Professional Documents
Culture Documents
December 2018
Approval of the Department of Electronic & Telecommunication
Engineering
……………………………………………….
Head, Department of Electronic &
Telecommunication Engineering
This is to certify that I/we have read this project and that in my/our opinion it is fully
adequate, in scope and quality, as an Undergraduate Graduation Project.
Signature: …………………………………
Date: ……………………………………….
ii
Declaration
……………………….. …………………………………
Date D.D.Denipitiya (140113L)
…………………………………
S.M.I.P.Gnanarathne(140178M)
…………………………………
P.G.T. Nuradha (140425U)
…………………………………
B.L.C. Perera (140448R)
iii
Declaration by Supervisor
I/We have supervised and accepted this dissertation for the submission of the degree.
……………………………………….. ………………………………
Prof./Dr. Dileeka Dias Date
iv
Abstract
With the increasing use of location-based services (LBS), the need of accurate
indoor position is increasing day by day. LBS can further enhance customer and
company relationship in order to engage in better business or can maintain inventories of
a warehouse more efficiently. It is evident that developing an indoor positioning system
opens doors to limitless opportunities. However, due to multipath losses and absence of
line of sight, it is not feasible to deploy the Global Positioning System (GPS) for indoor
environments. Therefore, our focus shifts to technologies that can be used indoors for
accurate positioning.
The scope of our project expands in the following areas. First, collection of data
from Wi-Fi access points and Bluetooth beacons is done where after a machine learning
model is developed to predict the position of the user. Analysis of the impact of a
malfunctioning Bluetooth or Wi-Fi access point is another area of study in this project.
Finally, the system accuracy in static and dynamic environments is analyzed in detail. As
an example for its usage in an LBS, a web app is developed to show the position of a user
within the ENTC building.
The purpose of this project is to develop a robust positioning system using Wi-Fi
and Bluetooth technologies. We focus on laying out an efficient an accurate platform on
which LBS can be implemented seamlessly. The system developed is fine-tuned with
different positioning algorithms using machine learning techniques. In the latter part of
the report, an analysis of the accuracy and the immunity of the system for malfunctioning
access points is discussed.
v
Using machine learning techniques for indoor localization, we have obtained
satisfactory and accurate positioning within the ENTC building, which was our test area.
The results attained prove satisfactory and shows improvement when compared with the
existing literature.
vi
DEDICATION
To our Parents
vii
Acknowledgments
We would like to take this opportunity to express our gratitude to
supervisor Prof. Dileeka Dias for the continuous guidance and immense support given
throughout the project. We also would like to take this opportunity to convey our sincere
gratitude to all the colleagues for the great endeavours taken in completing the project.
Furthermore, we would like to thank all the people who helped us in various aspects in
making this project a huge success.
viii
Approval of the Department of Electronic & Telecommunication Engineering
This is to certify that I/we have read this project and that in my/our opinion it is fully
adequate, in scope and quality, as an Undergraduate Graduation Project.
Signature: …………………………………
Date: ……………………………………….
ix
Table of Contents
Abstract ................................................................................................................................ v
1.INTRODUCTION ............................................................................................................ 1
2. METHODOLOGY ........................................................................................................ 16
Introduction .................................................................................................................... 16
3.RESULTS ....................................................................................................................... 35
xi
LIST OF FIGURES
Fig 1.1 Representation of a MIMO system .......................................................................... 3
Fig 1.2 AoA Computation considering 2 transmitters ......................................................... 4
Fig 1.3 Comparison of the Existing Systems for Indoor positioning.................................. 7
Fig 1.4 Typical architecture of an iBeacon .......................................................................... 8
Fig 1.5 Flow Diagram of Proposed Solution ..................................................................... 14
Fig 1.6 Basic Architecture of the System .......................................................................... 15
Fig 2.1 Python GUI for data collection ............................................................................. 16
Fig 2.2 Life cycle of an Activity ....................................................................................... 18
Fig 2.3 Life cycle of a service ........................................................................................... 19
Fig 2.4 User Interface of Wi-Fi scanner ........................................................................... 20
Fig 2.5 Performance of a BLE scanner ............................................................................. 22
Fig 2.6 Use Interface of BLE scanner ....................... 24
Fig 2.7 Breakdown Analysis ............................................................................................. 25
Fig 2.8 Outlier detection algorithm ................................................................................... 27
Fig 2.9 3D depiction of outlier .......................................................................................... 27
Fig 2.10 Removal of Outliers........................................................................................... 28
