sensors

Article
BLE Fingerprint Indoor Localization Algorithm Based
on Eight-Neighborhood Template Matching
Mingfeng Li 1, *, Lichen Zhao 1 , Ding Tan 2 and Xiaozhe Tong 1
1 School of Geomatics Science and Technology, Nanjing Tech University, Nanjing 211800, China;
Jennnings@njtech.edu.cn (L.Z.); xztong0810@163.com (X.T.)
2 School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, China; tanding-87@163.com
* Correspondence: njuter@163.com




Received: 19 September 2019; Accepted: 5 November 2019; Published: 7 November 2019


Abstract: Aiming at the problem of indoor environment, signal non-line-of-sight propagation and
other factors affect the accuracy of indoor locating, an algorithm of indoor fingerprint localization
based on the eight-neighborhood template is proposed. Based on the analysis of the signal strength of
adjacent reference points in the fingerprint database, the methods for the eight-neighborhood template
matching and generation were studied. In this study, the indoor environment was divided into four
quadrants for each access point and the expected values of the received signal strength indication
(RSSI) difference between the center points and their eight-neighborhoods in different quadrants were
chosen as the generation parameters. Then different templates were generated for different access
points, and the unknown point was located by the Euclidean distance for the correlation of RSSI
between each template and its coverage area in the fingerprint database. With the spatial correlation
of fingerprint data taken into account, the influence of abnormal fingerprint on locating accuracy is
reduced. The experimental results show that the locating error is 1.0 m, which is about 0.2 m less
than both K-nearest neighbor (KNN) and weighted K-nearest neighbor (WKNN) algorithms.

Keywords: indoor localization; fingerprint localization; eight-neighborhood; template matching;

bluetooth low energy

1. Introduction
Indoor localization has become more and more significant in location based service (LBS) [1,2].
Bluetooth [3–6], WIFI [7,8] and ultra-wide band (UWB) [9–11] are widely used in indoor localization
systems based on received signal strength indication (RSSI) from the known access points (APs).
Bluetooth low energy (BLE) has attracted increasing interests for its low-cost, low-power consumption
and ubiquitous availability in mobile devices [12]. Similar with other wireless technologies, the BLE
indoor localization based on RSSI can be divided into the trilateration algorithm and the method of
location fingerprint [13]. In trilateration, the distances between the unknown points and APs are
measured with acquired RSSI based on the signal propagation mode. Then the fingerprint localization
can be divided into the offline phase and the online phase. During the offline phase, the RSSI
distributions of different APs are determined by measuring the RSSI at different reference points (RPs),
and a fingerprint database is generated. During the online phase, the unknown points are located
with the best RPs, which are chosen based on the fingerprint database and the acquired RSSI from
different APs. Due to the complex indoor environment, signal non-of-sight propagation and other
factors, the RSSI does not always follow the signal propagation mode [14]. As a result, the accuracy
of distance between the unknown point and the AP may not satisfy the requirement. In contrast,
the fingerprint localization to distinguish the positions by the RSSI from different APs may be less

Sensors 2019, 19, 4859; doi:10.3390/s19224859 www.mdpi.com/journal/sensors

Sensors 2019, 19, 4859 2 of 13

affected by the environment. Therefore, the fingerprint localization may be more accurate than the
trilateration algorithm, especially in a complex environment [15].
The best RPs which are the most important for the fingerprint localization, are usually determined
with the K-nearest neighbor (KNN) algorithm, and the weighted K-nearest neighbor (WKNN) algorithm
or their improved algorithms. Bahl et al. determined the best RPs by the KNN algorithm and located
the unknown points with the error of 2–3 m [16]. Kukolj et al. matched the position fingerprint
through signal simplification and principal component analysis to improve the WKNN algorithm [17].
Li Hualiang et al. applied the method of kernel function extraction to train the original fingerprint
data, and matched the fingerprint data in the database with the error of 1.5 m based on an improved
WKNN algorithm [18]. Bi Jingxue et al. considered the influence of the human body on the signal
propagation and established an omnidirectional fingerprint database to locate the unknown points
in WKNN algorithm [19]. The fingerprint localization requires an established RSSI database to store
the coordinates and RSSI features from different APs of the RPs. For the low efficiency of the RSSI
database establishment, Liu Huiyuan et al. proposed a new method based on an improved Kriging
interpolation algorithm, and the experiments showed that the locating accuracy was kept and the
workload was reduced by 50% [20].
In the methods mentioned above, k RPs with the most similar RSSI features to the unknown points
are chosen. Because of different acquisition environments, signal strength is easily biased during RSSI
acquisition [21], and it may happen both in the offline phase and the online phase which usually leads
to unsuitable RPs. It is worth mentioning that if there are abnormal RSSI values in the original data, it is
easy to influence the reliability of the peripheral interpolation points. Most fingerprint techniques are
not satisfactorily performed due to the mismatch in offline training and online measurement caused
by these abnormal RSSI values. To avoid the abnormal RSSI’s influence on the locating accuracy,
we propose a method for fingerprint indoor localization based on eight-neighborhood template
matching (ENTM), which generates eight-neighborhood templates with the RSSI values acquired at
the unknown point, and the template matching is applied to choose the best RP as the estimation
of location.
This paper is organized as follows: in Section 2, the algorithms of KNN and WKNN and the
problems in application are introduced; the relationship of RSSI in fingerprint database is analyzed
and the proposed ENTM algorithm is introduced in Section 3. In Section 4, the experiment and the
analyzing results are presented; finally, the conclusions are presented in Section 5.

2. Related Works

2.1. Algorithms of KNN and WKNN

The KNN algorithm is one of the usual methods in data classification technology. First, Euclidean
distances of RSSI vectors at the unknown point and the i-th RP in the fingerprint database are
calculated as: v
u
M 
u
tX
   j 2
Di = Pr,dB d j − Pr,dB di , i = 1, 2, · · · , N, (1)
j=1

where M is the number of APs in the indoor environment and N is the number of RPs in the fingerprint
 j
database. Pr,dB (d j ) is the j-th AP’s RSSI acquired at the unknown points, and Pr,dB di is the RSSI of
the same AP at the i-th RP stored in the fingerprint database. After calculating N distances, k RPs
which has the smallest distances are selected to estimate the coordinates of the unknown point as:

k
1X
(x, y) = ( xi , yi ) , (2)
k
i=1

where (x, y) are the estimated coordinates and (xi , yi ) are the coordinates of the i-th RP.
Sensors
Sensors 2019,
2019, 19,
19, x4859
FOR PEER REVIEW 33 of
of 13
13

