Received September 14, 2021, accepted October 10, 2021, date of publication October 21, 2021, date of current

version October 27, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3122011

Enhanced Indoor Localization Based BLE Using

Gaussian Process Regression and
Improved Weighted kNN
DAI DUONG NGUYEN AND MINH THUY LE , (Member, IEEE)
School of Electrical Engineering, Hanoi University of Science and Technology, Hanoi 10000, Vietnam
Corresponding author: Minh Thuy Le ([email protected])
This research was funded by Hanoi University of Science and Technology under grant number T2021-SAHEP-007.

ABSTRACT Indoor positioning has attracted commercial developers and researchers in the last few decades.
Global positioning system (GPS) cannot well localize in indoor environment. For clinical applications such
as workflow analysis, Bluetooth Low Energy (BLE) beacons have been employed to map the positions of
individuals in indoor environments. The received signal strength indicator (RSSI) fingerprinting plays a key
role in the access point performance services. This paper deals with the issue of indoor localization based
RSSI fingerprint using only RSSI vectors without any prior knowledge of the pose. We proposed to use
machine learning (ML) combined with modified kNN algorithm to enhance the real-time performance and
propose method to detect uncertainty of estimated pose. The mentioned ML algorithm is Gaussian Process
Regression (GPR). In the online phase of our system, GPR gives a prediction for the location of the pose
using RSSI vector. This prediction helps kNN algorithm therefore limits the searching region leading to
reduce the computational cost. An analysis on the distribution of k nearest points is also presented which
aims to evaluate the confidence of the estimated pose for extracting a list of trustable points. Furthermore,
we also present an extrapolation process using this trust-list to have an optimized trajectory. The accuracy
and timing analysis of our proposal was realized on a challenging BBIL dataset which contains very noisy
RSSI signals due to a fast moving of the object. The experimental results show that our system exhibits a
much better performance than traditional kNN or WkNN algorithms. The RMSE of our optimal trajectory is
1.78 m in the room with a dimension of 10 m x 25 m, which is competitive in comparison to other methods
where initial pose is known.

INDEX TERMS Indoor localization, Bluetooth low energy, RSSI fingerprint, Gaussian process regression,
machine learning.

I. INTRODUCTION buildings, rooms, offices, factories, commercial centers or

Localization and positioning become increasingly impor- hospitals, the GPS technology is not pertinent due to the sharp
tant for many applications. Nowadays, the general indoor attenuation of GPS signals. To overcome this issue, some
positioning problem and localization based electromagnetic indoor positioning technologies have been proposed. Radio
wave technologies such as GPS, RFID, Wi-Fi, Zigbee, frequency identifier (RFID) positioning technology [1], [2]
Bluetooth Low Energy, Lora, NB-IoT has been focused by uses information exchange of radio frequency (RF) signals
many research groups around the world. Positioning in an for positioning. RFID is low-cost but the connection distance
outdoor environment with few obstacles, GPS technology is short so that we need a huge number of RFID tags in
is being used popularly with increasingly highly accurate environment. Ultra-wideband (UWB) [3] is a potential posi-
results. However, in an indoor environment such as corridors, tioning technology. UWB can obtain high accuracy but its
cost is very high by reason of specialized material support.
The associate editor coordinating the review of this manuscript and Visual positioning [4] is also a well-known technology when
approving it for publication was Theofanis P. Raptis . people use a camera equipped on object body. This technique

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 143795
 D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

