization
localization errors
errors of the two
of the two indoor
indoor trajectories
trajectories with
with densely
densely distributed
distributed BLE
BLE
beacons using the proposed method, PM+EKF, and FP + EKF: (a) Trajectory I; and (b) Trajectory
beacons using the proposed method, PM+EKF, and FP + EKF: (a) Trajectory I; and (b) Trajectory II. II.

Table 1. Summary of localization errors of two trajectories in the case of dense distribution of BLE
Table 1. Summary of localization errors of two trajectories in the case of dense distribution of BLE
beacons (m).
beacons (m).
Trajectory Algorithm 50% 90% Mean RMS
Trajectory Algorithm 50% 90%
PM + EKF 2.44 4.06 2.57Mean 2.81 RMS

I PM + EKF 2.44
FP + EKF 1.48 4.06
3.00 1.67 2.57 1.91 2.81
FP + EKF 1.48 3.00
I Proposed 1.46 2.57 1.59 1.67 1.74 1.91
Proposed 1.46 2.57 1.59 1.74
PM + EKF 2.49 3.92 2.59 2.76
PM + EKF 2.49 3.92 2.59 2.76
II FP + EKF 1.89 3.08 1.96 2.12
II FP + EKF 1.89 3.08 1.96 2.12
Proposed
Proposed 1.72 1.72 2.55
2.55 1.72 1.72 1.84 1.84

Then, we evaluated the localization performance with the sparse deployment of BLE beacons.
Then, we evaluated the localization performance with the sparse deployment of BLE beacons.
Figure 15 shows the estimated trajectories with sparsely distributed BLE beacons using the proposed
Figure 15 shows the estimated trajectories with sparsely distributed BLE beacons using the proposed
algorithm, PM + EKF, and FP + EKF. The proposed algorithm performs better than FP + EKF and PM
algorithm, PM + EKF, and FP + EKF. The proposed algorithm performs better than FP + EKF and
+ EKF and the improvement is more than the case of dense deployment. Figure 16 shows the
PM + EKF and the improvement is more than the case of dense deployment. Figure 16 shows the
localization errors of the two trajectories using the proposed method, PM + EKF, and FP + EKF in the
localization errors of the two trajectories using the proposed method, PM + EKF, and FP + EKF in
environment with the sparse distribution of BLE beacons. Table 2 summarizes the location errors of
the environment with the sparse distribution of BLE beacons. Table 2 summarizes the location errors
the two trajectories in this case. Figure 17 shows the CDF functions of localization errors of the two
of the two trajectories in this case. Figure 17 shows the CDF functions of localization errors of the
trajectories using the proposed method, PM + EKF, and FP + EKF. In Figure 17 and Table 2, the 90%
two trajectories using the proposed method, PM + EKF, and FP + EKF. In Figure 17 and Table 2, the
localization error of the proposed algorithm is 4.16 m in trajectory I, which is reduced by 37.72% over
90% localization error of the proposed algorithm is 4.16 m in trajectory I, which is reduced by 37.72%
PM + EKF (6.68 m) and 22.24% over FP + EKF (5.35 m). Figure 17 and Table 2 also show that the 90%
over PM + EKF (6.68 m) and 22.24% over FP + EKF (5.35 m). Figure 17 and Table 2 also show that
localization error of the proposed algorithm is 3.59 m in trajectory II, which is decreased by 61.44%
the 90% localization error of the proposed algorithm is 3.59 m in trajectory II, which is decreased by
over PM + EKF (9.31 m) and 20.58% over FP + EKF (4.52 m). These results demonstrate that the
61.44% over PM + EKF (9.31 m) and 20.58% over FP + EKF (4.52 m). These results demonstrate that
proposed algorithm achieves around 4.0 m 90% localization error in the two trajectories with the
the proposed algorithm achieves around 4.0 m 90% localization error in the two trajectories with the
sparse deployment of BLE beacons, which performs much better than the traditional PM + EKF and
sparse deployment of BLE beacons, which performs much better than the traditional PM + EKF and
FP + EKF. The improvement of the proposed method (37.72% and 61.44% for PM + EKF, and 22.24%
FP + EKF. The improvement of the proposed method (37.72% and 61.44% for PM + EKF, and 22.24%
and 20.58% for FP + EKF) in two trajectories with the sparse deployment is more than that (36.70%
and 20.58% for FP + EKF) in two trajectories with the sparse deployment is more than that (36.70%
and 34.95% for PM + EKF, and 14.33% and 17.21% for FP + EKF) with the dense deployment. This
and 34.95% for PM + EKF, and 14.33% and 17.21% for FP + EKF) with the dense deployment. This
outcome means that the proposed algorithm is especially useful to improve the localization accuracy
outcome means that the proposed algorithm is especially useful to improve the localization accuracy
in the environments with sparse beacon deployment.
in the environments with sparse beacon deployment.
Sensors 2016, 16, 596 17 of 20
Sensors 2016, 16, 596 17 of 20
Sensors 2016, 16, x 17 of 20
Sensors 2016, 16, x 17 of 20

