Indoor Localisation with Intelligent Luminaires for Home Monitoring

Iuliana Marin, Maria Iuliana Bocicor and Arthur-Jozsef Molnar

SC Info World SRL, Bucharest, Romania

Keywords:

Indoor Localization, Received Signal Strength, Trilateration, Kalman Filtering, Neural Network.

Abstract:

This paper presents the initial results of our experiments regarding accurate indoor localisation. The research

was carried out in the context of a European Union funded project targeting the development of a conﬁgurable,

cost-effective cyber-physical system for monitoring older adults in their homes. The system comprises a

number of hardware nodes deployed as intelligent luminaires that replace light bulbs present in the monitored

location. By measuring the strength of a Bluetooth Low Energy signal generated by a device on the monitored

person, a rough estimation of the person’s location is obtained. We show that the presence of walls, furniture

and other objects in typical indoor settings precludes accurate localisation. In order to improve accuracy, we

employ several software-based approaches, including Kalman ﬁltering and neural networks. We carry out an

initial experiment showing that additional software processing signiﬁcantly improves localisation accuracy.

1 INTRODUCTION

In the context of demographic and social changes

nowadays (World Health Organization, 2015) care for

older adults can be transformed according to their

own desires and necessities. Recent years have seen

a shift towards technology-driven, personalised solu-

tions designed to assist and monitor older adults living

in their own homes, with the support of their families,

friends and trained caregivers.

The work reported in this paper is part of a Eu-

ropean Union-funded research project during which a

technological platform for home monitoring and as-

sisted living was developed. This cyber-physical sys-

tem (Marin et al., 2018) is based on the use of energy-

efﬁcient intelligent luminaires equipped with embed-

ded sensing, indoor localisation and communication

systems. They enable a pervasive, seamless, and in-

expensive home monitoring scheme. These newly de-

veloped luminaires can replace existing light bulbs,

ensuring simple and straightforward deployment. The

platform removes some of the adoption barriers faced

by such systems by reducing costs and simplifying in-

stallation. Furthermore, it conforms to standards, and

using it does not require specially trained staff.

The main system components, together with pro-

posed innovations were described in previous papers

(Marin et al., 2018; Bocicor et al., 2017). One of

the important components is the one for indoor lo-

calisation. Existing satellite-based technologies such

as GPS perform very poorly indoors due to the pres-

ence of walls and furniture, which have a detrimen-

tal effect on signal strength, thus requiring alternative

approaches. Our system employs trilateration based

on the received strength of a Bluetooth Low Energy

(BLE) signal. Other approaches include visible light

communication (Haigh et al., 2014), acoustic back-

ground ﬁngerprinting (Tarzia et al., 2011) and user

movements (Tarrio et al., 2011). The main advan-

tages of BLE technology is the low cost of hardware

and its ubiquity in mobile devices and wearables. In

addition, Bluetooth signal is omni-directional and it

permeates walls, albeit imperfectly. We carry out an

experiment where we replace one light bulb in ev-

ery room with an intelligent luminaire incorporating

Bluetooth technology used both for localisation and

communication. This leads to a deployment that does

not require additional wires or devices and with no

components that look out of place. In addition, as

most lights are ceiling-mounted, there is less interfer-

ence from furniture or other equipment.

2 STATE OF THE ART

The last years we have witnessed a dramatic rise in the

number and diversity of intelligent devices deployed

within homes. Location-based applications play an

important role in identifying and studying user be-

haviours and their interests by analysing activities and

interactions with other people or objects.

464

Marin, I., Bocicor, M. and Molnar, A.

Indoor Localisation with Intelligent Luminaires for Home Monitoring.

