ArticlePDF Available

An Ensemble Filter for Indoor Positioning in a Retail Store Using Bluetooth Low Energy Beacons

Authors:

Abstract and Figures

This paper has developed and deployed a Bluetooth Low Energy (BLE) beacon-based indoor positioning system in a two-floor retail store. The ultimate purpose of this study was to compare the different indoor positioning techniques towards achieving efficient position determination of moving customers in the retail store. The innovation of this research lies in its context (the retail store) and the fact that this is not a laboratory, controlled experiment. Retail stores are challenging environments with multiple sources of noise (e.g., shoppers’ moving) that impede indoor localization. To the best of the authors’ knowledge, this is the first work concerning indoor localization of consumers in a real retail store. This study proposes an ensemble filter with lower absolute mean and root mean squared errors than the random forest. Moreover, the localization error is approximately 2 m, while for the random forest, it is 2.5 m. In retail environments, even a 0.5 m deviation is significant because consumers may be positioned in front of different store shelves and, thus, different product categories. The more accurate the consumer localization, the more accurate and rich insights on the customers’ shopping behavior. Consequently, retailers can offer more effective customer location-based services (e.g., personalized offers) and, overall, better consumer localization can improve decision making in retailing.
Content may be subject to copyright.
sensors
Article
An Ensemble Filter for Indoor Positioning in a Retail
Store Using Bluetooth Low Energy Beacons
Vasilis Stavrou 1, *, Cleopatra Bardaki 2, Dimitris Papakyriakopoulos 3and Katerina Pramatari 1
1Department of Management Science and Technology, Athens University of Economics and Business,
76 Patission Str., 10434 Athens, Greece; k.pramatari@aueb.gr
2Department of Informatics and Telematics, Harokopio University of Athens, 9 Omirou Str., 17778 Athens,
Greece; cleobar@hua.gr
3Department of Business Administration, University of West Attica, Campus 2, 250 Thivon & P. Ralli,
12241 Athens, Greece; dpapak@uniwa.gr
*Correspondence: stavrouv@aueb.gr
Received: 6 September 2019; Accepted: 18 October 2019; Published: 19 October 2019


Abstract:
This paper has developed and deployed a Bluetooth Low Energy (BLE) beacon-based indoor
positioning system in a two-floor retail store. The ultimate purpose of this study was to compare the
dierent indoor positioning techniques towards achieving ecient position determination of moving
customers in the retail store. The innovation of this research lies in its context (the retail store) and the
fact that this is not a laboratory, controlled experiment. Retail stores are challenging environments
with multiple sources of noise (e.g., shoppers’ moving) that impede indoor localization. To the best
of the authors’ knowledge, this is the first work concerning indoor localization of consumers in a
real retail store. This study proposes an ensemble filter with lower absolute mean and root mean
squared errors than the random forest. Moreover, the localization error is approximately 2 m, while
for the random forest, it is 2.5 m. In retail environments, even a 0.5 m deviation is significant because
consumers may be positioned in front of dierent store shelves and, thus, dierent product categories.
The more accurate the consumer localization, the more accurate and rich insights on the customers’
shopping behavior. Consequently, retailers can oer more eective customer location-based services
(e.g., personalized oers) and, overall, better consumer localization can improve decision making
in retailing.
Keywords:
Bluetooth Low Energy; indoor positioning; BLE Beacons; ensemble filter; fingerprinting;
retail store
1. Introduction
The advent of location-based technologies and the explosive growth of positioning equipment,
such as global positioning system (GPS) receivers, radio frequency identification (RFID), Wi-Fi and
Bluetooth, have facilitated tracking several objects moving through indoor or outdoor environments [
1
].
At the same time, the continuous evolution and usage of consumer electronics, such as mobile phones,
highlights a new infrastructure for location-based services, which are getting more and more attention
by academics and practitioners alike [
2
]. Location-based services identify the location of an object
(e.g., a person, a car, or a shopping cart) and utilize it to provide useful information and content to
the users of the service. These location-based services are applied to numerous fields and foster new
use-cases. For instance, they are used to provide navigation in shopping malls or public buildings, to
oer museum guidance, to ease product localization in supermarkets and indoor advertisements [3].
Thus, accurate object/human positioning is necessary for providing eective location-based
services. Although it is not so dicult and complex to locate the exact position of an object moving
Sensors 2019,19, 4550; doi:10.3390/s19204550 www.mdpi.com/journal/sensors
Sensors 2019,19, 4550 2 of 25
outdoors (e.g., using GPS signals and geofencing techniques [
4
]), this becomes a quite challenging
task in indoor environments. Indeed, the indoor positioning issue appears in the relevant literature as
a very challenging one and the achievement of high localization precision (accuracy) is a common
objective shared by various scholars [
5
,
6
]. Lymberopoulos et al. [
7
] argued that the indoor location
problem still remains unsolved and stresses the importance of the employment of a realistic approach
that would counterbalance the desired localization accuracy with low costs. Most studies have utilized
Wi-Fi technology for indoor localization experimenting with a vast variety of indoor positioning
techniques (e.g., [
8
11
]), while research on Bluetooth technology for indoor positioning is increasing
(e.g., [
4
,
12
15
]). However, most works are practically laboratory experiments (e.g., [
16
20
]) that have
not validated their indoor positioning approach in a real, fully operational environment.
This paper has developed a low-cost Bluetooth Low Energy (BLE) beacon-based indoor positioning
system and has put it into practice in a real grocery retail store with two floors of dierent layout
characteristics for shoppers’ positioning during their shopping trips. The ultimate purpose of this
study was to apply and compare dierent indoor positioning techniques towards achieving the most
possible ecient position determination of moving customers in the retail store. The key innovation
of this research lies in the context of this research (i.e., the retail store) and the fact that this is not a
laboratory experiment, namely a controlled context.
Retail stores are challenging environments that change often in terms of layout, product variety
etc. Moreover, there are multiple sources of noise in the stores (e.g., shoppers’ and stamoving,
products, mobile devices) that make indoor localization a challenging task. To the best of the authors’
knowledge, this is the first work concerning indoor localization of consumers in a fully operational
retail store.
Since indoor localization in retail environments is still unexplored, this study started by assessing
the performance of the most common techniques for indoor positioning (i.e., trilateration and
fingerprinting) and established classifiers. It was found that the random forest is the best classifier.
However, this study moved on to proposing and assessing an ensemble filter. The absolute mean
and root mean squared errors of the ensemble filter are significantly lower (40.7% and 18% lower,
respectively). More specifically, for approximately 70% of our cases (captured events of consumers), the
ensemble method results in a localization error of less than 1 m and in 80% of the cases, the localization
error is approximately 2 m. On the contrary, for the random forest, in 80% of the cases, the localization
error is approximately 2.5 m.
In retail environments in particular, such a deviation is significant, because even 0.5 m away from
the actual shopper’s position may lead to position him in a dierent shopping isle and in front of a
dierent store shelve, thus a dierent product category. Namely, the more accurate localization of
consumers, the more accurate and rich insights on the customers’ shopping behavior. Consequently,
the retailers and the marketing managers will be able to oer more eective customer location-based
services (e.g., personalized oers, coupons etc. reflecting their recent shopping trips). Overall, the
authors propose that it is worth moving forward from the random forest and proposing an ensemble
method because better consumer localization results in better decision making in the retail industry.
Next, in Section 2, the existing literature of indoor positioning techniques and technologies are
briefly reviewed, emphasizing on works that have applied Bluetooth technology and discussing the
gaps that this work attempts to address. Section 3presents the architecture of our BLE Beacon based
indoor positioning system and Section 4describes how it was deployed in the retail store. Section 5
presents the evaluation study of the indoor positioning techniques in the retail store, as well as the
authors’ proposed hybrid approach for determining the shoppers’ positions. In Section 6, the results
and the practical implications are briefly discussed, and in Section 7, the research limitations and plans
for further research are discussed.
Sensors 2019,19, 4550 3 of 25
2. Related Work
The popularization of location-based technologies, such as Wi-Fi, Bluetooth and radio-frequency
identification (RFID), has facilitated users to track and observe several objects moving into indoor
environments [
1
,
21
]. At the same time, massive amounts of data are generated while these objects
move around in various environments. However, several challenges may arise. A common challenge
regards how these data can be utilized to identify the position of an object within indoor environments.
Over the past years, researchers have proposed various techniques and approaches to tackle this issue
and eectively determine or track the location of an object in an indoor environment. These studies
can be classified into laboratory experiments (e.g., [
18
,
22
]) and real-world cases (e.g., [
9
,
23
]), where
indoor localization techniques are proposed for controlled and operational environments, respectively.
Although laboratory experiments achieve significant localization results, the proposed techniques do
not perform as eciently in real world scenarios. Thus, indoor positioning is a problem that remains
unsolved. In the following section, this study briefly presents the most common indoor positioning
techniques and provides an overview of the available studies that utilize wireless technologies for
localization purposes. We emphasize on works that have applied BLE beacons and the gaps that this
work attempts to address is discussed.
2.1. Indoor Positioning Techniques
Various techniques have been developed to determine the geometrical placement or position of
an object in an indoor environment. The location is estimated by taking into consideration the distance,
as well as the angle, between the transmitters and the tracked objects in the environment [
24
]. In
addition, the signal strength is used as a means of creating signatures for each location, in order to
identify the location of an object [
25
,
26
]. The most common techniques used for indoor localization are
triangulation and trilateration [
27
]. Albeit, techniques such as fingerprinting, proximity [
17
], dead
reckoning [
28
,
29
] and other machine learning-based approaches have been developed to mitigate
measurement errors. Each technique has certain advantages and disadvantages. Thus, the selection of
the most appropriate one highly depends on the application context, while in some cases, using more
than one technique and algorithms simultaneously may lead to better performance.
Fingerprinting is the most popular method of localization with high applicability in complex
environments due to its high accuracy and low complexity compared with other methods [
6
,
14
]. It
utilizes the received signal strength (RSS) as an approximate metric to determine the indoor location. To
apply the fingerprinting technique, first, the signal strengths are collected by the localization technology
(fingerprints) for each possible location of the environment in a database (oine phase) and, thus,
the reference fingerprinting map (RFM) [
30
] is formulated. Then, in the online phase, an object’s
signal collected from a location in real time is compared with the fingerprints in the RFM, to solve
for the location of the object [
18
,
30
]. In this phase, fingerprinting is based on classification algorithms
and methods, such as neural networks [
11
,
12
], decision trees [
23
], k-nearest neighbors [
31
], support
vector machines [
31
] and random forests [
32
], which predict the object’s current location based on the
fingerprint database [
33
,
34
]. Since fingerprinting relies on signal strength, the problems that may occur
are related to signal variations deriving from communication issues, such as fading, interference or
even from environmental factors.
This research began with tri- and multi-laterations, but the results were very inaccurate. This
confirms that they do not fit the BLE beacons due to the noise generated by context factors, hence,
signal strength is weaker than expected [
35
]. Next, fingerprinting along with the classification methods
are discussed and, finally, an ensemble method that improves indoor localization in the retail store
is proposed.
Sensors 2019,19, 4550 4 of 25
2.2. Wireless Technologies for Indoor Positioning
The technologies used to track an object while it is moving around in indoor environments are
mostly (a) Wi-Fi, a popular wireless networking technology [
36
]; (b) Bluetooth, a wireless technology
standard for communication over short distances [
37
]; (c) RFID, a technology based on radio-frequency
identification via electromagnetic fields that can identify and monitor tags that are attached to objects
of interest [
38
]. It is underlined that GPS technology is not appropriate for this purpose, as its signal
fails to reach indoor environments [27].
Indoor positioning with Wi-Fi faces the major problem of signal attenuations [
39
], which is
a common cause of faulty measurements. Apart from signal processing works [
40
], numerous
Wi-Fi-based studies have examined indoor positioning approaches [
31
], reporting either accuracy
improvement [8,41,42]
or applying machine learning techniques, such as decision trees [
23
],
unsupervised labelling on sequential data [
43
], unsupervised clustering for multi-floor environments [
9
]
and online sequential extreme learning [
44
,
45
]. Bayesian models have also been utilized eliminating
the problem of training data. Thus, they are more eective in terms of the cost of gathering data [
46
].
Other studies have applied fingerprinting [
33
,
47
] and classifiers to achieve eective and more accurate
Wi-Fi-based indoor positioning [11,48].
Table 1summarized the Wi-Fi studies with the employed technique.
Table 1. Overview of studies that utilize Wi-Fi technology for indoor positioning.
Authors Techniques
Lin and Lin (2005) Fingerprinting
Liu et al. (2007) Triangulation /Fingerprinting /Proximity
Yim J. (2008) Decision trees
Campos et al. (2014) Fingerprinting
Galván-Tejada et al. (2014) Random Forest /KNN/Neural Networks
Morales et al. (2014) Faulty measurements
Zou et al. (2014;2015) Online Sequential Extreme Learning
Torres-Sospedra et al. (2015) Fingerprinting
Perez et al. (2016) Unsupervised learning
Turgut et al. (2016) Trilateration/Fingerprinting/Proximity/PDR/Hybrid
Duque Domingo et al. (2017) Fingerprinting, Synchronized Euclidean distance
Cao et al. (2019) Fingerprinting
Wang et al. (2019) Fingerprinting/Clustering/signal weighted Euclidean distance
Zhang et al. (2019) Convolutional Neural Network/Gaussian Process Regression
Radio frequency identification (RFID) technology has also been utilized for indoor positioning
and involves curve fitting and location search [
18
], the hierarchical structure of classifiers [
3
] and the
Kalman-filter drift removal with Heron-bilateration location estimation [
49
]. In addition, algorithms
for higher localization accuracy [
5
,
50
,
51
] have been applied along with deployment innovations [
8
]
(i.e., a spinning RFID antenna).
Table 2summarized the Wi-Fi studies with the employed technique.
Table 2.
Overview of studies that utilize radio frequency identification (RFID) technology for
indoor positioning.
Authors Techniques
Zou et al. (2013) Extreme Learning
Keller et al. (2014) Classification
Montaser & Moselhi (2014) Triangulation/Proximity
Calderoni et al. (2015) Random Forest classifiers
Huang et al. (2015) Kalman-filter drift removal/Heron-bilateration
Wang et al. (2015) Curve fitting
Xu et al. (2017) Bayesian probability and K-Nearest Neighbor
Liu, F. and Zhong, D. (2018) Glowworm Swarm Optimization with
semi-supervised online sequential extreme learning
Shen et al. (2019) Angle of arrival method and spinning antenna
Sensors 2019,19, 4550 5 of 25
Apart from solely relying on one tracking technology, hybrid implementation approaches
have been found in the literature that combine dierent technologies, such as Wi-Fi and Bluetooth
technology [
52
,
53
]; Wi-Fi and RGB-D sensors [
54
] or Wi-Fi, Bluetooth, and long-term evolution (a 4G
wireless broadband technology) [55].