Fig 2.11 Actual Coverage Map ......................................................................................... 31
Fig 2.12 Calculation of path loss Component ................................................................... 32
Fig 2.13 Approximated coverage map .............................................................................. 33
Fig 2.14 Final beacon placement configuration ................................................................ 34
Fig 2.15 Actual heatmap of selected configuration .......................................................... 34
Fig 2.16 Approximated heatmap of selected configuration .............................................. 34
Fig 3.1 Cumulative Distribution Function (CDF) of positional error with different
filtering techniques ............................................................................................................. 35
Fig 3.2 Impact of applying post filtering techniques ........................................................ 36
Fig 3.3 Impact of self correcting algorithm for malfunctioning of Access Points............ 38
Fig 3.4 Impact when KNN is used .................................................................................... 39
xii
xiii
LIST OF TABLES
Table 1.1 Comparison of Wireless Technologies for indoor positioning ............................ 6
Table 1.2 Comparison of Hybrid Wireless Systems with Wi-Fi and Bluetooth ................ 11
Table 2.1 Observed RSSI values........................................................................................ 30
Table 2.2 Power level to Weight conversion ..................................................................... 31
Table 2.3 Path loss exponents for different environments ................................................. 32
Table 3.1 Accuracies of corresponding models ................................................................. 37
Table 3.2 Accuracy comparison of different configurations ............................................. 39
xiv
ACRONYMS AND ABBREVIATIONS
xv
Chapter 1
1. INTRODUCTION
1.1 Background
With the increase of extended uses of the internet, location-based services (LBS) have
been playing a major role in day-to-day life. These services provide practical information
that is useful in many ways. LBS can further enhance the customer and company
relationship in order to engage in better business, a consumer can easily identify an
available parking slot in a basement car park or an inventory of a warehouse can be
maintained efficiently. Now it is evident that developing indoor positioning systems open
doors to limitless opportunities.
Considering the fact that Global Positioning System (GPS) cannot be used indoors due to
multipath losses and absence of Line of Sight, it is not feasible to use GPS for indoor
positioning. Therefore, technologies used indoors can provide the solution for accurate
positioning.
In our project we focus on laying out an efficient and accurate platform on which
location-based services can be implemented seamlessly.
1
1. The accuracy of wireless positioning systems depends on various environmental
factors such as multipath propagation, environmental noise, environmental
dynamics etc.
2. The malfunctioning of access points deteriorates accuracy of the system.
The accuracy of the system in various environmental conditions and in situations where
the required data is missing has to be considered.
1.5 Scope
The scope of our project is as follows.
• Collect data from Wi-Fi access points and Bluetooth beacons
• Develop a Machine Learning model to predict the position of the user
• Analyse the impact of malfunctioning of BT and Wi-Fi access points
• Analyse the system accuracy in static and dynamic environments
• Develop a Web app to show the position of the user within ENTC building
2
1.6.1 Wireless Positioning Techniques
There are several positioning techniques proposed to be utilized as a means to do accurate
positioning. The Characteristics and operating methods of them are introduced in this
section.
1.6.1.1 Received Signal Strength Indicator (RSSI)
The power level of the received signal is denoted by the RSSI values. Therefore, the
higher the RSSI, the stronger the signal is, as denoted by 𝑅𝑆𝑆𝐼 = −10𝑛𝑙𝑜𝑔10 (𝑑) + 𝐴.
RSSI can be used to estimate the distance between transmitter and the receiver device.
According to the equation, higher the RSSI value the smaller the distance between
transmitter and the receiver. RSSI based localization needs at least three transmitter
points to the same receiver. Then calculate each absolute distance from the transmitter to
the receiver. Based on the calculations, the location can be obtained by using geometry
and trigonometry calculations.
1.6.1.2 Channel State Information CSI
The channel state information is introduced with the Multiple Inputs Multiple Outputs
antenna arrays (MIMO).
CSI is a measurement related to the quality of the channel between the transmitter and the
receiver. CSI can be written in a polar form as 𝐻(𝑓) = 𝐻(𝑓) ∨ 𝑒 𝑗𝑠𝑖𝑛𝑔(𝐻) where, |H(f)| is
the amplitude or the magnitude of the response and j sing(H) is the phase response of the
frequency f of the channel. As a general note, CSI describes how a signal propagates from
the transmitter to the receiver. Since the channel depends on the distance and the
3
environment, the CSI values also depends on those factors. Therefore CSI is utilized to
obtain the location of the transmitter.
1.6.1.3 Time of Arrival (ToA)
Measurement of the signal propagation time between the transmitter and the receiver can
be used to obtain the location of the receiver. The relative distance between the
transmitter and the receiver can be denoted by 𝑟 = 𝑡 ∗ 𝑐, where c is the speed of light and
the t is the signal propagation time from transmitter to the receiver. The receiver location
can be calculated using three transmitter base stations applying geometric mathematics.