DOI: 10.5220/0007751304640471

In Proceedings of the 14th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2019), pages 464-471

ISBN: 978-989-758-375-9

In outdoor environments, accurate location infor-

mation can be easily obtained using the Global Po-

sitioning System (GPS). While GPS positioning is

reliable outdoors, it cannot be used for accurate in-

door localisation due to poor signal strength (Ozsoy

et al., 2013). As such, efforts have been put into

designing accurate indoor positioning systems based

on acoustics, optics, Wi-Fi, Bluetooth, GSM and Ra-

dio Frequency Identiﬁcation (Ta, 2017; Lymberopou-

los et al., 2015; Xiao and Zhou, 2016). Some of

these systems employ Received Signal Strength In-

dex (RSSI) (Luo et al., 2011; Sadowski and Spa-

chos, 2018; Maduskar and Tapaswi, 2017), due to

the existence of the inverse-square law between the

strength of the received signal and the distance from

the source of the electromagnetic wave (Dong and

Dargie, 2012). This relation allows determining the

distance starting from signal strength. Unfortunately,

the RSSI varies over time due to multipath fading,

with the level of ﬂuctuation being higher in indoor en-

vironments (Pu et al., 2011). This instability degrades

localisation accuracy, keeping indoor localisation and

object detection an area of ongoing research.

One proposed solution to address RSSI measure-

ment variance is to ﬁlter RSSI values. The Simulta-

neous Localisation and Conﬁguration algorithm pro-

posed by Bulten et al. (Bulten et al., 2016) is based

on a factored solution to the simultaneous localisa-

tion and mapping problem (FastSLAM) for localisa-

tion based on RSSI values. Everything is done locally

on the user’s device. Another approach is the Kalman

ﬁlter (Welch and Bishop, 1995), which can be used to

smooth noisy data, i.e. RSSI value estimations. This

technique is amenable for use in many contexts given

its low computational requirements. The Kalman ﬁl-

ter is an efﬁcient recursive ﬁlter that factually eval-

uates the internal state of a system from a series of

measurements which are subject to noise. When a

new measurement occurs, the ﬁlter uses a weighted

average for prediction, with the goal of reducing un-

certainty. The weights are computed based on the co-

variance. The weighted average is a state estimate

which is between the predicted and measured states.

This process is repeated at a certain period of time,

computing the new average and covariance for every

iteration. The Kalman ﬁlter aims to remove signal

noise when multiple measurements are taken.

In (Robesaat et al., 2017), the authors showed that

Kalman ﬁltering increased the accuracy of the indoor

localisation of the monitored person. The Kalman

ﬁlter reduced accumulated errors by reducing noise

(Sung, 2016). It also diminished the energy consump-

tion and enhanced the stability of RSSI values.

Similar to GPS technology, indoor localisation re-

quires the use of several signal emitters. Each signal

emitter is placed at a known location, and RSSI val-

ues for each are determined to calculate the distance

between each signal emitter and the target of monitor-

ing. Using at least three signal emitters results in the

creation of three circles, the radii of which are deter-

mined by the RSSI values. The target of monitoring is

found at the point where these circles intersect. This

technique is called trilateration. Due to the issues de-

tailed above, RSSI values require post-processing and

ﬁltering in order to achieve accurate localisation.

The system we proposed uses custom developed

BLE-enabled devices, which are physically deployed

in the form of intelligent luminaires that replace exist-

ing light bulbs (Marin et al., 2018). While the physi-

cal design of the intelligent bulbs is beyond the scope

of this paper, care was taken to ensure luminaire de-

sign does not affect signal strength or directionality.

The light bulbs are typically ceiling mounted and thus

the signal is less prone to physical and electromag-

netic obstructions.

Our experiments employ a mobile phone to emit a

BLE signal that is received by intelligent luminaires

and transformed into localisation information. Any

smartwatch or smart life saving button can act as a

signal emitter, as long as it supports the BLE stan-

dard. The main advantage of our setup is cost, both

regarding the design of the intelligent luminaires, as

well as the possibility of using a wide range of exist-

ing BLE-enabled wearable devices

3 INDOOR LOCALISATION

COMPONENT

3.1 Intelligent Luminaires

Figure 1 presents a high-level view of the system’s

deployment architecture. The hardware side of the

cyber-physical system is represented by two types of

hardware devices, dubbed as dummy and smart nodes.

Both are implemented in the form of luminaires.

Dummy nodes are designed for small size and low

cost. They also work as luminaires, but lack the ad-

vanced processing and communication functions of

the smart nodes. Their main role, besides lighting,

is to perform BLE scans for other devices and report

the RSSI values to the smart node they are connected

to. As such, dummy nodes are employed to detect

the presence of monitored persons. This design re-

duces the overall system cost and improves its accu-

racy, by allowing several inexpensive nodes to be de-

ployed and connected to the same smart node. Smart

Indoor Localisation with Intelligent Luminaires for Home Monitoring

465

nodes provide several wireless interfaces and higher

computation power than is available to dummy nodes.