Next, the available studies that perform Bluetooth-based indoor positioning are discussed and
research gaps that this research addresses are identified.
2.3. Blueteooth Technology for Indoor Positioning
Bluetooth is a recent technology in the field of indoor positioning and has been applied both for
indoor localization, as well as for location proximity, presenting challenges that should be taken into
consideration [56].
More specifically, Bluetooth-based indoor positioning has been conducted by combining BLE
beacons with a pedestrian dead reckoning (PDR) technique to provide meter-level positioning and
estimate the current position by using a previously determined position [
57
]. Moreover, they have
developed a range-based localization system based on stigmergy that relies on the received signal
strength (RSS) of BLE beacon packets [
13
,
58
]. On the other hand, Diaz et al. [
59
] introduced an indoor
Bluetooth-based localization system (titled Bluepass) that achieves localization at a room-level based on
a signal coverage density method. Similarly, Bobek et al. [
60
] dealt with the issue of indoor positioning
by determining the location at room-level. Further, Liu et al. [
16
] designed and tested a smart home
indoor remote-control system that uses BLE beacons and geomagnetic sensors and proposed a learning
indoor location algorithm combining fingerprinting, geomagnetic sensing and PDR to guarantee the
accuracy of the location system.
Due to the nature of Bluetooth and the signals emitted, the most widely used and ecient
technique used for Bluetooth-based indoor positioning is fingerprinting (e.g., [
4
,
12
15
,
61
]). For
example, Subedi et al. [
14
] combined fingerprinting with weighted centroid localization aiming to
reduce the total number of collected fingerprints to improve the time required for the positioning
process. One of the most recent studies combined fingerprinting with geometric techniques to pinpoint
the movements of patients in a BLE beacon-enabled hospital room [15].
Moreover, when utilizing Bluetooth signals, a commonly used classification technique for indoor
positioning is neural networks. Mazan and Kovarova [
12
] collected fingerprints (signal strengths by
the beacon devices) and employed an artificial neural network to determine the user’s location. Thus,
employing multiple neural networks has achieved better accuracy by handling the variability of the
strength of the signals transmitted by Bluetooth devices [62].
Table 3summarized the Bluetooth studies with the employed technique for indoor positioning.
Overall, on the one hand, Wi-Fi technology has dominated the field of indoor positioning until
now utilizing a vast variety of techniques. On the other hand, Bluetooth technology has become
more mature and reliable for indoor positioning and the available studies are increasing. However,
most works are practically laboratory experiments (e.g., [1720]) that have not validated their indoor
positioning approach in a real, fully operational environment. Further, researchers have experimented
with indoor localization of subjects moving only in one-floor contexts (e.g., [
56
]). Moreover, most
studies employ a single positioning technique for performing indoor positioning (e.g., [
13
,
57
,
63
,
64
]).
They do not compare and evaluate dierent techniques or even combine them to aim for higher
eciency in indoor positioning. On the contrary, they have experimented with a hybrid of dierent
technologies (e.g., Galvan-Tejada et al. [53] proposed an algorithm to obtain the location of a receiver
combining Bluetooth and Wi-Fi technologies).
To fill these acknowledged gaps in the literature, this study developed and deployed a BLE
beacon-based indoor positioning system of shoppers moving in a retail store with two floors of dierent
layout characteristics. The authors experimented with dierent placement scenarios of the BLE beacons
and the ceiling of the store was identified as the most appropriate one. Next, the established localization
techniques were applied and evaluated with the purpose of achieving the most ecient position
Sensors 2019,19, 4550 6 of 25
determination of moving customers in the retail store. Tri- and multi-lateration was applied, but
the results were disappointing and, then, fingerprinting along with the classification methods were
assessed to determine the customer location at the store area-level. The authors did not consider the
two floors of the store split into equal smaller areas making a grid (i.e., fixed-length surface) to detect
the exact store area of the customer. Instead, the constraints of the store’s physical layout and the
retailer’s needs were identified and, thus, a more realistic, variable-length surface model was adopted
with store areas of dierent sizes and dimensions. Then, seven established classifiers were selected and
compared with their indoor localization performance in terms of the established metrics of accuracy,
precision, recall, F-measure (or F1 score) and Kappa statistic [
65
,
66
], as well as mean absolute error
and root mean squared error. The variance in parenthesis for each metric was also provided. The
random forest algorithm outperformed all the classifiers. A t-paired test was executed employing
the F-Measure and the random forest statistically outperformed the other classifiers. In essence, the
random forest combines a large number of classifiers targeting better classification results. Thus, it
inspired the authors to experiment with an ensemble classification approach targeting the improved
localization in the retail store. A hybrid approach was proposed that considers the three best classifiers
(random forest, K* and C4.5) based on the above assessment results and executes a voting process
taking a weighted vote of their predictions of the customers locations in the store. The absolute mean
and root mean squared errors of the ensemble method are significantly lower than the ones of the
random forest. Figure 1summarizes all the above phases of this study.
Sensors 2019, 19, x FOR PEER REVIEW 7 of 24
Figure 1. Methodology phases.
3.A BLE Beacon-Based Indoor Positioning System
In practice, the performance of current indoor positioning techniques was evaluated, as well as
the more efficient positioning approaches were pursued through an indoor positioning system that
the authors developed. The system developed relies on Bluetooth technology and, specifically, on
transmitter devices called beacons that emit radio signals to nearby devices. These signals contain
information regarding the identifier of the beacon and other data, such as its RSSI strength.
Figure 2 depicts the main modules of such indoor positioning systems that utilize BLE beacons:
(A) the BLE beacon environment; (B) the indoor positioning module that determines the user’s
position based on beacon data captured by the mobile application; (C) the indoor positioning data
analytics module that extracts and stores new knowledge from the collected data.
Figure 1. Methodology phases.
Sensors 2019,19, 4550 7 of 25
Table 3. Overview of studies that utilize Bluetooth technology for indoor positioning.
Authors Techniques
Altini et al. (2010) Neural Networks
Diaz et al. (2010) Signal Coverage Density Method
Subhan et al. (2011) Fingerprinting
Cabero et al. (2014) Proximity
Kim et al. (2015) Time windows and frequency
Li et al. (2015) Dead Reckoning
Palumbo et al. (2015) Stigmergy
Mazan and Kovarova (2015) Neural Networks
Bobek et al. (2015) Rule learning
Faragher and Harle (2015) Fingerprinting
Paek et al. (2016) Geometric Adjustment
Kriz et al. (2016) Fingerprinting
Castillo-Cara et al. (2017) Supervised learning
Liu et al. (2018) Fingerprinting, Dead Reckoning
Sung et al. (2018) Fingerprinting, Dead Reckoning, Kalman filter
Yohan et al. 2018 Trilateration, Least squares
Ferreira et al. (2018) Geometric Adjustment
Zuo et al. (2018) Graph optimization
Ke et al. (2018) Multilateration
Subedi & Pyun (2019) Fingerprinting
AL-Madani et al. (2019) Fuzzy logic
Mohsin et al. (2019) Fingerprinting
3. A BLE Beacon-Based Indoor Positioning System
In practice, the performance of current indoor positioning techniques was evaluated, as well as
the more ecient positioning approaches were pursued through an indoor positioning system that
the authors developed. The system developed relies on Bluetooth technology and, specifically, on
transmitter devices called beacons that emit radio signals to nearby devices. These signals contain
information regarding the identifier of the beacon and other data, such as its RSSI strength.
Figure 2depicts the main modules of such indoor positioning systems that utilize BLE beacons:
(A) the BLE beacon environment; (B) the indoor positioning module that determines the user’s position
based on beacon data captured by the mobile application; (C) the indoor positioning data analytics
module that extracts and stores new knowledge from the collected data.
Sensors 2019, 19, x FOR PEER REVIEW 7 of 24
Figure 1. Methodology phases.
3.A BLE Beacon-Based Indoor Positioning System
In practice, the performance of current indoor positioning techniques was evaluated, as well as
the more efficient positioning approaches were pursued through an indoor positioning system that
the authors developed. The system developed relies on Bluetooth technology and, specifically, on
transmitter devices called beacons that emit radio signals to nearby devices. These signals contain
information regarding the identifier of the beacon and other data, such as its RSSI strength.
Figure 2 depicts the main modules of such indoor positioning systems that utilize BLE beacons:
(A) the BLE beacon environment; (B) the indoor positioning module that determines the user’s
position based on beacon data captured by the mobile application; (C) the indoor positioning data
analytics module that extracts and stores new knowledge from the collected data.
Figure 2. Low-energy Bluetooth (BLE) beacon-based indoor positioning system overview.
Sensors 2019,19, 4550 8 of 25
Specifically, the first module (BLE Beacon environment) consists of the following two layers:
(I) The BLE beacon infrastructure layer concerns the beacons that are deployed in the indoor
environment (e.g., the beacons’ number and placement) and the data (e.g., beacon device identity)
that are transmitted and captured by the mobile applications of the moving users. It is practically the
setting upon which the system is deployed. The BLE beacons can be installed in various locations and
arrangements depending on the ultimate purpose of the system (e.g., indoor positioning, proximity
marketing etc.). Respectively, the BLE beacon manufacturer and the available sensors on the beacon
device prescribe the forms of data that the beacons provide, such as the identity (ID) of the beacon
device, the strength of the received signal (RSS), the distance between the beacon and the mobile device
of the user and more information, such as light strength and humidity or temperature.
(II) The mobile application layer refers to an application that captures the beacon data during the
indoor movement of the users. It transmits them via the internet to the back-end cloud infrastructure,
which stores the received data from the BLE beacons and the identity of the mobile application. Then,
the mobile application receives the user’s position in the indoor environment from the back-end
infrastructure each time that a request is sent with the captured information from the BLE beacons.
Specifically, each user’s visit in the indoor environment under study is recorded as a single user
session, which contains all the data gathered from the user’s mobile device while the user moves
indoors and passes through the installed beacons. The mobile application is programmed to capture
transmissions from nearby beacons in order to determine the user’s position. These snapshots are
called events and are contained in each session. Each captured event by the application contains (i)
the beacon ID (the unique identifier of the beacon), (ii) the signal strength (RSSI) that the application
captures from the specific beacon, in order to determine how close or how far the mobile device from
the transmitter is, and (iii) the distance between the beacon and the moving device. The application
can capture events from more than 6 beacons, which are then stored in the back-end’s database. The
data from the closest beacons to the mobile device are utilized to determine the user’s location.
Next, the indoor positioning module, implemented in the backend, first collects and stores the
pre-mentioned beacon data generated during the user’s sessions and the identity of the user ’s mobile
device. Then, the key component of this module, i.e., the indoor positioning technique, processes and
filters the beacon events’ data and determines the indoor position of the user’s device, namely the
user’s position.
Finally, the purpose of the indoor positioning data analytics module is to further analyze the
recorded data that reflects the users’ movements in the indoor environment as captured by the
beacon devices in collaboration with the mobile application. This module generates new insights and
knowledge describing the users’ indoor movements e.g., the users’ trac heat maps and the users’
navigation flows. The outcomes of this module can be exploited by other applications, such as a
security system of a building that needs to know the visitors trac patterns.
This research concerns the indoor positioning module and, specifically, the positioning technique
it executes. This study evaluates the available techniques and proposes ones that perform ecient
indoor positioning. To this end, a real BLE Beacon-based indoor positioning system in a grocery retail
store that monitors shoppers holding mobile devices was designed and deployed. The generated
beacon data was utilized to assess indoor positioning techniques and propose a hybrid approach that
improves indoor localization of consumers in the store.
4. Case Study: A BLE Beacon-Based Shopper Positioning System in a Grocery Retail Store
A BLE beacon-based indoor positioning system was deployed, such as the one described above,
in a grocery retail store. A retail store was selected as a test bed for evaluating the indoor positioning
techniques because a retail store is a challenging environment due to many factors that could aect the
localization accuracy, such as obstacles in the stores (e.g., walls or product signs), product materials
(e.g., liquids) that interact with the emitted signals or even the consumers moving in the store. The
more challenging the environment, the more generalizable and valid the evaluation results of the
Sensors 2019,19, 4550 9 of 25
indoor positioning techniques and the performance of the proposed positioning approach. Specifically,
the store application area consists of two floors (ground floor and first floor) with dierent layout
characteristics. This study installed 81 BLE Estimote beacon transmitters. The utilized Bluetooth
version was 4.2 LE standard, as it was supported by the deployed beacons. The more recent protocol
(i.e., Bluetooth 5) was not used as it was not supported by the beacon transmitters available.
Considering the mobile application layer of the shopper positioning system, a mobile application
was developed for shoppers to download in their Android smart phone devices. It interacts with the
beacons while the shopper moves in the store. During each shopper visit (i.e., a shopper session),
the mobile device captures all the data (events) from the beacons installed in the area into which
the shopper moves. Each event consists of the beacon ID, the beacon’s signal strength RSSI and the
distance between the beacon and the mobile phone. Several experiments were performed using event
data from dierent numbers of beacons to estimate the consumer’s in-store location. Utilizing the six
closest beacons to the consumer’s smart phone proved to be the most eective manner to identify
the shopper’s position. This is also aligned with the best practices described in Estimote’s forums
(https://forums.estimote.com).
Next, the system’s BLE beacon infrastructure layer is described, namely where the available BLE
beacons were placed in the store area. This is emphasized because it aects the accuracy of the captured
shopper data and, consequently, the performance of the indoor positioning technique that determines
the shopper’s indoor location
Map and BLE Beacons Placement
The system was deployed in a store area of two floors of 1200 m
2
each. On the one hand, these
two floors were selected due to the business purpose of the retailer, which concerned consumer trac
heatmaps showing how the consumers move within these two floors of the grocery store. Next, the
ground floor was divided into 23 areas and the first floor in 30 areas, based on the business purpose
and the technical eciency.
The literature proposes several approaches to determine an appropriate transmitter placement
within an indoor environment [
67
,
68
]. Drawing on them, this study aimed to place the available 81
beacon transmitters in such a manner that virtual perimeters were created around the consumers
per store area, in order for the location of the consumer to be determined more eciently. To this
end, dierent placement scenarios of the beacons were experimented with in order to detect the most
appropriate one.
On the one hand, various placements of the transmitters on the top of the store aisles were tested.
However, various obstacles (e.g., products that clerks placed over the transmitters during the shelf
replenishment process), which covered the beacons, prevented capturing the beacons’ signals. On the
other hand, alternative scenarios were experimented with by putting the transmitters on the ceiling.
Overall, the placement on the ceiling was more successful, as it was unlikely for beacons to be covered
by products, thus preventing their signal from being eciently emitted. Figures 3and 4present the
final beacons placement in the ground floor’s and the first floor’s ceiling, respectively. Each area in
a floor is indicated with a number (1–23 for ground floor, 1–30 for first floor) and each transmitter
is denoted with a red circle. The red circles filled with red color are the beacons in the elevator,
which are used to track the consumers’ transition between the two floors. This final deployment
facilitates the visibility of the moving shoppers from all the surrounding transmitters and ensures that
more accurate shopper event data are received by the indoor positioning module for determining the
shoppers’ location.