matches photo captured by camera with stored images associ- navigation proposed by Sun et al. [22] was another example
ated with known locations, then relative pose between current of applying ML technique. Besides, ML also exhibits as an
pose and reference pose are extracted. However, it requires a effective way to make a fusion of multi-dimensional data
creating of precise image database with large memory size collected from multiple positioning sensors, technologies and
to store information, which is very challenging in practice. methods. For example, Guo et al. [23] and Alhajri et al. [24]
In [5], the author proposed to use visible light technology. present the applying of both supervised and unsupervised
The system integrated with modulated white light-emitting learning for fusion weight generation. Furthermore, when
diode (LED) can provide information to obtain a high accu- there is a lack of labeled data, Liu et al. [25] proposed to use
racy localization. However, the system does not work if transfer learning in fingerprint-based localization to enhance
the light is shielded by obstacle. Therefore, visible light is system scalability without excessive site surveys while ensur-
not suitable to use in environment containing many objects. ing the accuracy.
Besides, many different alternative methods have been pro- In this paper, we focus on the accuracy of indoor local-
posed based on other electromagnetic waves in indoor envi- ization issue based on only RSSI vector and no-prior pose is
ronments such as Zigbee [6], Lora [7], Wifi [8], or Bluetooth known. Our distributions are: 1) To avoid the disadvantages
[9]. Recently, Bluetooth Low Energy (BLE) becomes more of previous work, our system makes a fusion of Gaussian
and more popular option for practical applications. BLE has process regression (GPR) and kNN methods. In practice,
the advantages of high security, low power consumption, and we propose to applying GPR to predict the searching area,
low cost. However the comparison between Wi-Fi and BLE then only perform kNN calculation in this area to help reduce
is still now debatable because Wi-Fi has a longer range than the number of calculations in pose estimation. 2) We propose
BLE. an evaluation model to determine whether estimated pose of
Machine learning (ML) is now a potential solution for kNN method is reliable or not. 3) We propose an optimization
indoor localization research. One of the most important processing based on reliable points to have a more accurate
advantage of ML is the ability to effectively and rapidly trajectory.
make decision in online-phase from observed data without The rest of this paper is structured as follows: Section II
time-consuming mathematical formulation. The widely use presents previous studies related to the research of our work.
of machine learning is due to the growing computational Section III introduces BBIL dataset used for experiments.
abilities of digital circuits with high speed microprocessors Our localization system is then described in Section IV.
and also the available of many datasets offered by the internet. Section V discusses the performance of our proposal on BBIL
ML algorithms were proposed in various applications like dataset. Finally, Section VI consists of the conclusions and the
health [10], [11], communication [12], [13], agriculture [14], future directions for our research.
energy [15], [16] and so forth. In recent years, in indoor
localization domain, machine learning (ML) algorithms also II. RELATED WORKS
achieve good success. For example, Bhatti et al. [17] has In 2010, Altini et al. [26] proposed to use a combination
developed a technique to detect outlier for Wi-Fi indoor of neural networks to locate users in the corridor based on
localization environment. The author analyzed RSSs using Bluetooth with an average error of 0.5 m. However, the author
the combination of supervised, unsupervised and ensem- does not mention the size of the corridor and the num-
ble machine learning methods. Supervised learning methods ber of broadcasting stations. Another study on BLE wave
were support vector machine (SVM), K-nearest neighbor based positioning by Sthapit et al. [27] used a classifier with
(KNN) and random forest (RF) classifiers. The evaluation two popular machine learning algorithms, Support Vector
of used ML method provides high accuracy of 97.8 per- and Logistic Regression and fingerprint technique to help
cent with proposed outlier detection methods and almost 2% determine the location. of the object using the BLE wave
improvement in the accuracy. Otherwise, Hsieh et al. [18] RSSI parameter. The study tested the results in the university
present a deep learning-based approach for indoor localiza- corridor, the size of the space is 4.5 m x 15 m, including
tion based on transmission channel quality metrics, including 14 BLE broadcasting modules. The average error is about
received signal strength (RSS) and channel state informa- 0.5 m. The idea of the method is to divide the space into
tion (CSI). In order to estimate the position of a subject 0.45m x 0.45m squares. A large number of samples will be
in a room, the authors use RSS and CSI and develop four collected in each plot to form a set to train the classifier. Thus,
deep neural networks with multi-layer perception (MLP) and for each signal to be positioned, the classifier calculates the
one-dimensional convolutional neural network (1D-CNN). probabilities that the signal belongs to the square regions.
On the other side, supervised and unsupervised learning- The position of the point will be calculated as the average
based dimension reduction techniques have been used to of the positions of the squares with the weighted to the calcu-
reduce the complexity and storage space of fingerprint data lated probabilities. This method of dividing cells is suitable
in [19] and [20]. More recently, ML techniques also have for empty spaces, free from large objects in the middle of
been used by Le et al. [21] to extract the hidden features of the room. In contrast, for spaces with many objects, this
the RSSI measurement to reduce the data size of fingerprint method of evenly dividing cells is no longer suitable. Another
collection. Object localization in the environment during the study by Alexander and Kusuma [28] in 2019 also used

143796 VOLUME 9, 2021

D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

FIGURE 1. BBIL dataset environment.

FIGURE 2. BBIL reference points of the environment constructed by

the fingerprint method and combined with machine learning 43 training trajectories.
for BLE wave based positioning. In that study, the author
uses two different neural networks to individually predict for problem is that when receiving waves, the fixed devices
each X and Y coordinate. The author tested with 4 machine have to sync the time with each other. The red circles rep-
learning algorithms: Artificial Neural Network Regression resent room-mounted BLE receivers for a total of 9 receivers.
(ANN), Multiple Linear Regression (MLR), Random For- Devices used in receiving stations are Raspberry Pi kits and
est Regression (RF), Support Vector Regression (SVR). The transmitters are BLE beacons. The exact locations of the
database was collected in a room with the size of 4m x 6m equipment were measured based on the landmarks attached
and the average positioning error was from 1.34m to 1.46m. to the room. The average distance between landmarks is
Most recently, Hoang et al. [29] used the Recurrent neural about 1.33m.
network (RNN) also with the fingerprint method to achieve To get sample points, transmitters were moved in the room
an average positioning error of 0.75m in a space with a ruler on random days and 3 times a day at any time. Figure 4 shows
of 16m x 21m. However, the RNN in the study assumes a summary of the RSSI sampling reference points after all.
that the object’s movement is smooth and the signal is not These points are spread across the positioning area and each
interrupted. The reason is that the RNN uses the current reference point gets an RSSI sample. The reference sampled
RSSI signal along with the entire distance the object has data set has 9211 points. In addition, the authors also built
traveled in the past to make predictions. The paper [30] pro- a test data set of 2107 points obtained after 11 moves with
posed the accuracy analysis of indoor Wi-Fi signal based on different orbits in the same room with the same equipment.
localization using two approach: Gaussian process regression However, in one trajectory, we found that there is a prob-
and K-nearest neighbor (KNN). According to the results, lem in connection of Raspberry Pi kits with 3 receivers
the author argues that the analysis of corridors using KNN so that the RSSI values do not change while moving so
could provide more accuracy than GPR for testing environ- that we considered only the 10 remaining trajectories in
ment. However, in terms of processing time, we note that evaluations.
GPR works much faster than KNN, especially for the dataset
containing huge number of RSSI reference points. This is the IV. LOCALIZATION ALGORITHM BASED MACHINE
reason for our research in this paper in combining the GPR LEARNING AND FINGERPRINT
and KNN to have a low computation cost why ensure the A. PROPOSED ALGORITHM
accuracy. For the earlier algorithms, where machine learning is not
used, the most common method for locating is kNN (k nearest
III. DATASET neighbor). The idea is to find k reference points that are
In order to compare our proposal to other methods with closest to the point to be located in terms of RSSI vector,
a fair assessment, we use a standard dataset with ground then the position is determined to be the weighted average
truth points measured accurately of other laboratory. The of the above points. A problem arises with this method for
selected dataset is BBIL (BLE Beacon Indoor Localization each location to be located, we have to scan all the reference
Dataset) [31]. points so the calculation time will be high when the reference
This database was collected from September 2018 to set is large. Alternatively, many authors recently use machine
May 2019. The size of the office space is roughly 11m x learning such as Recurrent Neural Network (RNN) [29] or
25m. In the experiment, the author group of the database Convolution Neural Network (CNN) with Gaussian Process
uses the model of fixed receiver and mobile transmitting Regression [32] to improve localization accuracy. RNN focus
station. This means that the device that needs to be located on the tracking problem so that it requires the first posi-
will broadcast, and the fixed device will receive. This role tion to be known. Otherwise, CNN with Gaussian regres-
change does not affect the positioning algorithm, only one sion requires a large training dataset with the reference

