Multi-radio Data Fusion for Indoor Localization using Bluetooth and

WiFi

Afaz Uddin Ahmed

1,2

, Reza Arablouei

1 b

, Frank de Hoog

1 c

Branislav Kusy

and Raja Jurdak

1,2 e

School of Information Technology and Electrical Engineering, University of Queensland, St. Lucia, QLD 4072, Australia

Data61 Group, CSIRO, QLD 4069, Australia

Keywords:

Data Fusion, Multi-radio, Indoor Localization, WiFi, Channel State Information, Bluetooth Low Energy.

Abstract:

Location estimation through fusing the information obtainable from multiple radio systems can reduce the

dependency on each system and improves the performance. Research on fusion-based indoor localization

using WiFi and Bluetooth-low-energy (BLE) beacons has mostly been limited to training-based approaches.

In this paper, we propose a training-free indoor localization technique using received signals from WiFi and

BLE device. The proposed technique estimates the position of the user device by fusing the information

that it gains regarding the position of the target from the WiFi channel state information (CSI) and the RSSI

measurements of BLE beacons. We use the WiFi CSI to estimate the angle of arrival (AoA), which we then use

in conjunction with the RSSI measurements from the BLE beacons to develop a multi-radio fusion framework

for indoor localization. We use a weighted centroid localization method to obtain an initial position estimate

from the RSSI measurements. The initial position estimation helps to resolve the ambiguities in the AoA.

The proposed technique is based on maximum-likelihood estimation (MLE) exploiting the probability density

functions of the estimated AoA and the RSSI-induced distances. Simulation results show that the proposed

technique improves the localization accuracy by 30% in a typical indoor environment compared with previous

approaches.

1 INTRODUCTION

Indoor localization using measurements from mul-

tiple wireless platforms offers better reliability for

location-based applications compared to a single

wireless platform (Kanaris et al., 2017). Fusing mea-

surements from multi-radio systems compensate for

the limitations of the individual platforms such as in-

consistency in the measurements, the randomness of

the measured data due to noise from the environment,

high computational complexity, and high duty cycle.

Integration of measurements from multiple platforms

improves the reliability and accuracy of the local-

ization system. It reduces the error in location esti-

mation and performs better than the stand-alone sys-

tem (Hilsenbeck et al., 2014).

https://orcid.org/0000-0002-9919-3126

https://orcid.org/0000-0002-6932-2900

https://orcid.org/0000-0002-4632-564X

https://orcid.org/0000-0001-9082-3243

https://orcid.org/0000-0001-7517-0782

Data fusion for user localization was initially in-

troduced by fusing the GPS and inertial sensors data

from the smart-phone for outdoor tracking (Qi and

Moore, 2002). The main purpose was to make the lo-

calization process more energy efﬁcient by reducing

the duty cycle of the GPS unit. This is feasible be-

cause even if the GPS signal gets interrupted at a reg-

ular interval, data collected from the inertial sensors

can aide the localization process through a data fusion

framework. Multi-radio-fusion-basedlocalization is a

new branch in indoor localization but is challenged by

the random ﬂuctuation of signal strength.

Past fusion based indoor localization tech-

niques were developed on platforms such as, Blue-

tooth (Mirowski et al., 2013), WiFi (Lepp¨akoski et al.,

2013; Lim et al., 2007), ultra-wide-band (Gezici et al.,

2005), or ultrasound (Kuban et al., 2005) in con-

junction with pedestrian dead reckoning (PDR) (Fang

et al., 2005; Chang et al., 2015). PDR is a pro-

cess to use smart-phones inertial sensors, accelerom-

eters, magnetometers, and gyroscopes to determine

the current position based on the previous positions

Ahmed, A., Arablouei, R., de Hoog, F., Kusy, B. and Jurdak, R.

Multi-radio Data Fusion for Indoor Localization using Bluetooth and WiFi.

DOI: 10.5220/0007954100130024

In Proceedings of the 9th International Conference on Pervasive and Embedded Computing and Communication Systems (PECCS 2019), pages 13-24

ISBN: 978-989-758-385-8

recorded. The initial data fusion for indoor localiza-

tion integrates the data from WiFi and PDR (Even-

nou and Marx, 2006). Fusion using data from inertial

measurement units (IMU) is widely used in recent in-

door fusion based techniques (Chen et al., 2014; Koo

et al., 2014; Li et al., 2017; Gallagher et al., 2012;

Lin et al., 2015). In (Liu et al., 2017), a multi-sensor

fusion method is proposed using cameras, WiFi, and

IMU measurements. Use of cameras for indoor local-

ization is also proposed in (Papaioannou et al., 2014;

Jiao et al., 2018; Gu et al., 2017). For indoor navi-

gation of robots, different heuristic fusion algorithms

have been developed using RFID tags (Choi et al.,

2011), ultrasonic sensors (Luo and Lai, 2012), piezo-

electric sensors (Luo and Chen, 2013), and position-

ing radar (Dobrev et al., 2016). Past researches tend

to characterize the PDR as a non-linear function and

leverage particle ﬁlters to tackle the non-linearity. In

the measurements of PDR, several parameters need

to be speciﬁed, including the initial point and walk-

ing direction (Tarr´ıo et al., 2011). However, the

high computational load of applying particle ﬁlters in

radio mapping techniques necessitates ofﬂine track-

ing (Gustafsson et al., 2002). In recent times, there

have been a few fusion-based techniques that use

both bluetooth low energy (BLE) and WiFi. These

techniques use the received signal strength indicator

(RSSI) measurements from both BLE and WiFi to

execute the training-based and training-free models.

Fusing RSSI from two different wireless devices im-

proves the localization performance, but the perfor-

mance is still limited by the random ﬂuctuation of the

RSSI measurements. Moreover, these techniques do

not take advantage of the recently available resource,

such as CSI measurement from the WiFi. In section

2, we brieﬂy discuss CSI based techniques.

With the growing need for location-based applica-

tions, there is a demand for a universal and easily de-

ployable indoor localization system. In this research,

we propose a multi-radio data fusion framework for

indoor localization using BLE and WiFi access point.