The next section presents the evaluation study of indoor positioning techniques using the beacon
data reflecting the shoppers’ movements in the retail store, as well as the proposed hybrid approach.
Sensors 2019,19, 4550 10 of 25
Sensors 2019, 19, x FOR PEER REVIEW 10 of 24
Figure 3. Map of the BLE beacon-enabled ground floor of the grocery store (store areas are labelled
with a number from 1 to 23, beacons are presented in red circles, elevator beacons are filled with red
color).
Figure 4. Map of the BLE beacon-enabled first floor of the grocery store (store areas are labelled with
a number from 1 to 30, beacons are presented in red circles).
The next section presents the evaluation study of indoor positioning techniques using the beacon
data reflecting the shoppers’ movements in the retail store, as well as the proposed hybrid approach.
5. Assessment of Indoor Positioning Techniques
This study tested and evaluated the most common and established localization techniques
referred to in Section 2.1, Indoor Positioning Techniques. At the beginning, tri- and multi-laterations
were applied, but they did not achieve satisfying results. Next, this study experimented with
fingerprinting. The following subsections present the evaluation results which motivated the
proposed approach that achieves more efficient position determination.
5.1. Trilateration
To perform trilateration for specifying the exact location of the consumer in the store, the exact
coordinates of all beacons/ transmitters were marked to calculate the consumer’s position.
Figure 3.
Map of the BLE beacon-enabled ground floor of the grocery store (store areas are labelled with
a number from 1 to 23, beacons are presented in red circles, elevator beacons are filled with red color).
Sensors 2019, 19, x FOR PEER REVIEW 10 of 24
Figure 3. Map of the BLE beacon-enabled ground floor of the grocery store (store areas are labelled
with a number from 1 to 23, beacons are presented in red circles, elevator beacons are filled with red
color).
Figure 4. Map of the BLE beacon-enabled first floor of the grocery store (store areas are labelled with
a number from 1 to 30, beacons are presented in red circles).
The next section presents the evaluation study of indoor positioning techniques using the beacon
data reflecting the shoppers’ movements in the retail store, as well as the proposed hybrid approach.
5. Assessment of Indoor Positioning Techniques
This study tested and evaluated the most common and established localization techniques
referred to in Section 2.1, Indoor Positioning Techniques. At the beginning, tri- and multi-laterations
were applied, but they did not achieve satisfying results. Next, this study experimented with
fingerprinting. The following subsections present the evaluation results which motivated the
proposed approach that achieves more efficient position determination.
5.1. Trilateration
To perform trilateration for specifying the exact location of the consumer in the store, the exact
coordinates of all beacons/ transmitters were marked to calculate the consumer’s position.
Figure 4.
Map of the BLE beacon-enabled first floor of the grocery store (store areas are labelled with a
number from 1 to 30, beacons are presented in red circles).
5. Assessment of Indoor Positioning Techniques
This study tested and evaluated the most common and established localization techniques referred
to in Section 2.1, Indoor Positioning Techniques. At the beginning, tri- and multi-laterations were
applied, but they did not achieve satisfying results. Next, this study experimented with fingerprinting.
The following subsections present the evaluation results which motivated the proposed approach that
achieves more ecient position determination.
Sensors 2019,19, 4550 11 of 25
5.1. Trilateration
To perform trilateration for specifying the exact location of the consumer in the store, the exact
coordinates of all beacons/transmitters were marked to calculate the consumer’s position. Subsequently,
based on the distance between the beacon and the shopper’s mobile phone indicated by each event,
the coordinates of the moving consumer were calculated.
More specifically, to calculate the position of the shopper, the following steps were performed:
(i) Bluetooth synchronization; (ii) time calculation; (iii) noise elimination. Regarding Bluetooth
synchronization, BLE beacons are to set to emit signals every 0.5 s. The emitted signal contains
information regarding the ID of the BLE beacon and when it is received by shopper’s mobile
application. The captured data include the ID, the received signal strength and the distance between
the mobile device and the BLE Beacon.
The mobile application synchronizes the received Bluetooth signals and records the top 10 BLE
beacons with the strongest signal strength that are closest to the shopper’s device. In order to avoid
outdated data, in the time calculation, the mobile application discards data that are older than the rest
of the captured.
This study began with event data from the three closest beacons in distance and more
beacons were experimented with until reaching six beacons (i.e., (multi)lateration) to calculate
the shopper’s coordinates.
The beacon transmitters used provided, via their software development kit (SDK) (https://
developer.estimote.com), the distance between the shopper’s device and the beacon. As a result,
trilateration was performed based on the respective distances.
As shown in Figure 5, the coordinates of each beacon (i.e.,
Beacon1
(
x1
,
y1
),
Beacon2
(
x2
,
y2
),
Beacon3
(
x3
,
y3
)) were utilized to determine the coordinates of the target point B, which indicates the
location of the shopper. In order to calculate the coordinates of B(
xb
), it is treated as the intersection of
3 circles, whose centers are the locations of the 3 beacons
Beaconi
(
xi
,
yi
), for i =1, 2, 3. The distances
between each beacon and the shopper are the radii of each circle. For each circle the equation is:
(xxi)2+(yyi)2=ri2.
By expanding out the squares for each one of the three equations:
x22x1x+x12+y22y1y+y12=r12(1)
x22x2x+x22+y22y2y+y22=r22(2)
x22x3x+x32+y22y3y+y32=r32(3)
Subtracting Equation (2) from Equation (1):
(2x1+2x2)x+ (2y1+2y2)y=r12r22x12+x22y12+y22(4)
Next, subtracting Equation (3) from Equation (2):
(2x2+2x3)x+ (2y2+2y3)y=r22r32x22+x32y22+y32(5)
Equations (4) and (5) form a system of two equations with two unknowns, x and y:
Ax +By =C
Dx +Ey =F
Sensors 2019,19, 4550 12 of 25
The solution of the system is:
x=CE FD
EA BD and y =CD AF
BD AE
Respectively for multilateration, a similar approach was followed for 6 beacons and, then, a linear
least squares solver was applied to get the (x,y) solution of the equations system.
Sensors 2019, 19, x FOR PEER REVIEW 12 of 24
Figure 5. Shopper position calculation using trilateration.
However, the results were very inaccurate. Trilateration and multilateration can hardly be used
with BLE beacons due to the noise generated by factors, such as product interference. Hence, the
signal strength, which can be easily affected, is weaker than expected [35]. The distance from a BLE
beacon is calculated based on a function that receives the signal strength as an input and provides
the estimated distance as the output. Thus, in most cases, the distance deviates from the expected and
it becomes impossible to have an accurate estimation.
Figure 6 presents two indicative evaluation cases of trilateration and multilateration in the store,
respectively. The blue arrow in both examples (a and b) highlights a shopper’s route that he moved
across aisle 6 and, then, returned to the point where the shopper initially started. In the first case
(Figure 6 (a)), the data from the closest (in distance) to the shopper’s mobile device three beacons was
utilized. Notably, the red dots, which are the outcome of the trilateration function, fall into an entirely
different aisle and even outside the boundaries of the floor (bottom-left red dots). The distance error
between the actual consumer’s position and the estimated varies from 5 to 10 m. In the second case
(Figure 6 (b)), the six closest beacons were used to identify the shopper’s location. Similarly, this study
found the consumer to be in entirely different areas than the actual ones (see the red dots). To improve
the results, additional logic was even employed for manipulating the trilateration data, such as
rejecting results (dots in the figures) that are not valid or have major deviations from the previous
estimations. Nonetheless, the results were not significantly improved, as they were already very
inaccurate. Thus, the authors realized that both trilateration and multilateration are not appropriate
approaches for accurate indoor positioning in a retail store.
Figure 5. Shopper position calculation using trilateration.
However, the results were very inaccurate. Trilateration and multilateration can hardly be used
with BLE beacons due to the noise generated by factors, such as product interference. Hence, the signal
strength, which can be easily aected, is weaker than expected [
35
]. The distance from a BLE beacon is
calculated based on a function that receives the signal strength as an input and provides the estimated
distance as the output. Thus, in most cases, the distance deviates from the expected and it becomes
impossible to have an accurate estimation.
Figure 6presents two indicative evaluation cases of trilateration and multilateration in the store,
respectively. The blue arrow in both examples (a and b) highlights a shopper’s route that he moved
across aisle 6 and, then, returned to the point where the shopper initially started. In the first case
(Figure 6a), the data from the closest (in distance) to the shopper’s mobile device three beacons was
utilized. Notably, the red dots, which are the outcome of the trilateration function, fall into an entirely
dierent aisle and even outside the boundaries of the floor (bottom-left red dots). The distance error
between the actual consumer’s position and the estimated varies from 5 to 10 m. In the second case
(Figure 6), the six closest beacons were used to identify the shopper’s location. Similarly, this study
found the consumer to be in entirely dierent areas than the actual ones (see the red dots). To improve
the results, additional logic was even employed for manipulating the trilateration data, such as rejecting
results (dots in the figures) that are not valid or have major deviations from the previous estimations.
Nonetheless, the results were not significantly improved, as they were already very inaccurate. Thus,
the authors realized that both trilateration and multilateration are not appropriate approaches for
accurate indoor positioning in a retail store.
Sensors 2019,19, 4550 13 of 25
Sensors 2019, 19, x FOR PEER REVIEW 13 of 24
Figure 6. Tri- and multi-lateration performance in the retail store with (a) 3 and (b) 6 beacons.
5.2. Fingerprinting
Then, the fingerprinting approach was applied by calculating the distance between the BLE
beacon and the tracked consumer, not on the x-y coordinates-level, but on area-level. To be more
specific, the beacons’ signal strength and the nearby beacon IDs were utilized to determine the
location of a consumer at the area-level, namely in the aforementioned store areas that were defined
with the beacons’ placements.
To formulate the reference fingerprinting map (RFM), two vectors were utilized that contain the
store areas and the BLE beacons. Vector A = { | i = 1,..,53} was used that contains the 53 store
areas were defined for both store floors and vector B = { | j = 1,..,81} that reflects the
exact locations of the deployed BLE beacons. For each of the areas, the RSS fingerprints were recorded
at time instances , with m = 1,..,M and were organized in the following fingerprinting map, that
contains the RSS values for all store areas from all BLE beacons at the time instant .
 = =


  


  
  


  
, m = 1,..,M , i = 1,..,23 , j = 1,..,81
The reference fingerprinting map RFM is a sparse array, as each area is not covered by all the
available BLE beacons. In addition, RFM contains the parameter of time. In addition, if the captures
of different timestamps are compared, how the shopper moved can be identified, based on the
captured RSS values. The matrix RFM was utilized for collecting the dataset of the offline
fingerprinting phase.
During the offline phase of fingerprinting, a dataset was collected that reflected the beacons’
signal strength in the store areas, as well as the distance of the tracked object (consumer’s mobile)
from a transmitter/ beacon. Two researchers were moving around in the two floors of the store for
two hours holding two mobile devices. The first device was a Samsung Galaxy J5 and the second one
was a Samsung A3 and both devices were using Android version 6.0.1. Further, 7.000 events
generated from the closest six beacons (in distance) to the researchers’ mobile device were gathered
during their simultaneous movements in the two floors of the store. The collected data was stored in
a table, where every row was a single event relating the store area where the researcher was with the
installed beacons (beacon IDs), the distance from each beacon/transmitter and the signal strength (see
Table 4).
Table 4. Structure of the collected dataset in the fingerprinting offline phase.
Beacon id data
Distance (meter)
Signal Strength (db)
Figure 6. Tri- and multi-lateration performance in the retail store with (a) 3 and (b) 6 beacons.
5.2. Fingerprinting
Then, the fingerprinting approach was applied by calculating the distance between the BLE
beacon and the tracked consumer, not on the x-y coordinates-level, but on area-level. To be more
specific, the beacons’ signal strength and the nearby beacon IDs were utilized to determine the location
of a consumer at the area-level, namely in the aforementioned store areas that were defined with the
beacons’ placements.
To formulate the reference fingerprinting map (RFM), two vectors were utilized that contain the
store areas and the BLE beacons. Vector A ={
areai
|i=1,
. . .
,53} was used that contains the 53 store
areas were defined for both store floors and vector B ={
beaconjxj,yj
|j=1,
. . .
,81} that reflects the
exact locations of the deployed BLE beacons. For each of the areas, the RSS fingerprints were recorded
at time instances
tm
, with m =1,
. . .
,M and were organized in the following fingerprinting map, that
contains the RSS values for all store areas from all BLE beacons at the time instant tm.
RFM(tm)= =
RSS1
1(tm)RSS1
2(tm)· · · RSS1
j(tm)
RSS2
1(tm)RSS2
2(tm)· · · RSS2
j(tm)
.
.
..
.
..
.
..
.
.
RSSi
1(tm)RSSi
2(tm)· · · RSSi
j(tm)
, m =1, . . . , M , i =1, . . . , 23 , j =1, . . . , 81
The reference fingerprinting map RFM is a sparse array, as each area is not covered by all the
available BLE beacons. In addition, RFM contains the parameter of time. In addition, if the captures of
dierent timestamps are compared, how the shopper moved can be identified, based on the captured
RSS values. The matrix RFM was utilized for collecting the dataset of the oine fingerprinting phase.
During the oine phase of fingerprinting, a dataset was collected that reflected the beacons’
signal strength in the store areas, as well as the distance of the tracked object (consumer’s mobile)
from a transmitter/beacon. Two researchers were moving around in the two floors of the store for two
hours holding two mobile devices. The first device was a Samsung Galaxy J5 and the second one was
a Samsung A3 and both devices were using Android version 6.0.1. Further, 7.000 events generated
from the closest six beacons (in distance) to the researchers’ mobile device were gathered during their
simultaneous movements in the two floors of the store. The collected data was stored in a table, where
every row was a single event relating the store area where the researcher was with the installed beacons
(beacon IDs), the distance from each beacon/transmitter and the signal strength (see Table 4).
Sensors 2019,19, 4550 14 of 25
Table 4. Structure of the collected dataset in the fingerprinting oine phase.
Beacon id Data Distance (meter) Signal Strength (db)
Area Beacon_id_1 . . . Beacon_id_6 Distance_1 . . . Distance_6 RSSI_1 . . . RSSI_6
12 139 . . . 138 2.24 . . . 4.92 90 . . . 100
24 164 . . . 203 1.06 . . . 4.87 81 . . . 101
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5 152 . . . 149 1.67 . . . 4.85 86 . . . 99
Then, the collected events were utilized as training data for training the classification algorithms
(classifiers) in predicting the store area where the consumer was located. The output of the algorithms
is the store area and the input are the afore-described collected data (beacon IDs, distance and signal
strength per store area) reflecting movements in the store.