WKNN is an improved algorithm of KNN and also selects k RPs to estimate the location of the
WKNN
unknown is an
point. improved
The impact ofalgorithm
differentofselected
KNN and RPsalso
whichselects have k RPs to estimate
different the location
distances from theof the
RSSI
unknown point. The impact of different selected RPs which have different
vector at the unknown point, and the smaller distance indicates that the RP is closer to the unknowndistances from the RSSI
vectorthan
point at the
theunknown point, and the
others theoretically. To smaller
express distance indicates
the relevance, that theweighting
a different RP is closer to the unknown
coefficient is used
point than the others theoretically. To express the relevance, a different
for RPs in WKNN instead of the middle point of the selected RPs. The estimated coordinates weighting coefficient is used
are
for RPs in WKNN
calculated as: instead of the middle point of the selected RPs. The estimated coordinates are
calculated as:
kk
(x,x, yy) =  wwi i (xxii,, yyii),,
X
(3)
(3)
i 1
i=1

where wwi iisisthe

theweighting
weightingcoefficient
coefficientofofthe
thei-thi -RP, RP, which
th which candetermined
can be be determined
by by

11/D
/ Dii ,
wii =
w  P
kk
, (4)

 1 / Dii 
( 1/D ) (4)
ii=
1
1

where D is the distance between the RSSI vectors of the unknown point and the i-th RP.
where Di i is the distance between the RSSI vectors of the unknown point and the i - th RP.
2.2. Problems in Application
2.2. Problems in Application
The key basis for the algorithm to select the RPs is the similarity (e.g., Euclidean distance)
The
between the keyRSSI
basisvectors
for theofalgorithm
RPs and the to unknown
select the point,
RPs isso the
thesimilarity
fingerprint(e.g., Euclidean
database distance)
is vital for the
between the RSSI vectors of RPs and the unknown point, so the fingerprint
accuracy of the localization. As shown in Figure 1, the chosen RPs should cluster around the unknown database is vital for the
accuracy of the it
point. However, localization.
is shown that Astheshown
signalinpropagation
Figure 1, the chosen
suffers fromRPs should
several noisycluster around like
characteristics the
unknown point. However, it is shown that the signal propagation suffers
temperature, the location of furniture, the multipath effect and even, the human body [22]. To solve from several noisy
characteristics
the problem, there like temperature,
are many outlier the location
suppression of furniture,
methodsthe multipath
in indoor effect and
localization sucheven, the median
as the human
body
filter, [22]. To solvefilter
the Kalman the problem,
and RPCA there are many
(Robust outlier
Principle suppression
Component methods[23].
Analysis) in indoor
In these localization
methods,
such
the single point reception RSSI value of the reference point is only taken into account, but there is[23].
as the median filter, the Kalman filter and RPCA (Robust Principle Component Analysis) still
In these methods,
correlation between thethesingle
featurepoint
RSSIreception
values ofRSSI value ofRPs.
the adjacent the reference
RPs which point
haveisabnormal
only taken into
feature
account,
RSSI values but may
therestill
is still correlation
exist between database.
in the fingerprint the featureFor RSSIthevalues of theof
algorithms adjacent
KNN and RPs.WKNNRPs which
that
have abnormal feature RSSI values may still exist in the fingerprint database.
select RPs by similarity, the abnormal RPs may affect the locating accuracy. It may be divided into two For the algorithms of
KNN and WKNN
categories: some of that
theselect RPsRPs
selected by similarity, the abnormal
are more similar RPs may affect
to the unknown pointthe
butlocating
further accuracy.
from it and It
may be divided into two categories: some of the selected RPs are more similar
all the selected RPs do not gather around the unknown point. The cluster generated by the selected to the unknown point
but
RPsfurther fromfrom
is unstable it andtheall the selected
perspective of RPs
clusterdo validation
not gather[24].around
As istheshown
unknownin thepoint. The cluster
following figure,
generated by the selected RPs is unstable from the perspective of cluster validation
the number beside the selected RP represents the similarity order between the RSSI vectors of the RP [24]. As is shown
in
andthethe following
unknownfigure,
point.the number beside the selected RP represents the similarity order between
the RSSI vectors of the RP and the unknown point.

(a) (b)
Figure1.1.The
Figure The normal
normal distribution of reference
reference points
points (RPs),
(RPs),where
whereselected
selectedRPs
RPsgather
gatheraround
aroundthe the
unknown
unknown point
point in (a)
in (a) andand none
none of the
of the selected
selected RPsRPs is significantly
is significantly awayaway from
from thethe cluster
cluster in (b).
in (b).
Sensors 2019, 19, 4859 4 of 13

Sensors2019,
Sensors 2019,19,
19,xxFOR
FORPEER
PEERREVIEW
REVIEW 44ofof13
13
In Figure 2a, Point 2 and Point 3 are more similar with the unknown Point A in the RSSI vector,
InInFigure
Figure
but further from 2a,Point
2a, PointA.
Point 22and
and
In Point2b,
Point
Figure 33are
are more
themore similar
similar
chosen with
RPs with
do the
the
not unknown
unknown
cluster Point
Point
around AAin
Point in the
B,the RSSI
andRSSI vector,
vector,
the middle
but
but further
further
point from
of thefrom Point
Point
chosen A. In
RPsA.isIn Figure
farFigure 2b, the chosen RPs do not cluster around Point B, and the
2b, theB.chosen RPs do not cluster around Point B, and the middle
from Point middle
pointof
point ofthe
thechosen
chosenRPsRPsisisfar
farfrom
fromPoint
PointB. B.

(a)
(a) (b)
(b)
Figure2.
Figure
Figure 2.2.The
Theabnormal
The abnormaldistribution
abnormal distributionof
distribution ofRPs,
of RPs,where
RPs, wheresome
where someof
some ofthe
of theselected
the selectedRPs
selected RPssimilar
RPs similarin
similar inRSSI
in RSSIto
RSSI tothe
to the
the
unknownpoint
unknown
unknown pointis
point isfurther
is furtherin
further in(a)
in (a)and
(a) andthe
and theselected
the selectedRPs
selected RPssimilar
RPs similarin
similar inRSSI
in RSSIto
RSSI tothe
to theunknown
the unknownpoint
unknown pointgather
point gatheron
gather on
on
one
one side
one side
side inin (b).
in (b).
(b).