Smart nodes are both Bluetooth and Wi-Fi enabled.

They also include a sensor module with a suite of sen-

sors used to monitor the indoor environment. They

receive dummy node readings via Bluetooth, process

them and transmit them to the software server using a

permanent Wi-Fi connection. In order to achieve this,

every monitored room typically includes between 1

and 3 luminaires, depending on its size and shape. By

replacing ceiling mounted light bulbs with intelligent

devices, homes and buildings can be monitored with

minimal installation and maintenance costs. In ad-

dition, when deploying the system in more personal

spaces, such as hospital wards, or in the homes of

older adults with the purpose of monitoring, they have

the advantage of being inconspicuous.

Figure 1: High-level architecture of the system.

Communication between smart and dummy bulbs

is achieved via BLE technology. To establish a BLE

connection, the smart bulb acts as a master, perform-

ing a scan to detect dummy bulbs in its proximity,

while the dummy acts as a slave, advertising itself.

On the other hand, the dummy must also be able to

scan the environment and detect additional devices,

such as those employed for localisation. As such, it

must also act as a master. As a result of these con-

straints, smart and dummy bulbs switch between roles

with regularity. For a certain period of time, each

node acts as a slave, emitting advertisement packets

detected by bulbs in a master role, and then switches

to slave mode, in order to detect and maintain infor-

mation about other devices.

The localisation algorithms based on the RSSI val-

ues are run on the system’s server several times ev-

ery minute, and the retrieved indoor coordinates are

recorded in a database. Considering current and re-

cent positions of the monitored person, the system is

able to detect possibly dangerous situations and emit

real time alerts, which are sent to designated care-

givers. As an example, assuming that the monitored

person has been in the bathroom for too long, using

a custom period deﬁned at system conﬁguration, the

platform will generate and send an alert to their care-

giver.

3.2 Methodology

In this section we present the direct and hybrid meth-

ods that we use to accurately identify the indoor po-

sition of the monitored person. Direct trilateration

and prediction of coordinates via an artiﬁcial neural

network are the two major techniques we are using.

Furthermore, we employ Kalman ﬁltering (Welch and

Bishop, 1995) to account for the inherent noise of

RSSI measurements.

3.2.1 Direct Trilateration

To convert RSSI values into distance measurements

we use the RSSI lognormal model described in (Dong

and Dargie, 2012), in which the distance can be com-

puted as shown in Formula 1:

d =

A−RSSI

10 ∗ n

(1)

In Formula 1 the signiﬁcance of variables is as fol-

lows:

• n represents the path-loss exponent and it ranges

between 2 to 6 for indoor environments.

• A is the signal strength expressed in dBm, mea-

sured from a distance of one meter. This parame-

ter is experimentally computed once for each type

of luminaire.

• RSSI is the received signal strength index. This is

the measure the system uses for trilateration.

• d is the computed distance. It is the distance be-

tween the luminaire that makes the measurement

and the person being monitored, or more speciﬁ-

cally the Bluetooth device that the person has on

them, such as a smartphone or smartwatch.

The distance is calculated relative to minimum 3

intelligent luminaires in order to achieve trilateration.

Following the calculations, at least three values are

deduced, for the three distances which represent the

three circles’ radii. The person’s location is computed

as the point of intersection of these circles. Because

of various environment interference caused by objects

blocking the signal, RSSI values are subject to noise.

We evaluate three different techniques to reduce noise

and improve the accuracy of localisation.

First, location at each recorded timestamp is not

computed using the RSSI value at that particular

ENASE 2019 - 14th International Conference on Evaluation of Novel Approaches to Software Engineering

466

timestamp, but considering an average of RSSI val-

ues. Let us consider n RSSI values for each moment

in time when signal strength was recorded, for each of

the considered luminaires (l

, l

, ··· , l

), where i ≥ 3.