VOLUME 9, 2021 143797

 D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

FIGURE 3. Offline phase and Online phase of proposed system.

point spread in the location environment to have a high

accuracy.
In our research, we focus on the case when no prior pose is
known. This is the more popular case in practice. Our system
tries to attain two milestones: the first is to track current point
as fast as possible to have a real-time response and the second
work is to reconstruct a optimal trajectory. The offline phase
and online phase of our algorithm are shown in figure 3.
In offline phase, the RSSI fingerprint collected from training
trajectories are divided into training set and validation set.
Training set is used to train GPR in order to have a optimal
model which exhibits the smallest error on validation set.
In online phase, there are two threads executed in parallel.
Tracking thread takes RSSI vector of unknown position (UP)
to estimate current pose. In order to have real-time perfor-
mance, the core processing of tracking thread is a fusion
of GPR and kNN method. RSSI vector will be passed the
optimal trained GPR to have a predicted pose. Then, instead
of searching in the whole reference points, we only apply our
kNN method to search k nearest neighbor located in the circle
area centered in the predicted pose with radius r. The output
tracking pose is the weighted average of these k nearest
neighbor. Concurrently, Optimization thread will evaluate all
tracked poses with stricter condition to reconstruct an optimal
trajectory. Due to the fact that RSSI is noisy so that the result
of kNN method is not reliable at some points, the block ‘‘Reli-
able point detection’’ examines the confidence at each pose
based on the distribution of the k nearest neighbors. If the
pose is trustable, it will be added to a trust-list. The block ‘‘tra-
jectory reconstruction’’ will use this trust-list to re-optimize
trajectory.

B. FINGERPRINT FILTERING
The receivers station in the environment are denoted AP1 ,
FIGURE 4. RSSI filtering along the trajectory.
AP2 ,. . . . APn where n is number of station. Information gath-
ered from the APs are the RSSI value and MAC address of
that device. Since at each point, the vector RSSI is the set of the vectors is obtained. To get high positioning accuracy,
values obtained from all the APs, it is easy to see that the we need to have a sufficient number of APs covering that
larger the number of n, the higher the distinction between space.

