Indoor Localization with lightweight RSS Fingerprint using BLE iBeacon on iOS platform44916

In this paper, we propose an BLE iBeacon-based indoor localization system using Fingerprinting method.. In order to reduce the computational cost and ensure the accuracy of the system, w

Indoor Localization with lightweight RSS

Fingerprint using BLE iBeacon on iOS platform

Ngoc-Son Duong and Thai-Mai Dinh

Department of Telecommunications Systems, Faculty of Electronics and Telecommunications

University of Engineering and Technology, Vietnam National University

Ha Noi, Viet Nam Email: duongson.vnu@gmail.com, dttmai@vnu.edu.vn

Abstract—In the near future, location-based services (LBSs)

will become an essential part of the smart city That’s why

indoor localization has been attracting researcher attention

A popular method for indoor localization is Fingerprinting

However, it is not easy to reach high accuracy in positioning

with only Fingerprinting Several current techniques could be

used to improve the accuracy of Fingerprinting such as Machine

Learning (ML) or Deep Learning (DL) In this paper, we

propose an BLE iBeacon-based indoor localization system using

Fingerprinting method The system applies k-nearest neighbors

(k-NN) - an ML algorithm to decide the user position In order

to reduce the computational cost and ensure the accuracy of

the system, we propose to use lightweight feature vectors that

include information of the nearest beacons and device azimuth

for training Machine Learning algorithm We performed some

experiments to verify the proposed system The results show that

the method provides a feasible indoor positioning solution with

high accuracy

Index Terms—Bluetooth Low Energy, iBeacon, indoor

position-ing system, iOS, Machine Learnposition-ing, K-nearest neighbors, kNN

I INTRODUCTION

As we know, Location Based Services (LBS) [1] requires

the exact location of the user in indoor or outdoor

environ-ments Global Positioning System (GPS) is a type of LBS for

outdoor environment For indoor environment, there are

sev-eral indoor positioning systems with the support of a number

of wireless technologies such as Wi-fi [2], Bluetooth [3], UWB

[4], RFID [5], etc Currently, the introduction of Bluetooth

Low Energy (BLE) iBeacon, released by Apple brings great

benefits to indoor positioning Some studies using iBeacon

for indoor positioning are presented in [6] [7] [19] Like other

wireless technologies, BLE iBeacon-based indoor localization

uses signal strength to find the location of the receiver An

RSS-based well-known method is Fingerprinting Due to the

nonlinear variation of the BLE signal that is primarily caused

by multipath fading, the application of Fingerprinting to indoor

positioning systems could not get high accuracy With the

development of Machine Learning, we can deal with these

nonlinear relationships through data set The data set of

ma-chine learning models attenuation of signals when transmitted

through structures which include walls, columns, plants, etc in

indoor environment It’s amazing that the database can change

(following a period of time in a day) to adapt to the variation of

infrastructure or human movement Along with Fingerprinting, most of the learning algorithms used are Supervised Learning (k-NN, Navie Bayes Classifier, neural network, etc) Machine Learning-based Fingerprinting was also presented in [9] [10]

An overview of the accuracy of machine learning algorithms using the RSSI fingerprint method is presented in [8] A re-cent study on accuracy improvement of iBeacon-based indoor fingerprinting positioning is presented in [11] The problem

of selecting vector features is addressed with the aim to improve the accuracy and to reduce the database size, thereby

to reduce the amount of information to be exchanged in the Online phase In addition, the authors used propagation affinity clustering to decrease the computational cost and introduce

an exponential averaging filter to smooth RSS values An other sizable indoor navigation system called SwiBluX is introduced in [12] This is a multi-layer system using Deep Neural Network and many layers of filters including a custom Gaussian Outliers Filter, weighted position estimation and particle filter Most of the factors affecting the accuracy of the system are presented including the RSSI variance, the wireless signal attenuation, the human absorption, and device orientation These elements are the basis to obtain a consistent feature vector for Machine/Deep Learning algorithms With the motivation to reduce the amount of data to be handled and ensure the accuracy of the system simultaneously, our study will introduce a Fingerprinting method based on a lightweight feature vector which includes ID and RSS of Nearest Beacon and smartphone orientation Then, we built up the propose system in reality and do experiments to verify the system All experiments were performed on the iPhone