To compute the location at time t (1 ≤ t ≤ n) an

option would be to use all RSSI values recorded at

time t (l

, l

), compute the distances according to

Formula 1 and identify the intersection point of the

three circles having as radii the computed distances

at each moment t. However, the noise present in

RSSI measurements precludes us from doing this. We

use a different approach, one that involves the last 10

recorded RSSI values, for each luminaire (l

t−10

, l

t−9

·· ·, l

t−1

). The reasoning is that single-reading RSSI

errors will be attenuated when taking multiple read-

ings over a longer time interval. Thus, for each intel-

ligent bulb, out of these ten values possible outliers

are eliminated and then the average avg

and stan-

dard deviation stdev

are computed. Out of all ten

values, we retain only those that are within the inter-

val [avg

−stdev

, avg

+stdev

] and use their average

to compute the distance.

The second technique used to account for noise is

Kalman ﬁltering. Starting from observed noisy mea-

surements and assuming a Gaussian distributed noise,

this algorithm infers the parameters of interest, taking

into account the history of measurements. For this

speciﬁc case the Kalman ﬁlter will be used to make

estimations of noise-free signal strength values. The

process is based on two equations that describe the

system’s state: one deﬁnes the observed measurement

in relation to the real measurement and noise and the

other deﬁnes the current estimate recursively, using

the previous estimate. There are two main steps to be

performed: a prediction about what the state should

be and an update to this prediction, based on the mea-

surement data and the old estimate. For each moment

in time we compute the RSSI estimates, using Kalman

ﬁltering. Then the trilateration is performed using the

distances obtained starting from these estimates.

The ﬁnal technique is a combination of the pre-

viously presented methods: initially, a Kalman ﬁlter

is applied on the raw RSSI values and the RSSI es-

timations are obtained. Then, we apply the heuris-

tic described previously, that we proposed as the ﬁrst

technique for noise reduction. We compute the lo-

cation at a certain moment by considering the RSSIs

associated with the past 10 measurements. The differ-

ence is that instead of employing the measured RSSI

values, we use the estimates. Given that calculations

are carried out on the server, further software-based

improvements can be integrated without affecting al-

ready deployed devices.

3.2.2 Using an Artiﬁcial Neural Network

In addition to the previous methods that employ di-

rect trilateration to obtain location, we also use an ar-

tiﬁcial neural network to predict the position of the

monitored person, starting from tuples of RSSI val-

ues. Such machine learning models have successfully

been employed in recent years for indoor positioning

(Zhanga et al., 2016; Lukito and Chrismanto, 2017;

Mittal et al., 2018). The network’s architecture is

not ﬁxed, as a suitable architecture is dependant on

the problem space and speciﬁc data. Therefore, we

experimented with various architectures. Still, cer-

tain aspects are established: the input layer contains

three neurons, one for each RSSI measurement, cor-

responding to the three intelligent luminaires used in

our experiment. The output layer contains two out-

puts, representing the resulting indoor coordinates in

a two-dimensional space. We experimented with a

variable number of hidden layers and report the ob-

tained results. The used activation function for all

hidden layers is rectiﬁed linear unit (ReLU) (Nair and

Hinton, 2010). Optimisation is achieved via stochas-

tic gradient descent enhanced with the adam opti-

miser (Kingma and Ba, 2014). Training was per-

formed during a variable number of epochs and the

network is evaluated using k-fold cross validation.

4 EXPERIMENTS AND RESULTS

The experiments were carried out in a three-room

dwelling in which the system was installed. The room

dimensions are: 2.50m x 3.29m (bedroom), 2.50m x

1.00m (study room), 2.34m x 2.21m (hallway). The

apartment also includes a kitchen and a bathroom, but

these were not considered in the evaluation. Each

room was ﬁtted with a ceiling-mounted smart node

connected to the software server over Wi-Fi. The

two-dimensional plan of the dwelling is represented

in Figure 2, in which the smart bulbs are illustrated

using yellow circles with black borders.

Both interior and exterior walls are made of brick

and cement. Exterior walls have a thickness of 35cm,

while the interior walls are 17cm thick. This is impor-

tant to note because these materials strongly inﬂuence

obtained RSSI values.