However, the ToA requires strict synchronization between transmitters and receivers as
the distance can be changed drastically, even if there is a slight change in time.
1.6.1.4 Time Difference of Arrival (TDoA)
This technique exploits the difference in signal propagation times from different
transmitters, measured at the receiver. These absolute time values are converted into
distance values by the ToA equation. The receiver is now located on the hyperboloid
given by
2 2 2
𝐿𝐷𝑖,𝑗 = √(𝑋𝑖 − 𝑥)2 − (𝑌𝑖 − 𝑦)2 − (𝑍𝑖 − 𝑧)2 − √(𝑋𝑗 − 𝑥) − (𝑌𝑗 − 𝑦) − (𝑍𝑗 − 𝑧)
where (𝑋𝑖 , 𝑌𝑖 , 𝑍𝑖 ) are the coordinates of the transmitter or reference node I and (x,y,z) are
the coordinates of the receiver. At least three transmitters are required for calculating the
exact location.
1.6.1.5 Angle of Arrival (AoA)
Angle of Arrival (AoA) mainly determines the source position through the measurement
of the base station signal to the object's azimuth, as shown in Fig 1- 2
4
This technique requires directional antennas for positioning. The main advantage is that
the location of the receiver can be estimated with two transmitters in 2D environment or
three transmitters in 3D environment. However, it requires complex hardware and careful
calibration compared to other techniques. Because, a slight error of angle can be
translated into a huge error in the distance calculation.
Positioning based service is increasing important technology that plays a vital role in
modern life. Positioning scenarios can be divided into two categories; outdoor positioning
and indoor positioning. GPS is used for outdoor positioning. GPS only works better in
open areas, since GPS signals can be blocked by buildings, thick forests and other types
of physical obstacles. Therefore, indoor positioning is introduced as an extension of
outdoor positioning.
Indoor positioning can be used for navigation in large malls or complex yet unfamiliar
place for users. Unlike outdoor positioning, indoor places might be small and easy to
deploy extra infrastructure. There are different technologies used for wireless positioning.
6
Zigbee is widely used for localization of wireless sensor networks (WSN). Another
wireless technology, RFID is primarily used for transferring and storing data using
electromagnetic transmission from a transmitter to RF compatible circuit. It consists of a
reader that can communicate with RFID tags. Both Zigbee and RFID are not favorable for
indoor positioning of users because, those technologies are not readily available in the
majority of portable user devices like smart phones and laptops.
Most of the portable user devices such as smart phones, laptops are Wi-Fi and Bluetooth
enabled. Since existing Wi-Fi access points can be also used as reference access points,
basic positioning systems can be built without the need of additional infrastructure.
Current BLE based positioning solutions rely on RSS based measurements and RSS
based systems are less complex. Therefore, a hybrid positioning system with Wi-Fi and
Bluetooth is considered here.
7
1.6.3.1 Wi-Fi RSSI
Received Signal Strength Indicator (RSSI) is defined as the indication of the power level
being received by the antenna of a user. Since this is a relative measure, the higher the
RSSI is the stronger the signal, and vice versa.
Wi-Fi RSSI is the indication of Wi-Fi signal strength captured by a device. Most of the
current work done in positioning systems using Wi-Fi, use RSSI as a benchmark for
comparison.
• Easy to implement
• Cost efficient
BLE can be used with different positioning techniques such as RSSI, AoA and ToF.
However, most of the existing BLE based systems are based on RSS inputs. Therefore, if
BLE in its original form is used in a positioning system, the reliance on RSS based inputs
limits its localization accuracy. However, its range, low cost and low energy consumption
motivate developers to use BLE as a positioning technology.
There are two BLE based protocols; iBeacon (by Apple Inc.) and Eddystone (by Google
Inc.) which are mainly used for proximity-based services.
8
iBeacon protocol allows iBeacon to transmit beacons or signals at periodic interval. The
beacon message consists of a mandatory 16-byte UUID (Universally Unique Identifier)
and optional 2-byte major and minor values. BLE enabled devices, which have
proprietary applications to listen to the beacons can pick up the beacon messages and use
RSSI to estimate the proximity between the iBeacon device and the user. After receiving
an iBeacon message, the device consults a server to identify the action affiliated with the
received beacon. For example to open a door to display some interactive content on a
monitor, based on the user’s proximity to some beacon or another entity. The typical
iBeacon architecture is shown in Fig 1-4
A fundamental constraint of iBeacon is that only the average RSSI value is reported to
the user device every one second, even though the beacons are transmitted at 50ms
intervals. This is to account for the variations in the instantaneous RSS values on the user
device. However, this RSS averaging and reporting delay can impose significant
challenges to real-time localization. [15]
1.6.3.3 Wi Fi CSI
Due to the simplicity and low hardware requirements, most available localization systems
use received signal strength(RSS) as fingerprints. The major disadvantage of this method
is that, RSS values usually have a high variability over the time for a fixed location, due
to the multipath effects in indoor environments. This causes large error in the location
estimation process. It is possible to use Channel State Information (CSI) as fingerprints.
