Indoor Real-Time 3-D Visible Light Positioning System Using Fingerprinting and Extreme Learning Machine

Abstract

Photodiode-based (PD-based) visible light positioning (VLP) has become a research focus of indoor positioning technology, while the existing VLP models rarely consider the anti-interference and positioning time of that. In this paper, indoor real-time three-dimensional visible light positioning system using fingerprinting and extreme learning machine (ELM) is proposed to make the system achieve not only high positioning accuracy and elevated anti-interference but also well-behaved real-time ability. In contrast to the positioning system based on K-Nearest Neighbor or Support Vector Machine, the proposed system achieves the highest positioning accuracy and the state-of-the-art positioning speed. Furthermore, the visible light positioning kernel is proposed as a method to reduce the size of the fingerprint database and thus reduce the training time exponentially. Both the simulation and the experiment results show that the proposed system achieves real-time 3-D positioning with high anti-interference. Therefore, this scheme can be considered as one of the effective methods for indoor 3-D positioning.

Full text

Received November 16, 2019, accepted December 16, 2019, date of publication December 24, 2019,
date of current version January 23, 2020.
Digital Object Identifier 10.1109/ACCESS.2019.2961939
Indoor Real-Time 3-D Visible Light Positioning
System Using Fingerprinting and Extreme
Learning Machine
YIRONG CHEN 1, WEIPENG GUAN 2, JINGYI LI 3, AND HONGZHAN SONG 3
1School of Electronic and Information Engineering, South China University of Technology, Guangzhou 510640, China
2Department of Information Engineering, The Chinese University of Hong Kong, Hong Kong
3School of Automation Science and Engineering, South China University of Technology, Guangzhou 510640, China
Corresponding author: Weipeng Guan (gwpscut@163.com)
This work was supported in part by the National Undergraduate Innovative and Entrepreneurial Training Program under Grant
201710561006 and Grant 201810561083, and in part by the Special Funds for the Cultivation of Guangdong College Students’ Scientific
and Technological Innovation (Climbing Program Special Funds) under Grant pdjh2017b0040 and Grant pdjha0028.
ABSTRACT Photodiode-based (PD-based) visible light positioning (VLP) has become a research focus
of indoor positioning technology, while the existing VLP models rarely consider the anti-interference and
positioning time of that. In this paper, indoor real-time three-dimensional visible light positioning system
using fingerprinting and extreme learning machine (ELM) is proposed to make the system achieve not only
high positioning accuracy and elevated anti-interference but also well-behaved real-time ability. In contrast
to the positioning system based on K-Nearest Neighbor or Support Vector Machine, the proposed system
achieves the highest positioning accuracy and the state-of-the-art positioning speed. Furthermore, the visible
light positioning kernel is proposed as a method to reduce the size of the fingerprint database and thus reduce
the training time exponentially. Both the simulation and the experiment results show that the proposed system
achieves real-time 3-D positioning with high anti-interference. Therefore, this scheme can be considered as
one of the effective methods for indoor 3-D positioning.
INDEX TERMS Extreme learning machine (ELM), photodiode (PD), positioning fingerprint, real-time
positioning, visible light positioning (VLP).
I. INTRODUCTION
With the development of artificial intelligence, indoor appli-
cations have become increasingly demanding for positioning
services. Performing high-speed indoor positioning on the
premise of ensuring accuracy is the basis for service robots
and drones to be used in indoor scenes. In the field of posi-
tioning, Global Positioning System (GPS) has been widely
used outdoors. However, GPS is difficult to work well in
indoor environments due to weak radio signal strength and the
inability to completely penetrate the walls of buildings and
houses [1]. In the field of indoor positioning technology, wire-
less local area networks (WLAN), infrared, radio frequency
identification (RFID), Zigbee, ultra-wide band (UWB) and
Bluetooth have been widely studied. However, these position-
ing technologies have not been widely applied on account
The associate editor coordinating the review of this manuscript and
approving it for publication was Nianqiang Li .
of additional wireless anchors, high cost, slow positioning
speed or low positioning accuracy [2]–[4]. Different from
the above-mentioned indoor positioning technology, VLP
technology is based on visible light communication (VLC).
On the one hand, indoor VLP technology does not need
to install additional transmitters in the room, only need
to change the Light Emitting Diode (LED) driving circuit.
On the other hand, the LED-based lighting provides illumi-
nation, positioning, and communication at little additional
cost [5].
According to the difference of sensors, indoor VLP can
be divided into two types: camera-based positioning and
photodiode-based (PD-based) positioning [6]–[8]. In the case
of camera-based, the indoor positioning system consists of
a LED array and the image sensor with high frame rate.
This positioning technique requires a stable camera place-
ment, as little jitter as possible [9]. What’s worse is that
it needs to repeatedly capture images and perform image
VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/ 13875