Formula 1 uses two parameters that need calibra-

tion: A - the signal strength at 1 meter from the lu-

minaire and n - the path loss exponent. The value

of A depends on the luminaire, and is computed once

for each type. In our case, we obtained a value of

A = −67. To obtain the most suitable value for n we

used a grid-search procedure and kept the value that

Indoor Localisation with Intelligent Luminaires for Home Monitoring

467

Figure 2: Plan of the dwelling used for experimentation

(icons from https://www.vecteezy.com).

lead to the most accurate results for the considered

indoor environment (n = 2.5). The result of the cal-

ibration step depends on deployment location, room

shapes and sizes, as well as wall placement and mate-

rials. Calibration must be undertaken once per loca-

tion, when the system is initially deployed.

The subsequent sections present the experiments

that were carried out to test and improve indoor lo-

calisation, starting from RSSI values. Mean squared

error (MSE) is the evaluation measure used and

it is computed considering the real indoor two-

dimensional coordinates as well as the coordinates

obtained by the positioning algorithms.

During the experiment the person is stationary in

the middle of each room for 15 minutes: bedroom

(02:33 - 02:48), study room (02:49 - 03:04) and hall-

way (03:07 - 03:22). Their mobile phone, which is

used for recording measurements, is kept on them.

RSSI value measurements are taken at an interval be-

tween 6 to 10 seconds. All doors between rooms are

open.

Figure 3 illustrates the RSSI values for all 3 inter-

vals of experimentation time, recorded by each of the

three intelligent bulbs. The Kalman estimates com-

puted starting from these RSSIs are also illustrated

using dashed lines. During the last 2 time periods (3b,

3c) we observe that the RSSIs do not suffer signiﬁcant

changes. On the other hand, for the ﬁrst time interval,

there seems to be more noise attenuated by Kalman

ﬁltering.

Table 1 shows the MSE for the three considered

time periods, for each of the techniques. Original val-

ues refer to measured, unprocessed RSSI values. The

second row presents the results for the heuristic pro-

posed in Section 3.2: the values are calibrated accord-

(a) First interval - person is located in the bedroom.

(b) Second interval - person is located in the study room.

Figure 3: Recorded RSSIs (continuous lines) vs. Kalman

estimates (dashed lines) for each time interval, correspond-

ing to each location.

ing to the described technique and use the previous

10 measurements. The third row shows the obtained

results when a Kalman ﬁlter is applied and the loca-

tions are computed using the obtained estimates. The

bottom row presents the errors for the hybrid proce-

dure - the heuristic is applied on the estimates ob-

tained from running Kalman ﬁlters. It can be ob-

served that when using trilateration applied directly

on the original RSSIs the results are not as good as

with the other techniques. On average, the proposed

heuristic leads to the next better outputs, followed by

Kalman ﬁlters, which result in even smaller errors.

The hybrid combination described in Section 3.2 sur-

passes all previous ones. Another observation is that

the results for the ﬁrst time interval accurately reﬂect

ENASE 2019 - 14th International Conference on Evaluation of Novel Approaches to Software Engineering

468

Table 1: Mean squared error using direct trilateration. Er-

rors reported in centimetres.

2:33-2:48 2:49-3:04 3:07-3:22 Avg. error

Original values 187.92 92.41 77.45 119.26

Heuristic 184.17 92.0377 77.69 117.97

Kalman ﬁlter 181.81 89.5987 77.50 116.30

Kalman ﬁlter +

178.11 88.87 77.47 114.81

heuristic

the higher noise reduction achieved using Kalman ﬁl-

ters, as shown in Figure 3a. The difference in error

between the original values and Kalman ﬁlters (Table

1, column 2:33-2:48), ﬁrst and third rows is signiﬁ-

cantly higher than that corresponding to the other two

time periods.

The neural network described in Section 3.2 has

been applied on the raw RSSI inputs measured during

our experiment. These results are presented in Table

2. Note that generally the increase in the number of

training epochs leads towards smaller errors, however

this is not always the case, as over-ﬁtting might arise.

The number of hidden layers also has an essential im-

pact on the network’s performance. Generally, while

our experiments show that the tendency of the output

error is to diminish as the number of epochs increases,

that was not always the case.

Table 2: Mean squared error obtained using the artiﬁcial

neural network. Errors reported in centimetres.