A BLE beacon is an easily-deployable low-cost wire-

less sensor device that supports a wide range of in-

door localization techniques. However, position esti-

mation using BLE beacon has a high error rate due

to RSSI ﬂuctuations. Nowadays, WiFi access points

are wide availability in indoor premises and have the

potential to resolve the ﬂuctuation of RSSI in fusion-

based localization. In recent times, the channel state

information (CSI) received at the WiFi access point is

easily accessible in some commercial network inter-

face card (NIC). CSI contains the information of the

channel impulse response, which is inﬂuenced by the

channel characteristics, including multipath propaga-

tion. CSI-based localization techniques are compli-

cated but more accurate and reliable than RSSI based

techniques (Kotaru et al., 2015). However, most CSI-

based techniques require the user device to connect

with multiple WiFi access points for localization pur-

poses. It is challenging to get connected to multi-

ple access point in residential places where access to

most of the networks is password protected. In the

case of indoor commercial space, proper RF site sur-

vey does not allow a densely overlapping zone of the

same network. Moreover, the user device periodi-

cally switches connection among WiFi access points

to provide CSI measurement to assist the localization

process. Not only this hinders the internet connec-

tivity of the device, but it also creates network insta-

bility as the devices will continuously disconnect and

reconnect.

In this work, we fuse the RSSI measurements

from BLE beacons and the CSI measurements from

WiFi to develop a new indoor localization technique.

The localization technique focuses on the easily de-

ployable feature of BLE beacon and CSI based angle

of arrival (AoA) estimation technique at WiFi access

points. The proposed fusion framework combined the

existing RSSI technique with AoA estimation from

WiFi and obviates the deployment of multiple WiFi

access points. The CSI data from the WiFi is fed into

the modiﬁed matrix pencil (MMP) method developed

in (Ahmed et al., 2018). The MMP is a sub-space

based high-resolution direction-of-arrival (DoA) es-

timation technique which works with the CSI data.

It estimates the AoA and time of ﬂight (ToF) of the

incoming signal. ToF estimation from CSI data is

not reliable. Previous works have used multiple WiFi

access points to estimate the position using triangu-

lation (Kotaru et al., 2015). In our fusion frame-

work, we use only one WiFi access point. Use of

just one WiFi access point to implement the existing

techniques is challenging. The estimated AoA con-

tains three ambiguities. In the paper, we eliminate

these ambiguities by estimating an initial position us-

ing weighted centroid localization (WCL). WCL is

a weight based range-free technique that estimates a

position using weights which are propositional to the

magnitude of the RSSI (Blumenthal et al., 2007; Chan

and Soong, 2016; Laurendeau and Barbeau, 2009).

The range-free technique does not calculate the dis-

tance between the transmitters and receivers (Karim

et al., 2010). We use the lognormal shadowing

pathloss model to estimate the distances using the

BLE RSSI and then develop a fusion technique using

the estimated-distance, and estimated-AoA. The pro-

posed fusion technique is called distance-AoA fusion,

and it fuses the probability density functions (PDF) of

PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems

estimated-distance and estimated-AoA. We also com-

pare the results with a similar fusion technique us-

ing BLE and WiFi proposed in (Galv´an-Tejada et al.,

2013). The technique uses multilateration technique

using the RSSI from both BLE and WiFi without us-

ing radio-mapping.

2 RELATED WORKS

A BLE and WiFi-based fusion technique is presented

in (Aparicio et al., 2008) that uses RSSI measurement

from both platforms. The proposed method creates

two sets of radio-maps using the RSSI measurement

from BLE and WiFi. The tracking area is divided into

crown-shaped regionsby considering differentthresh-

olds for RSSI values. In the fusion part, a crown-

shaped region is selected initially using the Bluetooth

radio-map. The WiFi radio-map is used for the ﬁnal

position estimation inside the target zone. The aver-

age error reported is approximately 2.32 meters. A

similar RSSI based technique is proposed in (Apari-

cio et al., 2009) that uses RSSI from both WiFi and

BLE for ﬁngerprinting. The technique calculates the

absolute difference of WiFi RSSI and BLE RSSI for

each position. It calculates the mini-max distance

between the points. Then the distance is calculated

by averaging the three nearest neighbors co-ordinates

and weights which are inversely proportional of the

distances among them. The reported error in the tech-

nique is 2.21 meter. Another technique has been pre-

sented in (Kanaris et al., 2017) that uses RSSI from

the WiFi to create the radio-map. The RSSI from the

BLE is used to calculate the distance from a particu-

lar BLE. Then, adding and subtracting a ﬁxed toler-

ance with the distance, two circles are drawn from

BLE position. It creates a doughnuts shape region

inside the mapping area. The ﬁnal position is esti-

mated by looking at the WiFi ﬁngerprints inside the

doughnut-shaped region. The average position error

is reported to be 4.05 meters for 12 test points. A

training-freefusion technique is presented in (Galvan-

Tejada et al., 2012) which performs multilateration

using RSSI-distance equations. RSSI from WiFi and

BLE gives an estimation of distance which is used in

multilateration. In the analysis of the simulations, 400

iterations are performed with setups having 3 WiFi

and 1 BLE, 3 WiFi - 2 BLE, and 3 WiFi-5 BLE. The

result shows that 45% of the estimation have a po-

sition error of less than 1.839 meters. A similar ap-

proach is used in (Galv´an-Tejada et al., 2013) where

a regression model is developed using RSSI from 10

points ranging distances from 0.5 to 5 meter. The

model is used to estimate the distance from the RSSI.

The fusion technique uses a minimum of 8 equations

using RSSI from WiFi and BLE, and then the least

square solution is performed to estimate the ﬁnal po-

sition. For 30 measurements in three locations, the

lowest total error achieved is 0.87 meters. Table 1

brieﬂy shows work discussed above.

In previous research, WiFi and BLE are used for

RSSI based radio mapping/ﬁngerprinting. Some of

the training-free techniques use the pathloss model to

estimate the distance. Fingerprinting techniques are

conﬁned to indoor location, and the resolution of the

mapping dataset has an impact on the estimation per-

formance (Evennou and Marx, 2006). The process

requires a pre-proﬁling stage that requires an ofﬂine

session to create a dataset for indoor-locations. The

training process limits the easily deployable feature

of the training-based system. Moreover, any change

or movement of the furniture invalidates the ofﬂine

training data. On the other hand, the training-free

technique calculates distances using a path loss model

to estimate location using multilateration. The tech-

nique is independent of pre-proﬁling stage and uses

the RSSI from both BLE and WiFi. To the best of our

knowledge, till to-date, no fusion based framework

has used the CSI measurements. In this research, we

present an indoor location system that fuses the CSI

collected from a WiFi access point and the RSSI mea-

surements from multiple BLE beacons. Our goal is to

develop a cost-effective system using easily deploy-

able wireless device and BLE beacons. The clos-

est comparable technique discussed to ours is pre-

sented (Galv´an-Tejada et al., 2013), uses RSSI mea-

surements from both WiFi and BLE. It is a training-

free technique that calculates the distance from RSSI.