143798 VOLUME 9, 2021

D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

Given that RSSI is noisy and it is collected while mobile that x is multi-dimensions but the value of y in formula 6 is
device is moving, the raw values of RSSI have a huge scalar so that we apply GPR separately for x-axis and y-axis
variation. The figure 4 show the variation of RSSI val- of 2D-pose. The GP defines posterior distributions over func-
ues of one receiver station along one of the training tra- tion from input data to output data f (x1 ), f (x2 ), f (x3 ), , f (xn )
jectories. We can see that those values have some critical obeys a multivariate Gaussian distribution shown in
changes even though two positions are very close to each formulas (7)-(9).
other. In many applications, to have a reliable RSSI value
at one location, the mobile device must be set at that posi- [f (x1), f (x2 ), f (x3 ), . . . , f (xn )] ∼ N (M , Kij ) (7)
tion for along time before being located. However, it is not M = E[f (x1 ), f (x2 ), . . . , f (xn )] (8)
suitable for application where mobile devices move con- Kij = k(xi , xj ) (9)
tinuously. In our implementation, to avoid critical changes
of RSSI, we propose the step of filtering. It means that where M is the mean and K is the kernel function of covari-
RSSI set of each training trajectory will be passed a low- ance. In the concept of kernel, there are three kinds of kernel
pass filter. We evaluated two kinds of filter: mean and that could be suitable for RSSI vectors such as Squared
median. Exponential (SE), Periodic (PER) and Matern. These kernel
functions are shown in formulas (10)-(13).
t
RSSIAPk t−w
= mean(RSSIAPk , . . . , RSSIAPk
t
, . . . , RSSIAPk
t+w
) (1)
t
RSSIAPk t−w
= median(RSSIAPk , . . . , RSSIAPk
t
, . . . , RSSIAPk
t+w
) ||x − x 0 ||
kse (x, x 0 ) = s2f exp(− ) (10)
(2) 2l 2
0)
2sin2 (π (x−x p )
Formula (1) and (2) illustrate the filtering applied for one kper (x, x 0 ) = s2f exp(− ) (11)
√ l2 √
element of RSSI vector and it is similar to other elements.
t , RSSI t−w and RSSI t+w corresponds to the value 3d(x, x 0 ) 3d(x, x 0 )
RSSIAPk APk APk kmatern3/2 (x, x 0 ) = (1+ )exp(− ) (12)
received by receiver station APk at time t, t − w and t + w √ l l
respectively. We also call w as ‘‘window size’’ of filtering. 5d(x, x 0 )
kmatern5/2 = (1 +
The value of w must be selected so that all points in the l √
filter are located geographically close to each other. It is 5d(x, x 0 )2 5d(x, x 0 )
obviously that the value of w depends on the moving speed + )exp(− ) (13)
3l 2 l
of mobile device. In BBIL dataset, with maximum speed is
around 2m/s and sampling time-step is 0.5 second, we choose where s2f is the signal variance and is also considered as an
the value of w equal to 2. Figure 4b and 4c show the result output-scale amplitude and parameter l is the input length-
of median and mean filtering for the raw RSSI of figure 4a. scale that determines the strength of the correlation between
We can note that critical changes are decreased and RSSI inputs. Before training the GPR model, the signal variance
values are smoother. In practice, during the experiments, is set to 1.0, period is set to 1.0 and the length-scale is
mean filter exhibits slightly more precise than median filter set to 0.2. Since our system uses RSSI vector to give the
but we found that localization performances are not much first predicted 2D-pose, the training set for regression was
different. built:
• Input data: N elements of m-dimensional RSSI vectors.
C. GAUSSIAN PROCESS REGRESSION N is number of reference points. In BBIL dataset, N gets
Generally, a Gaussian process (GP) is usually specified by the value of 9211 and m is 9.
its mean function and covariance function [33]. The mean • Output data: N elements of 2D-pose corresponding
function µ(x) and the covariance function k(x, x 0 ) of a real respectively to RSSI vectors.
process f (x) and target value y were defined as in formulas The input data is high dimensional vectors leading to a com-
(3)-(6). plex nonlinear relationship with the output data. In previous
researches, GPR proved the natural advantage of solving this
f (x) ∼ GP(µ(x), k(x, x 0 )) (3) problem [32], [34], [35]. In practice, we also tried many
µ(x) = E[f (x)] (4) other regression methods such as Linear and Support vector
k(x, x 0 ) = cov(f (x), f (x 0 )) (5) machine to make a comparison. The training error shown
y = f (x) + ω (6) in figure 5 proved that GPR algorithm is a the best choice
when the problem is a non-linear model. We can note that on
A GP is defined as a collection of random variables. In this the filtered training dataset, the GPR with Rational Quadratic
paper, x is RSSI vector and y is the location coordinates kernel function has the best performance for X-axis. Oth-
in R. ω is additive Gaussian noise with zero mean and a erwise, for Y-axis, GPR with exponential kernel function is
known variance denoted as σn2 . In BBIL dataset, we have n = the best. Therefore, we use these two models to predict the
9211 pairs of training data, where each input data x is a signal corresponding coordinate axis. The result on figure 5 is also
vector and each output data y is 2D-localization pose. Noting a proof of the fingerprint filtering mentioned in IV-B when all