Number of hidden layers

Epochs 1 2 3 4 5

500 48.14 47.67 26.18 47.39 16.45

1000 34.54 14.53 12.80 6.25 16.41

1500 40.12 6.32 7.58 5.20 4.18

2000 44.57 13.53 4.12 2.22 4.07

2500 9.31 14.24 4.94 6.02 2.83

3000 27.90 14.50 17.01 3.83 3.13

The highest average error obtained by the net-

work (48.14 centimeters) is substantially lower than

the best results reported in Table 1. In the best case,

the network’s error is less than 3 centimetres and on

average it is less than 20 centimetres, for all under-

taken tests. The 95% conﬁdence interval for all runs

is 16.87 ± 5.46 (centimetres). Thus, for the consid-

ered data, the network performs considerably better

when compared to trilateration-based methods. How-

ever, a disadvantage of the network is that it has to be

trained using a considerable amount of data collected

in various settings and from various positions in order

to be suitable for real-time localisation.

Figure 4 illustrates the positions obtained by

the best performing trained neural network. This

plot represents the two-dimensional ﬂoor plan of the

dwelling, as seen in Figure 2. For more context, we

mention that the left-most wall’s x coordinate is 167

(left wall of the hallway) and the entrance door’s y

Figure 4: Indoor positions obtained by the neural net-

work, represented in the two-dimensional ﬂoor plan of the

dwelling (Figure 2). The units of measure in this plot are

centimetres.

coordinate is 40; the measurements are in centime-

tres. For a certain location, many of the predicted po-

sitions are identical and thus, even though the number

of predictions is the same with the number of recorded

RSSI values, they are all superimposed in the image.

We notice that the accuracy is very high, particularly

for the ﬁrst and the last periods of time (bedroom and

hallway). In the case of the study (second period

of time) the predicted locations are not all identical,

however the error is low and the predictions are very

close to reality.

The location of the person during the second time

interval is illustrated in Figure 5. The image is ex-

tracted from the cyber-physical system’s software ap-

plication, which allows displaying the computed in-

door locations of the monitored person in real-time,

both in a two- and and three-dimensional space. The

image illustrates the positions of the smart bulbs as

well as the person’s estimated location. This applica-

tion feature allows caregivers to supervise the moni-

tored person’s location in real time.

Figure 5: Illustration of the person’s indoor location within

the system’s interface.

Indoor Localisation with Intelligent Luminaires for Home Monitoring

469

5 COMPARISON WITH

RELATED WORK

We brieﬂy compare relevant existing systems with

ours. The RADAR system (Bahl and Padmanabhan,

2000) determines the person’s current position by

considering the signal strength of its k nearest neigh-

bours. The signal is inﬂuenced by noise, building ma-

terials and other environment-dependant factors. The

localisation error was between 2 and 3 meters. For

the system where a neural network was used for de-

termining the location of a person based on wireless

LANs, the accuracy is of about 3 meters (Battiti et al.,

2002). For the localisation based on trilateration and

Kalman ﬁlters (Fariz et al., 2018), the error varied be-

tween 14.82 and 3.58 meters. The extended Kalman

ﬁlter for indoor localisation based on fusing Wi-Fi

(Deng et al., 2015) has a positioning error of 2.83

meters. Another paper targeting indoor positioning

(Li et al., 2018) reported a maximum positioning er-

ror of 3.29 meters when using direct trilateration with

Bluetooth and 1.61 meters when using a back propa-

gation neural network. As opposed to these reports,

our techniques lead to smaller errors for the consid-

ered experiment, having a maximum average error of

1.19 metres. However, a thorough comparison cannot

be undertaken given differences in experiment setup,

measurement devices and procedures.

Indoor localisation based on RSSI data gathered

via Wi-Fi was tested within two environments and

triggered lower positioning errors (Sadowski and Spa-

chos, 2018). The ﬁrst environment was one big room

in which there were several BLE devices and multi-

ple Wi-Fi networks causing interference. The second

environment comprised a smaller room that only con-

tained furniture with no other devices present, lead-

ing to a low level of noise. Inside the ﬁrst environ-

ment the average error for positioning was 0.84 me-

ters, while for the second environment it was 0.48

meters. The nodes which measured the RSSI values

have been placed on tables of the same height for lim-

iting the number of signal reﬂections. In our work,

the nodes were ceiling-mounted and three rooms were

considered, having eight Wi-Fi networks and seven

BLE connections causing interference. The phone

was placed at about 1 meter above the ﬂoor. The walls

made of brick and cement caused reﬂections and due

to the multiple path signal propagation, a higher than

desired positioning error was obtained. Transmission

signals are also blocked due to surrounding furniture

and equipment, as well as persons who move inside

the environment (Wang et al., 2012), including the

monitored person itself.