Table 4exhibits the structure of the final data input of the dierent classification algorithms. The
store area is the dependent variable (output of the algorithm) and has 53 alternative values (class labels)
that correspond to the areas/regions which divided the two store floors (as Figures 3and 4depicts).
To ensure high internal validity of the experimentation, an ecient sampling process was needed.
Thus, sucient events were captured for all store areas based on their surface in m
2
. The larger areas
were represented by more events as seen in Table 5. The area size varies from relatively large (57 m
2
)
to small (9 m
2
) with the first floor having a grid layout, while the ground floor is a mix of grid and
open layouts. Next, the classification algorithms used are presented and their classification capabilities
are compared.
Table 5. Fingerprinting data sampling.
Area Size (m2)Number of Areas Store Areas Floor Captured Events
57.5 1 20 Ground 154
50 1 7 Ground 150
44 1 2 Ground 165
33 2 15,16 Ground 143
28 1 30 First 162
19.5 1 29 First 148
19 1 23 Ground 80
16 16 5-12,17-24 First 146
15.5 6 8-10,17-19 Ground 115
12.5 8 3-6,11-14 Ground 123
9.5 12 1-4,13-16,25-28 First 75
5 2 21,22 Ground 71
5.3. Fingerprinting: Classifiers’ Assessment and Comparison
The most commonly used classifiers were selected, which have already been mentioned. Indeed,
the following seven approaches were executed: (i) Naïve Bayes (NB) [
69
]; (ii) support vector machines
(SVM) [
70
]; (iii) logistic regression (LR) [
71
]; (iv) decision trees (C4.5) [
72
]; (v) multilayer perceptron
neural networks model (MLP) [
73
]; (vi) KStar (K*) [
74
] and (vii) random forests (RF) [
32
]. Naïve Bayes
was selected as it is one of the simplest classification algorithms with strong independence assumptions
between the features. Additionally, logistic regression was selected to examine the classification
problem as a generalized linear model and KStar, an entropy-based algorithm for investigating their
performance in such classification cases. Weka software was used for our experiments [75].
Sensors 2019,19, 4550 15 of 25
Specifically, this study assessed and compared the performance of the classifiers based on the
established metrics of accuracy, precision, recall and f-measure (or F1 score) [
65
]. Accuracy measures
the number of correct classifications performed by the classifier.
Accuracy =True Positive +True Negative
True Positive +True Negative +False Positive +False Negative
Precision indicates the exactness of the classifier, meaning that higher and lower precision leads to
less and more false positive classifications, respectively.
Precision =True Positive
True Positive +False Positive
Recall measures the classifier’s completeness. Higher and lower recall means less and more false
negative classifications (the captured events are not assigned as related to a store area, although they
should be), respectively.
Recall =True Positive
True Positive +False Negative
Precision and recall are increased at the expense of each other. Thus, they were combined to
produce the weighted harmonic mean of both metrics, which is the F-measure.
FMeasure =2Recall Precision
Recall +Precision
Further, the Kappa statistic metric was utilized. This measures how closely the captured events
classified by the algorithm match the true events, controlling for the accuracy of a random classifier
as measured by the expected accuracy [
66
]. In other words, it shows how much better the assessed
classifier is performing over a classifier that simply guesses at random according to the frequency of
each class.
The evaluation was performed via a 10-fold cross-validation [
76
], where the original
aforementioned collected events/data were randomly divided into ten equal subsets. Of these
ten sub-sets, one is retained as the validation test of the classifier and the remaining nine ones are
used as training data. Table 6includes the assessment results in terms of the above metrics and mean
absolute error and root mean squared error. The values in Table 6are the weighted averages of all
classes (i.e., the store areas) in order to acquire a more compact interpretation of the performance of
each classifier. The variance in parenthesis for each metric was also provided to depict how it fluctuates.
The results show that the random forest algorithm outperforms all the other algorithms in most of
the metrics.
Table 6. Assessment results of classifiers.
Classification Algorithms (Classifiers)
Assessment Metrics NB SVM LR C4.5 MLP K* RF
Accuracy 74.08%
(1.73)
85.11%
(1.48)
85.92%
(1.70)
86.84%
(1.36)
90.87%
(2.01)
93.51%
(1.33)
95.95%
(1.06)
F-Measure
0.744 (0.10) 0.850 (0.08) 0.859 (0.09) 0.868 (0.06) 0.908 (0.13) 0.935 (0.05) 0.959 (0.03)
Kappa statistic 0.7341
(0.02)
0.8477
(0.01)
0.8564
(0.02)
0.8651
(0.01)
0.9062
(0.01)
0.9336
(0.01)
0.9586
(0.01)
Mean Absolute Error 0.0101
(0.03)
0.0371
(0.02)
0.0056
(0.02)
0.0056
(0.01)
0.0043
(0.02)
0.0026
(0.01)
0.0086
(0.01)
Root Mean Squared
Error
0.08891
(0.03)
0.1351
(0.02)
0.0592
(0.02)
0.0662
(0.01)
0.0543
(0.02)
0.0439
(0.01)
0.0478
(0.01)
Sensors 2019,19, 4550 16 of 25
Moreover, a t-paired test was utilized to explore which classifiers were significantly better than
the others (a =0.05 level) [
77
]. The F-Measure was employed to compare them because it is more
reliable and provides a good trade-obetween precision and recall. Again, random forest statistically
outperforms the other classifiers, followed by K*, C4.5, multilayer perceptron, logistic regression, Naïve
Bayes and support vector machines in descending order. Random forest is by design an ensemble
learning method, meaning that it combines a large number of classifiers targeting better classification
results. Therefore, the performance of random forest inspired the authors to explore the potential of an
ensemble classification approach that follows the design principles of random forest.
5.4. Proposed Hybrid Approach: An Ensemble Filter
Instead of relying solely on the most ecient classifier for our case (as shown, that is the
random forest), a hybrid approach was proposed that executes a voting process involving a set of
multiple classifiers. In essence, this study suggests an ensemble method [
78
,
79
] that runs the three
significantly better classifiers based on their afore-assessed F-measure metric, assessed in the previous
step (Section 5.3), and takes a weighted vote of their predictions (see Figure 7).
Sensors 2019, 19, x FOR PEER REVIEW 16 of 24
Figure 7. The Proposed ensemble classifier.
More specifically, the proposed ensemble filter is a meta-classifier that performs weighted
majority voting among the three selected classifiers (i.e., C4.5, K* and random forest). The ensemble
filter predicts the class label y. To do so, a weight is associated with each classifier . The filter
formula is the following:
y = 
 
where is the classification function  and A is the set of unique class labels
(i.e., the store areas). The outcome of the formula is the class with the arguments with the greater
weight (i.e., argmax). In this case, the weights are assigned automatically via the Weka software.
Similar to the previous step, the three more efficient classifiers were utilized and a new classifier
was devised. The evaluation is performed as before via a 10-fold cross-validation [76] and the
performance of the suggested ensemble filter is presented in Table 7 compared to the three classifiers
used.
To assess the indoor positioning capability of our ensemble method, this study compared the
actual locations of the consumers in the store that the aforementioned collected data show with their
final predicted locations (estimations of store areas) by the proposed ensemble. Each classifier
predicts its output and then weighted majority voting is performed to predict the area of the shopper.
Table 7. Assessment results of the proposed ensemble method.
Classifiers
Assessment Metrics
C4.5
K*
RF
Ensemble
Accuracy
86.84% (1.36)
93.51% (1.33)
95.95% (1.06)
95.78% (1.00)
F-Measure
0.868 (0.06)
0.935 (0.05)
0.959 (0.03)
0.957 (0.02)
Kappa statistic
0.8651 (0.01)
0.9336 (0.01)
0.9586 (0.01)
0.9569 (0.01)
Mean absolute error
0.0056 (0.01)
0.0026 (0.01)
0.0086 (0.01)
0.0051 (0.01)
Root mean squared error
0.0662 (0.01)
0.0439 (0.01)
0.0478 (0.01)
0.0392 (0.01)
Table 7 shows that the ensemble method achieves a slightly lower accuracy than the best
classifier (i.e., random forest). Most importantly, the absolute mean and root mean squared errors are
significantly lower (40.7% and 18% lower, respectively). Lower errors mean that the result is closer
to the actual one, thus leading to more efficient location determination.
The absolute mean error (Figure 8) and root mean square error (Figure 9) of the three classifiers
and our filter were plotted to show their variations. Boxplots were used to depict the error variance
at each step of the 10-fold cross validation.
Figure 7. The Proposed ensemble classifier.
More specifically, the proposed ensemble filter is a meta-classifier that performs weighted majority
voting among the three selected classifiers (i.e., C4.5, K* and random forest). The ensemble filter
predicts the class label y. To do so, a weight
wj
is associated with each classifier
Cj
. The filter formula
is the following:
y=argmaxiXm
j=1wjXA(Cj(x) = i),
where
A
is the classification function
hCj(x)=iAi
and A is the set of unique class labels (i.e., the
store areas). The outcome of the formula is the class with the arguments with the greater weight (i.e.,
argmax). In this case, the weights are assigned automatically via the Weka software.
Similar to the previous step, the three more ecient classifiers were utilized and a new classifier was
devised. The evaluation is performed as before via a 10-fold cross-validation [
76
] and the performance
of the suggested ensemble filter is presented in Table 7compared to the three classifiers used.
To assess the indoor positioning capability of our ensemble method, this study compared the
actual locations of the consumers in the store that the aforementioned collected data show with their
final predicted locations (estimations of store areas) by the proposed ensemble. Each classifier predicts
its output and then weighted majority voting is performed to predict the area of the shopper.
Table 7shows that the ensemble method achieves a slightly lower accuracy than the best classifier
(i.e., random forest). Most importantly, the absolute mean and root mean squared errors are significantly
lower (40.7% and 18% lower, respectively). Lower errors mean that the result is closer to the actual one,
thus leading to more ecient location determination.
Sensors 2019,19, 4550 17 of 25
The absolute mean error (Figure 8) and root mean square error (Figure 9) of the three classifiers
and our filter were plotted to show their variations. Boxplots were used to depict the error variance at
each step of the 10-fold cross validation.
For a better visualization of the performance of all classifiers examined, this study decided to
utilize receiver operating characteristic (ROC) graphs, which is a technique usually adopted in machine
learning and data mining [
80
]. Figure 10 presents the trade-obetween the true positive and false
positive rate for each classifier. The important points in this graph are (a) point (0,0), which means that
the classifier commits no false positive errors, but also gains no true positives, (b) point (1,1) where the
classifier commits positive classifications unconditionally and (c) point (0,1), representing a perfect
classification [
80
]. A classifier is considered to perform better the more it moves to the north-west part
of the graph and also makes positive classifications only with strong evidence, as the false positive rate
gets close to the Y axis [80].
Table 7. Assessment results of the proposed ensemble method.
Classifiers
Assessment Metrics C4.5 K* RF Ensemble
Accuracy 86.84% (1.36) 93.51% (1.33) 95.95% (1.06) 95.78% (1.00)
F-Measure 0.868 (0.06) 0.935 (0.05) 0.959 (0.03) 0.957 (0.02)
Kappa statistic 0.8651 (0.01) 0.9336 (0.01) 0.9586 (0.01) 0.9569 (0.01)
Mean absolute error 0.0056 (0.01) 0.0026 (0.01) 0.0086 (0.01) 0.0051 (0.01)
Root mean squared error 0.0662 (0.01) 0.0439 (0.01) 0.0478 (0.01) 0.0392 (0.01)
Sensors 2019, 19, x FOR PEER REVIEW 17 of 24
Figure 8. Absolute mean error of the ensemble method and the three classifiers.
Figure 9. Root mean squared error of the ensemble method and the three classifiers.
For a better visualization of the performance of all classifiers examined, this study decided to
utilize receiver operating characteristic (ROC) graphs, which is a technique usually adopted in
machine learning and data mining [80]. Figure 10 presents the trade-off between the true positive and
false positive rate for each classifier. The important points in this graph are (a) point (0,0), which
means that the classifier commits no false positive errors, but also gains no true positives, (b) point
(1,1) where the classifier commits positive classifications unconditionally and (c) point (0,1),
representing a perfect classification [80]. A classifier is considered to perform better the more it moves
to the north-west part of the graph and also makes positive classifications only with strong evidence,
as the false positive rate gets close to the Y axis [80].
As shown, the random forest classifier and the ensemble approach are close to the upper left
point of the graph, namely they make positive predictions based on strong evidence, followed by K*
and C4.5. The rest of the assessed classifiers (i.e., multilayer perceptron, simple logistic, Naïve Bayes
Figure 8. Absolute mean error of the ensemble method and the three classifiers.
Sensors 2019,19, 4550 18 of 25
Sensors 2019, 19, x FOR PEER REVIEW 17 of 24
Figure 8. Absolute mean error of the ensemble method and the three classifiers.
Figure 9. Root mean squared error of the ensemble method and the three classifiers.
For a better visualization of the performance of all classifiers examined, this study decided to
utilize receiver operating characteristic (ROC) graphs, which is a technique usually adopted in
machine learning and data mining [80]. Figure 10 presents the trade-off between the true positive and
false positive rate for each classifier. The important points in this graph are (a) point (0,0), which
means that the classifier commits no false positive errors, but also gains no true positives, (b) point
(1,1) where the classifier commits positive classifications unconditionally and (c) point (0,1),
representing a perfect classification [80]. A classifier is considered to perform better the more it moves
to the north-west part of the graph and also makes positive classifications only with strong evidence,
as the false positive rate gets close to the Y axis [80].
As shown, the random forest classifier and the ensemble approach are close to the upper left
point of the graph, namely they make positive predictions based on strong evidence, followed by K*
and C4.5. The rest of the assessed classifiers (i.e., multilayer perceptron, simple logistic, Naïve Bayes
Figure 9. Root mean squared error of the ensemble method and the three classifiers.
Sensors 2019, 19, x FOR PEER REVIEW 18 of 24
and support vector machines) tend to have a lower performance than the first ones as their distance
from point (0,1) increases.
Figure 10. Classifiers receiver operating characteristic (ROC) space graph.
For further comparison between the two best methods (random forest and our ensemble
classifier), the multiclass classification of the consumer’s area was transformed into a binary
classification to evaluate the capability of only these two classifiers to correctly predict the consumer’s
position. Figure 11 presents the ROC curve of the random forest and ensemble classifiers and
demonstrates their training speed. The ensemble classifier is characterized by a better learning ability
as the training instances increase up to 0.65 of the true positive rate. From 0.65 to 0.80, both algorithms
tend to behave the same and above 0.8, the random forest is slightly better. When both algorithms
reach 0.95 of the true positive rate, they have a similar behavior, with the random forest being slightly
superior by 0.01.
Figure 11. ROC curve of random forest and the proposed ensemble method.