VOLUME 9, 2021 143799

 D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

the training errors applied on filtered RSSI dataset are lower Otherwise, when K nearest-points are in widespread distri-
than that applied on raw RSSI dataset. bution as shown in figure 6b, weighted pose is not reliable.
The block ‘‘Reliable point detection’’ realizes this evalua-
D. TRAJECTORY RECONSTRUCTION tion at each tracking point. A nearest-point is considered as
In localization based fingerprint, to locate unknown point inlier when the distance from it to weighted pose is smaller
(UP), the traditional kNN algorithm computes the simi- than a threshold rc . If the number of inliers is bigger than
larity distance of RSSI vector between UP and reference σ K (σ =0.5 in our experiments), weighted pose is added to
point (RP). The similarity distance can be calculated by trustable-list.
formula (14). The trajectory is then optimized based on this trustable-list.
r The idea is that when a reliable-point j is detected, the sub-
Xm
DR−>U = (RSSIiRP − RSSIiUP )2 (14) trajectory from it to reliable-point j − 1 will be re-computed.
i=1 We suppose that there are h poses from reliable-point j − 1
(h−1)
where RSSIiRP is the ith component of the RSSI vector of the to reliable-point j denoted as P0j−1,j , P1j−1,j , P2j−1,j , . . . ., Pj−1,j
(h−1)
reference point and RSSIiUP is the ith element of the RSSI where P0j−1,j is the pose of reliable-point j − 1 and Pj−1,j is
vector of the point to be located. Then k RPs with the smallest the pose of reliable-point j. Noting that all poses of h point
similarity distances are chosen and the UP coordinates is are estimated from WkNN methods by searching in the local
calculated by the average of these RPs’ coordinates shown reference-point set of GPR prediction (radius r). We call
in formula (15): it the ‘‘large WkNN’’. To make the extrapolation in sub-
Xk trajectory, we apply two-directions ‘‘narrow WkNN’’. For
(x i , yi ) forward direction, we take the initial point at P0j−1,j then re-
i=1 RP RP
(xUP , yUP ) = (15) compute P1j−1,j by searching k nearest-points located around
k
In the WkNN algorithm, K RPs with the smallest spatial P0j−1,j within a radius rc (smaller than r) using RSSI vector.
1f
distances are selected as well, but with weighted average. The new value of P1j−1,j is denoted as Pj−1,j . The following
Considering that the distances from the each RP to the UP 2f
pose Pj−1,j is then computed by applying narrow WkNN
are different, the weights assigned to the reference points 1f
based on Pj−1,j . This forward extrapolation process is done
are different. The weights of the selected reference points (h−1)
can be calculated by formula (16); the weight is inversely until Pj−1,j . On the other side, we do the same work for
proportional to the distance DR−>U . The coordinate of the UP backward direction but from the initial point considered at
is calculated by formula (17). The WkNN algorithm normally Ph−1 0
j−1,j to last point Pj−1,j . The optimal pose at each point is
gives a better accuracy than the KNN so that it is frequently the mean of three estimations as shown in formula (18).
used in matching algorithm. if
Pij−1,j = mean(Pij−1,j , Pj−1,j , Pib
j−1,j ) (18)
1
DR−>U
ωiR−>U = Xk i
(16) if
where Pij−1,j , Pj−1,j , Pib
j−1,j are respectively the result of
DR−>U ‘‘large WkNN’’, ‘‘forward narrow WkNN’’ and ‘‘backward
i=1 i
Xk narrow WkNN’’.
(xUP , yUP ) = ωR−>U (xRP
i
, yiRP ) (17)
i=1 i
This traditional weighting is based on the similar- V. EXPERIMENT
ity of RSSI vector. Lately, Peng et al. [36] proposed an We evaluate our system on 10 test trajectories of BBIL
enhanced method by adding the physical distance to weight- dataset. All parameters are not changed for all trajectories.
ing. This physical distance is also computed from RSSI We use root mean square error (RMSE) and its 90th percentile
by the signal attenuation model. However, with noisy for comparison because it is commonly used by many other
signal due to fast moving, physical distance does not researches in the state of the art. In order to validate presented
work. It is even worst than traditional weighting in our proposals, different test scenarios are:
experiments. • kNN(raw signal): Applying traditional kNN algorithm
Our work uses traditional WkNN algorithm but we propose on raw RSSI signal available in dataset.
novel processing that taking into account the distribution of • WkNN(raw signal): Implementation of Weighted kNN
K nearest neighbours to evaluate the estimated UP pose. algorithm also on raw RSSI signals.
The reason is that in some case, some points of K nearest • kNN(filtered signal): Applying traditional kNN algo-
points are located quite far from weighted position which rithm on RSSI signals after filtering proposed in IV-B.
can be called outliers but some others (inliers) are located • WkNN(filtered signal): Weighted kNN on filtered RSSI
closely. Let’s consider two distributions example illustrated signals.
on figure 6, we found that the weighted pose will have • GPR+WkNN(filtered signal): The fusion of Gaussian
high possibility to be a good estimation if the majority of Process Regression and WkNN on filtered RSSI signals
K nearest-points located close to it as shown in figure 6a. without reconstruction.

143800 VOLUME 9, 2021

D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

FIGURE 5. RMSE error (in meter) of different regression models in training phase using raw RSSI dataset and filtered RSSI datasets.

prediction and WkNN is realized only on reference points

located in a limited region around this prediction instead of
searching in the whole reference set. We do not mention the
kNN in this table because it has a same algorithm complexity
with WkNN so it has a similar timing performance.
When trajectory reconstruction is applied, RMSE is even
better. The optimization helps RMSE to be decreased
between 10% and up to 38% in almost all the trajectories
except that in trajectory 6, RMSE of the reconstruction is
slightly bigger but it is not too severe (2.14m compared
to 2.08m). Figure 7 shows the number of trustable points
and untrustable points in each trajectory based on the eval-
uating condition presented in IV-D. Following this, table 3
presents a RMSE of two sub-sets: Trustable-points set and
Untrustable-points set in each trajectory. By observing table 1
and figure 7, the issue of accuracy degeneracy after recon-
struction in trajectory 6 can be explained by the reasons
that: 1) the result of WkNN is good enough in this trajec-
tory before reconstruction and 2) the number of trustable
FIGURE 6. Example of trustable (a) and Un-trustable (b) kNN distribution. points attains a high value of 75%. Therefore, doing opti-
mization on an already good result could lead to a small
degeneracy. On the other side, for other 9 trajectories, we can
• GPR+WkNN+Reconstruct: trajectory reconstruction easily see the advantage of optimization process. We can
after localization by GPR+WkNN. note that trajectory reconstruction is done in parallel thread
Table 1 shows a RMSE comparison of different scenarios. as illustrated in figure 3 hence it does not increase the
We can note that WkNN is slightly more precise than kNN execution time comparing to the version of GPR+WkNN
on all raw and filtered signals. Otherwise, filtering method only.
brings an obviously enhancement in accuracy when RMSE in In table 4, we group the result of all 10 trajectories and fig-
all trajectories are significantly reduced. The fusion of GPR ure 8 shows the CDF of Euclidean distance error. By observ-
and WkNN does not help to get a lower RMSE but it helps ing the CDF, we can notice that ∼ 90% of poses have the
to accelerate the processing. As shown in table 2, execution error smaller than 3m, 70% of poses having error within 2m
time for estimating location of one pose is reduced by half with the version of GPR+WkNN+Reconstruct. Other ver-
from 350ms to 180ms. The reason is that GPR gives a location sions have lower performance. In table 4, we make a com-