Some Network Interface Cards provide their CSI details. The Intel Wi-Fi link 5300 NIC
provides CSI data from 30 different subcarriers.
Different approaches have been proposed using phase and amplitude of CS data.
• Consider the amplitude responses for fingerprinting
CSI values are stable at fixed positions, but those are varying at adjacent locations and the
multipath effect on CSI values causes various clusters of subcarriers with respect to the
attenuation experienced by the subcarriers. Therefore, pre-processing techniques are used
to generate input values. After the calibration process, in the offline phase, machine
learning approaches can be used to generate clusters for different locations. Another
approach is generating feature-based fingerprints. This utilizes many weights obtained by
deep learning to denote different locations. In the online localization phase, the
9
probabilistic method or the position estimation based on data fusion, which normalizes
the magnitudes of CSI values using weights for different positions to obtain its estimated
location.
• Using the calibrated phase information of CSI
Most of the literature ignored the phase information of CSI data due to the randomness
and the unavailability of the raw phase information. Some of the literature uses phase
information without calibrating while some of them are using linear transformation to
remove the offset of phase information. Then they use deep networks with hidden layers
to train the calibrated phase data and use weights to represent their fingerprints. Greedy
algorithms have been used to reduce the computational overhead. The online stage uses a
probabilistic method (Bayes method) for location estimation.
Some of the approaches incorporate both magnitude and phase information of CSI from
each subcarrier for fingerprinting.
1.6.3.4 Hybrid Systems
Two approaches have been proposed in the literature to combine Wi-Fi and BT into one
positioning system.
1. Employing Scene Analysis for both infrastructure. [1,2,5]
A combined Radio map is created for both Wi-Fi and BT signals at each reference
point in the offline phase and that combined map is utilized in the online phase to
estimate user position.
2. Applying Scene Analysis and Proximity Analysis to Wi-Fi and BT respectively in
one positioning system [3,4]
When Wi-Fi is deployed as the main infrastructure, a limited no of BT Access points are
deployed as add-ons to divide indoor space into disjoint partitions. As a result, the
original, large radio map is divided into small ones which result in reducing the
computation cost of online position estimation.
10
Table 1.2 Comparison of Hybrid Wireless Systems with Wi-Fi and Bluetooth
[1] Wi-Fi - RSSI Wi-Fi APs- RSS K- NN for both The average rate
BT - RSSI Provide Fingerprint the selection of of correct
Fingerprinting Umbrella Data Wi-Fi and estimation of Wi-
Coverage Collection iBeacon AP Fi, iBeacon and
iBeacons - (Noise Hybrid system are
Precisely Filtering) 54%, 72% and
localize the Online 81%respectively.
mobile users phase - The performance
Position is also evaluated
Selection by the running
- Signal time. The
selection averaged
(Wi-Fi & positioning time
iBeacon) of Wi-Fi, iBeacon
and Hybrid
system are 326
ms, 504 ms, and
430 ms
respectively.
11
approach the reference assets in a large
asset. points into building"
database.
12
information spatial to the mobile which allows to
on their distribution device. Then obtain almost
accuracy of location do either K-NN perfect location
errors in the matching or K- estimation
offline phase means (99.74%) with as
and use it to clustering few as 10
correct principal
estimation. components.
13
1.7 Proposed Solution
The proposed solution mainly consists of three phases. The first phase is collecting Wi-Fi
RSSI data, Wi-Fi CSI data and Bluetooth RSSI data. As the next phase, developing five
different machine learning models.
• Wi-Fi RSSI
• BT RSSI
• Hybrid Wi-Fi and Bluetooth RSSI
• Wi-fi CSI
Final phase of the proposed solution is analysing the models based on the malfunctioning
of access points and system accuracy in static and dynamic environments. Malfunctioning
of access points is further subdivided into the breakdown of Wi-Fi access points, BT
beacons and the analysis of access points that are not in range. The analysis of the system
accuracy in various environments considering environmental changes will be carried
down based on error CDF.
14
1.8 Basic Architecture
The indoor positioning system with Wi-Fi and Bluetooth has two phases; offline phase
and online phase. At the offline phase, the proposed space will be divided into a grid.