The lowest mean position errors it reports at three dif-

ferent points in an area surrounded by the BLE and

WiFi are 0.869 meters, 1.0112 meters, and 2.326 me-

ters. In the result section, we implement the technique

in the simulation and compare it with our proposed

fusion techniques.

3 SYSTEM OVERVIEW

In the proposed multi-radio fusion-based localization,

we use two properties of the incoming signals from

two different wireless platforms the CSI of the re-

ceived signal at the WiFi access point and RSSI mea-

surements at the user device. The CSI and RSSI val-

ues are sent to the cloud through the WiFi access-

point to perform the indoor localization. The user

learns the position through a smart-phone app con-

nected to the cloud.

Multi-radio Data Fusion for Indoor Localization using Bluetooth and WiFi

Table 1: Related works.

Article Year Bluetooth WiFi Radio

Map

Fusion Error

A Fusion Based on Bluetooth and

WLAN Technologies for Indoor

location(Aparicio et al., 2008).

2008 Radio-map

using

RSSI.

Radio-map

using

RSSI.

Required. Fingerprint

matching.

2.32 meter.

An Indoor Location Method Based

on a Fusion Map using Bluetooth

and WLAN Technologies (Apari-

cio et al., 2009).

2009 Radio-map

using

RSSI.

Radio-map

using

RSSI.

Required. Fingerprint

matching.

2.21 meter.

WiFi Bluetooth based combined

positioning algorithm (Galvan-

Tejada et al., 2012).

2012 Distance

form

RSSI.

Distance

form

RSSI.

Not re-

quired.

Multilateration. 45% results

has less than

1.839 meter

error.

Bluetooth-WiFi based combined

positioning algorithm, implemen-

tation and experimental evalua-

tion (Galv´an-Tejada et al., 2013).

2013 Distance

from

RSSI.

Distance

from

RSSI.

Not re-

quired.

Least square

solution.

Lowest total

error is 0.87

meter.

Fusing Bluetooth Data with WiFi

Radio-maps for Improved Indoor

Localization (Kanaris et al., 2017).

2017 Distance

form

RSSI.

Radio-map

using

RSSI.

Required. Fingerprint

matching.

Average 4.05

meter for 12

test points.

3.1 WiFi

In a previous study, a fast algorithm has been de-

veloped to estimate the AoA and ToF of the multi-

path signals using the CSI values at the WiFi access

point (Ahmed et al., 2018). The algorithm uses an up-

dated version of the high-resolution DoA estimation

algorithm, called the modiﬁed matrix pencil (MMP).

It signiﬁcantly outperforms the 2D MUSIC (multiple

signal classiﬁcation) algorithms used in (Kotaru et al.,

2015) by estimating the AoA and ToF at least 200

times faster. In both techniques, the high-resolution

algorithms are fed with CSI measurements in a modi-

ﬁed matrix format. However, in 2D MUSIC it creates

a 2D grid of MUSIC spectrum to ﬁnd the peaks that

correspond to AoA-ToF pairs of the multipath sig-

nals. While in MMP, it performs two 1-dimensional

solutions for AoA and ToF without creating any 2D

grid. It also pairs the AoA and ToF by using eigen

vector multiplication process. Both techniques pro-

vide an accurate estimation of the AoA with unreli-

able ToF. In the presence of noise and indoor multi-

path propagation, the estimation of ToF in nanosec-

ond scale is very challenging mainly because of the

presence of sampling-offsets and frequency-offsetsin

the commercial NICs. The differences between the

subcarrier frequencies are very small, and distance

traveled in indoor is also very short and do not create

a signiﬁcant difference in the phases of the subcarri-

ers. However, the crude estimation of ToF helps to

determine the correct AoA. The smallest ToF repre-

sents the signal traveling in the direct path, although

it is not equal to the travel time of the signal (Kotaru

et al., 2015). In a uniformarray antenna, the estimated

AoA is called the projected AoA. If the antennas are

placed horizontally, the estimated AoA becomes the

vertical projection. The estimated AoA on a uniform

antenna array contains an array ambiguity (Gazzah

and Delmas, 2012). At the receiver, the carrier phase

measurements are generated using a frequency lock

loop (FLL). The FLL matches the frequency of the re-

ceived and the transmitted signals. The measurement

of the phase in CSI has a modulo of π, which leads to

another type of ambiguity, called the phase ambigu-

ity (Jacobs and Ralston, 1981; Teunissen, 2003). In

our fusion-based localization, we resolve these ambi-

guities in the estimated AoA by using the initial posi-

tion estimate given by the WCL.

3.2 Bluetooth

In the BLE network, we perform an initial position

estimation using the RSSI measurements from mul-

tiple BLE beacons. The initial position of the user

is estimated using the WCL technique. For the fu-

sion technique, we estimate the distance between the

receiver and the BLE beacons using the log-normal

shadowing pathloss model (B¨uy¨ukc¸orak et al., 2015).

In the log-normal shadowing model, the distribution

of RSSI is approximated by a Gaussian in the loga-

rithmic domain (Zanca et al., 2008; Ergen et al., 2014;

Yiu et al., 2016).

PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems

Figure 1: The true AoA and target position, the AoA es-

timated from WiFi CSI, and the target position estimated

from BLE beacon RSSIs.

3.3 Fusion

We use the estimated AoA and RSSI measurements in

the fusion model development. The steps in the fusion

model development are mentioned below.

3.3.1 Step 1

In the ﬁrst step, we resolve the array and phase am-

biguities in the estimated AoA. The estimated AoA

has three ambiguities. For an estimated AoA, θ,

the phase-periodic ambiguity is π ±θ and the phase-

symmetric ambiguity is ±θ. The phase-periodic am-

biguity also has its symmetric ambiguity. These AoA

ambiguities are resolved with the help of the initial

position estimation using WCL. We select the angle

for AoA that creates the least difference with the an-

gle of the initial position estimate.