(a) ( b)
(a) (b)
Figure 15. Estimated trajectories with sparse deployment of BLE beacons(b)
(a) using the proposed
Figure Estimated
15.15.
Figure trajectories
Estimated with sparse
trajectories deployment of BLE ofbeacons using the proposed algorithm,
algorithm,
Figure PM + EKF and FP + EKF: with
(a) sparse
Trajectory deployment BLE beacons
I; and (b) Trajectory
of BLE II.
using the proposed
+ EKF15.
PMalgorithm,
and Estimated
FP + EKF:trajectories
(a) with
Trajectory I; sparse
and (b) deployment
Trajectory II.
PM + EKF and FP + EKF: (a) Trajectory I; and (b) Trajectory II.
beacons using the proposed
algorithm, PM + EKF and FP + EKF: (a) Trajectory I; and (b) Trajectory II.

(a) ( b)
(a) (b)
(a) (b) BLE beacons using
Figure 16. Localization errors of two indoor trajectories with sparsely distributed
Figure
the 16. Localization
proposed method, PM errors
+ EKFof and
two FP
indoor trajectories
+ EKF: with I;
(a) Trajectory sparsely
and (b)distributed
Trajectory BLE beacons using
II.
Figure Localization
16.16.
Figure errors
Localization errorsofoftwo
twoindoor
indoor trajectories
trajectories withsparsely
with sparsely distributed
distributed BLEBLE beacons
beacons using
using
the proposed method, PM + EKF and FP + EKF: (a) Trajectory I; and (b) Trajectory II.
thethe
proposed
proposed method,
method,PMPM++EKF
EKFand andFP
FP++ EKF: (a) Trajectory
TrajectoryI;I;and
and(b)
(b)Trajectory
TrajectoryII.II.

(a) ( b)
(a) (b)
(a) (b)
Figure 17. CDFs of localization errors of the two indoor trajectories with sparsely distributed BLE
Figure 17.
beacons CDFs
using the of localization
proposed errors
method, PMof the two
andindoor trajectories with sparsely
I; and (b)distributed BLE
Figure 17. CDFs of localization errors of+ EKF
the two FP + EKF:
indoor (a) Trajectory
trajectories with sparsely Trajectory II.
distributed BLE
Figure CDFsthe
17. using
beacons of proposed
localization errorsPM
method, of +the
EKFtwo indoor
and trajectories
FP + EKF: with sparsely
(a) Trajectory I; and (b) distributed
Trajectory II.BLE
beacons using the proposed method, PM + EKF and FP + EKF: (a) Trajectory I; and (b) Trajectory II.
beacons using the proposed method, PM + EKF and FP + EKF: (a) Trajectory I; and (b) Trajectory II.
Sensors 2016, 16, 596 18 of 20

Table 2. Summary of localization errors of the two trajectories in the case of sparse distribution of BLE
beacons (m).