Then, the BLE RSS data and Wi-Fi RSS/CSI data will be collected at each grid point
using existing Wi-Fi access points in the ENTC building and the iBeacons placed.
Collected data will be sent to a server for processing and an optimum machine learning
model will be developed. Then at the online phase, test data (both BLE RSS and Wi-Fi
RSS/CSI) will be collected and fed to a location algorithm. Finally, the estimated location
will be displayed on a web application.
15
Chapter 2
METHODOLOGY
Introduction
This section covers the methods and approaches we have used with respective to the
proposed methodology based on the literature review conducted.
• Designing a python GUI and Android application to scan for Wi-Fi and BT
fingerprints to collect RSSI data
16
2.1.2 Android development overview
The mobile application development has several major approaches such as Google
android, apple iPhone, RIM Blackberry and Microsoft Windows 8. Android is a Linux
based operating system. The android systems support rich user interface libraries and
background processing. The applications were developed using Android Studio IDE and
it provides graphical development environment based on the Eclipse IDE to develop
applications. The Android Developer tool (ADT) is a plugin for Eclipse IDE that is
designed to give a powerful integrated environment in which to build android
applications. ADT has capabilities to quickly setup new android projects, debug
applications using Android SDK tools and export .apk files in order to distribute the
application.
17
Android applications are primarily written in Java programming language. The compiled
Java code is compressed with an archive file with a suffix of .apk. this file can be
distributed for installing applications on mobile devices. The android applications use
activities as central features because it has rich visual user interface for the user to
undertake. A simplified lifecycle of an Activity is shown in the Fig 2.2.
18
Service has a simple life cycle than activity lifecycle. Generally, an activity can start or
stop a service. Therefore, the service lifecycle can be represented as follows.
21
Finally, when the specified time is finished, the thread will die, Bluetooth
scanning process will be stopped. Then the application is again available for a new
scan.
6. Main activity
The Main activity clearly defines the basic flow and its phases of the application.
“onCreate” phase performs several activities. First, it checks whether the phone is
BLE capable. The app will be terminated if the device is not BLE supported.
Broadcast Receiver, Scanner, “ArrayList” and Hash map are initialized at the
“onCreate” phase. Also, the Array adapter will update the user interface according
to the changes of access points. “onClick” activity is defined here which will
provide options for saving a text file with the scanned data within the phone
storage whenever, the scan period is finished.
“onStart” phase registers the Broadcast Receiver. “onStop” phase unregisters the
Broadcast Receiver. Therefore, these two phases deal with the Bluetooth state of
the device.
“onClick” activity is defined for starting BLE scan whenever the “Scan” button is
pressed, if the scanner is not in the scanning state.
22
Fig 2.5 Performance of a BLE scanner
23
The above components altogether provide BLE scanning functionality. The basic
scanning process is as follows.
24
The above table shows the RSSI data of the access point at various reference locations. In
addition, a number of positions in each place and number of Wi-FI scans in each
reference location. The laptop with the Wi-FI adapter wirelessly obtains the MAC address
and RSSI value of the AP at various locations.
Let 𝑃(𝛼) represents the matrix of RSSI values measured from different APs at reference
point α.
𝑅𝑆𝑆𝐼1,1 ⋯ 𝑅𝑆𝑆𝐼1,𝑁
𝑃(𝛼) = [ ⋮ ⋱ ⋮ ]
𝑅𝑆𝑆𝐼𝑀,1 ⋯ 𝑅𝑆𝑆𝐼𝑀,𝑁
M represents the number of Wi-Fi scans at reference point α and N is the total number of
available access points. After selecting APs which covers all three places, RSS time series
are measured from each AP at each RP, which is utilized to create the location fingerprint
database. After the completion of the fingerprint collection, different machine learning
classification algorithms were used to predict the position. But the accuracies are not up
to the expected level with the raw data. As a solution, we implemented an outlier removal
mechanism to increase accuracy. After choosing most suitable algorithm, analysis of
access point breaks down has been performed. Fig 2.14 shows the overall hierarchy of the
proposed system. First analyse the raw data using different machine learning algorithms.
After choosing the most suitable one anomaly detection technique was used to remove the
noise in the radio map. Finally, room level classification and breakdown analysis were
done.
25
26
on the mean squared error of the machine learning model which was trained from raw
data.
The outlier detection system based on clustering algorithms and dimension reduction
technique. Collected dataset consists of 11 features, which can be considered as high
dimension. Moreover, the dataset has a sparse distribution over the high dimensional
space. Since clustering is based on the distance between two data points we used
dimension reduction such as PCA to prepare data for outlier detection using clustering.
The PCA applied on the power values which are converted from RSSI values in the
dataset.