VOLUME 9, 2021 143801

 D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

TABLE 1. RMSE in meter of different scenarios using raw signal and filtered signal for each test trajectories.

TABLE 2. Mean of estimation time for each pose.

FIGURE 7. Report of the number of trustable-points and

untrustable-points in each trajectory.
FIGURE 8. CDF of Euclidean distance error.

TABLE 3. Mean RMSE comparison between trustable-points set and

untrustable-point set for each trajectory. TABLE 4. RMSE and 90% percentile for whole dataset. The result of
Trilateration, EN and LSTM methods (initial pose must be known) are
collected from author of BBIL dataset [31] on website.

parison to 3 others algorithm: Trilateration, Elman network

(EN) and Long Short Term Memory (LSTM). The result
of these 3 algorithms is presented by Kennedy et al. [31]-
author of BBIL dataset. Trilateration method used path loss
model to estimate distance from unknown pose to each BLE
beacon and make a multi-trilateration to obtain its loca- RSSI fingerprint. The RMSE of these 2 latters are low at
tion. This method is strongly influenced by noise so that respectively 1.62m and 1.54m. However, the problem of these
RMSE has a very high value at 6.32m. Otherwise, EN and 2 methods is that for each trajectory, the truth location for
LSTM are two algorithms using machine learning applied on the initial pose must be known which is not a popular case

143802 VOLUME 9, 2021

D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

FIGURE 9. Trajectories comparison.

VOLUME 9, 2021 143803

 D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

FIGURE 9. (Continued.) Trajectories comparison.