3. Proposed
3.Proposed
3. ENTM
ProposedENTM Algorithm
ENTMAlgorithm
Algorithm

3.1. Template
3.1.Template Matching
TemplateMatching
Matchingandand Eight-Neighborhood
andEight-Neighborhood
Eight-Neighborhood
3.1.
Template matching
Templatematching
matchingisis is an
an effective
effectivepattern
patternrecognition
recognitiontechnology
technology in
inimage image
image processing
processing [25].[25].
By
Template an effective pattern recognition technology in processing [25]. By
By calculating
calculating thethe correlation
correlation between
between template
template andand region,
region, the
the similarity
similarity between
between matching
matching template
template
calculating the correlation between template and region, the similarity between matching template
and
and matching
matchingregionregionmay
may bebe moremoredirectly
directlyreflected. As shown
reflected. As shown in Figure
shown 3, the3,template
in Figure
Figure T is a known
the template
template
and matching region may be more directly reflected. As in 3, the TT isis aa
small image
knownsmall which
smallimage has
imagewhich a size
whichhas of l ×
hasaasize w.
sizeof The template
of l l  ww. .The matching
Thetemplate is to
templatematching search
matchingisistofor a target
tosearch
searchforin a large
foraatarget
targetimage
inaa
known in
which has a size of L × W ( L > l,
 W > w ) , and the target is similar to the template in size and
large image which has a size of L  W ( L > l, W > w) , and the target is similar to the template in
large image which has a size of L W ( L > l, W > w ) , and the target is similar to the template in
direction. During the matching process, T moves on the image, and the region where T is overlaid
sizeand
size anddirection.
direction.During
Duringthe thematching
matchingprocess,
process, T moves moveson onthe
theimage,
image,and andthe theregion
regionwhere
where T isis
(Si in Figure 3) is called the subimage. The relatedT coefficient between T and Si is calculated, andT the
overlaid (SSi in in Figure
Figure 3) 3) isis called
called the
the subimage.
subimage. The relatedrelated coefficient between and S is
overlaid
subimage( withi the greatest correlation is considered The as the target.coefficient between TT and Si i is
calculated,and
calculated, andthe thesubimage
subimagewith withthethegreatest
greatestcorrelation
correlationisisconsidered
consideredas asthethetarget.
target.

Figure 3.
Figure3. Schematic
3.Schematic of
Schematicof template
oftemplate matching.
templatematching.
matching.
Figure
Because the RSSI values of the adjacent RPs are highly correlated which will be explained in
Because the
Because theRSSI
RSSIvalues
valuesofofthe
theadjacent
adjacentRPs
RPs are
arehighly
highlycorrelated
correlated which
whichwill
willbe
beexplained
explainedin in
Sections 3.2 and 3.3, the density of RPs may be increased by spatial interpolation and the distribution of
Sections 3.2 and 3.3, the density of RPs may be increased by spatial interpolation and the distribution
Sections 3.2 and 3.3, the density of RPs may be increased by spatial interpolation and the distribution
the RPs may be converted into a regular grid form which is similar to raster image. Therefore, the method
ofthe
of theRPs
RPsmay
maybebeconverted
convertedinto
intoaaregular
regulargrid
gridform
formwhich
whichisissimilar
similarto
toraster
rasterimage.
image.Therefore,
Therefore,the
the
method of template matching may be used to select the best RPs with the RSSI correlation
method of template matching may be used to select the best RPs with the RSSI correlation taken into taken into
account.In
account. Inorder
orderto
toreduce
reducethe
theinfluence
influenceofofabnormal
abnormalRPs RPsand
andtototake
takethe
thecorrelation
correlationof
ofsignal
signalstrength,
strength,
Sensors 2019, 19, 4859 5 of 13
Sensors 2019, 19, x FOR PEER REVIEW 5 of 13

of 3  3 matching
thetemplate template or maythebeeight-neighborhood
used to select the best RPs with
template is the RSSI correlation
selected taken template.
as the matching into account.
As
In order
Sensors to
2019, reduce
19, x FORthe influence
PEER REVIEW of abnormal RPs and to take the correlation of signal
shown in Figure 4, the eight-neighborhood refers to the organization of data in a regular grid, which strength, the 3
5 ×
of 3
13
template
is centered oron
theGrid
eight-neighborhood template
No. 0 and neighboring GridisNo. selected
1, 2 as8the inmatching template.
all directions [26]. As shown in
the 3 4, 3thetemplate
Figure or the eight-neighborhood
eight-neighborhood template of
refers to the organization is data
selected as the matching
in a regular template.
grid, which As
is centered
shown
on GridinNo.
Figure
0 and4, neighboring
the eight-neighborhood 2 ·1· · 8to
Grid No. 1, refers 2 in 3 organization
the
all directions [26].of data in a regular grid, which
is centered on Grid No. 0 and neighboring Grid 8 No. 0 1,4 2 8 in all directions [26].

17 26 35

8 0 4
Figure 4. Schematic of the eight neighborhood.
7 6 5

3.2. Spatial Resolution of BLE RSSI

Figure
Figure 4. Schematic of
4. Schematic of the
the eight
eight neighborhood.
neighborhood.
The RSSI distribution in the AP is susceptible to the indoor environment, but still satisfies the
3.2. Spatial
general Resolution
rule of
of signalof BLE RSSI
propagation.
3.2. Spatial Resolution BLE RSSI Because the RSSI of the signal is continuously decreasing with the
distance from distribution
The RSSI the center AP, it isAP
in the necessary to discuss
is susceptible to the theindoor
decrease of RSSI with
environment, butthe
stilldistance.
satisfiesBLE
the
The
signal RSSI
intensitydistribution
attenuationin the
model AP is
can susceptible
be described to
as the
the indoor environment,
log-normal but
distribution still
[27] satisfies
which is: the
general rule of signal propagation. Because the RSSI of the signal is continuously decreasing with the
general rule of signal propagation. Because the RSSI of the signal is continuously decreasing with the
distance from the center AP, it is necessary to discuss the decrease of RSSI with the distance. BLE signal
ddecrease
distance from the center AP, it is necessary
 RSSI dto discuss the  X , of RSSI[27]
with the distance. BLE
intensity attenuation model canRSSI be described as 10
the n log (
log-normal
10
) distribution which is: (5)
0
signal intensity attenuation model can be described as the dlog-normal
0 distribution [27] which is:
d
where RSSI d0 is the measured RSSI
RSSI =
reference RSSI
 RSSI d0− 10
value 10n
at n log
the
log (d ) ) +
distance
10( XX,dσ0, , RSSI is the measured value
of (5)
d0 10 d (5)
d0 0
at the distance of d , and n is the path loss exponent, which indicates the rate of increasing path loss
relatedRSSI
where to dthe
0
is the measured
distance. X reference
denotes value at the distance
a Gaussian randomofvariabled0 , RSSI with
is thezero
measured
meanvaluecausedat the
by
where RSSI
distance of d,d0andis nthe measured
is the path loss reference
exponent, value
which at the distance
indicates theofrated0of, increasing
RSSI is the measured
path value
loss related to
shadowing [28].
at the distance of d , and n is the path loss exponent, which indicates
the distance. Xσ denotes a Gaussian random variable with zero mean caused by shadowing [28]. the rate of increasing path loss
As shown in Figure 5, an iBeacon is taken as a known AP and set at the certain height of 1 m
As shown
related to the in Figure 5, X
distance. an iBeacon
denotesis taken as a known
a Gaussian random AP and set at the
variable with certain
zero height
mean of 1 m from
caused by
from the ground, and the AP’s RSSI is acquired with a mobile phone at the horizontal intervals of 1
the ground,
shadowing [28].and the AP’s RSSI is acquired with a mobile phone at the horizontal intervals of 1 m at the
m at the same inheight in an open the area. In the experiment,
d0 aisknown d0 is tochosen to be 1 m, RSSI d is To−66.88. To
same Asheight
shown an open
in Figure area.
5, anIniBeacon experiment,
is taken as chosen AP be
and setRSSI
1 m, d0 is
at the −66.88.
certain
0
height simplify
of 1 m
the
from calculation,
simplify
the ground, Xσandis set
the calculation, toAP’s
the Xbe

0isdueset to
RSSI isthe
to be 0unencumbered
due to
acquired experiment
theaunencumbered
with mobile phoneenvironment
experiment
at [29].
the horizontal And the path
environment
intervals [29].
of 1
loss
And exponent
the path n is
loss 1.263 which
exponent n isisdetermined
1.263 which byisfitting with
determined the
by acquired
fitting data.
with the acquired data.
m at the same height in an open area. In the experiment, d0 is chosen to be 1 m, RSSI d is −66.88. To
0

simplify the calculation, X  is set to be 0 due to the unencumbered experiment environment [29].
And the path loss exponent n is 1.263 which is determined by fitting with the acquired data.