Principle Component Analysis is mainly used in an unsupervised manner and use to
detect the inherent structure of data. In addition, it uses to reduce the dimensionality in
feature space which contains most of the information in a larger set. Each feature of input
data represents one dimension in feature space. PCA is transforming these correlated
features into an uncorrelated variable set. The first principal component accounts for as
much of the variance in the data as possible, and each succeeding component accounts for
as much of the remaining variance as possible.
27
The next step was clustering using K means algorithm. K means aims to partition n
samples into k clusters and each observation belongs to the cluster which has nearest
mean. As shown in the Fig , all identified outliers were removed from the data set. Then
different classification algorithms were performed for the outlier removed dataset.
28
• In the self-correcting system, the position estimation model is rebuilt
without considering the malfunctioning AP
Similar approach has been presented in the literature using Random Forest algorithm
[17]. However, they have applied the self-correcting algorithm to the complete model.
We are proposing a method to apply the self-correcting algorithm only for relevant radio
maps based on the room level classification.
When room level classification is done, we can only focus on a sub section of the radio
map without considering the entire radio map when the online fingerprints are matched
with offline fingerprints included in the radio map. This also reduces the complexity of
the positioning system without comprising the accuracy.
29
converted into its received power (in mW) and scaled to assign a weight for each level.
The variability of the measured data was very high as shown in the table.
The coverage map was generated as follows.
The Table 3 shows the calculated mean RSSI values at each cell point considering 10
scans and the variances at each cell.
Table 2.1 Observed RSSI values 60
cm
-82.83 -79.45 -78.32 -77.51 -76.83 -74.54 -72.67 -80.55 -72.96 -79.54 -77.03 -77.98 -78.23 -81.75 -83.26
3.01 2.64 3.9 3.28 4.01 2.66 1.09 3.65 1.25 3.94 3.53 3.52 2.49 1.76 2.47 60 cm
-84.56 -77.96 -79.56 -76.52 -76.95 -70.27 -70.79 -77.56 -74.38 -71.19 -79.83 -75.81 -77.46 -79.18 -84.27
2.83 1.26 4.2 3.15 3.56 1.95 2.87 3.8 3.56 1.06 2.12 4.25 2.48 3.77 1.79
-81.46 -79.84 -78.36 -77.3 -77.3 -72.68 -69.18 -72.98 -74.48 -72.0 -74.81 -75.4 -76.8 -78.94 -81.84
2.41 3.7 3.5 2.34 4.07 3.0 3.66 2.46 2.62 1.11 2.86 2.16 1.48 1.76 2.53
-82.65 -79.78 -76.9 -75.55 -78.4 -74.3 -78.4 -79.33 -78.51 -74.0 -74.83 -75.89 -75.55 -79.77 -82.71
1.45 2.74 3.06 1.83 1.99 2.75 3.25 2.61 2.32 2.45 2.71 2.37 3.15 2.76 2.4
-85.12 -81.49 -75.84 -78.22 -75.4 -76.86 -76.7 -75.39 -75.76 -77.74 -76.17 -76.34 -78.3 -78.13 -83.74
2.13 3.29 2.47 3.2 2.63 3.64 3.52 3.29 3.87 2.95 2.65 4.69 2.06 1.88 2.81
-82.43 -84.72 -79.6 -78.66 -79.0 -78.53 -82.5 -81.31 -79.8 -82.27 -78.53 -78.31 -79.78 -84.3 -85.7
3.11 2.39 4.28 3.94 4.73 2.91 4.0 4.27 1.22 4.58 4.21 2.16 3.67 4.37 3.2
-85.59 -84.28 -82.64 -80.8 -79.5 -80.75 -82.58 -77.25 -79.27 -82.43 -77.64 -81.7 -83.64 -81.2 -84.9
1.28 1.6 3.47 2.85 2.01 1.88 1.32 4.29 2.38 1.66 2.72 4.3 3.77 2.4 1.98
-85.37 -82.46 -80.79 -77.93 -78.44 -78.66 -82.79 -79.19 -78.64 -79.86 -79.1 -78.37 -81.94 -80.76 -83.17
2.04 1.83 3.58 2.87 2.27 3.43 2.1 2.19 3.77 4.0 2.24 1.7 3.65 2.84 1.97
-84.36 -82.19 -80.73 -85.29 -77.48 -78.76 -77.35 -85.01 -85.78 -82.03 -82.95 -80.78 -83.28 -83.91 -85.21
1.58 3.24 2.16 1.33 1.88 3.61 1.06 2.33 1.89 3.63 3.38 2.87 3.77 2.84 2.57
-84.95 -83.74 -82.61 -80.35 -82.49 -81.7 -83.49 -85.75 -84.9 -81.47 -80.57 -83.69 -85.9 -81.76 -84.6
2.76 3.1 4.16 3.4 1.42 2.73 3.09 2.81 3.27 1.75 2.83 2.79 3.43 2.82 1.3
-85.1 -83.45 -81.34 -84.27 -82.76 -85.98 -83.76 -84.19 -82.47 -82.67 -82.76 -84.9 -81.16 -85.43 -83.74
2.1 2.46 3.28 3.17 2.03 2.49 3.25 2.9 3.07 2.56 1.93 2.31 1.28 1.76 2.15
1. RSSI value at each cell (in dBm) was converted into linear power level in
milliwatt (mW).