143804 VOLUME 9, 2021

D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

in reality. Moreover, the result of EN and LSTM depends [5] H. Q. Tran and C. Ha, ‘‘Fingerprint-based indoor positioning system using
strongly on hyper parameters of network model so it is not visible light communication—A novel method for multipath reflections,’’
Electronics, vol. 8, no. 1, p. 63, Jan. 2019.
stable for each launching. [6] Z. Y. Dong, W. M. Xu, and H. Zhuang, ‘‘Research on ZigBee indoor
Finally, figure 9 (at the end of the paper) shows result technology positioning based on RSSI,’’ Proc. Comput. Sci., vol. 154,
of trajectory built by our algorithm compared to ground- pp. 424–429, Jun. 2019.
[7] K.-H. Lam, C.-C. Cheung, and W.-C. Lee, ‘‘RSSI-based Lora localization
truth (orange curve). The result of our whole system is in
systems for large-scale indoor and outdoor environments,’’ IEEE Trans.
blue while the implementation of WkNN with raw signals Veh. Technol., vol. 68, no. 12, pp. 11778–11791, Dec. 2019.
is in red and the combination of GPR+WkNN on filtered [8] F. Liu, J. Liu, Y. Yin, W. Wang, D. Hu, P. Chen, and Q. Niu, ‘‘Survey on
signals (without optimization) is in green. From left-side WiFi-based indoor positioning techniques,’’ IET Commun., vol. 14, no. 9,
pp. 1372–1383, Jun. 2020.
to right-side, we can observe easily the enhancement of [9] L. Ruan, L. Zhang, T. Zhou, and Y. Long, ‘‘An improved Bluetooth indoor
our proposal by using signal filtering and then optimization positioning method using dynamic fingerprint window,’’ Sensors, vol. 20,
process. The trajectory after optimization is closer to the no. 24, p. 7269, Dec. 2020.
ground-truth. [10] K. Shailaja, B. Seetharamulu, and M. A. Jabbar, ‘‘Machine learning in
healthcare: A review,’’ in Proc. 2nd Int. Conf. Electron., Commun. Aerosp.
Technol. (ICECA), Mar. 2018, pp. 910–914.
VI. CONCLUSION [11] H. Yin and N. K. Jha, ‘‘A health decision support system for disease
diagnosis based on wearable medical sensors and machine learning ensem-
Experiments on the challenging BBIL database (room dimen- bles,’’ IEEE Trans. Multi-Scale Comput. Syst., vol. 3, no. 4, pp. 228–241,
sion is 10m x 25m) show that the implementation of machine Oct./Dec. 2017.
learning in positioning helps us to reduces the computation [12] D. E. Ruíz-Guirola, C. A. Rodríguez-López, S. Montejo-Sánchez,
R. D. Souza, and M. A. Imran, ‘‘DRX-based energy-efficient supervised
time of the kNN algorithm when the searching region is machine learning algorithm for mobile communication networks,’’ IET
narrowed. The computation time is hence decreased by half Commun., vol. 15, no. 7, pp. 1000–1013, Apr. 2021.
in estimating pose location from RSSI vector. Furthermore, [13] G. Yang, Y. Zhang, Z. He, J. Wen, Z. Ji, and Y. Li, ‘‘Machine-learning-
based prediction methods for path loss and delay spread in air-to-ground
the analysis in the distribution of k-nearest points will give millimetre-wave channels,’’ IET Microw., Antennas Propag., vol. 13, no. 8,
an effective evaluation for pose confidence. When mobile pp. 1113–1121, Jul. 2019.
device is moving continuously, the proposal of signal fil- [14] A. Sharma, A. Jain, P. Gupta, and V. Chowdary, ‘‘Machine learning appli-
tering is also a good method to reduce noise effect. The cations for precision agriculture: A comprehensive review,’’ IEEE Access,
vol. 9, pp. 4843–4873, 2021.
results indicated on 10 different test sequences of BBIL [15] M. N. Akhter, S. Mekhilef, H. Mokhlis, and N. M. Shah, ‘‘Review on
showed that even before trajectory reconstruction, the accu- forecasting of photovoltaic power generation based on machine learning
racy is significantly enhanced where RMSE is reduced from and Metaheuristic techniques,’’ IET Renew. Power Gener., vol. 13, no. 7,
pp. 1009–1023, May 2019.
3.64m to 2.54m when using filtered RSSI signals instead
[16] D. Schweizer, M. Zehnder, H. Wache, H.-F. Witschel, D. Zanatta, and
of the raw RSSI signals. Moreover, our proposal of opti- M. Rodriguez, ‘‘Using consumer behavior data to reduce energy consump-
mizing trajectory based on high confident poses gives an tion in smart homes: Applying machine learning to save energy without
lowering comfort of inhabitants,’’ in Proc. IEEE 14th Int. Conf. Mach.
attractive enhancement in accuracy. The RMSE after opti-
Learn. Appl. (ICMLA), Dec. 2015, pp. 1123–1129.
mization attains even a lower value at 1.78m. This error [17] M. A. Bhatti, R. Riaz, S. S. Rizvi, S. Shokat, F. Riaz, and S. J. Kwon,
is competitive with other tracking algorithms (EN, LSTM ‘‘Outlier detection in indoor localization and Internet of Things (IoT)
or GRU) which require the initial pose of trajectory to be using machine learning,’’ J. Commun. Netw., vol. 22, no. 3, pp. 236–243,
Jun. 2020.
known. [18] C.-H. Hsieh, J.-Y. Chen, and B.-H. Nien, ‘‘Deep learning-based indoor
In the future work, we intend to validate our system in other localization using received signal strength and channel state information,’’
environments with other configurations of RSSI fingerprint. IEEE Access, vol. 7, pp. 33256–33267, 2019.
Moreover, we will research the possibility of using Recurrent [19] B. Jia, B. Huang, H. Gao, and W. Li, ‘‘On the dimension reduction of
radio maps with a supervised approach,’’ in Proc. IEEE 42nd Conf. Local
Neural Network algorithm such as EN or LSTM in the extrap- Comput. Netw. (LCN), Oct. 2017, pp. 199–202.
olation step to improve the timing performance of trajectory [20] A. H. Salamah, M. Tamazin, M. A. Sharkas, and M. Khedr, ‘‘An enhanced
reconstruction. WiFi indoor localization system based on machine learning,’’ in Proc. Int.
Conf. Indoor Positioning Indoor Navigat. (IPIN), Oct. 2016, pp. 1–8.
[21] D. V. Le, N. Meratnia, and P. J. M. Havinga, ‘‘Unsupervised deep feature
REFERENCES learning to reduce the collection of fingerprints for indoor localization
using deep belief networks,’’ in Proc. Int. Conf. Indoor Positioning Indoor
[1] C.-H. Huang, L.-H. Lee, C. C. Ho, L.-L. Wu, and Z.-H. Lai, ‘‘Real-time Navigat. (IPIN), Sep. 2018, pp. 1–7.
RFID indoor positioning system based on Kalman-filter drift removal [22] D. Sun, E. Wei, Z. Ma, C. Wu, and S. Xu, ‘‘Optimized CNNs to indoor
and Heron-bilateration location estimation,’’ IEEE Trans. Instrum. Meas., localization through BLE sensors using improved PSO,’’ Sensors, vol. 21,
vol. 64, no. 3, pp. 728–739, Mar. 2015. no. 6, p. 1995, Mar. 2021.
[2] T.-H. Dao, M.-T. Le, and Q.-C. Nguyen, ‘‘Indoor localization system [23] X. Guo, L. Li, N. Ansari, and B. Liao, ‘‘Knowledge aided adaptive local-
using passive UHF RFID tag and multi-antennas,’’ in Proc. Int. Conf. Adv. ization via global fusion profile,’’ IEEE Internet Things J., vol. 5, no. 2,
Technol. Commun. (ATC), Oct. 2014, pp. 405–410. pp. 1081–1089, Apr. 2018.
[3] X. Yu, Q. Li, J. Peña Queralta, J. Heikkonen, and T. Westerlund, [24] M. AlHajri, N. Ali, and R. Shubair, ‘‘Indoor localization for IoT using
‘‘Applications of UWB networks and positioning to autonomous robots adaptive feature selection: A cascaded machine learning approach,’’ IEEE
and industrial systems,’’ 2021, arXiv:2103.13488. [Online]. Available: Antennas Wireless Propag. Lett., vol. 18, pp. 2306–2310, 2019.
http://arxiv.org/abs/2103.13488 [25] K. Liu, H. Zhang, J. K.-Y. Ng, Y. Xia, L. Feng, V. C. Lee, and A. S. H. Son,
[4] Ø. Volden, A. Stahl, and T. I. Fossen, ‘‘Vision-based positioning system for ‘‘Toward low-overhead fingerprint-based indoor localization via transfer
auto-docking of unmanned surface vehicles (USVs),’’ Int. J. Intell. Robot. learning: Design, implementation, and evaluation,’’ IEEE Trans. Ind. Infor-
Appl., vol. 2021, pp. 1–18, Aug. 2021. mat., vol. 14, no. 3, pp. 898–908, Mar. 2018.