Figure 5. Schematic
Figure 5. Schematic of
of data
data acquisition
acquisition experiment.
experiment.

Table
Table 11 displays
displays the
the RSSI
RSSI values
values atat the
the interval
interval of of 0.5
0.5 m,
m, where
where aa part
part of
of the
the RSSI
RSSI values
values are
are
obtained by fitting. The differences
Figure 5. of RSSI
Schematic values
of data between
acquisitioneach neighboring
experiment.
obtained by fitting. The differences of RSSI values between each neighboring pair are also shown,pair are also shown,
which
which arearecalculated
calculatedby:by:
Table 1 displays the RSSI values ∆RSSI
at the=interval
RSSIi+1of−0.5 RSSI
m,i ,where a part of the RSSI values are (6)
obtained by fitting. The differencesof RSSI values
RSSIi 1between
- RSSIi ,each neighboring pair are also shown,
where RSSIi is the estimated value of a certain
RSSI distance and ∆RSSI is the difference at (6) the
which are calculated by:
neighboring distances.
It isRSSI
found isthat theestimated
RSSI is weaker and
of athe difference is more subtle with the increasing interval
where i the
of distance. At the distance of 3.5 m,
value
RSSI certain
RSSI
the difference RSSI , and RSSI is the difference at the
-distance
1 RSSI iis less than 1 dBm, which means that(6)
i of
the
neighboring distances.
difference in RSSI values is little. Considering that the APs are normally placed at a higher position such
It is found that the RSSI is weaker and the difference is more subtle with the increasing interval
as
where RSSIi the
the ceiling, is RSSI values of adjacent
the estimated value ofRPsa are close,distance
certain especiallyand RSSI
when the RPsisare
thedensely distributed.
difference at the
of distance. At the distance of 3.5 m, the difference of RSSI is less than 1 dBm, which means that the
neighboring
difference indistances.
RSSI values is little. Considering that the APs are normally placed at a higher position
It is found that the RSSI is weaker and the difference is more subtle with the increasing interval
of distance. At the distance of 3.5 m, the difference of RSSI is less than 1 dBm, which means that the
difference in RSSI values is little. Considering that the APs are normally placed at a higher position
 No. d (m) RSSI (dBm) RSSI (dBm)
1 0.5 −63.08 /
2 1.0 −66.88 3.80
3 1.5 −69.10 2.22
Sensors 2019, 19, 4859 4 2.0 −70.68 1.58 6 of 13

5 2.5 −71.91 1.23

6 bluetooth
Table 1. Relationship between 3.0 low energy
−72.91 (BLE) received
1 signal strength indication (RSSI)
and physical distance. 7 3.5 −73.75 0.84
8 4.0 −74.48 0.73
No.
9 d (m)
4.5 RSSI (dBm) ∆RSSI
−75.13 0.65(dBm)
101 5.0
0.5 −75.71
−63.08 /
0.58
112 1.0
5.5 −66.88
−76.23 3.80
0.52
−69.10
123 1.5
6.0 −76.71 2.22
0.48
4 2.0 −70.68 1.58
135 6.5
2.5 −77.15
−71.91 0.44
1.23
146 7.0
3.0 −77.55
−72.91 0.40
1
157 7.5
3.5 −77.93
−73.75 0.38
0.84
168 4.0
8.0 −74.48
−78.27 0.73
0.34
9 4.5 −75.13 0.65
10 5.0 −75.71 0.58
3.3. Correlation of Spatial RSSI Fingerprint
11 5.5 −76.23 0.52
12 6.0 −76.71
The indoor fingerprint database organized by the regular0.48 grid is similar to the raster image.
13 6.5 −77.15 0.44
Although there might be abnormal 14
RPs,
7.0
the RSSI values
−77.55
of the adjacent
0.40
RPs are highly correlated in
the most parts of the fingerprint15 database.
7.5 The−77.93
value of each0.38
grid in the eight-neighbor template
indicates the RSSI of a certain16AP signal
8.0 that can be measured0.34
−78.27 in the indoor environment, and the
RSSI of the center grid is the measured RSSI data at the unknown point. Due to the fact that RSSI
vectors
3.3. of the eight
Correlation neighboring
of Spatial grids cannot be obtained in a short time during the online phase, the
RSSI Fingerprint
RSSI data in the fingerprint database should be analyzed to find out the relationship between the
The indoor fingerprint database organized by the regular grid is similar to the raster image.
central grid and its neighbors.
Although there might be abnormal RPs, the RSSI values of the adjacent RPs are highly correlated in the
In order to discuss the distribution of the RSSI value of a single AP signal indoors, an AP is
most parts of the fingerprint database. The value of each grid in the eight-neighbor template indicates
arranged at a certain point in the 8 m  8 m room, and a separate coordinate system is established.
the RSSI of a certain AP signal that can be measured in the indoor environment, and the RSSI of the
The RPs are chosen at interval of 1 m, and the RSSI feature of each RP is determined by mean filtering.
center grid is the measured RSSI data at the unknown point. Due to the fact that RSSI vectors of the
Considering the influence of the interpolation on the locating effect of the abnormal RP, the data is
eight neighboring grids cannot be obtained in a short time during the online phase, the RSSI data in
interpolated at an interval of 0.5 m by Kriging interpolation. As shown in Figure 6, the eight-
the fingerprint database should be analyzed to find out the relationship between the central grid and
neighborhood can be divided into four directions.
its neighbors.
The differences between the RPs and its neighbor RPs in the four directions are calculated and
In order to discuss the distribution of the RSSI value of a single AP signal indoors, an AP is
recorded. Since the RSSI difference is described as a floating number, the results are processed as:
arranged at a certain point in the 8 m × 8 m room, and a separate coordinate system is established.
The RPs are chosen at interval of 1  m,R and
 floor R feature
the (RSSI 0.5) , of each RP is determined by mean (7)
filtering. Considering the influence of the interpolation on the locating effect of the abnormal RP,
where R is the raw difference of the adjacent RPs in different directions, and R is an integer (
the data is interpolated at an interval of 0.5 m by Kriging interpolation. As shown in Figure 6,
..., 3, 2, 1,0,1,2,3,... ). The occurrences of each integer are recorded as n . In Figure 6, the histogram
the eight-neighborhood can be divided into four directions.
of RSSI differences in the direction of AA is shown.
B
C D

A A

D C
B

Figure 6. The four directions of the center point.