3.3.2 Step 2

In the second step, we use the AoA and RSSI mea-

surements to develop a fusion technique based on

maximum-likelihood estimation. To this end, we

maximize the product of the PDFs of the RSSI-

induced distances and the estimated AoA.

We model the PDF of the AoA as Gaussian. The

approximation is based on a simulation analysis dis-

cussed in the next section. To evaluate the fusion tech-

nique, we consider two choices for the variance of the

AoA, the theoretical value based on the Cramer-Rao

lower bound (CRLB), and the measured value. The

CRLB givesthe minimum achievable variance for any

unbiased estimator. The unbiased variance is least

noisy and in this study we use the CRLB variance to

test the performance of the fusion technique. We also

calculate the measured variance through simulations

with known position and AoA.

4 MULTI-RADIO FUSION

ARCHITECTURE

4.1 PDF of BLE RSSI

Let us consider q BLE beacons deployed in an area

having the same height as the target node. The target

node and the BLE beacons have isotropic antennas.

The set of BLE beacons is U = {A

,··· ,A

}. The

2-D locations of the BLE beacons are known and the

location set is u = {a

,··· ,a

}, where a

= (x

Let the position of the target node be b = (x,y). Let

(dBm) be the RSSI of the ith beacon at the target

node without any noise, i.e. R

(dBm) = 10log

The symbol R

(dBm) represents the RSSI in decibel-

milliwatt and R

in milliwatt. At any given time, the

target node receives signals from p number of BLE

beacons. The active set of BLE beacons is V ⊆ U.

The goal is to estimate the position of the target node

The measured RSSI from the i

beacon is

(dBm), which includes noise. The noise in

(dBm)

is additive zero-mean Gaussian with standard devia-

tion (STD) σ

. Therefore, we have

(dBm) = R

(dBm) + n

∼ N (0,σ

RSSI distance measurement uses the logarithmic dis-

tance path-loss model (Golestani et al., 2014; Nowak

et al., 2016; Nguyen et al., 2011). It is expressed as

(dBm) = R

(dBm) −10η

log

) + n

(1)

here d

(x−x

)

+ (y −y

)

, and R

(dBm) is the

power received at a reference distance d

. The param-

eter η

is the pathloss co-efﬁcient that depends on the

environment. In most cases, η

is estimated from the

prior measurements. The PDF of the RSSI measure-

ment from the i

BLE is:

(dBm)) =

√

2πσ

−

(

(dBm)−R

(dBm)

)

2σ

(2)

Substituting (1) into (2) gives

) =

√

2πσ

−



log



. (3)

here ξ = 10/ln10 = 4.3429, v =



10η

ln10



is the

distance calculated from

(dBm).

Multi-radio Data Fusion for Indoor Localization using Bluetooth and WiFi

4.2 Weighted Centroid Localization

We use the WCL to calculate the initial values of x

and y. The WCL is a localization technique that esti-

mates the position by assigning weights to each BLE

beacon and taking the weighted average of their RSSI

values. The weightes are assigned based on the tech-

nique discussed in the literature review (Blumenthal

et al., 2007). The estimated position of the target node

is given by the WCL algorithm considered:

∑

i=1

∑

i=1

∑

i=1

∑

i=1

here w

is the assigned weight for the ith BLE beacon.

4.3 PDF of WiFi AoA

Let us consider that the target node is connected to

a WiFi access point and sends a signal in the uplink.

Let τ and θ be the ToF and AoA of the uplink signal.

The WiFi access point has three antennas separated

by a distance l, and each of the antennas receives n

narrow-band signals, one for each of the subcarriers

whose frequencies are separated by δ. WiFi uses

OFDM (orthogonal frequency-division multiplexing)

modulation, where the signal gets divided into mul-

tiple narrow-band signals and transmitted over the

channel. Let us denote the CSI value of the subcarrier

with the lowest frequency on the ﬁrst antenna by

the complex number γ. It then follows that the CSI

value of the s

subcarrier received at the a

antenna

is γg

a−1

s−1

where g = exp



2π jlλ

−1

sinθ



and

h = exp(2πjδτ). Let f

= lλ

−1

sinθ, which gives

g = exp(2π j f

). Let ρ and ψ be the amplitude and

phase of γ, respectively. Let us assume that there

is only one line-of-sight signal, and the multipath

signals are weak and do not effect the CSI mea-

surements. The work on CSI based AoA estimation

proposed earlier shows the procedure of ﬁnding the

real AoA of the target user in case there a multipath

dominating multi paths affecting the CSI measure-

ments (Ahmed et al., 2018). Here, for simplicity,

we assume there is only one existing AoA in the CSI

measurements. Using MMP, we estimate the AoA

of the uplink signal. The CSI values associated with

all three antennas and n sub-carriers for only one

incoming signal can be collected in the CSI matrix as

′











1 h ··· h

n−1



Let us assume N

be the environment noise. N

has

a dimension of (3 ×n). Therefore, noisy CSI can be

expressed as

C = C

′

+ N

In the previous research in (Ahmed et al., 2018),

this C

′

matrix is used to estimate the AoA and ToF

of the incoming signals, including multipath. The ar-

ticle also discusses the technique to differentiate the

line of sight or direct path signal using the ToF esti-

mation. The presence of multipath signals inside the

CSI dataset results in multiple AoAs. The shortest

ToF represents the line-of-sight or the direct-path sig-

nal.

We assume that the estimation of AoA is unbiased.

In the next section, it is shown that the empirical dis-

tribution of estimated AoA in a noisy environment is

very close to Gaussian. Therefore, the PDF of the es-

timated AoA is

p(θ) =

√

2πσ

exp

−

(θ−

θ)

2σ

. (4)

here σ

is the STD and

θ is the estimated mean of

AoA. The STD of the AoA is calculated from a num-

ber of estimated AoAs at known angles. We assume

the WiFi access point is positioned at the origin and

θ = tan

−1





4.4 Fusion Technique

Let us assume that the errors for AoA and distance are

independent. The log likelihood function of the prod-

uct of distance and AoA PDFs given in (3) and (4)

is,

L(x,y) =ln

√

2πσ

exp

−



tan

−1





−



2σ

∑

i=1

2π((x−x

)

+ (y −y

)

exp

−



log



L(x,y) is twice continuously differentiable with re-

spect to the variables. We estimate the values of (x, y)

PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems

by maximizing the function. Therefore, the optimiza-

tion problem can be expressed as

max

x,y

√

2πσ

exp

−



tan

−1





−



2σ

∑

i=1







√

2πσ

−



log



#





The equation above is a nonlinear optimization prob-

lem without any constraints. We calculate ﬁrst and

second order derivatives to ﬁnd the value of x and y

that maximizes the equation.