Trajectory Algorithm 50% 90% Mean RMS

PM + EKF 3.72 6.68 3.93 4.46
I FP + EKF 2.47 5.35 2.83 3.40
Proposed 1.70 4.16 2.07 2.44
PM + EKF 4.60 9.31 5.59 6.20
II FP + EKF 1.80 4.52 2.27 2.75
Proposed 1.63 3.59 1.89 2.27

Table 3 compares the localization errors of the proposed method in two cases: (a) dense
deployment and (b) sparse deployment. It shows that the 90% localization errors with dense
deployment are 2.57 m and 2.55 m in two trajectories, which are less than the 4.16 m and 3.59 m
of the sparse deployment. The 90% localization errors with sparse deployment are increased by 61.87%
and 40.78% over that with dense deployment in the two trajectories, respectively. This result illustrates
that the proposed method in the dense deployment performs better than the sparse deployment.
However, the dense deployment has more expense. Therefore, it is suggested that the deployment
of BLE beacons should comprehensively consider both requirement of the localization accuracy and
budget in specific indoor positioning and navigation applications.

Table 3. Summary of localization errors of the two trajectories using the proposed algorithm with both
dense and sparse distribution of BLE beacons (m).

Trajectory Algorithm 50% 90% Mean RMS

Dense Distribution 1.46 2.57 1.59 1.74
I
Sparse Distribution 1.70 4.16 2.07 2.44
Desnse Distribution 1.72 2.55 1.72 1.84
II
Sparse Distribution 1.63 3.59 1.89 2.27

5. Conclusions
This paper proposed an innovative algorithm based on the integration of channel-separate
polynomial regression model (PRM), channel-separate fingerprinting (FP), extended Kalman filtering
(EKF), and outlier detection for indoor localization using Bluetooth Low Energy (BLE) beacons. Field
experiments showed that the proposed PRM for distance estimation achieved the accuracies of <3.1 m
at 90% of the time, which was reduced by 18.42% over the 3.8 m of traditional propagation model
(PM). Also, the proposed algorithm provided the accuracies of <2.56 m at 90% of the time (average of
two trajectories) with dense deployment of BLE beacons (1 beacon per 9 m), which performed better
than <3.99 m of the classical PM + EKF algorithm and <3.04 m of the classical FP + EKF algorithm.
With sparse deployment of BLE beacons (1 beacon per 18 m), the proposed algorithm achieved the
accuracies of <3.88 m at 90% of the time (average of two trajectories), which performed better than
<8.00 m of the classical PM + EKF algorithm and <4.94 m of the classical FP + EKF algorithm. The
improvement of the proposed algorithm over the classical PM + EKF and FP + EKF with sparse
deployment of BLE beacons was more significant than the case of dense deployment. This proposed
algorithm can be easily implemented without extra hardware costs and promote the development of
robust smartphone-based indoor localization systems with BLE beacons.

Acknowledgments: This work was supported by National Natural Science Foundation of China under
Grant 61474022. Publishing costs are covered by National Natural Science Foundation of China under
Grant 61474022.
Sensors 2016, 16, 596 19 of 20