The differences between the RPs and its neighbor RPs in the four directions are calculated and
recorded. Since the RSSI difference is described as a floating number, the results are processed as:

∆R0 = f loor (∆R + 0.5), (7)

where ∆R is the raw difference of the adjacent RPs in different directions, and ∆R0 is an integer
(. . . , −3, −2, −1, 0, 1, 2, 3, . . .). The occurrences of each integer are recorded as n. In Figure 6, the histogram
of RSSI differences in the direction of AA0 is shown.
Sensors 2019, 19, x FOR PEER REVIEW 7 of 13

Figure 6. The four directions of the center point.

As shown in Figure 7, the RSSI difference of adjacent RPs is less than 3 dBm, which is consistent
Sensors 2019, 19, 4859 7 of 13
with the spatial resolution of RSSI. It is obviously that there exist still a few differences larger than 3
dBm, which may be regarded as the abnormal RPs. Therefore, the impact of abnormal RPs on the
template generation
As shown parameters
in Figure needs
7, the RSSI to be reduced
difference duringRPs
of adjacent theisgeneration
less than 3ofdBm,the eight-neighborhood
which is consistent
template. Although
with the spatial the occurrences
resolution of RSSI. It of RSSI differences
is obviously that there throughout
exist still thea fewindoor environment
differences larger thanare
similar to the normal distribution, the differences of RSSI between
3 dBm, which may be regarded as the abnormal RPs. Therefore, the impact of abnormal RPs on the the center grid and its eight
Sensors 2019, 19, x FOR PEER REVIEW 7 of 13
neighboring grids areparameters
template generation related to both
needsthe to position
be reduced of the center
during thegrid and thatofofthe
generation theeight-neighborhood
AP. Since the RSSI
value of the
template. AP cannot
Although be accurately
the occurrences of collected
RSSI during
differences offline
throughout phase,
the
Figure 6. The four directions of the center point. the
indoor maximum
environmentposition are of the
similar
RSSI feature after the interpolation is selected instead of the AP position.
to the normal distribution, the differences of RSSI between the center grid and its eight neighboring With the selected point
As shown in Figure 7, the RSSI difference of adjacent RPs is less than 3 dBm, which is consistent
taken as the
grids are origin
related toand
boththe thehorizontal
position of andthevertical coordinates
center grid and thatasofthe thedividing
AP. Since lines, the indoor
the RSSI value space
of the
with the spatial resolution of RSSI. It is obviously that there exist still a few differences larger than 3
can be divided
AP cannotdBm, into
be accurately four quadrants.
which may collected
be regarded In
during Figure 4, the
offline phase,
as the abnormal frequency
the maximum
RPs. Therefore, of each
the impact RSSI
position difference
of the
of abnormal RPsRSSI in
on the the fourth
feature after
quadrant istemplate
shown.
the interpolation is selectedparameters
generation instead of thetoAP
needs position.
be reduced With
during thethe selected
generation point
of the taken as the origin and
eight-neighborhood
template.
As shown
the horizontal and Although
in Figure the
theoccurrences
vertical8,coordinates asofthe
RSSI difference RSSI differences
of the lines,
dividing throughout
neighborhood
the indoorthegrid
indoor
space environment
which
can is divided
be closer aretointo
the four
AP
similar to the normal distribution, the differences of RSSI between the center grid and its eight
and the center grid
quadrants.neighboring is
In Figuregrids biased
4, the to the positive direction. Therefore, the quadrant
frequency of each RSSI difference in the fourth quadrant is shown. and the location of the
are related to both the position of the center grid and that of the AP. Since the RSSI
neighborhoodvaluegrid
of theshould
AP cannotbe betaken into consideration
accurately collected during to generate
offline phase, the template.position of the
the maximum
RSSI feature after the interpolation is selected instead of the AP position. With the selected point
taken as the origin and the horizontal and vertical coordinates as the dividing lines, the indoor space
can be divided into four quadrants. In Figure 4, the frequency of each RSSI difference in the fourth
quadrant is shown.
As shown in Figure 8, the RSSI difference of the neighborhood grid which is closer to the AP
and the center grid is biased to the positive direction. Therefore, the quadrant and the location of the
neighborhood grid should be taken into consideration to generate the template.

Figure 7. Histogram of RSSI difference in the direction of AA0 .

Figure 7. Histogram of RSSI difference in the direction of AA .

As shown in Figure 8, the RSSI difference of the neighborhood grid which is closer to the AP
and the center grid is biased to the positive direction. Therefore, the quadrant and the location of the
neighborhood grid should be taken into consideration to generate the template.
Figure 7. Histogram of RSSI difference in the direction of AA .

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

(c) (d)

Figure 8. Histogram of (c)∆R in the fourth quadrant and (a) Histogram of(d) ∆R in the direction of AA0 ,
(b) Histogram of ∆R in the direction of BB , (c) Histogram of ∆R in the direction of CC0 , (d) Histogram
0

of ∆R in the direction of D0 D.
Sensors 2019, 19, 4859 8 of 13

3.4. ENTM Algorithm

The proposed ENTM algorithm may be divided into the three steps, i.e., the generation of
fingerprint database, the generation of the matching template and the location estimation.
(1) The generation of the fingerprint database
Assuming that there are M APs in the indoor environment, the RSSI of any point may be received
as (r1 , r2 , . . . , rM ). With the m-th AP taken as an example, the data is organized into a regular grid
by interpolation to make the fingerprinting database similar to a raster image in structure. The RSSI
distribution of the m-th AP may be written as:
 m
rm rm

 r(0, 0) (0, 1)
··· (0, V ) 

 rm rm ··· rm


 (1, 0) (1, 1) (1, V ) 
Rm =  , (8)

.. .. ..
rm


 . . (u, v)
. 