We determine the Hessian matrix of L(x,y). The

Hessian matrix is (see Appendix A)

∇

L =

∂

∂x

∂

∂y∂x

∂

∂x∂y

∂

∂y

We use the Newton’s algorithm to solve the above

problem. At each iteration of the algorithm, we solve

the following system of linear equations:

∂

∂x

∂

∂y∂x

∂

∂x∂y

∂

∂y



δx

δy



∂L

∂x

∂L

∂y

here, L denotes the log-likelihood (objective) func-

tion. Then, the estimates are updated as ( ˆx, ˆy) →

( ˆx + δx, ˆy + δy). The iterations are continued until

(δx,δy) becomes small.

4.5 CRLB Calculation

The STDs σ

and σ

in L(x,y) are the parameter

known from prior measurements. It also can be cal-

culated using some initial measurements. Change of

STDs value also changes the accuracy of the loca-

tion estimation. In this sub-section, we calculate the

CRLB of the STDs using the known values of dis-

tances and AoA. The CRLB variance of (3) is the

same as the minimum variance of noise n

. There-

fore, we write σ

,cr

is the CRLB STD.

4.5.1 CRLB of WiFi AoA

In WiFi, the unknown parameters are γ, φ, θ and τ

(see section 4.3). In practice, we only need the value

of θ but for that we need to calculate the Fisher in-

formation matrix (FIM), which contains all the other

unknown parameters. For the AoA estimation, the un-

known parameter vector κ is:

κ = [γ, φ,θ,τ]

Therefore, the PDF of C

′

as in (Hua, 1992) is

p(C

′

|κ) =

(2πσ

)

exp



−

||C

′

−C||



here ||.|| donates the 2-norm, σ

is the variance of

the noise, and κ is a 4I ×1 vector of unknown, I is

the number 2-D sinusoids in the CSI dataset, M is the

number of antenna elements, and N is the number of

subcarriers. Hence, the log-likelihood function is:

L(κ|C

′

) = ln(p(C

′

|ρ,φ,θ,τ)).

Therefore, the corresponding FIM is:

i, j

= −E



∂

L(C

′

|κ)

∂κ





∂C

′H

∂κ

∂C

∂κ



here Re[.] denotes the real part, F(κ) is a 4I ×4I ma-

trix and, F

is the (i, j)th entry of F.

In (Hua, 1992), the FIM has been derived for m

antenna arrays and n subcarriers as (see Appendix B)

F =







ρρ

ρφ

ρθ

ρτ

ρφ

φφ

φθ

φτ

ρθ

φθ

θθ

θτ

ρτ

φτ

θτ

ττ







. (5)

The square root of the third diagonal element of F

−1

corresponds to the CRLB of AoA. Let σ

θ,cr

be the

CRLB STD of the estimated AoA.

Replacing σ

with σ

θ,cr

in (4) gives:

p(x,y) =

√

2πσ

θ,cr

exp

−



tan

−1





−



2σ

θ,cr

. (6)

5 SIMULATION AND RESULTS

We develop a 2D model simulation model using

MATLAB. The simulation model considers an area

of 20m x 20m with 20 BLE beacons and one WiFi

access point. The current setup of the simulation is

considering 2D localization in a 2D plane. Azimuth

angle for indoor localization is not that important and

it does not have any signiﬁcant effect in indoor posi-

tioning (Seong et al., 2019).

The WiFi access point is positioned in the cen-

ter of the simulation area. The BLE beacons are ran-

domly positioned with in the simulation area. The

frequency band in WiFi is channel one in 2.4 GHz

band. In an indoor environment, signal has 5-8 strong

multipaths(Kotaru et al., 2015). In our simulation,

we consider 5 multipaths in indoor environment. The

Multi-radio Data Fusion for Indoor Localization using Bluetooth and WiFi

-10 -8 -6 -4 -2 0 2 4 6 8 10

-10

-8

-6

-4

-2

Real position

BLE position

Real AoA

Array ambiguity

Phase ambiguity

Array-Phase ambiguity

Figure 2: WiFi AoA estimation with ambiguity in the func-

tioning area.

magnitude of the multipath signal is obtained using

3D wireless prediction software - Wireless insite(wir,

). The details of the simulation using Wireless-Insite

is shown earlier research in (Ahmed et al., 2018). For

every initial position estimation using WCL, we select

six BLEs with the highest RSSI values. The simula-

tion parameters are given in Table 2. Figure 2 shows

the ambiguities of the estimated AoA and initial esti-

mated position inside the simulation area.

Table 2: System Parameters.

Parameters/Variables Properties

Area 20m x 20m

Number of BLE beacons 20

Number of BLE beacons used

for localization

Frequency 2403.0225 MHz

Number of subcarrier 30

Number of WiFi antenna 3

Subcarrier spacing 0.630 MHz

SNR (WiFi) 15 dB

Number of Multi-path 5

Transmission Power -12 dBm

Received power at 1 m distance -77 dBm

Pathloss co-efﬁcient, η

1.8

BLE std 3 dB

5.1 Distribution of Estimated AoA

Error

We calculate the AoA estimation error for 10000 trails

with a 15 dB signal to noise ratio (SNR). In each sim-

ulation, we randomly distribute the target node and

calculate the noisy CSI data. All the simulations we

performed having the target node at a 30-degree an-

gle from the WiFi access point. Figure 3 QQ plot

-4 -3 -2 -1 0 1 2 3 4

Standard Normal Quantiles

27.5

28.5

29.5

30.5

31.5

32.5

Quantiles of Input Sample

standard line

measured data

Figure 3: QQplot test.

is a graphical method for comparing two probability

distributions by plotting their quantiles against each

other. Figure 3 illustrates that the distribution of the

AoA is in good agreement with a normal distribution.