Author Contributions: All authors contributed considerably to this article. The article was conceived and
structured by all the authors. Yuan Zhuang, Jun Yang, and Naser El-Sheimy conceived and designed the
experiments, Yuan Zhuang, You Li, and Longning Qi performed the experiments, analyzed the data, and wrote
the paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Zhuang, Y.; Syed, Z.; Georgy, J.; El-Sheimy, N. Autonomous smartphone-based WiFi positioning system by
using access points localization and crowdsourcing. Pervasive Mob. Comput. 2015, 18, 118–136. [CrossRef]
2. Kim, S.J.; Kim, B.K. Dynamic ultrasonic hybrid localization system for indoor mobile robots. IEEE Trans.
Ind. Electron. 2013, 60, 4562–4573. [CrossRef]
3. De Angelis, A.; Dwivedi, S.; Handel, P. Characterization of a flexible UWB sensor for indoor localization.
IEEE Trans. Instrum. Meas. 2013, 62, 905–913. [CrossRef]
4. Faragher, R.; Harle, R. Location fingerprinting with bluetooth low energy beacons. IEEE J. Sel. Areas Commun.
2015, 33, 2418–2428. [CrossRef]
5. Zhuang, Y.; El-Sheimy, N. Tightly-coupled integration of WiFi and mems sensors on handheld devices for
indoor pedestrian navigation. IEEE Sens. J. 2016, 16, 224–234. [CrossRef]
6. Colombo, A.; Fontanelli, D.; Macii, D.; Palopoli, L. Flexible indoor localization and tracking based on a
wearable platform and sensor data fusion. IEEE Trans. Instrum. Meas. 2014, 63, 864–876. [CrossRef]
7. Li, Y.; Georgy, J.; Niu, X.; Li, Q.; El-Sheimy, N. Autonomous calibration of MEMS Gyros in consumer portable
devices. IEEE Sens. J. 2015, 15, 4062–4072. [CrossRef]
8. Dabove, P.; Ghinamo, G.; Lingua, A.M. Inertial sensors for smartphones navigation. SpringerPlus 2015, 4,
1–18. [CrossRef] [PubMed]
9. Piras, M.; Lingua, A.; Dabove, P.; Aicardi, I. Indoor navigation using smartphone technology: A future
challenge or an actual possibility? In Proceedings of the IEEE/ION Position, Location and Navigation
Symposium, PLANS 2014, Monterey, CA, USA, 5–8 May 2014; pp. 1343–1352.
10. Wang, J.; Hu, A.; Li, X.; Wang, Y. An improved PDR/magnetometer/floor map integration algorithm for
ubiquitous positioning using the adaptive unscented kalman filter. ISPRS Int. J. Geo-Inf. 2015, 4, 2638–2659.
[CrossRef]
11. Lan, H.; Yu, C.; Zhuang, Y.; Li, Y.; El-Sheimy, N. A novel kalman filter with state constraint approach for the
integration of multiple pedestrian navigation systems. Micromachines 2015, 6, 926–952. [CrossRef]
12. Zhuang, Y.; Lan, H.; Li, Y.; El-Sheimy, N. PDR/INS/WiFi integration based on handheld devices for indoor
pedestrian navigation. Micromachines 2015, 6, 793–812. [CrossRef]
13. Jongbae, K.; Heesung, J. Vision-based location positioning using augmented reality for indoor navigation.
IEEE Trans. Consum. Electron. 2008, 54, 954–962.
14. Dae Hee, W.; Eunsung, L.; Moonbeom, H.; Seung-Woo, L.; Jiyun, L.; Jeongrae, K.; Sangkyung, S.; Young Jae, L.
Selective integration of GNSS, vision sensor, and INS using weighted DOP under GNSS-challenged
environments. EEE Trans. Instrum. Meas. 2014, 63, 2288–2298.
15. Ghinamo, G.; Corbi, C.; Lovisolo, P.; Lingua, A.; Aicardi, I.; Grasso, N. Accurate positioning and orientation
estimation in urban environment based on 3D models. In Proceedings of the New Trends in Image Analysis
and Processing—ICIAP 2015 Workshops, Genoa, Italy, 7–8 September 2015; Springer: New York, NY, USA;
pp. 185–192.
16. Aicardi, I.; Dabove, P.; Lingua, A.; Piras, M. Sensors integration for smartphone navigation: Performances
and future challenges. Int. Arch. Photogram. Remote Sens. Spat. Inf. Sci. 2014, 40, 9–16. [CrossRef]
17. Ghinamo, G.; Corbi, C.; Francini, G.; Lepsoy, S.; Lovisolo, P.; Lingua, A.; Aicardi, I. The MPEG7 visual search
solution for image recognition based positioning using 3D models. In Proceedings of the 27th International
Technical Meeting of the Satellite Division of the Institute of Navigation, Tampa, FL, USA, 8–12 September
2014; pp. 8–12.
18. Zhuang, Y.; Shen, Z.; Syed, Z.; Georgy, J.; Syed, H.; El-Sheimy, N. Autonomous wlan heading and position
for smartphones. In Proceedings of the 2014 IEEE/ION Position, Location and Navigation Symposium,
PLANS 2014, Monterey, CA, USA, 5–8 May 2014; pp. 1113–1121.
19. Li, Y.; Zhuang, Y.; Lan, H.; Zhou, Q.; Niu, X.; El-Sheimy, N. A hybrid WiFi/magnetic matching/PDR
approach for indoor navigation with smartphone sensors. IEEE Commun. Lett. 2016, 20, 169–172. [CrossRef]
Sensors 2016, 16, 596 20 of 20