 rm rm ··· rm

(U, 0) (U, 1) (U, V )

where Rm is the RSSI distribution of the m-th AP, u (0 ≤ u ≤ U ) and v (0 ≤ v ≤ V ) are the indices of
the fingerprint, U and V are the maximum values of the indices which are determined by interpolation
intervals of different axes, and rm
(u,v)
is the feature RSSI of the RP determined by interpolation. With the
interpolation intervals set, the coordinates of each RP may be described by:
(
xu = x0 + u∆x
, (9)
yv = y0 + v∆y

where ∆x and ∆y are the intervals of the grid, and (x0 , y0 ) is the coordinate of the starting point in
the grid.
(2) The generation of the matching template
The generation of eight neighborhood matching template needs to reduce the influence of the
abnormal reference point and to follow the AP signal propagation law. Therefore, different matching
templates need to be generated according to the position of the template center point. The expected
values of the RSSI differences in different directions of each quadrant are used as the parameters to
generate the template. The expected values may be determined by:
X
Eq (h) = ∆R • p(∆R), (10)

where ∆R is the RSSI differences processed from Equation (7), p(∆R) is the probability of occurrence of
∆R, q is the number of quadrant, and h is the position of the neighborhood.
The RSSI value of each neighborhood point may be expressed as:

rhq = ro + Eq (h), (11)

where r0 is the RSSI vector acquired at the unknown point. The eight-neighborhood template in the
q-th quadrant may be described as:
 1 2 3 
 rq rq rq 
Tq =  r8q ro r4q . (12)
 
 7 6 5 
rq rq rq

(3) The location estimation

In the location estimation, the matching template Tqm of the m-th AP is moved on Rm , and the
quadrant q where the template is located is determined by the direction from the m-th AP to the center
 Sensors 2019, 19, x FOR PEER REVIEW 9 of 13

 rq1 rq2 rq3 

 
Tq  rq8 ro rq4  . (12)
 7 
Sensors 2019, 19, 4859 rq rq6 rq5  9 of 13

(3) The location estimation

point of the template. The estimation,
In the location coverage the
area of Rm is
matching Sm , and
template Tqmthe
of center point
the m - th Sm is on
AP isofmoved Rm , andwith (u, v)
indexed
in the fingerprint similarity D between template T and sub-area S
the quadrant q where the template is located is determined by the direction from the m - thisAP
database. The calculated
to by
Euclidean the
distance as: of the template. The coverage area of R is Sm and the center point of Sm is
center point m ,
indexed with (u, v) in the fingerprintv
udatabase. The similarity D between template T and sub-
M tX 3 X 3 
u
X
area S is calculated by Euclidean distance as:
2
D(u, v) = Sm [φ][ϕ] − Tqm [φ][ϕ] , (13)
m=1 M φ3=03 ϕ=0

D  u, v      Sm [ ][ ]  Tqm [ ][ ] 
2
, (13)
m 1  0  0

where φ and ϕ are the element in matrix Sm

and Tqm . Therefore, the position of the RP with the
where  and  are the element in matrix S m and Tqm . Therefore, the position of the RP with the
minimum D is chosen as the estimated position. The flowchart of the proposed ENTM is shown in
minimum D is chosen as the estimated position. The flowchart of the proposed ENTM is shown in
Figure 9. Figure 9.

Figure 9. Figure
The9.flow
The flow chartofof fingerprint
chart fingerprint indoor localization
indoor method based
localization methodon the proposed
based on eight-
the proposed
neighborhood template matching (ENTM) method.
eight-neighborhood template matching (ENTM) method.
4. Experiment and Analysis of Its Result
4. Experiment and Analysis of Its Result
4.1. The Environment of Experiment
4.1. The Environment of Experiment
In order to evaluate the performance of the proposed algorithm, experiments were conducted in
In order to evaluate
a room which wasthe usedperformance
as the reading of theinproposed
room Tian Gong algorithm, experiments
Building of Nanjing were conducted
Tech University. The in a
panel in Figure 10 shows the room of dimensions of 8 m  8 m, where 4 iBeacon
room which was used as the reading room in Tian Gong Building of Nanjing Tech University. The panelAPs are laid. In Figure
9, the star symbols represent the position of the APs, the shadow symbols represent furniture in the
in Figure 10 shows the room of dimensions of 8 m × 8 m, where 4 iBeacon APs are laid. In Figure 9,
room and the dot symbols represent the RSSI sampling points with the 1 m interval. Due to the indoor
the star symbols represent
environment the position
restrictions, of the
the sampling APs, the
intervals wereshadow symbols
not strictly 1 m. Therepresent
RSSI wasfurniture
respectively in the room
and the dot symbols
collected with represent
the rate of 1 sthe
for RSSI sampling
about 100 s at each points
samplingwith
point the 1 m the
and with interval.
rate of 0.5Due
s for to
20 sthe indoor
environment at each testing point.the
restrictions, An independent
sampling coordinate
intervalssystem
werewas notestablished
strictly 1inm.
the room
The for convenience.
RSSI was respectively
collected with the rate of 1 s for about 100 s at each sampling point and with the rate of 0.5 s for 20 s at
eachSensors
testing point.
2019, AnPEER
19, x FOR independent
REVIEW coordinate system was established in the room for convenience. 10 of 13

Figure10.
Figure 10.The
The plan
plan of
ofthe
theexperiments
experimentsarea.
area.

4.2. Comparison of Locating Accuracy

The fingerprint database was obtained by interpolating the original fingerprint sequence, and
32 testing points were randomly selected in the indoor environment. KNN, WKNN and ENTM
algorithms were used to achieve static location. To compare the results of the experiment, the best
locating results based on KNN and WKNN were respectively obtained by changing k . With the
 Sensors 2019, 19, 4859 10 of 13

Figure 10. The plan of the experiments area.

4.2. Comparison of Locating Accuracy
4.2. Comparison of Locating Accuracy
The fingerprint database was obtained by interpolating the original fingerprint sequence,
andThe fingerprint
32 testing database
points was obtained
were randomly by interpolating
selected in the indoorthe environment.
original fingerprintKNN, sequence,
WKNNand and
32 testing points were randomly selected in the indoor environment. KNN, WKNN
ENTM algorithms were used to achieve static location. To compare the results of the experiment, and ENTM
algorithms were used
the best locating resultstobased
achieve static and
on KNN location.
WKNN Towere
compare the results
respectively of thebyexperiment,
obtained changing k. the best
With the
locating
Euclideanresults based
distance on KNN
between and WKNN
the locating resultwere respectively
and the obtained
real position by changing
of the point . With
taken askthe the
accuracy
Euclidean
standard,distance
the best between
locating the locating
results wereresult and the
compared realeach
with position
other.of It
the pointthat
shows taken aserror
the the accuracy
between
standard,
them wasthethebest locating
smallest whenresults
k = 7.were
Thencompared with each
the best locating other.
results It shows
were comparedthat the
witherror
that between
of ENTM
them was the
algorithm. Thesmallest
comparisonwhenof kthe=locating
7 . Thenaccuracy
the bestoflocating results
each point were incompared
is shown Figure 11.with that of
ENTM algorithm. The comparison of the locating accuracy of each point is shown in Figure 11.

Figure 11. Comparison of the locating error of each testing point.