The mean estimated AoA is 29.93-degree.

5.2 Performance of the Proposed

Techniques

We ran 10000 independent trials using both the the-

oretical (CRLB) and measured variances. In each

trial, the BLE beacon and target node positions are

arbitrarily selected. We consider zero-mean Gaus-

sian noise in the RSSI. Figure 4 shows the cumulative

distribution function (CDF) of the position error us-

ing PDFs having four possible combination of CRLB

and measured variances. From the ﬁgure, we can

see that the proposed technique performs best with

both PDFs having variances as measured variances.

The root-mean-square-error (RMSE) of position with

measured variances of 3.35-degree is 1.89 meters. Ta-

ble 3 shows the estimated position error in meters as

RMSE for different combinations of CRLB and mea-

sured variances in the PDFs. However, if we con-

sider 80% of the estimation with higher accuracy, the

fusion technique using CRLB variance in BLE PDF

shows a better performance. It also performs better

for position estimation with 3-meter error tolerance.

CRLB variance of AoA requires the values of ρ, φ,

θ, and τ. In practice, it is not feasible to estimate the

values of all these variables. Estimation of τ is possi-

ble by using precise and sophisticated measurements

of the signal’s travel-time. However, the values of ρ

and φ cannot be calculated using a commercial NIC.

Therefore, CRLB variance cannot be used in prac-

tice. The proposed fusion technique uses the variance

of the RSSI and variance of the AoA in the PDFs.

Change of the AoA variance does not have a signiﬁ-

PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems

Table 3: Estimated position error for different variance (RMSE).

AoA CRLB variance, σ

θ,cr

AoA variance, σ

RSSI CRLB variance, σ

R,cr

2.4176 meter 2.4970 meter

RSSI variance, σ

2.0320 meter 1.8884 meter

0 2 4 6 8 10 12

estimated position error (m)

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

CDF

CRLB (

)

EST. (

)

CRLB ( ), EST. (

)

EST. ( ), CRLB. (

)

Figure 4: CDF of position error of two techniques using

CRLB variances.

1 2 3 4 5 6 7 8 9 10

est.

(dB)

estimated position error (m)

distance-AoA (EST. [

])

Figure 5: estimated σ

vs estimated position error.

cant effect on the estimation performance. However,

changing the variances of RSSI from CRLB to ’mea-

sured’ improves the position error by 40 cm, approx-

imately. In Figure 5, we show an estimation perfor-

mance for a range of σ

(RSSI STD) values. The

range of σ

is between 0.5 dB and 10 dB with 0.5

dB steps. The position errors with σ

are measured

as the average of RMSE for 10000 trials. For the val-

ues below 2 dB, the position error is large while it

decreases rapidly at 2 dB point. After 2dB, the error

graph remains almost ﬂat until the end. In our calcu-

lations, the value of CRLB RSSI STD, σ

R,cr

, is fre-

quently lower than 2 dB while estimated STD, σ

, is

frequently higher than the 2 dB. It leads to a higher

accuracy in position estimation using the estimated

variance of RSSI. In practical scenarios, σ

can be

measure by taking multiple RSSI measurements at a

0 2 4 6 8 10 12

estimated position error (m)

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

CDF

distance-AoA [EST. (

)]

multilateration

Figure 6: CDF of position error of distance-AoA fusion and

multilateration.

0 5 10 15

estimated position error (m)

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

CDF

Empirical CDF

distance-AoA [EST. (

)]

multilateration

Figure 7: CDF of position error of distance-AoA fusion and

multilateration for 30m×30m scenario.

ﬁxed position and σ

can be measured by taking mul-

tiple AoA estimation at a ﬁxed angle.

5.3 Comparison

The most similar technique discussed earlier esti-

mates the location of the target node using the RSSI

from both BLE and WiFi access point (Galv´an-Tejada

et al., 2013). We implement this technique us-

ing RSSI at the WiFi access point along with RSSI

received from BLE beacons. We use the multi-

lateration technique on the calculated distance from

the RSSI measurements and compare the estimation

performance with our proposed fusion technique. We

use the best performing combination of the variances

for the proposed fusion technique. The performance

Multi-radio Data Fusion for Indoor Localization using Bluetooth and WiFi

comparison is shown in Figure 6. The CDF of esti-

mated position error reduces the RMSE of the esti-

mated position by almost 30% from 2.7450 meters to

1.8884 meters. Figure 7 shows the CDF of estimated

position error when the simulation area has been in-

creased to 30m×30m and the number of BLE beacon

has been increased to 30. The mean error of the fu-

sion technique is 3.0883 meters and for multilatera-

tion based technique its 4.2658 meters.

6 CONCLUSION

Position estimation using RSSI has high estimation

error due to its random ﬂuctuation in the indoor envi-

ronment. On the other hand, estimation of AoA using

CSI data has ambiguities. In our proposed multi-radio

fusion framework, we combine RSSI and CSI mea-

surements to develop a technique that overcomes the

limitation of these two wireless platforms. We used

the CSI from WiFi for the ﬁrst time in the fusion-

based location estimation technique. CSI based tech-

niques are complex but accurate, while RSSI based

techniques are less scalable, and the localization per-

formance vastly depends on the number and strategy

of beacon deployment. From the BLE RSSI measure-

ments, we estimate a position using the MLE solu-

tion considering the log-normal shadowing pathloss

model. In WiFi, we use MMP algorithm on the CSI

data to estimate the AoA. We propose a fusion tech-

nique using the RSSI and CSI measurements. We es-

timate an initial position using WCL to remove the

ambiguities of the estimated AoA. In the fusion tech-

nique, we use the PDFs of the estimated distance and

AoA. The distances are calculated using the RSSI

measurements. We present a performance analysis of

the fusion technique using a different combination of

CRLB variances and measured variances for both dis-

tance and AoA in the PDFs. The performance anal-

ysis shows that the fusion technique performs better

with both PDF using the measured variances. More-

over, the technique performs 30% better than the most

similar fusion based technique using BLE and WiFi.

In future work, we will investigate the prospect of in-

tegrating the information available from the neighbor-

ing WiFi access points to improve the localization ac-

curacy.

REFERENCES

Wireless InSite - 3D Wireless Prediction Software,

https://www.remcom.com/wireless-insite-em-

propagation-software/. Accessed 10 Sep. 2016.