The remainder of this paper is organized as follows After this short introduction, we describe the model of the proposed system in Section II Section III will present in details the Fingerprinting method in part A, in part B of this section, we will show what feature will be chosen to using in the classifica-tion problem Secclassifica-tion IV provides the system parameters and experimental results Finally, section V concludes this paper

II PROPOSED SYSTEM MODEL

A Proposed system

The overall architecture is depicted in Fig 1 At first, the smartphone scan RSS values of beacons around by using BLE

chip Meanwhile, the magnetometer sensor provides the device

heading The positioning method that we use in our proposed

system is Fingerprinting The proposed system consists of two

parts: Offline/Training phase and Online/Execution phase In

the Offline phase, when the information of nearest beacon

and user heading are available, the smartphone record these

parameters to the local database as feature vectors In Online

phase, K-nearest neighbors is used to estimate the user position

through comparing measurement vector and stored database

BLE Chip Magnetometer

Informtion of nearest beacon Device’s Heading

Location (Class) Major RSS Heading

1

C

Measurement Vector

Offine

Phase

Online Phase

.QHDUHVW

QHLJKERUV Location Smart Phone

'

#1

m ' ' '

#1 , #2 , #3

r r r h'

1 1 1

#1 , #2 , #3

r r r

#1C, #2C, #3C

r r r

1

#1

m

#1

C m

1

h

C h

Fig 1 System overview and architecture.

B iBeacon and iOS Device-based indoor localization

In our system, we use iBeacons as transmitters and iOS

device (iPhone) as receiver iBeacon protocol is released by

Apple in WWDC 2013 [20] The physical device run on

the iBeacon protocol and powered by Bluetooth Low Energy

technology is called BLE Beacon The beacon hardware itself

broadcasts a message that contains 4 crucial fields which

are Unique Universal Identifier (UUID), Major, Minor -

sub-segments of Major and TxPower The receiver will capture

BLE messages and then analyze information in the message

according to usage purpose With the advent of the iBeacon

protocol, Apple provides a CoreLocation framework to help

developers read parameters from iBeacon messages

Specifi-cally, in our system, by using the CLBeacon class, iOS device

can read information related to UUID, Major, Minor and

RSSI With built-in magnetometer, iOS device can determine

a device orientation [13] This information can be obtained by

using CLHeading class.

III INDOORLOCALIZATIONBASEDONK-NEAREST

NEIGHBORS

A Fingerprinting Method

Fingerprinting is by the far the most favored form when researching on indoor localization based on radio frequency

In Fingerprinting method, the position obtained from Finger-printing must be pass through two phases: Offline/Training phase and Online/ Execution phase In the Offline phase, the area is divided into grid point or reference point, then collect data for each one These data are then saved in the form of vectors and stored in a database Suppose we haveQ beacons,

andP reference points (RPs) in the area We need to collect

data for a location at RPs in the shape of a vector Each vector normally has the form as following:

{x p , y p ; RSS p1 , , RSS pq , , RSS pQ }, p ∈ [1, P ] , q ∈ [1, Q]

(1)

in which, {x p , y p } is coordinate of p th position and RSS p1 , , RSS pq , , RSS pQ are RSS value of beacon

re-spectively All of these reference points will contribute to creating a radio map The problem of Fingerprinting is es-timating a position when RSS value measurement on-the-fly

of an unknown position is available:

{RSS 

1, , RSS q  , , RSS Q  }, q ∈ [1, Q] (2) This task will be done in Online phase by using a matching algorithm such as k-NN, SVM, Neural Network and so on Fingerprinting itself has its advantages and disadvantages Compare to Tri-lateration, it would be difficult to determine the spatial propagation model for BLE signals due to the vari-ability of the terrains Fingerprint’s advantage is not dependent

on converting RSSI to distance The stability of Fingerprinting

is higher than Tri-lateration because the position is fixed into reference points Besides, the major problem of Fingerprinting

is several reference points have similar data vectors This lead

to decrease the localization accuracy

B K-nearest neighbors-Based Fingerprinting 1) Feature Analysis: In general, when working with actual

iBeacon-based Fingerprinting problems, we only have raw data

of beacons such as ID (UUID, Major value and Minor value)

or RSS value Some of them have not been handled or filtered

We need to choose reliable data and eliminate noise data to find a feature vector This feature vector must ensure that specific pieces of information are kept for the original data

In this section, we will choose the factors that have a high impact on the decision of position as the features on the K-nearest neighbors problem

Inference 1: In literature, by using the Voronoi diagram,

we can use nearest beacon information (determined via RSSI)

to divide the fingerprinting point into clusters (Fig 2) Thus,

in the case that coordinate of beacons is known, knowledge about nearest beacon informs a position relatively, in the sense that indicates what region the user is in Therefore, ID of the nearest beacon (Major or Minor) classifies a group of RPs within a certain Voronoi unit

Voronoi Unit Reference Point

Beacon Node

Fig 2 Division by cluster using Voronoi diagram.