𝑃𝑜𝑤𝑒𝑟 (𝑚𝑊)
𝑅𝑆𝑆𝐼 (𝑑𝐵𝑚) = 10 log10
1 𝑚𝑊
2. A weight is assigned for each power level as shown in the Table 4.
30
Table 2.2 Power level to Weight conversion
Since we wanted a generalized coverage map which can be applied for any beacon for
analyzing the coverage of floor map, first we generated a coverage map for free space.
Since it highly deviated with the observed coverage map, we generated an approximate
coverage map by considering the path loss component (n)
𝑑
𝑅𝑆𝑆𝐼(𝑑) = 𝑅𝑆𝑆𝐼(𝑑0 ) + 10𝑛 log 𝑑 where, 𝑑0 = 1𝑚
0
31
Measured RSSI values of grid size 11*15 were fitted into a linear regression line which
minimizes the least square error in order to calculate the path loss component. The
concluded path loss component (n) in our scenario is obtained as 3.25 and RSSI at 1m
distance is obtained as -76.42 Dbm according to the Fig .
Table 2.3 Path loss exponents for different environments
Then, the approximated RSSI values were calculated at same cell locations by
considering the calculated path loss component and generated an approximate beacon
coverage map for a single beacon.
32
Fig 2.13 Approximated coverage map
The necessity of a generalized beacon coverage map was coming up because, the
variability of Bluetooth signals differs from beacon to beacon. When the observed and
approximated coverage maps are compared, it can be seen that the coverage distribution
of the observed map can be approximated to the distribution of a calculated one. Since we
were going to introduce a better beacon placement algorithm, the generalized
approximated coverage map was used as the main entity for checking the signal coverage
of a selected building.
• Divide the lab into a grid and initialize all the cell values as 0.
• Select distances between two beacons which is greater than 200cm (both
horizontally and vertically) and start placing beacons from a corner of a wall.
• The placement height was selected as 2m which exceeds the height of an average
man a little bit
33
• Place beacons using the placement algorithm in both symmetrically and
asymmetrically by varying horizontal and vertical distances between beacons.
• While placing, each cell value accumulates with the value of overlapped cell from
the coverage map.
• Optimum beacon placement configuration is selected in order to maximize the
beacon coverage with a limited number of beacons.
• Compare the accuracy of rejected configurations the with same number of
beacons with the selected configuration.
• Verify the selected configuration as the best configuration.
34
Chapter 3
3. RESULTS
This section covers the results obtained from the following different experiments
discussed in the Methodology section. The results are analysed based on Wi-Fi RSSI
system and Bluetooth RSSI system
35
Classification accuracy of different section (classes) in the Common Room and Computer
Lab also increases with a higher margin with the filtering of raw data.
This shows that data preprocessing improves the accuracy of the positioning system
reducing the impact of the variation of signal strengths.
The impact of applying Kalman filtering for position estimation is well depicted from the
improvement of the positional estimation as shown in the diagram below.
Datasets were collected from ENTC1 and the Common room to perform localization
using finger printing method. Approaches for localization are mainly two-fold and can be
categorized as classification models and regression models. After considering the uniform
geometry, ENTC1 dataset was selected for the experiments and kept other datasets for the
validation purposes. At the beginning, experiments were conducted using raw data from
APs for classification approach. Support Vector Machine (SVM) and Artificial Neural
Network (ANN) were used as supervised classifiers and KNN, K-means clustering were
used as unsupervised cluster-based classifiers for the experiments. However,
classifications accuracies are not up to the expected level and implemented an outlier
detection as a data pre-processing method to enhance the accuracies of models. Outlier
detection was done based on a clustering approach using K-means algorithm. Since the
input data was high dimensional, Principal Component Analysis (PCA) was applied to the
inputs, which are converted from decibel scale to the linear scale to reduce dimensionality
to prevent from curse of dimensionality effect. Thereafter, K-means clustering was
applied, the outliers were detected and filtered out those detected outliers from the data
36
set. This dataset was again used in the classification stage and accuracy was improved as
expected.