Finally, Figure 12 presents the overall cumulative probability and error correlation of the
proposed method and the random forest. Based on our experimentation, it was concluded that for
approximately 70% of the cases (captured events of consumers), the ensemble method results in a
localization less than 1 m and in 80% of the cases, the localization error is approximately 2 m. The
results confirm prior similar studies that try to improve the detection accuracy when using BLE
beacons. On the contrary, for the random forest classifier, in 80% of the cases the localization error is
approximately 2.5 m. In retail environments in particular, such deviation is significant, because even
0.5 m away from the actual shopper’s position may lead to position him in a different shopping isle
and in front of a different store shelve, thus a different product category. Considering the need of
Figure 10. Classifiers receiver operating characteristic (ROC) space graph.
As shown, the random forest classifier and the ensemble approach are close to the upper left point
of the graph, namely they make positive predictions based on strong evidence, followed by K* and
C4.5. The rest of the assessed classifiers (i.e., multilayer perceptron, simple logistic, Naïve Bayes and
support vector machines) tend to have a lower performance than the first ones as their distance from
point (0,1) increases.
For further comparison between the two best methods (random forest and our ensemble classifier),
the multiclass classification of the consumer’s area was transformed into a binary classification to
evaluate the capability of only these two classifiers to correctly predict the consumer’s position.
Figure 11 presents the ROC curve of the random forest and ensemble classifiers and demonstrates
their training speed. The ensemble classifier is characterized by a better learning ability as the training
instances increase up to 0.65 of the true positive rate. From 0.65 to 0.80, both algorithms tend to behave
Sensors 2019,19, 4550 19 of 25
the same and above 0.8, the random forest is slightly better. When both algorithms reach 0.95 of the
true positive rate, they have a similar behavior, with the random forest being slightly superior by 0.01.
Sensors 2019, 19, x FOR PEER REVIEW 18 of 24
and support vector machines) tend to have a lower performance than the first ones as their distance
from point (0,1) increases.
Figure 10. Classifiers receiver operating characteristic (ROC) space graph.
For further comparison between the two best methods (random forest and our ensemble
classifier), the multiclass classification of the consumer’s area was transformed into a binary
classification to evaluate the capability of only these two classifiers to correctly predict the consumer’s
position. Figure 11 presents the ROC curve of the random forest and ensemble classifiers and
demonstrates their training speed. The ensemble classifier is characterized by a better learning ability
as the training instances increase up to 0.65 of the true positive rate. From 0.65 to 0.80, both algorithms
tend to behave the same and above 0.8, the random forest is slightly better. When both algorithms
reach 0.95 of the true positive rate, they have a similar behavior, with the random forest being slightly
superior by 0.01.
Figure 11. ROC curve of random forest and the proposed ensemble method.
Finally, Figure 12 presents the overall cumulative probability and error correlation of the
proposed method and the random forest. Based on our experimentation, it was concluded that for
approximately 70% of the cases (captured events of consumers), the ensemble method results in a
localization less than 1 m and in 80% of the cases, the localization error is approximately 2 m. The
results confirm prior similar studies that try to improve the detection accuracy when using BLE
beacons. On the contrary, for the random forest classifier, in 80% of the cases the localization error is
approximately 2.5 m. In retail environments in particular, such deviation is significant, because even
0.5 m away from the actual shopper’s position may lead to position him in a different shopping isle
and in front of a different store shelve, thus a different product category. Considering the need of
Figure 11. ROC curve of random forest and the proposed ensemble method.
Finally, Figure 12 presents the overall cumulative probability and error correlation of the proposed
method and the random forest. Based on our experimentation, it was concluded that for approximately
70% of the cases (captured events of consumers), the ensemble method results in a localization less
than 1 m and in 80% of the cases, the localization error is approximately 2 m. The results confirm
prior similar studies that try to improve the detection accuracy when using BLE beacons. On the
contrary, for the random forest classifier, in 80% of the cases the localization error is approximately
2.5 m. In retail environments in particular, such deviation is significant, because even 0.5 m away
from the actual shopper’s position may lead to position him in a dierent shopping isle and in front
of a dierent store shelve, thus a dierent product category. Considering the need of retailers to use
such indoor positioning systems for knowing the actual shopping trips of customers and oering
them personalized services (e.g., promotions designed based on their route in the store), the lower
positioning error of our ensemble filter compared to the random forest is significant even though they
do not have significant dierences in terms of accuracy and F-measure.
Sensors 2019, 19, x FOR PEER REVIEW 19 of 24
retailers to use such indoor positioning systems for knowing the actual shopping trips of customers
and offering them personalized services (e.g., promotions designed based on their route in the store),
the lower positioning error of our ensemble filter compared to the random forest is significant even
though they do not have significant differences in terms of accuracy and F-measure.
Figure 12. Positioning error for the proposed ensemble filter.
6. Discussion
Most indoor localization studies are laboratory experiments (e.g., [5,19,20]). For example,
Subhan et al. [13] achieved a localization error to 2.67 m [13], while broader experimental
environments achieved 2.6 m [81]. However, high performance is not guaranteed in the respective
real, fully operational environment. However, this study’s approach concerns a real-world case with
much more BLE beacon transmitters and achieves better localization that varies from 2 to 2.5 m. These
results are encouraging considering such environments as retail stores impose several restrictions to
experimentation. For example, there are various sources of noise that affect the signal quality and, as
a result, the efficiency of indoor positioning (e.g., the moving shoppers and the density of shoppers
in the store affect the emitted signals). In addition, the mobile devices of the shoppers have different
kind of wireless cards and interpret signal strength differently. This may lead to misreading the signal
strength and, thus, to inaccurate data. Finally, the store layout and the contained products can affect
signal quality.
To the best of the authors’ knowledge, this is the first work concerning indoor localization of
consumers in a fully operational retail store with two floors. Yohan et al. [64] applied BLE technology
and proposed a secure indoor positioning-based mobile system for a retail store. They contributed
with an indoor positioning-based mobile payment authentication protocol and evaluated its security
strength. The proposed system applies two well-known indoor positioning algorithms (least squares
and trilateration) and they do not work on the improvement of indoor positioning of the customers.
On the contrary, this study focused on the achievement of the best possible indoor localization of
consumers. To this end, different indoor positioning techniques were evaluated and compared and,
finally, a hybrid approach was proposed that offers better indoor localization results.
Further, for indoor positioning, researchers usually split the initial surface into equal smaller
areas (fixed-length surface), thus forming a grid which can then be used to detect the exact area where
the object is located. However, as Wang et al. [18] suggested an important limitation of fingerprinting
which relies on the assumption that indoor (sub) areas are regularly formed (e.g., constitute a perfect
square). Thus, this study took under consideration the technology limitations of BLE beacons, as well
as the challenging physical layout of a retail store with different sized areas, and selected store areas
of different sizes and dimensions (variable-length surface). The special business needs of the retailer
guided the final segmentation of the areas and their dimensions. For example, the retailer is more
interested to know whether the consumers pass often and stay in store areas with popular, fast-selling
products.
Figure 12. Positioning error for the proposed ensemble filter.
Sensors 2019,19, 4550 20 of 25
6. Discussion
Most indoor localization studies are laboratory experiments (e.g., [
5
,
19
,
20
]). For example,
Subhan et al. [
13
] achieved a localization error to 2.67 m [
13
], while broader experimental environments
achieved 2.6 m [
81
]. However, high performance is not guaranteed in the respective real, fully
operational environment. However, this study’s approach concerns a real-world case with much
more BLE beacon transmitters and achieves better localization that varies from 2 to 2.5 m. These
results are encouraging considering such environments as retail stores impose several restrictions to
experimentation. For example, there are various sources of noise that aect the signal quality and, as a
result, the eciency of indoor positioning (e.g., the moving shoppers and the density of shoppers in
the store aect the emitted signals). In addition, the mobile devices of the shoppers have dierent
kind of wireless cards and interpret signal strength dierently. This may lead to misreading the signal
strength and, thus, to inaccurate data. Finally, the store layout and the contained products can aect
signal quality.
To the best of the authors’ knowledge, this is the first work concerning indoor localization of
consumers in a fully operational retail store with two floors. Yohan et al. [
64
] applied BLE technology
and proposed a secure indoor positioning-based mobile system for a retail store. They contributed
with an indoor positioning-based mobile payment authentication protocol and evaluated its security
strength. The proposed system applies two well-known indoor positioning algorithms (least squares
and trilateration) and they do not work on the improvement of indoor positioning of the customers.
On the contrary, this study focused on the achievement of the best possible indoor localization of
consumers. To this end, dierent indoor positioning techniques were evaluated and compared and,
finally, a hybrid approach was proposed that oers better indoor localization results.
Further, for indoor positioning, researchers usually split the initial surface into equal smaller areas
(fixed-length surface), thus forming a grid which can then be used to detect the exact area where the
object is located. However, as Wang et al. [
18
] suggested an important limitation of fingerprinting
which relies on the assumption that indoor (sub) areas are regularly formed (e.g., constitute a perfect
square). Thus, this study took under consideration the technology limitations of BLE beacons, as
well as the challenging physical layout of a retail store with dierent sized areas, and selected store
areas of dierent sizes and dimensions (variable-length surface). The special business needs of the
retailer guided the final segmentation of the areas and their dimensions. For example, the retailer
is more interested to know whether the consumers pass often and stay in store areas with popular,
fast-selling products.
Regarding the practical implications of this study, the proposed indoor positioning system can
be the cornerstone of various location-based services for consumers in retail stores. For example, the
ability of one-to-one customer tracking and identification, which was not available with technologies
such as Wi-Fi and RFID, can enable personalization and, thus, improved shopper experience by
providing personalized oers, coupons, or even content. Moreover, the analysis of the recorded
consumers’ movements in the store can provide new, valuable insights on the consumers’ behavior.
For example, heatmaps depicting the areas that the shoppers spend the most time, and/or the areas
that the most shoppers pass by can be very helpful for marketing and store managers for designing
eective marketing actions and in-store advertisements, or even identifying selling gaps by monitoring
the areas that shoppers spend considerable time, but finally do not purchase the products displayed
in them.
7. Conclusions
This paper addresses customers’ indoor positioning in a real grocery retail store with two
floors via a BLE beacon-based indoor positioning system. Targeting to achieve the most ecient
position determination of the moving customers, fingerprinting was first applied and, then, the
indoor localization performance of seven established classifiers was compared to determine the
customer location at the store area-level. The random forest algorithm outperformed all the classifiers
Sensors 2019,19, 4550 21 of 25
and inspired the authors to build an ensemble classification approach that can achieve even better
localization in the retail store. The absolute mean and root mean squared errors of this ensemble
method are significantly lower than the ones of the random forest (40.7% and 18% better, respectively).
This study is not without limitations. For example, the authors started by assessing the
performance of the most common techniques for indoor positioning (i.e., trilateration and fingerprinting).
Fingerprinting may be a widely used technique, however it faces several limitations when applied on
dynamically changing environments [
82
,
83
], such as retail stores. One of the major disadvantages of
fingerprinting is the maintenance and updating of the reference fingerprinting map, which is a labor
intensive and time-consuming task. Further disadvantages are the number of access points required to
perform ecient fingerprinting (in order to cover the store surface) and the eects of mobile devices
on the received signal strength, as the signal measurement at the same location by dierent devices
may vary significantly.
Inspired by the fingerprinting disadvantages, future research may concern the assessment of more
advanced approaches (e.g., [
30
,
61
]) that aim to eliminate noise and provide more ecient results. The
authors also plan to test a hybrid approach that relies both on BLE beacons and Wi-Fi technology with
the purpose of exploring the impact on the indoor localization eciency.
Author Contributions:
Conceptualization, V.S., D.P. and C.B.; formal analysis, V.S. and D.P.; methodology,
V.S., D.P. and C.B.; resources, V.S.; supervision, K.P.; visualization, V.S.; writing—original draft, V.S., C.B., D.P.;
writing—review & editing, C.B., V.S., D.P. and K.P.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
Jin, B.; Zhuo, W.; Hu, J.; Chen, H.; Yang, Y. Specifying and Detecting Spatio-Temporal Events in the Internet
of Things. Decis. Support Syst. 2013,55, 256–269. [CrossRef]
2.
ˇ
Civilis, A.; Jensen, C.S.; Pakalnis, S. Techniques for Ecient Road-Network-Based Tracking of Moving
Objects. IEEE Trans. Knowl. Data Eng. 2005,17, 698–712. [CrossRef]
3.
Calderoni, L.; Ferrara, M.; Franco, A.; Maio, D. Indoor Localization in a Hospital Environment Using Random
Forest Classifiers. Expert Syst. Appl. 2015,42, 125–134. [CrossRef]
4.
Kriz, P.; Maly, F.; Kozel, T. Improving Indoor Localization Using Bluetooth Low Energy Beacons. Mob. Inf.
Syst. 2016,2016, 2083094. [CrossRef]
5.
Liu, F.; Zhong, D. GSOS-ELM: An RFID-Based Indoor Localization System Using GSO Method and
Semi-Supervised Online Sequential ELM. Sensors 2018,18, 1995. [CrossRef]
6.
He, S.; Chan, S.H.G. Wi-Fi Fingerprint-Based Indoor Positioning: Recent Advances and Comparisons. IEEE
Commun. Surv. Tutor. 2016,18, 466–490. [CrossRef]
7.
Lymberopoulos, D.; Liu, J.; Yang, X.; Choudhury, R.R.; Handziski, V.; Sen, S.; Lemic, F.; Busch, J.; Jiang, Z.;
Zou, H.; et al. A Realistic Evaluation and Comparison of Indoor Location Technologies: Experiences and
Lessons Learned. In Proceedings of the 14th International Conference on Information Processing in Sensor
Networks, Seattle, WA, USA, 13–16 April 2015; ACM: New York, NY, USA, 2015; pp. 178–189. [CrossRef]
8.
Shen, L.; Zhang, Q.; Pang, J.; Xu, H.; Li, P.; Xue, D. ANTspin: Ecient Absolute Localization Method of RFID
Tags via Spinning Antenna. Sensors 2019,19, 2194. [CrossRef]
9.
Campos, R.S.; Lovisolo, L.; De Campos, M.L.R. Wi-Fi Multi-Floor Indoor Positioning Considering
Architectural Aspects and Controlled Computational Complexity. Expert Syst. Appl.
2014
,41, 6211–6223.
[CrossRef]
10.
Cao, J.; Sun, Y.; Kong, Y.; Qian, W. The Sensitivity of Grating-Based SPR Sensors with Wavelength Interrogation.
Sensors 2019,19, 405. [CrossRef]
11.
Galv
á
n-Tejada, C.E.; Garc
í
a-V
á
zquez, J.P.; Garc
í
a-Ceja, E.; Carrasco-Jim
é
nez, J.C.; Brena, R.F. Evaluation
of Four Classifiers as Cost Function for Indoor Location Systems. Procedia Comput. Sci.
2014
,32, 453–460.
[CrossRef]
12.