Figure 11. Comparison of the locating error of each testing point.
It is shown in Figure 10 that the locating accuracy of the improved algorithm was superior to that
It is shown in Figure 10 that the locating accuracy of the improved algorithm was superior to
of KNN and WKNN at points 5, 6, 14, 22 and 29. The maximum locating errors of the three algorithms
that of KNN and WKNN at points 5, 6, 14, 22 and 29. The maximum locating errors of the three
were at the same point located at the edge of the room, where AP signal fluctuated greatly due to the
algorithms were at the same point located at the edge of the room, where AP signal fluctuated greatly
nearby bookshelves. Further analysis indicated that the abnormal RPs were unavoidable to be selected
due to the nearby bookshelves. Further analysis indicated that the abnormal RPs were unavoidable
in KNN and WKNN leading to lower locating accuracy, but the influence of the abnormal RPs may be
to be selected in KNN and WKNN leading to lower locating accuracy, but the influence of the
filtered to a certain extent with the ENTM algorithm.
abnormal RPs may be filtered to a certain extent with the ENTM algorithm.
Table 2 is the comparison of the locating errors with the three algorithms, including the RMSE
Table 2 is the comparison of the locating errors with the three algorithms, including the RMSE
(Root of Mean Square Error) and the 80th percentile error. It may be seen that ENTM algorithm
(Root of Mean Square Error) and the 80th percentile error. It may be seen that ENTM algorithm was
was 16.7% improved as compared with KNN and WKNN algorithms, and ENTM produced an 80th
16.7% improved as compared with KNN and WKNN algorithms, and ENTM produced an 80th
percentile error of 1.1 m which was 21.4% improved as compared with KNN and WKNN. The locating
percentile error of 1.1 m which was 21.4% improved as compared with KNN and WKNN. The
accuracy of ENTM was better than both KNN and WKNN.
locating accuracy of ENTM was better than both KNN and WKNN.
Table 2. Locating error statistics.

Algorithm RMSE (m) Improved (%) 80th (m) Improved (%)

K-nearest neighbor (KNN) 1.2 0 1.4 0
Weighted K-nearest neighbor (WKNN) 1.2 0 1.4 0
ENTM 1.0 16.7 1.1 21.4

Figure 12 shows the cumulative probability distribution of the locating error of the three algorithms.
It can be seen in the figure that the probability of the locating error of less than 1.2 m with ENTM
algorithm was 81.2%, but the probabilities with KNN and WKNN algorithms were both 71.9%.
 Weighted K-nearest neighbor (WKNN) 1.2 0 1.4 0
ENTM 1.0 16.7 1.1 21.4

Figure 12 shows the cumulative probability distribution of the locating error of the three
algorithms. It 4859
Sensors 2019, 19, can be seen in the figure that the probability of the locating error of less than 1.2 m11with
of 13
ENTM algorithm was 81.2%, but the probabilities with KNN and WKNN algorithms were both
71.9%.

Figure 12. Comparison of locating error with three algorithms in terms of CDF (Cumulative Distribution
Figure 12. Comparison of locating error with three algorithms in terms of CDF (Cumulative
Function).
Distribution Function).
5. Conclusions and Future Work
5. Conclusions and Future Work
This paper proposes a BLE fingerprint indoor localization algorithm based on eight-neighborhood
This matching.
template paper proposes With thea BLE
spatialfingerprint
resolutionindoor
of BLE localization
RSSI and the algorithm
correlation based on eight-
of spatial RSSI
neighborhood template matching. With the spatial resolution of BLE RSSI
fingerprint in the database analyzed, a method for the generation of eight-neighborhood template and the correlation of
spatial RSSI fingerprint in the database analyzed, a method for the generation
was proposed, which divided the indoor environment into four quadrants for each AP and chose of eight-neighborhood
template was proposed,
the expected values of the which
RSSIdivided the indoor
difference between environment into four
the center points andquadrants for each AP and
their eight-neighborhood
chose the expected values of the RSSI difference between the
in different quadrants as the generation parameters. RP was selected through template center points and their eight-
matching,
neighborhood in different quadrants as the generation parameters. RP was selected
which took the correlation of fingerprint data into account and may have filtered the influence of through template
matching,
the abnormal whichRPs. took
Thethe correlation
ENTM of fingerprint
algorithm was testeddata into an
through account and may
experiment in ahave filtered
complex the
indoor
influence of the abnormal RPs. The ENTM algorithm was tested through an
environment. It showed that the locating error was 1.0 m, which was about 0.2 m less than that withexperiment in a complex
indoor
KNN and environment. It showed that the locating error was 1.0 m, which was about 0.2 m less than
WKNN algorithms.
that with
There KNN
are aand
fewWKNN
researchesalgorithms.
on spatial fingerprint correlation in the field of fingerprint localization,
and we believe that our work on
There are a few researches mayspatial
inspirefingerprint correlation
researchers to improve in the field ofaccuracy
locating fingerprint localization,
based on RSSI
and we believe that our work may inspire researchers to improve locating
correlation. Since the acceptable localization results are illustrated with a relatively small room,accuracy based on RSSI
correlation. Since the acceptable localization results are illustrated with a relatively
we will expand the experimental area to the whole floor. The increase of the location area means that small room, we
will expand the experimental area to the whole floor. The increase of
there would be more data in the fingerprint database, and the number of RPs with abnormal RSSIthe location area means that
there
would would be more
increase. We plandatatoindesign
the fingerprint
a solutiondatabase,
based onand the the numberof
techniques ofdata
RPs mining
with abnormal RSSI
and machine
would increase. We plan to design a solution based on the techniques of data mining
learning to evaluate the continuity of fingerprint data [30], and to establish a fingerprint database with and machine
learning to evaluate the
MARS (Multivariate continuity
Adaptive of fingerprint
Regression dataor[30],
Spline) [31] CMARS and to establish aMultivariate
(Continuous fingerprint Adaptive
database
with MARSSpline)
Regression (Multivariate Adaptive
[32] to improve theRegression
robustnessSpline) [31] or system.
of the locating CMARS (Continuous Multivariate
Adaptive Regression Spline) [32] to improve the robustness of the locating system.
Author Contributions: All authors contributed equally to this work. M.L. and L.Z. designed the main concept of
Author Contributions: All authors contributed equally to this work. M.L. and L.Z. designed the main concept
the algorithm and wrote the paper. L.Z., D.T. and X.T. performed the experiments and analyzed the data.
of the algorithm and wrote the paper. L.Z., D.T. and X.T. performed the experiments and analyzed the data.
Funding: This research was funded by National Natural Science Foundation of China, grant number 41974214.
Funding: This research was funded by National Natural Science Foundation of China, grant number 41974214.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Liu, H.; Darabi, H.; Banerjee, P.; Liu, J. Survey of wireless indoor positioning techniques and systems.
IEEE Trans. Syst. Man Cybern. Part. C Appl. Rev. 2007, 37, 1067–1080. [CrossRef]
2. Brimicombe, A.; Li, C. Location-Based Services and Geo-Information Engineering; John Wiley Sons: Hoboken, NJ,
USA, 2009; Volume 21.
3. Huh, J.-H.; Seo, K. An indoor location-based control system using bluetooth beacons for IoT systems.
Sensors 2017, 17, 2917. [CrossRef] [PubMed]
Sensors 2019, 19, 4859 12 of 13