Inference 2: The ID information is not enough to determine

which RP the user belongs to The received signal strength

from the beacon of each RPs in a Voronoi unit is different

Therefore, the RSS value of each RPs should be known In

this study, we do not include far away beacon information as a

feature because, in fact, their signals are easily interfered due

to change of indoor environment such as the movement of

people To ensure that the duplicate of RPs in the database is

negligible, we use the RSSI information of the three beacons

which have smallest RSS value

Inference 3: Device azimuth is selected as one feature

because of the direction of the mobile’s wave reception and

human body-blocking Some studies have shown that antenna

orientation affects the RSSI [14] [15] [16] In a Voronoi unit

and under LOS condition, several RPs have the same ID of

the nearest beacon and RSS value but device orientation is

not in a similar fashion In order to match the actual context

that the user holds the phone in a position parallel to the

horizontal plane The presence of human should be included

for consideration [12] [17] are two of the studies that consider

this factor The results show that the RSSI level of attenuation

is greater in the presence of human occlusions Figure 3 shows

the signal attenuation when considering the influence of the

two above factors

From the above 3 inferences, the position class is a function

depending on the three feature variables: Major of nearest

beacon, RSSI of 3 beacons with the smallest RSS and device

orientation Table I is the database for radio map

TABLE I

Position

Major of

#1 beacon

RSSI of

#1,#2 and #3 beacon

Device Azimuth

1j r1

1j , r2

1j , r3

1j h 1j

2j r1

2j , r2

2j , r3

2j h 2j

Cj , r2

Cj , r3

Cj h Cj

Feature vector is saved in the form:

c i = {x i , y i ; m1

ij , r ij1, r ij2, r ij3, h ij },

i ∈ [1, 2, , C] , j ∈ [1, 2, , N ] (3)

-78

-78 -76

-76 -74

-74 -72

-72 -70

-70 -68

-68 -66

-66 -64

-64 -62

-62 -60 dBm

-60 dBm

0 o

 o

0 o

0 o

0 o

0 o

0 o

0 o

0 o

0 o

0 o

0 o

Fig 3 Attenuation of received signal strength at 2 m of distance 0° corresponds to the situation that the user stands and holds the phone the opposite side to the beacon (LOS) 180° corresponds to the case that the user turns away from the beacon (non-LOS).

whereC is the number of RP need to be classified, {x i , y i } is

coordinate ofi th position and N is the number of data points

for each RP

2) The use of K-nearest neighbors in Fingerprinting ap-proach: When we determine what feature will be used Our

problem is to classify the position in a fingerprinting radio map has C reference points Assume that we received a

measurement vectorx at time k:

x k = {m1

k , r k1, r k2, r3k , h k } (4) The measurement vector was then compared to the database

by calculating the Euclidean distance:

d =





|m1− m1

ij |2+3

γ=1

|r k γ − r ij γ |2+ |h k − h ij |2 (5)

The position with the smallest Euclidean distance will be selected as the user position

IV EVALUATION

A Experiment Setup

The proposed indoor positioning system was tested in a semi-open indoor environment on the 1st floor of G2 building, University of Engineering and Technology, Vietnam National University Beacons were placed on the same plane at a height

of 1.8 m where not to be covered by other structures To ensure the reliability of the signal, the broadcast interval of beacons were set to 100 ms (or 10 broadcast messages per 1 second) The minimum distance between two adjacent beacons was 6

m The experimental area was divided into evenly distributed grid points, the two adjacent points were 1.6 m apart We collect data for 15 RPs in the experimental area To ensure

no position change, sampling is done by keeping the phone

fixed on a tripod and placed parallel to the horizontal plane

at a height of 1.4 m The user then turns the device and

moves around the tripod according to the change of the device

heading to get the sample Because the difference of RSSI in

consecutive angle group is negligible, we just sample at 8 fixed

angle which started by 0◦ and next one is increased by45◦.