VOLUME 9, 2021 143805

 D. D. Nguyen, M. T. Le: Enhanced Indoor Localization Based BLE Using GPR and Improved Weighted kNN

[26] M. Altini, D. Brunelli, E. Farella, and L. Benini, ‘‘Bluetooth indoor DAI DUONG NGUYEN received the Electrical
localization with multiple neural networks,’’ in Proc. IEEE 5th Int. Symp. Engineering degree in industrial information from
Wireless Pervasive Comput., May 2010, pp. 295–300. the Hanoi University of Science and Technology,
[27] P. Sthapit, H.-S. Gang, and J.-Y. Pyun, ‘‘Bluetooth based indoor positioning Vietnam, in 2014, and the M.S. degree in informa-
using machine learning algorithms,’’ in Proc. IEEE Int. Conf. Consum. tion system and technology and the Ph.D. degree
Electron., Jun. 2018, pp. 206–212. in robotics from Paris-Sud University, France,
[28] I. Alexander and G. P. Kusuma, ‘‘Predicting indoor position using Blue- in 2015 and 2018, respectively. He is currently a
tooth low energy and machine learning,’’ Int. J. Sci. Technol. Res., vol. 8,
Lecturer with the Department of Instrumentation
pp. 1661–1667, Jun. 2019.
and Industrial Informatics (3I), School of Electri-
[29] M. T. Hoang, B. Yuen, X. Dong, T. Lu, R. Westendorp, and
K. Reddy, ‘‘Recurrent neural networks for accurate RSSI indoor local- cal Engineering (SEE), Hanoi University of Sci-
ization,’’ IEEE Internet Things J., vol. 6, no. 6, pp. 10639–10651, ence and Technology (HUST). His research activities include localization-
Sep. 2019. based RF, visual SLAM, and real-time applications on embedded systems.
[30] M. M. Maw, H. M. N. M. Tint, and S. Duangsuwan, ‘‘Analysis of indoor wi-
fi localization using Gaussian process regression and K-nearest neighbor
algorithms,’’ UTK Res. J., vol. 14, no. 1, pp. 30–39, 2020.
[31] M. Kennedy, P. Spachos, and W. Graham Taylor, ‘‘BLE beacon indoor
localization dataset,’’ Univ. Guelph, Guelph, ON, Canada, Tech. Rep.,
2019. [Online]. Available: https://dataverse.scholarsportal.info/
dataset.xhtml?persistentId=doi:10.5683/SP2/UTZTFT
[32] G. Zhang, P. Wang, H. Chen, and L. Zhang, ‘‘Wireless indoor localization MINH THUY LE (Member, IEEE) received the
using convolutional neural network and Gaussian process regression,’’ Engineer and M.S. degrees in electrical engineer-
Sensors, vol. 19, no. 11, p. 2508, May 2019. ing from the Hanoi University of Science and
[33] C. K. Williams and C. E. Rasmussen, Gaussian Processes for Machine Technology (HUST), 2006 and 2008, respectively,
Learning, vol. 2. Cambridge, MA, USA: MIT Press, 2006. and the Ph.D. degree in optics and radio fre-
[34] Y. Zhao, C. Liu, L. S. Mihaylova, and F. Gunnarsson, ‘‘Gaussian processes
quency from the Grenoble Institute of Technol-
for RSS fingerprints construction in indoor localization,’’ in Proc. 21st Int.
ogy, France, in 2013. She is currently a Lecturer
Conf. Inf. Fusion (FUSION), Jul. 2018, pp. 1377–1384.
[35] X. Wang, X. Wang, S. Mao, J. Zhang, S. C. G. Periaswamy, and J. Patton, and also a Group Leader of the Radio Frequency
‘‘Indoor radio map construction and localization with deep Gaussian Group, Department of Instrumentation and Indus-
processes,’’ IEEE Internet Things J., vol. 7, no. 11, pp. 11238–11249, trial Informatics (3I), School of Electrical Engi-
Nov. 2020. neering (SEE), HUST. Her current research interests include built-in antenna,
[36] X. Peng, R. Chen, K. Yu, F. Ye, and W. Xue, ‘‘An improved weighted antenna array, beamforming, metamaterials, indoor localization and RF
K-nearest neighbor algorithm for indoor localization,’’ Electronics, vol. 9, energy harvesting, wireless power transfer, and autonomous wireless sensor.
no. 12, p. 2117, Dec. 2020.

143806 VOLUME 9, 2021