Accuracy (With
Accuracy (Without Outlier
Model Outlier
Detection)
Detection)
37
With Random Forest Regressor
Fig 3.3 Impact of self correcting algorithm for malfunctioning of Access Points
38
Fig 3.4 Impact when KNN is used
According to the selected configuration the whole computer lab was able to cover using
four beacons. The accuracy verification table shows the accuracies of Random Forest and
39
K-nearest-neighbor models for the configurations generated from the algorithm which has
exactly four beacons. Only the configurations which has four beacons were compared for
accuracy because, by increasing the number of beacons, the accuracy will inherently
increase and by decreasing the number of beacons the accuracy will inherently decrease.
We found seven symmetrical and asymmetrical beacon configurations from the
algorithm, each provide different positions for beacons. As described in the placement
algorithm, the best configuration for beacon placement was chosen by the algorithm by
considering the maximum coverage with minimum number of beacons which minimizes
the interference between them. The verification table proves that the selected
configuration (Config 6) provide higher positioning accuracy than the rejected ones.
40
Chapter 4
4. DISCUSSIONS AND CONCLUSION
4.1 Discussion
With the results we have obtained so far, it is clear that raw data should be pre-processed
to eliminate the effect of environmental noise to a certain extent for both the Wi-Fi RSSI
system and BT RSSI system. Both the data filtering techniques and outlier detection
approach has proved to be applicable for improving the performance of the indoor
positioning system.
The analysis carried down based on the malfunctioning of access points and the self-
correcting algorithm with the detection of AP breakdown have given substantial
improvements to the performance of the system in unexpected scenarios.
We used existing Wi-Fi infrastructure for the positioning purpose. So we do not have
controllability over them. With the BT placement algorithm employed, we can add
controllability to the hybrid positioning system with the use of good features of both Wi-
Fi and BT systems.
Conclusion
We have considered about Wi-Fi and BT positioning systems separately deciding the
optimum placements for BT beacons. We have also analyzed the performance of the
system under different considerations such as malfunctioning of APs.
We are planning to complete the following tasks in the future.
• Studying about Channel State Information (CSI) of Wi-Fi and positioning models
using CSI
• Combining Wi-Fi RSSI and BT RSSI and coming up with a hybrid model for
wireless positioning
• Carrying out further analysis on the system performance considering the
malfunctioning of Access points, environmental changes, device diversity in the
receiving device used for positioning
• Developing a mobile application to predict the location in the online phase.
41
BIBLIOGRAPHICAL REFERENCES
42
[9] Y. Xie, Y. Wang, A. Nallanathan and L. Wang, "An Improved K-Nearest-Neighbor
Indoor Localization Method Based on Spearman Distance," in IEEE Signal Processing
Letters, vol. 23, no. 3, pp. 351-355, March 2016.
[10] Y. Xie, Y. Wang, A. Nallanathan and L. Wang, "An Improved K-Nearest-Neighbor
Indoor Localization Method Based on Spearman Distance," in IEEE Signal Processing
Letters, vol. 23, no. 3, pp. 351-355, March 2016.
[11] I. Gorovyi, A. Roenko, A. Pitertsev, I. Chervonyak and V. Vovk, "Real-time system
for indoor user localization an navigation using bluetooth beacons," 2017 IEEE First
Ukraine Conference on Electrical and Computer Engineering (UKRCON), Kiev, 2017,
pp. 1025-1030.
[12] S. S. Chawathe, "Beacon Placement for Indoor Localization using Bluetooth," 2008
11th International IEEE Conference on Intelligent Transportation Systems, Beijing, 2008,
pp. 980-985.
[13] S. P. Subramanian, J. Sommer, F. P. Zeh, S. Schmitt and W. Rosenstiel, "PBIL PDR
for scalable Bluetooth Indoor Localization," 2009 Third International Conference on Next
Generation Mobile Applications, Services and Technologies, Cardiff, Wales, 2009, pp.
170-175.
[14] P. K. Yoon, S. Zihajehzadeh, B. S. Kang and E. J. Park, "Adaptive Kalman filter for
indoor localization using Bluetooth Low Energy and inertial measurement unit," 2015
37th Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC), Milan, 2015, pp. 825-828.
[15] F. Zafari, A. Gkelias, and K. K. Leung, “A Survey of Indoor Localization Systems
and Technologies”, arXiv:1709.01015, 2017
[16] Online Reference - https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5461075/
[17] G ́orak, R., Luckner, M., "Malfunction immune wi-fi localisation method"
In:Computational Collective Intelligence, Lecture Notes in Computer Science, vol. 9329,
pp.328–337. Springer International Publishing (2015)
43