Each angle is sampled in 30 seconds With a sampling rate of

about 1 sample per 1 second, each RP will be represented by

240 data vectors Thus, the entire radio map data set contains

3600 vectors

15m

1.6m

x y

O

Fig 4 Experimental platform and the RPs distribution

TABLE II

User’s device iPhone SE

Opera System iOS 12.1

Beacon 4 Estimote beacons

Bluetooth Interface BLE v4.2/ 2.4 GHz

Advertising Interval 100 ms

Broadcasting Power 0 dBm

Broadcasting Range 50 m

B Experiment Results

1) Size of the database with lightweight RSS Fingerprint

map: Database size problem was mentioned in [11]

Accord-ingly, to reduce the database size, some components that are

considered redundant should be removed [18] [11] concluded

that it would be redundant to store all the RSS information

of far away beacons even when data was not obtained They

suggest storing data for nearby beacon and set a weight on

them With our proposal, by using the information of the

closest beacon and device heading, with the same number of

vectors, our database size decreases by 77.5% compared to the

conventional method and 30.77% compared to the proposal in

[11]

2) Performance of the proposed method: To verify the

accuracy, we performed two experiments to compare The first

one has done for our proposed method The second one carried

out for conventional RSS-based fingerprinting (not include

ID of nearest beacon and device heading) using kNN The

experiment was executed by allowing users to hold mobile

phones and move around the experimental area At each step,

RP

No. X Y R1 R 2 R 3 R 4 R B

Coordinate of RP RSS (dBm) of all

deployed beacons

Coordinate of RP RSS (dBm) and rank of

nearby u in B beacons

Weight

RP

No. X Y M#1 R #1 R #2

RP

No. X Y XW Y W R 1 #1 R 2 #2 R u #u

Coordinate of RP Major, RSS of closest

beacon and heading

(a)

(b)

(c)

R #3 H

Fig 5 Inside of Fingerprinting database (a) Typical RSS-based finger-printing (b) Santosh’s proposed method [11] (c) Our proposed database fingerprinting method.

position on the true path and the estimated position is recorded

to calculate the error

According to the results shown in Fig 6, with the proposed method, kNN gives better result than conventional method The system reaches 50 percent of the error of less than 1 m and the maximum error is not more than 3 m

Localization Error [m]

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

kNN with Proposed Method kNN with Conventional Approach

Fig 6 Cumulative probability of localization error.

V CONCLUSION ANDFUTUREWORK

In this paper, we introduced ID (Major), RSSI of nearest beacon and device orientation as a feature that defines a reference point in Fingerprint-based indoor localization Since only the information of the nearest beacons is selected, the size

of the database is actually greatly reduced compared to other methods The average error of the system is approximately

1 meter With the choice of using BLE iBeacon, this is a

low-cost, low power consumption solution with high accuracy

Following the conventional Fingerprint approach, this system

is to divide the map into a thick grid of reference points

Therefore, data collection for reference points is a

time-consuming task This is considered as the biggest challenge for

Fingerprint method in actual implementation In the future, our

goal is to reduce the number of reference points to a minimum

while ensuring high accuracy for the system

ACKNOWLEDGEMENT This work has been supported/partly supported by

Viet-nam National University, Hanoi (VNU), under Project No

QG.19.25

REFERENCES [1] Anind Dey, Jeffrey Hightower, Eyal de Lara, Nigel Davies,

“Location-Based Services,” IEEE Pervasive Computing, vol 9, pp 11–12, Mar

2010.

[2] Chouchang Yang, Huai-rong Shao, “WiFi-based indoor positioning,”

IEEE Communications Magazine, vol 53, pp 150–157, Mar 2015.

[3] Adam Satan, “Bluetooth-based indoor navigation mobile system,” in

19th International Carpathian Control Conference (ICCC), May 2018.

[4] Haohao Yin, Weiwei Xia, Yueyue Zhang, Lianfeng Shen, “UWB-based

indoor high precision localization system with robust unscented Kalman

filter,” in IEEE International Conference on Communication Systems

(ICCS), Dec 2016.