Mazan, F.; Kovarova, A. A Study of Devising Neural Network Based Indoor Localization Using Beacons:
First Results. Comput. Inf. Syst. J. 2015,19, 15–20.
Sensors 2019,19, 4550 22 of 25
13.
Subhan, F.; Hasbullah, H.; Rozyyev, A.; Bakhsh, S.T. Indoor Positioning in Bluetooth Networks Using
Fingerprinting and Lateration Approach. In Proceedings of the International Conference on Information
Science and Applications, Jeju Island, Korea, 26–29 April 2011; pp. 1–9. [CrossRef]
14.
Subedi, S.; Pyun, J.-Y. Practical Fingerprinting Localization for Indoor Positioning System by Using Beacons.
J. Sens. 2017,2017, 9742170. [CrossRef]
15.
Mohsin, N.; Payandeh, S.; Ho, D.; Gelinas, J.P. Study of Activity Tracking through Bluetooth Low Energy-Based
Network. J. Sens. 2019,2019, 6876925. [CrossRef]
16.
Liu, Q.; Yang, X.; Deng, L. An IBeacon-Based Location System for Smart Home Control. Sensors
2018
,18,
1897. [CrossRef] [PubMed]
17.
Yoon, S.; Kim, J.; Connolly, D.J. Understanding Motivations and Acceptance of Location-Based Services.
Int. J. Hosp. Tour. Adm. 2018,19, 187–209. [CrossRef]
18.
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]
19.
Paek, J.; Ko, J.; Shin, H. A Measurement Study of BLE IBeacon and Geometric Adjustment Scheme for Indoor
Location-Based Mobile Applications. Mob. Inf. Syst. 2016,2016, 8367638. [CrossRef]
20.
Ke, C.; Wu, M.; Chan, Y.; Lu, K. Developing a BLE Beacon-Based Location System Using Location Fingerprint
Positioning for Smart Home Power Management. Energies 2018,11, 3464. [CrossRef]
21.
Jovicic, A.; Li, J.; Richardson, T. Visible Light Communication: Opportunities, Challenges and the Path to
Market. IEEE Commun. Mag. 2013,51, 26–32. [CrossRef]
22.
Palipana, S.; Pietropaoli, B.; Pesch, D. Recent Advances in RF-Based Passive Device-Free Localisation for
Indoor Applications. Ad Hoc Netw. 2017,64, 80–98. [CrossRef]
23.
Yim, J. Introducing a Decision Tree-Based Indoor Positioning Technique. Expert Syst. Appl.
2008
,34,
1296–1302. [CrossRef]
24.
Curran, K.; Furey, E.; Lunney, T.; Santos, J.; Woods, D.; McCaughey, A. An Evaluation of Indoor Location
Determination Technologies. J. Locat. Based Serv. 2011,5, 61–78. [CrossRef]
25.
Lassabe, F.; Canalda, P.; Chatonnay, P.; Spies, F.; Baala, O. A Friis-Based Calibrated Model for WiFi Terminals
Positioning. In Proceedings of the Sixth IEEE International Symposium on a World of Wireless Mobile and
Multimedia Networks, Taormina-Giardini Naxos, Italy, 16 June 2005; pp. 382–387. [CrossRef]
26.
Youssef, A.M.; Youssef, M. A Taxonomy of Localization Schemes for Wireless Sensor Networks. In Proceedings
of the 2007 International Conference on Wireless Networks (ICWN), Las Vegas, NV, USA, 25–28 June 2007;
pp. 444–450.
27.
Turgut, Z.; Aydin, G.Z.G.; Sertbas, A. Indoor Localization Techniques for Smart Building Environment.
Procedia Comput. Sci. 2016,83, 1176–1181. [CrossRef]
28.
Pratama, A.R.; Hidayat, R. Smartphone-Based Pedestrian Dead Reckoning as an Indoor Positioning System.
In Proceedings of the 2012 International Conference on System Engineering and Technology (ICSET),
Bandung, Indonesia, 11–12 September 2012; pp. 1–6. [CrossRef]
29.
Sharp, I.; Yu, K. Sensor-Based Dead-Reckoning for Indoor Positioning. Phys. Commun.
2014
,13, 4–16.
[CrossRef]
30.
Zuo, Z.; Liu, L.; Zhang, L.; Fang, Y. Indoor Positioning Based on Bluetooth Low-Energy Beacons Adopting
Graph Optimization. Sensors 2018,18, 3736. [CrossRef]
31.
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]
32. Breiman, L. Random Forests, Machine Learning 45. J. Clin. Microbiol. 2001,2, 199–228.
33.
Li, J.; Gao, X.; Hu, Z.; Wang, H.; Cao, T.; Yu, L. Indoor Localization Method Based on Regional Division with
IFCM. Electronics 2019,8, 559. [CrossRef]
34.
Carbonell, J.G.; Michalski, R.S.; Mitchell, T.M. An Overview of Machine Learning. Mach. Learn.
1983
,1, 3–23.
[CrossRef]
35.
Dong, Q.; Dargie, W. Evaluation of the Reliability of RSSI for Indoor Localization. In Proceedings of
the 2012 International Conference on Wireless Communications in Underground and Confined Areas,
Clermont Ferrand, France, 28–30 August 2012; pp. 1–6.
36.
Youssef, M.A.; Agrawala, A.; Udaya Shankar, A. WLAN Location Determination via Clustering and
Probability Distributions. In Proceedings of the First IEEE International Conference on Pervasive Computing
and Communications, Fort Worth, TX, USA, 26 March 2003; pp. 143–150. [CrossRef]
Sensors 2019,19, 4550 23 of 25
37. Haartsen, J.C. The Bleutooth Radio System. IEEE Pers. Commun. 2000,7, 28–36. [CrossRef]
38.
Srivastava, B. Radio Frequency ID Technology: The next Revolution in SCM. Bus. Horiz.
2004
,47, 60–68.
[CrossRef]
39.
Morales, J.A.; Akopian, D.; Agaian, S. Faulty Measurements Impact on Wireless Local Area Network
Positioning Performance. IET Radar Sonar Navig. 2014,9, 501–508. [CrossRef]
40.
Dardari, D.; Closas, P.; Djuric, P.M. Indoor Tracking: Theory, Methods, and Technologies. IEEE Trans.
Veh. Technol. 2015,64, 1263–1278. [CrossRef]
41.
Wang, B.; Liu, X.; Yu, B.; Jia, R.; Gan, X. An Improved WiFi Positioning Method Based on Fingerprint
Clustering and Signal Weighted Euclidean Distance. Sensors 2019,19, 2300. [CrossRef] [PubMed]
42.
Torres-Sospedra, J.; Montoliu, R.; Trilles, S.; Belmonte,
Ó
.; Huerta, J. Comprehensive Analysis of Distance
and Similarity Measures for Wi-Fi Fingerprinting Indoor Positioning Systems. Expert Syst. Appl.
2015
,42,
9263–9278. [CrossRef]
43.
Perez, I.; Pinchin, J.; Brown, M.; Blum, J.; Sharples, S. Unsupervised Labelling of Sequential Data for Location
Identification in Indoor Environments. Expert Syst. Appl. 2016,61, 386–393. [CrossRef]
44.
Zou, H.; Jiang, H.; Lu, X.; Xie, L. An Online Sequential Extreme Learning Machine Approach to WiFi Based
Indoor Positioning. In Proceedings of the IEEE World Forum Internet Things, Seoul, Korea, 6–8 March 2014;
pp. 111–116. [CrossRef]
45.
Zou, H.; Lu, X.; Jiang, H.; Xie, L. A Fast and Precise Indoor Localization Algorithm Based on an Online
Sequential Extreme Learning Machine. Sensors 2015,15, 1804–1824. [CrossRef]
46.
Madigan, D.; Elnahrawy, E.; Martin, R.P.; Ju, W.H.; Krishnan, P.; Krishnakumar, A.S. Bayesian Indoor
Positioning Systems. In Proceedings of the IEEE 24th Annual Joint Conference of the IEEE Computer and
Communications Societies, Miami, FL, USA, 13–17 March 2005; Volume 2, pp. 1217–1227. [CrossRef]
47.
Lin, T.N.; Lin, P.C. Performance Comparison of Indoor Positioning Techniques Based on Location
Fingerprinting in Wireless Networks. In Proceedings of the 2005 International Conference on Wireless
Networks, Communications and Mobile Computing, Maui, HI, USA, 13–16 June 2005; Volume 2,
pp. 1569–1574. [CrossRef]
48.
Villarubia, G.; Rubio, F.; De Paz, J.F.; Bajo, J.; Zato, C. Applying Classifiers in Indoor Location System. In
Trends in Practical Applications of Agents and Multiagent Systems; Springer: Cham, Switzerland, 2013; pp. 53–58.
49.
Huang, C.H.; Lee, L.H.; Ho, C.C.; Wu, L.L.; Lai, Z.H. Real-Time RFID Indoor Positioning System Based on
Kalman-Filter Drift Removal and Heron-Bilateration Location Estimation. IEEE Trans. Instrum. Meas.
2015
,
64, 728–739. [CrossRef]
50.
Xu, H.; Ding, Y.; Li, P.; Wang, R.; Li, Y. An RFID Indoor Positioning Algorithm Based on Bayesian Probability
and K-Nearest Neighbor. Sensors 2017,17, 1806. [CrossRef]
51.
Zou, H.; Wang, H.; Xie, L.; Jia, Q.S. An RFID Indoor Positioning System by Using Weighted Path Loss and
Extreme Learning Machine. In Proceedings of the 2013 IEEE 1st International Conference on Cyber-Physical
Systems, Networks, and Applications, Taipei, Taiwan, 19–20 August 2013; pp. 66–71. [CrossRef]
52.
Vu, L.; Do, Q.; Nahrstedt, K. Jyotish: A Novel Framework for Constructing Predictive Model of People
Movement from Joint Wifi/Bluetooth Trace. In Proceedings of the 2011 IEEE International Conference on
Pervasive Computing and Communications, Seattle, WA, USA, 21–25 March 2011; pp. 54–62. [CrossRef]
53.
Galv
á
n-Tejada, C.E.; Carrasco-Jim
é
nez, J.C.; Brena, R.F. Bluetooth-WiFi Based Combined Positioning
Algorithm, Implementation and Experimental Evaluation. Procedia Technol. 2013,7, 37–45. [CrossRef]
54.
Domingo, J.D.; Cerrada, C.; Valero, E.; Cerrada, J.A. An Improved Indoor Positioning System Using RGB-D
Cameras and Wireless Networks for Use in Complex Environments. Sensors
2017
,17, 2391. [CrossRef]
[PubMed]
55.
Mirowski, P.; Ho, T.K.; Yi, S.; MacDonald, M. SignalSLAM: Simultaneous Localization and Mapping with
Mixed WiFi, Bluetooth, LTE and Magnetic Signals. In Proceedings of the International Conference on Indoor
Positioning and Indoor Navigation, Montbeliard-Belfort, France, 28–31 October 2013; pp. 28–31. [CrossRef]
56.
Cabero, J.M.; Molina, V.; Urteaga, I.; Liberal, F.; Mart
í
n, J.L. Acquisition of Human Traces with Bluetooth
Technology: Challenges and Proposals. Ad Hoc Netw. 2014,12, 2–16. [CrossRef]
57.
Li, X.; Wang, J.; Liu, C. A Bluetooth/PDR Integration Algorithm for an Indoor Positioning System. Sensors
2015,15, 24862–24885. [CrossRef] [PubMed]
Sensors 2019,19, 4550 24 of 25
58.
Palumbo, F.; Barsocchi, P.; Chessa, S.; Augusto, J.C. A Stigmergic Approach to Indoor Localization Using
Bluetooth Low Energy Beacons. In Proceedings of the 2015 12th IEEE International Conference on Advanced
Video and Signal Based Surveillance, Karlsruhe, Germany, 25–28 August 2015. [CrossRef]
59.
Diaz, J.J.M.; De Maues, R.A.; Soares, R.B.; Nakamura, E.F.; Figueiredo, C.M.S. Bluepass: An Indoor
Bluetooth-Based Localization System for Mobile Applications. In Proceedings of the IEEE Symposium on
Computers and Communications, Riccione, Italy, 22–25 June 2010; pp. 778–783. [CrossRef]
60.
Bobek, S.; Grodzki, O.; Nalepa, G.J. Indoor Microlocation with BLE Beacons and Incremental Rule Learning.
In Proceedings of the 2015 IEEE 2nd International Conference on Cybernetics, Gdynia, Poland, 24–26 June
2015; pp. 91–96. [CrossRef]
61.
Sung, K.; Lee, D.K.R.; Kim, H. Indoor Pedestrian Localization Using Ibeacon and Improved Kalman Filter.
Sensors 2018,18, 1722. [CrossRef] [PubMed]
62.
Altini, M.; Brunelli, D.; Farella, E.; Benini, L. Bluetooth Indoor Localization with Multiple Neural Networks.
In Proceedings of the IEEE 5th International Symposium on Wireless Pervasive Computing, Modena, Italy,
5–7 May 2010; pp. 295–300. [CrossRef]
63.
Al-Madani, B.; Orujov, F.; Maskeli
¯
unas, R.; Damaševiˇcius, R.; Venˇckauskas, A. Fuzzy Logic Type-2 Based
Wireless Indoor Localization System for Navigation of Visually Impaired People in Buildings. Sensors
2019
,
19, 2114. [CrossRef]
64.
Yohan, A.; Lo, N.W.; Winata, D. An Indoor Positioning-Based Mobile Payment System Using Bluetooth Low
Energy Technology. Sensors 2018,18, 974. [CrossRef]
65.
Manning, C.D.; Raghavan, P.; Schütze, H. Introduction to Information Retrieval. Nat. Lang. Eng.
2010
,16,
100–103. [CrossRef]
66.
Nisbet, R.; Elder, J.; Miner, G. Handbook of Statistical Analysis and Data Mining Applications; Academic Press:
Cambridge, MA, USA, 2009.
67.
Chawathe, S.S. Beacon Placement for Indoor Localization Using Bluetooth. In Proceedings of the 11th
International IEEE Conference on Intelligent Transportation Systems, Beijing, China, 12–15 October 2008;
pp. 980–985. [CrossRef]
68.
Domingo-Perez, F.; Lazaro-Galilea, J.L.; Wieser, A.; Martin-Gorostiza, E.; Salido-Monzu, D.; de la Llana, A.
Sensor Placement Determination for Range-Dierence Positioning Using Evolutionary Multi-Objective
Optimization. Expert Syst. Appl. 2016,47, 95–105. [CrossRef]
69.
Lewis, D.D. Naive (Bayes) at Forty: The Independence Assumption in Information Retrieval. In European
Conference on Machine Learning; Springer: Berlin/Heidelberg, Germany, 1998; pp. 4–15. [CrossRef]
70.