Ahmed, A., Arablouei, R., de Hoog, F., Kusy, B., Ju-

rdak, R., and Bergmann, N. (2018). Estimating

angle-of-arrival and time-of-ﬂight for multipath com-

ponents using wiﬁ channel state information. Sensors,

18(6):1753.

Aparicio, S., P´erez, J., Bernardos, A. M., and Casar, J. R.

(2008). A fusion method based on bluetooth and wlan

technologies for indoor location. In IEEE Interna-

tional Conference on Multisensor Fusion and Integra-

tion for Intelligent Systems, pages 487–491. IEEE.

Aparicio, S., P´erez, J., Tarr´ıo, P., Bernardos, A. M., and

Casar, J. R. (2009). An indoor location method

based on a fusion map using bluetooth and wlan tech-

nologies. In International Symposium on Distributed

Computing and Artiﬁcial Intelligence, pages 702–710.

Springer.

Blumenthal, J., Grossmann, R., Golatowski, F., and Tim-

mermann, D. (2007). Weighted centroid localization

in zigbee-based sensor networks. In IEEE Interna-

tional Symposium on Intelligent Signal Processing,

pages 1–6. IEEE.

B¨uy¨ukc¸orak, S., Vural, M., and Kurt, G. K. (2015). Lognor-

mal mixture shadowing. IEEE Transactions on Vehic-

ular Technology, 64(10):4386–4398.

Chan, Y. W. E. and Soong, B.-H. (2016). Discrete weighted

centroid localization (dwcl): Performance analysis

and optimization. IEEE Access, 4:6283–6294.

Chang, Q., Van de Velde, S., Wang, W., Li, Q., Hou, H., and

Heidi, S. (2015). Wi-ﬁ ﬁngerprint positioning updated

by pedestrian dead reckoning for mobile phone indoor

localization. In China Satellite Navigation Conference

(CSNC) 2015 Proceedings: Volume III, pages 729–

739. Springer.

Chen, L.-H., Wu, E. H.-K., Jin, M.-H., and Chen, G.-H.

(2014). Intelligent fusion of wi-ﬁ and inertial sensor-

based positioning systems for indoor pedestrian navi-

gation. IEEE Sensors Journal, 14(11):4034–4042.

Choi, B.-S., Lee, J.-W., Lee, J.-J., and Park, K.-T. (2011).

A hierarchical algorithm for indoor mobile robot lo-

calization using rﬁd sensor fusion. IEEE Transactions

on Industrial Electronics, 58(6):2226–2235.

Dobrev, Y., Flores, S., and Vossiek, M. (2016). Multi-modal

sensor fusion for indoor mobile robot pose estimation.

In IEEE/ION Position, Location and Navigation Sym-

posium (PLANS), pages 553–556. IEEE.

Ergen, S. C., Tetikol, H. S., Kontik, M., Sevlian, R.,

Rajagopal, R., and Varaiya, P. (2014). Rssi-

ﬁngerprinting-based mobile phone localization with

route constraints. IEEE Transactions on Vehicular

Technology, 63(1):423–428.

Evennou, F. and Marx, F. (2006). Advanced integration of

wiﬁ and inertial navigation systems for indoor mobile

positioning. EURASIP journal on applied signal pro-

cessing, 2006:164–164.

Fang, L., Antsaklis, P. J., Montestruque, L. A., McMickell,

M. B., Lemmon, M., Sun, Y., Fang, H., Koutroulis,

I., Haenggi, M., Xie, M., et al. (2005). Design of a

wireless assisted pedestrian dead reckoning system-

the navmote experience. IEEE transactions on Instru-

mentation and Measurement, 54(6):2342–2358.

PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems

Gallagher, T., Wise, E., Li, B., Dempster, A. G., Rizos, C.,

and Ramsey-Stewart, E. (2012). Indoor positioning

system based on sensor fusion for the blind and visu-

ally impaired. In International Conference on Indoor

Positioning and Indoor Navigation (IPIN), pages 1–9.

IEEE.

Galv´an-Tejada, C. E., Carrasco-Jim´enez, J. C., and Brena,

R. F. (2013). Bluetooth-wiﬁ based combined position-

ing algorithm, implementation and experimental eval-

uation. Procedia Technology, 7:37–45.

Galvan-Tejada, I., Sandoval, E. I., Brena, R., et al. (2012).

Wiﬁ bluetooth based combined positioning algorithm.

Procedia Engineering, 35:101–108.

Gazzah, H. and Delmas, J.-P. (2012). Direction ﬁnding an-

tenna arrays for the randomly located source. IEEE

Transactions on Signal Processing, 60(11):6063–

6068.

Gezici, S., Tian, Z., Giannakis, G. B., Kobayashi, H.,

Molisch, A. F., Poor, H. V., and Sahinoglu, Z. (2005).

Localization via ultra-wideband radios: a look at posi-

tioning aspects for future sensor networks. IEEE sig-

nal processing magazine, 22(4):70–84.

Golestani, A., Petreska, N., Wilfert, D., and Zimmer, C.

(2014). Improving the precision of rssi-based low-

energy localization using path loss exponent estima-

tion. In 2014 11th Workshop on Positioning, Naviga-

tion and Communication (WPNC), pages 1–6. IEEE.

Gu, F., Niu, J., and Duan, L. (2017). Waipo: A fusion-based

collaborative indoor localization system on smart-

phones. IEEE/ACM Transactions on Networking,

25(4):2267–2280.

Gustafsson, F., Gunnarsson, F., Bergman, N., Forssell,

U., Jansson, J., Karlsson, R., and Nordlund, P.-J.

(2002). Particle ﬁlters for positioning, navigation, and

tracking. IEEE Transactions on signal processing,

50(2):425–437.

Hilsenbeck, S., Bobkov, D., Schroth, G., Huitl, R., and

Steinbach, E. (2014). Graph-based data fusion of pe-

dometer and wiﬁ measurements for mobile indoor po-

sitioning. In Proceedings of the 2014 ACM interna-

tional joint conference on pervasive and ubiquitous

computing, pages 147–158. ACM.

Hua, Y. (1992). Estimating two-dimensional frequencies by

matrix enhancement and matrix pencil. IEEE Trans-

actions on Signal Processing, 40(9):2267–2280.