6 CONCLUSION

During the past decades there has been a dramatic in-

crease in the world’s older population and this trend is

expected to continue. Population ageing has a series

of major effects in various sectors such as housing, lo-

cal communities and healthcare. As a consequence, a

variety of technology-based solutions have emerged,

which aim to facilitate maintaining an independent

lifestyle for older adults within their communities and

to ensure their safety. This study presents the results

of our experiments regarding the indoor localisation

component of a cyber-physical system designed to aid

in monitoring older adults.

The system is built around a network of intel-

ligent luminaires equipped with embedded sensing

for environment monitoring, indoor localisation and

communication. The system measures received sig-

nal strength from nearby BLE devices and is able to

compute their indoor location with actionable accu-

racy. Two techniques were used for indoor position-

ing: one using direct trilateration and one based on

an artiﬁcial neural network. We proposed a heuristic

for more accurate localisation that also incorporates

a noise reduction technique. Our experiments show

that the neural network is superior to direct trilatera-

tion. With regard to trilateration, results indicate that

heuristics combined with Kalman ﬁltering for noise

reduction lead to smaller localisation errors.

Compared to other indoor positioning systems that

require ﬁngerprinting or additional wires or devices,

our system has a signiﬁcant advantage: no complex

indoor measurements are needed. Furthermore, de-

ployment is straightforward, as existing bulbs can

be replaced by intelligent luminaires using the same

electrical infrastructure.

Future work will be carried out towards further

improving localisation accuracy. Further experimen-

tation will be performed within the same setup, with

both opened and closed doors and considering user

movement to augment static localisation results. As a

consequence of more tests, additional data will be col-

lected, which will also enable a more thorough train-

ing process for the artiﬁcial neural network and enable

other approaches.

ACKNOWLEDGEMENT

This work was supported by a grant of the Ro-

manian National Authority for Scientiﬁc Research

and Innovation, CCCDI UEFISCDI, project number

46E/2015, i-Light - A pervasive home monitoring sys-

tem based on intelligent luminaires.

ENASE 2019 - 14th International Conference on Evaluation of Novel Approaches to Software Engineering

470

REFERENCES

Bahl, P. and Padmanabhan, V. N. (2000). Radar: An in-

building RF-based user location and tracking system.

In Proc. IEEE INFOCOM 2000, volume 2, pages

775–784.

Battiti, R., Thang Le, N., and Villani, A. (2002). Location-

aware computing: A neural network model for de-

termining location in wireless LANs. Int. Semicond.

Conf., 4.

Bocicor, M. I. et al. (2017). Cyber-physical system for as-

sisted living and home monitoring. In 13th IEEE In-

ternational Conference on Intelligent Computer Com-

munication and Processing, pages 487–493.

Bulten, W., Rossum, A. C. V., and Haselager, W. F. G.

(2016). Human slam, indoor localisation of devices

and users. In IEEE First International Conference on

Internet-of-Things Design and Implementation.

Deng, Z. A., Hu, Y., Yu, J., and Na, Z. (2015). Extended

kalman ﬁlter for real time indoor localization by fus-

ing wiﬁ and smartphone inertial sensors. Microma-

chines, 6:523–543.

Dong, Q. and Dargie, W. (2012). Evaluation of the reliabil-

ity of RSSI for indoor localization. In Proceedings of

the IEEE International Conference on Wireless Com-

munications in Unusual and Conﬁned Areas.

Fariz, N. et al. (2018). An improved indoor location tech-

nique using kalman ﬁlter. International Journal of En-

gineering & Technology, 7.

Haigh, P. A. et al. (2014). Visible light communications:

real time 10 mb/s link with a low bandwidth poly-

mer light-emitting diode. Optics express, 22(3):2830–

2838.

Kingma, D. P. and Ba, J. (2014). Adam: A method for

stochastic optimization. CoRR, abs/1412.6980.

Li, G. et al. (2018). Indoor positioning algorithm based on

the improved RSSI distance model. Sensors, 18(9).