[5] Farhan Manzoor, Yi Huang, Karsten Menzel, “Passive RFID-based

Indoor Positioning System, An Algorithmic Approach,” in IEEE

Inter-national Conference on RFID-Technology and Applications, Jun 2010.

[6] Xin-Yu Lin, Te-Wei Ho, Cheng-Chung Fang, Zui-Shen Yen, Bey-Jing

Yang, Feipei Lai, “A mobile indoor positioning system based on iBeacon

technology,” in 37th Annual International Conference of the IEEE

Engineering in Medicine and Biology Society (EMBC), Aug 2015.

[7] Xiangjie Li, Dan Xu, Xuzhi Wang, Rizwan Muhammad, “Design and

implementation of indoor positioning system based on iBeacon,” in

International Conference on Audio, Language and Image Processing

(ICALIP), Jul 2016.

[8] Hanen Ahmadi, Ridha Bouallegue, “Exploiting machine learning

strate-gies and RSSI for localization in wireless sensor networks: A survey,”

in 13th International Wireless Communications and Mobile Computing

Conference (IWCMC), Jun 2017.

[9] Jesus Lovon-Melgarejo, Manuel Castillo-Cara, Luis Orozco-Barbosa,

Ismael Garc´ıa-Varea, “Supervised learning algorithms for indoor

local-ization fingerprinting using BLE4.0 beacons,” in IEEE Latin American

Conference on Computational Intelligence (LA-CCI), Nov 2017.

[10] Pranesh Sthapit, Hui-Seon Gang, Jae-Young Pyun, “Bluetooth Based

Indoor Positioning Using Machine Learning Algorithms,” in IEEE

International Conference on Consumer Electronics - Asia (ICCE-Asia),

Jun 2018.

[11] Santosh Subedi, Hui-Seon Gang, Nak Yong Ko, Suk-Seung Hwang,

Jae-Young Pyun, “Improving Indoor Fingerpriting Positioning With

Affinity Propagation Clustering and Weighted Centroid Fingerprint,”

IEEE Access, vol 7, pp 31738 - 31750, Mar 2019.

[12] Alberto Belmonte-Hern´andez, Gustavo Hern´andez-Pe˜naloza, David

Mart´ın Guti´errez, Federico ´ Alvarez, “SWiBluX: Multi-Sensor Deep

Learning Fingerprint for Precise Real-Time Indoor Tracking,” IEEE

Sensors Journal, vol 19, pp 3473 - 3486, Jan 2019.

[13] “Getting Heading and Course Information” Apple [online] Available:

https://bitly.vn/2ntm

[14] Manish Wadhwa, Min Song, Vinay Rali, Sachin Shetty, “The impact

of antenna orientation on wireless sensor network performance,” in 2nd

IEEE International Conference on Computer Science and Information

Technology, Aug 2009.

[15] Zhi-An Deng, Zhiyu Qu, Changbo Hou, Weijian Si, Chunjie Zhang,

“WiFi Positioning Based on User Orientation Estimation and

Smart-phone Carrying Position Recognition,” Hindawi Wireless

Communica-tions and Mobile Computing, vol 2018, Sep 2018.

[16] Munesh Singh, Pabitra Mohan Khilar, “Actuating Sensor For

Deter-mining The Direction Of Arrival Using Maximal RSSI,” International

Journal of Scientific and Technology Research, vol 3, Sep 2018.

[17] Syed Hassan Ahmed, Safdar H Bouk, N Javaid, Iwao Sasase, “Com-bined Human, Antenna Orientation in Elevation Direction and Ground

Effect on RSSI in Wireless Sensor Networks,” in 10th International

Conference on Frontiers of Information Technology, Dec 2012.

[18] J Luo, L Fu, “A smartphone indoor localization algorithm based on WLAN location fingerprinting with feature extraction and clustering,”

Sensors, vol 17, p 1339, Jun 2017.

[19] Xin-Yu Lin, Te-Wei Ho, Cheng-Chung Fang, Zui-Shen Yen, Bey-Jing Yang, Feipei Lai, “A mobile indoor positioning system based on iBeacon

technology,” in 37th Annual International Conference of the IEEE

Engineering in Medicine and Biology Society (EMBC), Aug 2015.

[20] “What’s New in Core Location” Apple [online] Available: https://developer.apple.com/videos/play/wwdc2013/307/