Hearst, M.A.; Dumais, S.T.; Osuna, E.; Platt, J.; Scholkopf, B. Support Vector Machines. IEEE Intell. Syst.
Their Appl. 1998,13, 18–28. [CrossRef]
71.
Osborne, J. Best practices in logistic regression; Sage Publications: Thousand Oaks, CA, USA, 2014. [CrossRef]
72. Quinlan, J.R. C4. 5: Programs for Machine Learning; Elsevier: Amsterdam, the Netherlands, 2014.
73.
Kosko, B. Neural Networks and Fuzzy Systems: A Dynamical Systems Approach to Machine Intelligence/Book and
Disk; Prentice Hall: Upper Saddle River, NJ, USA, 1992.
74.
Cleary, J.; Trigg, L. K*: An Instance-Based Learner Using an Entropic Distance Measure. In Proceedings of
the 12th International Conference on Machine Learning, Tahoe City, CA, USA, 9–12 July 1995; pp. 108–114.
[CrossRef]
75.
Witten, I.H.; Frank, E.; Hall, M.A.; Pal, C.J. Data Mining: Practical Machine Learning Tools and Techniques;
Morgan Kaufmann: Burlington, MA, USA, 2016.
76.
Kohavi, R. A study of cross-validation and bootstrap for accuracy estimation and model selection. Ijcai
1995
,
14, 1137–1145.
77.
Salzberg, S.L. On Comparing Classifiers: Pitfalls to Avoid and a Recommended Approach. Data Min.
Knowl. Discov. 1997,1, 317–328. [CrossRef]
78.
Dietterich, T.G. Ensemble Methods in Machine Learning. In International Workshop on Multiple Classifier
Systems; Springer: Berlin/Heidelberg, Germany, 2000; pp. 1–15. [CrossRef]
79.
Kotsiantis, S.B.; Zaharakis, I.D.; Pintelas, P.E. Machine Learning: A Review of Classification and Combining
Techniques. Artif. Intell. Rev. 2006,26, 159–190. [CrossRef]
Sensors 2019,19, 4550 25 of 25
80. Fawcett, T. An Introduction to ROC Analysis. Pattern Recognit. Lett. 2006,27, 861–874. [CrossRef]
81.
Faragher, R.; Harle, R. Location fingerprinting with Bluetooth low energy beacons. IEEE J. Sel. Areas Commun.
2015,33, 2418–24248. [CrossRef]
82.
P
é
rez-Navarro, A.; Torres-Sospedra, J.; Montoliu, R.; Conesa, J.; Berkvens, R.; Caso, G.; Costa, C.; Dorigatti, N.;
Hern
á
ndez, N.; Knauth, S. Challenges of Fingerprinting in Indoor Positioning and Navigation. In Geographical
and Fingerprinting Data to Create Systems for Indoor Positioning and Indoor/Outdoor Navigation; Academic Press:
Cambridge, MA, USA, 2019; pp. 1–20. [CrossRef]
83.
Khalajmehrabadi, A.; Gatsis, N.; Akopian, D. Modern WLAN fingerprinting indoor positioning methods
and deployment challenges. IEEE Commun. Surv. Tutor. 2017,19, 1974–2002. [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/).
... In existing research, several innovative solutions based on BLE-based localization systems have been proposed and evaluated, including indoor guidance and location [12], indoor positioning in a retail store [13], driver assistance in indoor parking areas [14], navigation systems for car searching in indoor parking garages [15], the real-time tracking of construction laborers [16], location-enhanced activity recognition in indoor environments [17], multi-residential activity labeling for smart homes solutions [18], location system for smart home control [19], the navigation of visually impaired people in buildings [20], and elderly people monitoring [21]. ...
... Artificial intelligence-based localization techniques include methods such as artificial neural networks, neural fuzzy inference systems, and particle swarm optimization. Each technique has certain advantages and disadvantages, and the selection of the most appropriate one depends on the application context [13]. In existing research, studies have often combined different techniques (e.g., [19,[31][32][33]) that can provide better performance. ...
... The existing literature has provided a vast plethora of studies that have dealt with the performance evaluation or performance improvement of RSSI-based localization methods. BLE localization performance studies have mostly reported results from measurements in laboratory settings (e.g., [37,40,41]), although several studies have reported performance evaluation in real-world environments, such as museums [27], office environments (e.g., [14,[42][43][44]), a construction site [14], a retail store [13], and underground parking [28]. ...
Article
Full-text available
Advancements in protocols, computing paradigms, and electronics have enabled the development of wireless sensor networks (WSNs) with high potential for various location-based applications in different fields. One of the most important topics in WSNs is the localization in environments with sensor nodes being scattered randomly over a region. Localization techniques are often challenged by localization latency, efficient energy consumption, accuracy, environmental factors, and others. The objective of this study was to improve the technique for detecting the nearest Bluetooth Low Energy sensor, which would enable the development of more efficient mobile applications for location advertising at fairs, exhibitions, and museums. The technique proposed in this study was based on the iBeacon protocol, and it was tested in a controlled room with three environmental settings regarding the density of obstacles, as well as in a real-world setting at the Expo Museum at Postojna in Slovenia. The results of several independent measures, conducted in the controlled room and in the real-world environment, showed that the proposed algorithm outperformed the standard algorithm, especially in the environments with a medium or high densities of obstacles. The results of this study can be used for the more effective planning of placing beacons in space and for optimizing the algorithms for detecting transmitters in mobile location-based applications that provide users with contextual information based on their current location.
... Therefore, machine learning algorithms have been added in many studies to estimate the final position with improved accuracy. For example, Stavrou et al. [20] used a BLE beacon as an ensemble filter for indoor positioning. In the study, a fingerprint positioning combined with random forest algorithm was used in a retail store positioning application, and the positioning error results were 2 m-2.5 m. ...
... The authors noted that interference in the environment should be evaluated when performing indoor positioning, and the interference should be eliminated to achieve more satisfactory results. In the future, WiFi technology can be integrated to explore the effectiveness of positioning [20]. Ho and Chan [21] proposed a decentralized BLE-based positioning protocol. ...
Article
Full-text available
A rapid increase in the number of patients with dementia, particularly memory decline or impairment, has led to the loss of self-care ability in more individuals and increases in medical and social costs. Numerous studies, and clinical service experience, have revealed that the intervention of nonpharmacological management for people with dementia is effective in delaying the degeneration caused by dementia. Due to recent rapid developments in information and communications technology, many innovative research and development and cross-domain applications have been effectively used in the dementia care environment. This study proposed a new short-term memory support and cognitive training application technology, a “positioning and shadowing system,” to delay short-term memory degeneration in dementia. Training courses that integrate physical and digital technologies for the indoor location of patients with dementia were constructed using technologies such as Bluetooth Low Energy, fingerprint location algorithm, and short-range wireless communication. The Internet of Things was effectively applied to a clinical training environment for short-term memory. A pilot test verified that the results demonstrated learning effects in cognitive training and that the system can assist medical personnel in training and nursing work. Participants responded with favorable feedback regarding course satisfaction and system usability. This study can be used as a reference for future digital smart cognitive training that allows observation of the performance of patients with dementia in activities of daily living.
... The proposed system is based on the well-known fingerprint technique using commercial BLE beacons deployed at fixed positions in the scenario, and a person carrying a BLE scanner (which is the target to be localized) but, instead of localizing the target at grid point inside a room on a 2D plane [18,19], our system estimates the room where the person is. The BLE scanner periodically scans for advertisements of the surroundings BLE beacons and generates a fingerprint from a time interval. ...
... Then, based on these considerations, we heuristically set the BLE beacons' locations on the map, trying to establish intersection zones formed by at least three BLE beacons. In addition, unlike other works in which the beacons are placed on the ceiling [19,24], we consider placing the beacons only on the wall (which implies a restriction in the deployment). ...
Article
Full-text available
The ability to locate an object or a person at room-level inside a building or a house could have multiple applications. In this study, we adapt the fingerprint technique using Bluetooth Low Energy to locate the exact room of a person, seeking a simple and low-cost solution. The system is based on BLE beacons deployed at fixed positions and a person carrying a BLE scanner that generates fingerprints from the BLE beacons in coverage. We formulate it as a classification problem where each room is a class; the objective is to estimate the exact room, trying to maximize the area and number of rooms, but also trying to minimize the number of BLE beacons. The room estimation engine is based on a kNN (k-nearest neighbors) classifier. We evaluate the accuracy in two real scenarios and empirically measure the room estimation success related to the number of BLE beacons. As a proof-of-concept, a laptop and a Raspberry Pi are used as BLE scanners to test different hardware. We follow a measurement campaign for several days at different times to evaluate the stability and repeatability of the system. With just a few beacons an accuracy between 70 and 90% is achieved for house and university scenarios.
... At present, there are some mature methods for the study of plane localization: the global navigation satellite system (GNSS) [10,11] and China's BeiDou navigation satellite system (BDS) [12], which play a major role in outdoor positioning; the pseudolite indoor positioning [13,14] and the indoor positioning system using the beacon [15,16] and Bluetooth [17]; in order to achieve seamless indoor and outdoor positioning services, Chinese researchers are developing a BDS/GPS indoor positioning pseudo-satellite system; precise positioning technology using real-time kinematics (RTK) [18]; using natural or artificial landmarks for positioning, such as two-dimensional code positioning [19]; signal guidance and positioning based on the sensor signals of visual camera [20], RFID [21], ultrasonic [22] and LiDAR [23]; positioning systems using Wi-Fi [24,25] such as (OS-ELM) [26]; inertial navigation and positioning (MEMS) [27]; odometer positioning [28]; etc. ...
Article
Full-text available
This paper introduces a planar positioning sensing system based on orthogonal wheels and encoders for some surfaces that may float (such as ship decks). The positioning sensing system can obtain the desired position and angle information on any such ground that floats. In view of the current method of using the IMU gyroscope for positioning, the odometer data on these floating grounds are not consistent with the real-time data in the world coordinate system. The system takes advantage of the characteristic of the orthogonal wheel, using four vertical omnidirectional wheels and encoders to position on the floating ground. We design a new structure and obtain the position and angle information of a mobile robot by solving the encoder installed on four sets of omnidirectional wheels. Each orthogonal wheel is provided with a sliding mechanism. This is a good solution to the problem of irregular motion of the system facing the floating grounds. In the experiment, it is found that under the condition that the parameters of the four omnidirectional wheels are obtained by the encoder, the influence of the angle change of the robot in the world coordinate system caused by the flotation of the ground can be ignored, and the position and pose of the robot on the fluctuating ground can be well obtained. Regardless of straight or curved motion, the error can reach the centimeter level. In the mobile floating platform experiment, the maximum error of irregular movement process is 2.43 (±0.075) cm and the RMSE is 1.51 cm.
... For example, it can only discriminate between two classes where the data points are categorised by two classes in n-dimensional space, where n corresponds to the number of metabolites [100]. RF belongs to the family of classification trees and is found to be the best classifier [101,102]. RF has the strength to deal with large datasets without variable deletion, to provide a feature importance measure of the metabolite (mean decrease in accuracy) and a measure of the internal structure of the data (mean decrease in Gini index), and to handle missing values [84]. ...
Article
Full-text available
In the past decades, as an emerging omic, metabolomics has been widely used in meat science research, showing promise in meat quality analysis and meat authentication. This review first provides a brief overview of the concept, analytical techniques, and analysis workflow of metabolomics. Additionally, the metabolomics research in quality analysis and authentication of meat is comprehensively described. Finally, the limitations, challenges, and future trends of metabolomics application in meat quality analysis and meat authentication are critically discussed. We hope to provide valuable insights for further research in meat quality.
... The development and popularization of communication technology have allowed the integration of indoor positioning technologies in smartphones [1][2][3][4]. The positioning services can be used for multiple purposes such as finding cars in underground parking lots [5] and shopping in retail stores [6]. However, owing to complex indoor environments, using a single technology to achieve positioning in various places is a challenging concept. ...
Article
Full-text available
As communication technology and smartphones develop, many indoor positioning applications based on Bluetooth low energy (BLE) beacons have emerged. However, in a complex BLE network, it can be challenging to select the optimal reference beacon, and accurate positioning becomes difficult. Fortunately, if the BLE network is displayed on a map, we can intuitively grasp the structure and density of the beacons in each area, which is important information for accurate positioning. Therefore, in this study we developed a map-aided indoor positioning algorithm to model the relationship between beacons in the positioning area in a parking lot. Specifically, the algorithm split all beacons into multiple cell areas to find the optimal reference beacon in that area. Then, the optimal reference beacon is used to find the preferred reference beacons among the real-time beacons. Finally, the positioning results were calculated and evaluated according to the preferred beacons. According to the results, our method can optimize the selection of reference beacons in different areas. The average positioning accuracy was 2.09 m and the results can be scored accurately. The results verify that our algorithm can effectively use map information to guide the selection of reference beacons in complex environments.
... Enabled by the progress in indoor localization technologies and related electronic devices [2], the demand for precise localization systems is in the spotlight of many industrial fields. The number of use cases where indoor tracking or localization of an object is essential has been continuously increasing [3]. To name a few, the applications may target healthcare (e.g., patient localization), office (e.g., access to printer based on location), or home environment (e.g., automatic light control). ...
Article
Full-text available
The fingerprinting technique is a popular approach to reveal location of persons, instruments or devices in an indoor environment. Typically based on signal strength measurement, a power level map is created first in the learning phase to align with measured values in the inference. Second, the location is determined by taking the point for which the recorded received power level is closest to the power level actually measured. The biggest limit of this technique is the reliability of power measurements, which may lack accuracy in many wireless systems. To this end, this work extends the power level measurement by using multiple anchors and multiple radio channels and, consequently, considers different approaches to aligning the actual measurements with the recorded values. The dataset is available online. This article focuses on the very popular radio technology Bluetooth Low Energy to explore the possible improvement of the system accuracy through different machine learning approaches. It shows how the accuracy–complexity trade-off influences the possible candidate algorithms on an example of three-channel Bluetooth received signal strength based fingerprinting in a one dimensional environment with four static anchors and in a two dimensional environment with the same set of anchors. We provide a literature survey to identify the machine learning algorithms applied in the literature to show that the studies available can not be compared directly. Then, we implement and analyze the performance of four most popular supervised learning techniques, namely k Nearest Neighbors, Support Vector Machines, Random Forest, and Artificial Neural Network. In our scenario, the most promising machine learning technique being the Random Forest with classification accuracy over 99%.
... Considering that people spend more than 80% of their lifetime indoors [4], indoor positioning is at least as important as outdoor positioning. Additionally, different applications may benefit from indoor positioning, such as marketing and sales [5], health [6], or security and emergency solutions [7]. Thus, a number of indoor positioning technologies have been developed and are currently the subjects of further research. ...