Lukito, Y. and Chrismanto, A. R. (2017). Recurrent neu-

ral networks model for WiFi-based indoor positioning

system. In 2017 International Conference on Smart

Cities, Automation & Intelligent Computing Systems

(ICON-SONICS).

Luo, X., O’Brien, W. J., and Julien, C. (2011). Comparative

evaluation of received signal strength index (RSSI)-

based indoor localization techniques for construction

jobsites. Adv. Eng. Inform., 25:355–363.

Lymberopoulos, D. et al. (2015). A realistic evaluation and

comparison of indoor location technologies: Experi-

ences and lessons learned. In The 14th ACM/IEEE

Conference on Information Processing in Sensor Net-

works (IPSN ’15).

Maduskar, D. and Tapaswi, S. (2017). RSSI based adap-

tive indoor location tracker. Scientiﬁc Phone Apps and

Mobile Devices, 3.

Marin, I. et al. (2018). i-Light - Intelligent Luminaire Based

Platform for Home Monitoring and Assisted Living.

Electronics, 7(220).

Mittal, A., Tiku, S., and Pasricha, S. (2018). Adapting Con-

volutional Neural Networks for Indoor Localization

with Smart Mobile Devices. In Proceedings of the

2018 on Great Lakes Symposium on VLSI.

Nair, V. and Hinton, G. E. (2010). Rectiﬁed linear units im-

prove restricted boltzmann machines. In Proceedings

of the 27th International Conference on International

Conference on Machine Learning, ICML’10, pages

807–814, USA. Omnipress.

Ozsoy, K., Bozkurt, A., and Tekin, I. (2013). Indoor po-

sitioning based on global positioning system signals.

Microwave and Optical Technology Letters, 55:1091–

1097.

Pu, C. C., Pu, C.-H., and Lee, H.-J. (2011). Indoor location

tracking using received signal strength indicator. In

Emerging Communications for Wireless Sensor Net-

works.

Robesaat, J., Zhang, P., Abdelaal, M., and Theel, O. (2017).

An improved ble indoor localization with kalman-

based fusion: An experimental study. Sensors, 17.

Sadowski, S. and Spachos, P. (2018). RSSI-based indoor

localization with the internet of things. In IEEE Ac-

cess.

Sung, Y. (2016). RSSI-based distance estimation frame-

work using a kalman ﬁlter for sustainable indoorcom-

puting environments. Sustainability, 8.

Ta, V.-C. (2017). Smartphone-based indoor positioning

using Wi-Fi, inertial sensors and Bluetooth. https://

tel.archives-ouvertes.fr/tel-01883828/document. Ma-

chine Learning, Universit

e Grenoble Alpes.

Tarrio, P. et al. (2011). An energy-efﬁcient strategy for

combined rss-pdr indoor localization. In IEEE In-

ternational Conference on Pervasive Computing and

Communications Workshops (PERCOM Workshops),

pages 619–624.

Tarzia, S. P. et al. (2011). Indoor localization without in-

frastructure using the acoustic background spectrum.

In Proceedings of the 9th international conference on

Mobile systems, applications, and services, page 155.

Wang, J.-Y. et al. (2012). High-precision RSSI-based in-

door localization using a transmission power adjust-

ment strategy for wireless sensor networks. In 2012

IEEE 14th International Conference on High Perfor-

mance Computing and Communication, 2012 IEEE

9th International Conference on Embedded Software

and Systems.

Welch, G. and Bishop, G. (1995). An introduction to the

kalman ﬁlter. Technical report, Department of Com-

puter Science, University of North Carolina at Chapel

Hill, Chapel Hill, NC, USA.

World Health Organization (2015). World report on age-

ing and health. http://apps.who.int/iris/bitstream/

10665/186463/1/9789240694811 eng.pdf?ua=1.

Xiao, J. and Zhou, Z. (2016). A survey on wireless indoor

localization from the device perspective. ACM Com-

puting Surveys, 49.

Zhanga, W., Liua, K., Zhang, W., Zhang, Y., and Gu, J.

(2016). Deep neural networks for wireless localiza-

tion in indoor and outdoor environments. Neurocom-

puting, 194:279–287.

Indoor Localisation with Intelligent Luminaires for Home Monitoring

471