4. Zou, H.; Jiang, H.; Luo, Y.; Zhu, J.; Lu, X.; Xie, L. BlueDetect: An iBeacon-enabled scheme for accurate
and energy-efficient indoor-outdoor detection and seamless location-based service. Sensors 2016, 16, 268.
[CrossRef] [PubMed]
5. Zhuang, Y.; Yang, J.; Li, Y.; Qi, L.; El-Sheimy, N. Smartphone-based indoor localization with bluetooth low
energy beacons. Sensors 2016, 16, 596. [CrossRef] [PubMed]
6. Zhou, S.; Pollard, J.K. Position measurement using bluetooth. IEEE Trans. Consum. Electron. 2006, 52, 555–558.
[CrossRef]
7. Yang, C.; Shao, H.-R. WiFi-based indoor positioning. IEEE Commun. Mag. 2015, 53, 150–157. [CrossRef]
8. Deng, Z.-A.; Hu, Y.; Yu, J.; Na, Z. Extended kalman filter for real time indoor localization by fusing WiFi and
smartphone inertial sensors. Micromachines 2015, 6, 523–543. [CrossRef]
9. Tiemann, J.; Schweikowski, F.; Wietfeld, C. Design of an UWB indoor-positioning system for UAV navigation
in GNSS-denied environments. In Proceedings of the 2015 International Conference on Indoor Positioning
and Indoor Navigation (IPIN), Calgary, AB, Canada, 13–16 October 2015; IEEE: Piscataway, NJ, USA, 2015;
pp. 1–7.
10. Niu, Q.F.; Cao, Y.F.; Wang, L.; Hui, Y.B. Study of UWB indoor positioning technology and optimization
algorithm based on TW-TOF. Autom. Instrum. 2018, 11, 93–99.
11. Shi, G.; Ming, Y. Survey of indoor positioning systems based on ultra-wideband (UWB) technology. In Wireless
Communications, Networking and Applications; Springer: Berlin, Germany, 2016; pp. 1269–1278.
12. Xiao, C.; Yang, D.; Chen, Z.; Tan, G. 3-D BLE indoor localization based on denoising autoencoder. IEEE Access
2017, 5, 12751–12760. [CrossRef]
13. Xi, R.; Li, Y.; Hou, M. Survey on indoor localization. Comput. Sci. 2016, 43, 1–6, 32.
14. Fang, S.H.; Lin, T.N. Projection-based location system via multiple discriminant analysis in wireless local
area networks. IEEE Trans. Veh. Technol. 2009, 58, 5009–5019. [CrossRef]
15. Zhao, P.; Chen, J. A study on localization algorithm using existing wireless local area network. J. Signal
Process. 2014, 30, 1413–1418.
16. Bahl, P.; Padmanabhan, V.N. RADAR: An in-building RF-based user location and tracking system.
In Proceedings of the IEEE INFOCOM 2000, Conference on Computer Communications, Nineteenth
Annual Joint Conference of the IEEE Computer and Communications Societies (Cat. No.00CH37064),
Tel Aviv, Israel, 26–30 March 2000; IEEE: Piscataway, NJ, USA, 2000; Volume 2, pp. 775–784.
17. Kukolj, D.; Vuckovic, M.; Pletl, S. Indoor Location Fingerprinting Based on Data Reduction. In Proceedings
of the International Conference on Broadband Wireless Computing, Barcelona, Spain, 26–28 October 2011.
18. Li, H.L.; Qian, Z.H.; Tian, H.L. Research on indoor localization algorithm based on kernel principal component
analgsis. J. Commun. 2017, 38, 158–167.
19. Bi, J.X.; Wang, Y.J.; Cao, H.J.; Wang, Y.K. A method of WiFi indoor positioning based on omnidirectional
fingerprint database. Bull. Surv. Mapp. 2018, 44, 25–29.
20. Liu, H.; Ma, J.; Huang, Q. Construction method of fingerprint databased on improved Kriging interpalation
for indoor location. J. Chongqing Univ. Posts Telecommun. Sci. Ed. 2017, 29, 751–757.
21. Krishnakumar, A.S.; Krishnan, P. On the accuracy of signal strength-based location estimation techniques.
In Proceedings of the IEEE 24th Annual Joint Conference of the IEEE Computer and Communications
Societies, Miami, FL, USA, 13–17 March 2005; IEEE: Piscataway, NJ, USA, 2005; Volume 1, pp. 642–650.
22. Yousef, N.R.; Sayed, A.H.; Jalloul, L.M.A. Robust wireless location over fading channels. IEEE Trans. Veh.
Technol. 2003, 52, 117–126. [CrossRef]
23. Zhang, L.; Tan, T.; Gong, Y.; Yang, W. Fingerprint database reconstruction based on robust PCA for indoor
localization. Sensors 2019, 19, 2537. [CrossRef]
24. Zeev, B.; Zeev, V.; Akteke-Öztürk, B.; Weber, G.-W. On minimal spanning tree approach in the cluster
validation problem. Informatica 2009, 20, 187–202.
25. Liang, L.H.; Ai, H.Z.; Xiao, X.P.; Ye, H.J.; Xu, G.Y.; Zhang, B. Face detection based on template matching and
support vector machines. Chin. J. Comput. 2002, 25, 22–29.
26. Qu, Y.; Cheng, Y.; Qiu, L. An improved method of extracting 3D features of buildings based on eight
neighborhood search method. J. Air Force Eng. Univ. Nat. Sci. Ed. 2015, 16, 66–69.
27. 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]
Sensors 2019, 19, 4859 13 of 13

28. Salo, J.; Vuokko, L.; El-Sallabi, H.M.; Vainikaine, P. An additive model as a physical basis for shadow fading.
IEEE Trans. Veh. Technol. 2007, 56, 13–26. [CrossRef]
29. Röbesaat, J.; Zhang, P.; Abdelaal, M.; Theel, O. An improved BLE indoor localization with kalman-based
fusion: An experimental study. Sensors 2017, 17, 951. [CrossRef] [PubMed]
30. Weber, G.-W.; Özöğür-Akyüz, S.; Kropat, E. A review on data mining and continuous optimization
applications in computational biology and medicine. Birth Defects Res. 2009, 87, 165–181. [CrossRef]
[PubMed]
31. Wang, S.H.; Zhang, T.T. Analysis of influence factors of electric geothermal indoor temperature based on
multivariate adaptive regression spline. Sci. Technol. Eng. 2018, 18, 200–206.
32. Weber, G.-W.; Batmaz, İ.; Köksal, G.; Taylan, P.; Yerlikaya-Özkurt, F. CMARS: A new contribution
to nonparametric regression with multivariate adaptive regression splines supported by continuous
optimization. Inverse Probl. Sci. Eng. 2012, 20, 371–400. [CrossRef]

© 2019 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/).