Article
Full-text available
In this paper, we present a novel, low-cost approach to indoor localization that is capable of performing localization processes in real indoor environments and does not require calibration or recalibration procedures. To this end, we propose a single-anchor architecture and design based on an electronically steerable parasitic array radiator (ESPAR) antenna and Nordic Semiconductor nRF52840 utilizing Bluetooth Low Energy (BLE) protocol. The proposed algorithm relies on received signal strength (RSS) values measured by the receiver equipped with the ESPAR antenna for every considered antenna radiation pattern. The calibration-free concept is achieved by using inexpensive BLE nodes installed in known positions on the walls of the test room and acting as reference nodes for the positioning algorithm. Measurements performed in the indoor environment show that the proposed approach can successfully provide positioning results better than those previously reported for single-anchor ESPAR antenna localization systems employing the classical fingerprinting method and relying on time-consuming calibration procedures.
... The purpose of a trilateration technique is to calculate the coordinates (x, y) i.e., the intersection point of the three circles. Least Square (LSQ) method is an efficient way to determine the position of the mobile device [7]. The coordinates of each ESP32 receiver (i.e., ESP-1(x1, y1), ESP-2(x2, y2), ESP-3(x3, y3)) as shown in Fig. 1 are utilized to determine the coordinates of the BLE device (x, y), which indicates the location of the user. ...
... The purpose of a trilateration technique is to calculate the coordinates (x, y) i.e., the intersection point of the three circles. Least Square (LSQ) method is an efficient way to determine the position of the mobile device [7]. The coordinates of each ESP32 receiver (i.e., ESP-1(x1, y1), ESP-2(x2, y2), ESP-3(x3, y3)) as shown in Fig. 1 are utilized to determine the coordinates of the BLE device (x, y), which indicates the location of the user. ...
Article
Full-text available
Indoor positioning and navigation-based applications are emerging rapidly. Indoor navigation systems play a very useful role in huge buildings and offices when one needs to locate a room, a shop, or a part of the premises. The indoor routing mechanism gives directions towards selected destinations within a building with the help of an indoor navigation application. The positioning works on a "client-based application" which requires a smartphone with the mobile application used to determine the position directly. Outdoor navigation uses global positioning system (GPS) technology, however using GPS for tracking within a building is not feasible as reception rate is low and accuracy is in the range 30 meters due to indoor environmental factors. Thus, Bluetooth Low Energy (BLE)being a wireless technology is used for short range communication. Here its battery life is preferred over its high data transfer speeds. The position coordinates of the users are estimated using trilateration method where the relative distance from each receiver is calculated based on Received Signal Strength Indicator (RSSI) value. A very lightweight publish/subscribe messaging transport protocol: Message Queuing Telemetry Transport (MQTT) and a communication protocol showing the establishment and maintenance of network conversations by which application programs exchange data: Transmission Control Protocol (TCP) are useful for establishing wireless connections within an indoor location for sending packets of data. Using these technologies indoor navigation can be successfully achieved at a greater accuracy compared to GPS. Keywords: Global Positioning System (GPS), Received Signal Strength Indicator (RSSI), Bluetooth low energy (BLE), Message Queuing Telemetry Transport (MQTT) and Transmission Control Protocol (TCP) I. INTRODUCTION Location based services are rapidly evolving around the globe. Presently we have highly efficient outdoor location and navigation system based on Global Positioning System (GPS) technology. GPS technology cannot be implemented for positioning within indoor locations due to various shortcomings. It relies on uninterrupted satellite reception from the user, this cannot be achieved while the user is inside an enclosed building. This limitation of GPS calls for a new positioning system for indoor environment. Indoor positioning system has numerous sets of applications in the market where it can be used to navigate inside a huge building, track assets in a warehouse, improve security of a building, data analytics, etc. Variety of indoor positioning systems has been proposed in the past which use technologies like WIFI, RFID, NFC, and Bluetooth. Bluetooth low energy (BLE) is preferred over other technologies due to its low energy consumptions, higher accuracy, and low latency. Combination of BLE, Message Queuing Telemetry Transport (MQTT) with hardware modules make a good alternative for the situation addressed.Mobile devices with Bluetooth 5.0 chip support the advertising of small packets of data using Bluetooth low energy. This feature enables us to calculate the position of a device inside a building with the help of multiple Bluetooth receivers placed at strategic places in the building. The proposed work aims to achieve the same using path loss models of wireless communication and trilateration algorithm. This paper presents an Indoor navigation System where a mobile application helps the users to navigate inside a building. It provides a user friendly and reliable navigation system for indoor facilities using which the user can navigate through the whole facility without relying on any external maps or personnel. With the use of BLE and by processing most of the data on a single central processing unit, workload on user mobile devices is reduced hence mobile devices consume less energy making this system more energy efficient and user friendly. The paper is organized such that Section II discusses the background and related works, Section III describes the concepts and techniques used in detail, Section IV describes the block diagram and implementation, Section V showcases the obtained results, and the conclusions and future works are conferred in Section VI.
Article
Full-text available
With the development of wireless technology, indoor localization has gained wide attention. The fingerprint localization method is proposed in this paper, which is divided into two phases: offline training and online positioning. In offline training phase, the Improved Fuzzy C-means (IFCM) algorithm is proposed for regional division. The Between-Within Proportion (BWP) index is selected to divide fingerprint database, which can ensure the result of regional division consistent with the building plane structure. Moreover, the Agglomerative Nesting (AGNES) algorithm is applied to accomplish Access Point (AP) optimization. In the online positioning phase, sub-region selection is performed by nearest neighbor algorithm, then the Weighted K-nearest Neighbor (WKNN) algorithm based on Pearson Correlation Coefficient (PCC) is utilized to locate the target point. After the evaluation on the effect of regional division and AP optimization of location precision and time, the experiments show that the average positioning error is 2.53 m and the average computation time of the localization algorithm based on PCC reduced by 94.13%.
Article
Full-text available
WiFi fingerprint positioning has been widely used in the indoor positioning field. The weighed K-nearest neighbor (WKNN) algorithm is one of the most widely used deterministic algorithms. The traditional WKNN algorithm uses Euclidean distance or Manhattan distance between the received signal strengths (RSS) as the distance measure to judge the physical distance between points. However, the relationship between the RSS and the physical distance is nonlinear, using the traditional Euclidean distance or Manhattan distance to measure the physical distance will lead to errors in positioning. In addition, the traditional RSS-based clustering algorithm only takes the signal distance between the RSS as the clustering criterion without considering the position distribution of reference points (RPs). Therefore, to improve the positioning accuracy, we propose an improved WiFi positioning method based on fingerprint clustering and signal weighted Euclidean distance (SWED). The proposed algorithm is tested by experiments conducted in two experimental fields. The results indicate that compared with the traditional methods, the proposed position label-assisted (PL-assisted) clustering result can reflect the position distribution of RPs and the proposed SWED-based WKNN (SWED-WKNN) algorithm can significantly improve the positioning accuracy.
Article
Full-text available
The Global Positioning System (GPS) has been widely applied in outdoor positioning, but it cannot meet the accuracy requirements of indoor positioning. Comprising an important part of the Internet of Things perception layer, Radio Frequency Identification (RFID) plays an important role in indoor positioning. We propose a novel localization scheme aiming at the defects of existing RFID localization technology in localization accuracy and deployment cost, called ANTspin: Efficient Absolute Localization Method of RFID Tags via Spinning Antenna, which introduces a rotary table in the experiment. The reader antenna is fixed on the rotary table to continuously collect dynamic data. When compared with static acquisition, there is more information for localization. After that, the relative incident angle and distance between tags and the antenna can be analyzed for localization with characteristics of Received Signal Strength Indication (RSSI) data. We implement ANTspin using COTS RFID devices and the experimental results show that it achieves a mean accuracy of 9.34 cm in 2D and mean accuracy of 13.01 cm in three-dimensions (3D) with high efficiency and low deployment cost.
Article
Full-text available
The ability to precisely locate and navigate a partially impaired or a blind person within a building is increasingly important for a wide variety of public safety and localization services. In this paper, we explore indoor localization algorithms using Bluetooth Low Energy (BLE) beacons. We propose using the BLE beacon's received signal strength indication (RSSI) and the geometric distance from the current beacon to the fingerprint point in the framework of fuzzy logic for calculating the Euclidean distance for the subsequent determination of location. According to our results, the fingerprinting algorithm with fuzzy logic type-2 (hesitant fuzzy sets) is fit for use as an indoor localization method with BLE beacons. The average error of localization is only 0.43 m, and the algorithm obtains a navigation precision of 98.2 ± 1%. This precision confirms that the algorithms provide great aid to a visually impaired person in unknown spaces, especially those designed without physical tactile guides, as confirmed by low Fréchet and Hausdorff distance values and high navigation efficiency index (NEI) scores.
Article
Full-text available
This paper proposes a proof-of-concept, low-cost, and easily deployable Bluetooth low energy- (BLE-) based localization system which actively scans and localizes BLE beacons attached to mobile subjects in a room. Using the received signal strength (RSS) of a BLE signal and the uniqueness of BLE hardware addresses, mobile subjects can be identified and localized within the hospital room. The RSS measurement of the BLE signal from a wearable BLE beacon varies with distance to the wall-anchored BLE scanner. In order to understand and demonstrate the practicality of the relationship between RSS of a BLE beacon and the distance of a beacon from a scanner, the first part of the paper presents the analysis of the experiments conducted in a low-noise and nonreflective environment. Based on the analysis conducted in an ideal environment, the second half of the paper proposes a data-driven localization process for pinpointing the movements of the subject within the experimental room. In order to ensure higher accuracy like fingerprinting techniques and handle the increased number of BLE-anchored scanners like geometric techniques, the proposed algorithm was designed to combine the best aspects of these two techniques for better localization. The paper evaluates the effects of the number of BLE wall-mounted scanners and the number of packets on the performance of the proposed algorithm. The proposed algorithm locates the patient within the room with error less than 1.8 m. It also performs better than other classical approaches used in localization.
Article
Full-text available
In this paper, we derive the analytical expression for the sensitivity of grating-based surface plasmon resonance (SPR) sensors working in wavelength interrogation. The theoretical analysis shows that the sensitivity increases with increasing wavelength and is saturated beyond a certain wavelength for Au and Ag gratings, while it is almost constant for Al gratings in the wavelength range of 500 to 1000 nm. More importantly, the grating period (P) and the diffraction order (m) dominate the value of sensitivity. Higher sensitivity is possible for SPR sensors with a larger grating period and lower diffraction order. At long wavelengths, a simple expression of P/|m| can be used to estimate the sensor sensitivity. Moreover, we perform experimental measurements of the sensitivity of an SPR sensor based on an Al grating to confirm the theoretical calculations.
Article
Full-text available
In recent years, smart homes have begun to use various sensors to detect the location of users indoors. However, such sensors may not be stable, resulting in high detection error rates. Thus, how to improve indoor positioning accuracy has become an important topic. This study explored Bluetooth Low Energy (BLE) Beacon indoor positioning for smart home power management. A novel system framework using BLE Beacon was proposed to detect the user location, and to perform power management in the home through a mobile device application. Since the BLE Beacon may produce a multipath effect, this study used the positioning algorithm and hardware configuration to reduce the error rate. Location fingerprint positioning algorithm and filter modification were used to establish a positioning method for facilitating deployment, and to reduce the required computing resources. The experiments included an observation of the Received Signal Strength Indicators (RSSI) and selecting filters and a discussion of the relationship between the characteristics of the BLE Beacon signal accuracy and the number of the BLE Beacons deployed in the observation space. The BLE Beacon multilateration positioning was combined with the In-Snergy intelligent energy management system for smart home power management. The contribution of this study is to allow users to enjoy smart home services based on their location within the home using a mobile device application.
Article
Full-text available
Bluetooth Low-Energy (BLE) beacons-based indoor positioning is a promising method for indoor positioning, especially in applications of position-based services (PbS). It has low deployment cost and it is suitable for a wide range of mobile devices. Existing BLE beacon-based positioning methods can be categorized as range-based methods and fingerprinting-based methods. For range-based methods, the positions of the beacons should be known before positioning. For fingerprinting-based methods, a pre-requisite is the reference fingerprinting map (RFM). Many existing methods focus on how to perform the positioning assuming the beacon positions or RFM are known. However, in practical applications, determining the beacon positions or RFM in the indoor environment is normally a difficult task. This paper proposed an efficient and graph optimization-based way for estimating the beacon positions and the RFM, which combines the range-based method and the fingerprinting-based method. The method exists without need for any dedicated surveying instruments. A user equipped with a BLE-enabled mobile device walks in the region collecting inertial readings and BLE received signal strength indication (RSSI) readings. The inertial measurements are processed through the pedestrian dead reckoning (PDR) method to generate the constraints at adjacent poses. In addition, the BLE fingerprints are adopted to generate constraints between poses (with similar fingerprints) and the RSSIs are adopted to generate distance constraints between the poses and the beacon positions (according to a pre-defined path-loss model). The constraints are then adopted to form a cost function with a least square structure. By minimizing the cost function, the optimal user poses at different times and the beacon positions are estimated. In addition, the RFM can be generated through the pose estimations. Experiments are carried out, which validates that the proposed method for estimating the pre-requisites (including beacon positions and the RFM). These estimated pre-requisites are of sufficient quality for both range-based and fingerprinting-based positioning.
Article
Full-text available
With the rapid development of indoor positioning technology, radio frequency identification (RFID) technology has become the preferred solution due to its advantages of non-line-of-sight, non-contact and rapid identification. However, the accuracy of existing RFID indoor positioning algorithms is easily affected by the tag density and algorithm efficiency, and their environmental robustness is not strong enough. In this paper, we have introduced an RFID positioning algorithm based on the Glowworm Swarm Optimization (GSO) fused with semi-supervised online sequential extreme learning machine (SOS-ELM), which is called the GSOS-ELM algorithm. The GSOS-ELM algorithm automatically adjusts the regularization weights of the SOS-ELM algorithm through the GSO algorithm, so that it can quickly obtain the optimal regularization weights under different initial conditions; at the same time, the semi-supervised characteristics of the GSOS-ELM algorithm can significantly reduce the number of labeled reference tags and reduce the cost of positioning systems. In addition, the online learning phase of the GSOS-ELM algorithm can continuously update the system to perceive changes in the environment and resist the environmental interference. We have carried out experiments to study the influence factors and validate the performance, both the simulation and testbed experiment results show that compared with other algorithms, our proposed GSOS-ELM localization system can achieve more accurate positioning results and has certain adaptability to the changes of the environment.
Chapter
The symposium “Challenges of Fingerprinting in Indoor Positioning and Navigation” took place at Barcelona (Spain) on May 3rd and 4th, 2016. The audience comprised academic, scientists, engineers, company representatives, and institutional members. The program offered reports on the state of the art for indoor positioning based on fingerprinting, as well as discussions of challenges of the technology for the near future. Discussions gave potential users of indoor positioning technology the opportunity to expose real indoor location problems that need to be solved. This chapter gives a summary of the topics dealt in the symposium and constitutes a brief introduction to the chapters that appear in the entire book.