20. Hui, L.; Darabi, H.; Banerjee, P.; Jing, L. Survey of wireless indoor positioning techniques and systems.
IEEE Trans. Syst. Man Cybern. C Appl. Rev. 2007, 37, 1067–1080.
21. Lohan, E.S.; Talvitie, J.; Figueiredo e Silva, P.; Nurminen, H.; Ali-Loytty, S.; Piche, R. Received signal
strength models for wlan and ble-based indoor positioning in multi-floor buildings. In Proceedings of the
International Conference on Localization and GNSS (ICL-GNSS), Gothenburg, Sweden, 22–24 June 2015;
pp. 1–6.
22. Feng, Y.; Yuxin, Z.; Gunnarsson, F. Proximity report triggering threshold optimization for network-based
indoor positioning. In Proceedings of the 18th International Conference on Information Fusion (Fusion),
Washington, DC, USA, 6–9 July 2015; pp. 1061–1069.
23. Zhao, Y.; Feng, Y.; Gunnarsson, F.; Amirijoo, M.; Ozkan, E.; Gustafsson, F. Particle filtering for positioning
based on proximity reports. In Proceedings of the 18th International Conference on Information Fusion
(Fusion), Washington, DC, USA, 6–9 July 2015; pp. 1046–1052.
24. Thaljaoui, A.; Val, T.; Nasri, N.; Brulin, D. BLE localization using RSSI measurements and iRingLA. In
Proceedings of the IEEE International Conference on Industrial Technology (ICIT), Seville, Spain, 17–19
March 2015; pp. 2178–2183.
25. Palumbo, F.; Barsocchi, P.; Chessa, S.; Augusto, J.C. A stigmergic approach to indoor localization using
bluetooth low energy beacons. In Proceedings of the 12th IEEE International Conference on Advanced Video
and Signal Based Surveillance (AVSS), Karlsruhe, Germany, 25–28 August 2015; pp. 1–6.
26. Zhao, X.; Xiao, Z.; Markham, A.; Trigoni, N.; Ren, Y. Does BTLE measure up against WifI? A comparison
of indoor location performance. In Proceedings of the 20th European Wireless Conference on European
Wireless, Barcelona, Spain, 14–16 May 2014; pp. 1–6.
27. Zhang, L.; Liu, X.; Song, J.; Gurrin, C.; Zhu, Z. A comprehensive study of bluetooth fingerprinting-based
algorithms for localization. In Proceedings of the 27th International Conference on Advanced Information
Networking and Applications Workshops, Barcelona, Spain, 25–28 March 2013; pp. 300–305.
28. Pei, L.; Chen, R.; Liu, J.; Kuusniemi, H.; Tenhunen, T.; Chen, Y. Using inquiry-based bluetooth RSSI probability
distributions for indoor positioning. J. Glob. Position. Syst. 2010, 9, 122–130.
29. Cabarkapa, D.; Gruji, I.; Pavlovi, P. Comparative analysis of the bluetooth low-energy indoor positioning
systems. In Proceedings of the 12th International Conference on Telecommunication in Modern Satellite,
Cable and Broadcasting Services (TELSIKS), Nis, Serbia, 14–17 October 2015; pp. 76–79.
30. Wang, B.; Zhou, S.; Liu, W.; Mo, Y. Indoor localization based on curve fitting and location search using
received signal strength. IEEE Trans. Ind. Electron. 2015, 62, 572–582. [CrossRef]
31. Mazuelas, S.; Bahillo, A.; Lorenzo, R.M.; Fernandez, P.; Lago, F.A.; Garcia, E.; Blas, J.; Abril, E.J. Robust indoor
positioning provided by real-time RSSI values in unmodified wlan networks. IEEE J. Sel. Top. Signal Process.
2009, 3, 821–831. [CrossRef]
32. Kuang, S. Geodetic Network Analysis and optimal Design: Concepts and Applications; Ann Arbor PressInc.:
Chelsea, MI, USA, 1996.
33. Teunissen, P.J.G. Quality control in integrated navigation systems. In Proceedings of the IEEE PLANS ’90,
Position Location and Navigation Symposium Record. The 1990’s—A Decade of Excellence in the Navigation
Sciences, Las Vegas, NV, USA, 20–23 March 1990; pp. 158–165.

© 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).