Jacobs, E. and Ralston, E. W. (1981). Ambiguity resolution

in interferometry. IEEE Transactions on Aerospace

and Electronic Systems, (6):766–780.

Jiao, J., Li, F., Tang, W., Deng, Z., and Cao, J. (2018). A

hybrid fusion of wireless signals and rgb image for in-

door positioning. International Journal of Distributed

Sensor Networks, 14(2):1550147718757664.

Kanaris, L., Kokkinis, A., Liotta, A., and Stavrou, S.

(2017). Fusing bluetooth beacon data with wi-ﬁ ra-

diomaps for improved indoor localization. Sensors,

17(4):812.

Karim, L., Nasser, N., and El Salti, T. (2010). Relma:

a range free localization approach using mobile an-

chor node for wireless sensor networks. In Global

Telecommunications Conference (IEEE GLOBECOM

2010, pages 1–5. IEEE.

Koo, B., Lee, S., Lee, M., Lee, D., Lee, S., and Kim,

S. (2014). Pdr/ﬁngerprinting fusion indoor location

tracking using rss recovery and clustering. In Interna-

tional Conference on Indoor Positioning and Indoor

Navigation (IPIN), pages 699–704. IEEE.

Kotaru, M., Joshi, K., Bharadia, D., and Katti, S. (2015).

Spotﬁ: Decimeter level localization using wiﬁ. In

ACM SIGCOMM Computer Communication Review,

volume 45, pages 269–282. ACM.

Kuban, D. A., Dong, L., Cheung, R., Strom, E., and

De Crevoisier, R. (2005). Ultrasound-based localiza-

tion. In Seminars in radiation oncology, volume 15,

pages 180–191. Elsevier.

Laurendeau, C. and Barbeau, M. (2009). Centroid local-

ization of uncooperative nodes in wireless networks

using a relative span weighting method. EURASIP

Journal on Wireless Communications and Network-

ing, 2010(1):567040.

Lepp¨akoski, H., Collin, J., and Takala, J. (2013). Pedestrian

navigation based on inertial sensors, indoor map, and

wlan signals. Journal of Signal Processing Systems,

71(3):287–296.

Li, Y., Zhuang, Y., Zhang, P., Lan, H., Niu, X.,

and El-Sheimy, N. (2017). An improved iner-

tial/wiﬁ/magnetic fusion structure for indoor naviga-

tion. Information Fusion, 34:101–119.

Lim, C.-H., Wan, Y., Ng, B.-P., and See, C.-M. S. (2007).

A real-time indoor wiﬁ localization system utilizing

smart antennas. IEEE Transactions on Consumer

Electronics, 53(2).

Lin, T., Li, L., and Lachapelle, G. (2015). Multiple sen-

sors integration for pedestrian indoor navigation. In

International Conference on Indoor Positioning and

Indoor Navigation (IPIN), pages 1–9. IEEE.

Liu, M., Chen, R., Li, D., Chen, Y., Guo, G., Cao, Z., and

Pan, Y. (2017). Scene recognition for indoor localiza-

tion using a multi-sensor fusion approach. Sensors,

17(12):2847.

Luo, R. C. and Chen, O. (2013). Wireless and pyroelectric

sensory fusion system for indoor human/robot local-

ization and monitoring. IEEE/ASME Transactions on

mechatronics, 18(3):845–853.

Luo, R. C. and Lai, C. C. (2012). Enriched indoor map

construction based on multisensor fusion approach for

intelligent service robot. IEEE Transactions on Indus-

trial Electronics, 59(8):3135–3145.

Mirowski, P., Ho, T. K., Yi, S., and MacDonald, M. (2013).

Signalslam: Simultaneous localization and mapping

with mixed wiﬁ, bluetooth, lte and magnetic signals.

In International Conference on Indoor Positioning

and Indoor Navigation (IPIN), pages 1–10. IEEE.

Nguyen, H. A., Guo, H., and Low, K.-S. (2011). Real-

time estimation of sensor node’s position using par-

ticle swarm optimization with log-barrier constraint.

IEEE Transactions on Instrumentation and Measure-

ment, 60(11):3619–3628.

Nowak, T., Hartmann, M., Zech, T., and Thielecke, J.

(2016). A path loss and fading model for rssi-based

Multi-radio Data Fusion for Indoor Localization using Bluetooth and WiFi

localization in forested areas. In2016 IEEE-APS Topi-

cal Conference on Antennas and Propagation in Wire-

less Communications (APWC), pages 110–113. IEEE.

Papaioannou, S., Wen, H., Markham, A., and Trigoni, N.

(2014). Fusion of radio and camera sensor data for

accurate indoor positioning. In IEEE 11th Interna-

tional Conference on Mobile Ad Hoc and Sensor Sys-

tems (MASS), pages 109–117. IEEE.

Qi, H. and Moore, J. B. (2002). Direct kalman ﬁltering

approach for gps/ins integration. IEEE Transactions

on Aerospace and Electronic Systems, 38(2):687–693.

Seong, J.-H., Lee, S.-H., Yoon, K.-K., and Seo, D.-H.

(2019). Ellipse coefﬁcient map-based geomagnetic

ﬁngerprint considering azimuth angles. Symmetry,

11(5):708.

Tarr´ıo, P., Cesana, M., Tagliasacchi, M., Redondi, A., Bor-

sani, L., and Casar, J. R. (2011). An energy-efﬁcient

strategy for combined rss-pdr indoor localization. In

IEEE International Conference on Pervasive Com-

puting and Communications Workshops (PERCOM

Workshops), pages 619–624. IEEE.

Teunissen, P. (2003). Theory of carrier phase ambiguity

resolution. Wuhan University Journal of Natural Sci-

ences, 8(2):471.

Yiu, S., Dashti, M., Claussen, H., and Perez-Cruz, F. (2016).

Locating user equipments and access points using rssi

ﬁngerprints: A gaussian process approach. In IEEE

International Conference on Communications (ICC),

pages 1–6. IEEE.

Zanca, G., Zorzi, F., Zanella, A., and Zorzi, M. (2008). Ex-

perimental comparison of rssi-based localization al-

gorithms for indoor wireless sensor networks. In Pro-

ceedings of the workshop on Real-world wireless sen-

sor networks, pages 1–5. ACM.

PECCS 2019 - 9th International Conference on Pervasive and Embedded Computing and Communication Systems