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
processing which means that the memory consumption of
the system is extremely large and it is difficult to achieve
real-time positioning due to the computational complexity.
Different from the camera-based positioning technology, PD-
based positioning technology obtains the position ID and
the received power of the LED through the PD detecting
the LED signal. So far, the PD-based positioning technol-
ogy has been deeply explored, and there are some models
to calculate the position of the receiver, such as the angle
of arrival (AOA), time of arrival (TOA), time difference
of arrival (TDOA) or received signal strength (RSS) [10].
Taking into account of the difficulty, accuracy and cost of
indoor positioning, PD-based indoor positioning using RSS
algorithm is preferred due to its high accuracy and low cost.
Of course, before applying the RSS algorithm, the prob-
lem of multiple LEDs transmission signals causing inter-cell
interference needs to be solved. Therefore, in our previous
study [11], [13], the signal from LEDs installed on the ceil-
ing is modulated in code division multiple access (CDMA)
to reduce inter-cell interference caused by the presence of
multiple LEDs.
So far, indoor three-dimensional positioning based on VLC
focuses more on algorithm research. Heidi Steendam put for-
ward a 3-D positioning algorithm based on AOA and the max-
imum likelihood principle, achieving an accuracy of 10 cm
but with high complexity [12]. In [13], Hao Chen et al.
proposed a reversed three-dimensional VLC-based position-
ing system using RSS and genetic algorithm (GA) to real-
ize precise positioning service at high computational costs
operation. In [11], Ye Cai et al. proposed a 3-D VLC-based
positioning system based on the modified particle swarm
optimization (PSO) algorithm and RSS with the average
error of 3.9 mm, using only when LEDs’ luminous power is
extremely stable. In [14], DANIEL KONINGS et al. devel-
oped a VLP system using Convolutional Neural Network-
based (CNN-based) wireless localization with the mean
error 0.12 m. However, the proposed system was only real-
ized in simulation with no experimental verification. The
above-mentioned three-dimensional positioning algorithms
all achieve a very high positioning accuracy. However, since
the above positioning method involves a high amount of
computation, the positioning time is more than 0.1 s and it is
difficult to apply to the actual system. Alternately, Bingcheng
Zhu et al. considered both positioning time and positioning
accuracy and proposed a three-dimensional VLC-based posi-
tioning algorithm based on the method of exhaustion (MEX).
The proposed algorithm achieved an average positioning
error of 3.20 cm with a positioning time cost of 0.36 s, thus
making challenging to achieve real-time position [15]. Ming
Xu et al. combined the fingerprint positioning algorithm
with the VLP algorithm, and use K-Nearest Neighbor (KNN)
algorithm to achieve three-dimensional positioning [16]. The
disadvantage of KNN is that the amount of calculation is
large, because each sample to be located must calculate its
distance to all known samples in order to obtain its K nearest
neighbors [17].
FIGURE 1. Indoor optical wireless positioning system: the left side of the
figure is the actual scene, and the right side is the system model.
In this paper, we propose an indoor real-time 3-D
VLC-based positioning system using fingerprinting [18], [19]
and extreme learning machine (ELM) technology [20]–[22].
The positioning fingerprint is originally applied in the
PADAR positioning system [19]. This method effectively
avoids the shortcomings of the traditional indoor positioning
system relying on a single RF signal. In addition, in 3-D posi-
tioning algorithm, the receiving intensity of the PD is affected
by both the distance and the angle. Fortunately, fingerprinting
positioning is more practical in this case. ELM is a simple and
effective single hidden layer feedforward networks (SLFNs)
learning algorithm that can obtain the global optimal solu-
tion with high learning speed and strong adaptability to new
samples, compared with support vector machine (SVM) [22],
[23]. Therefore, using ELM and fingerprinting based on VLC
can quickly obtain the three-dimensional coordinates of the
target point.
The rest of this article is organized as follows. Section II
introduced the proposed three-dimensional positioning algo-
rithm in detail. Section III provided positioning accuracy and
time in simulated experiment. Section IV provided position-
ing accuracy and time in real-world experiment. Section V
concluded the article.
II. SYSTEM PRINCIPLE AND ALGORITHM
A. INDOOR OPTICAL WIRELESS CHANNEL MODEL
As shown in Fig. 1, a large number of LEDs are installed on
the ceiling to meet the lighting requirements. PD receivers
are installed on indoor objects such as unmanned aerial vehi-
cle (UAV) and robots to receive signals from LEDs. The
absolute coordinates of these LEDs can be represented by
Cij =[Xij,Yij ,Zij]T, (i=1,2,3, ... and j=1,2,3, ... is the
LEDs’ distribution number), respectively. Similarly, the abso-
lute coordinates of different PDs can also be represented by
ck=[xk,yk,zk]T, (k=1,2,3, ... is the positioning num-
ber), respectively. The radiant intensity of a LED is usually
assumed to follow a Lambertian radiation pattern. Relation
between the received optical power of the PD located at ck
and the emitted optical power of the LED located at Cij can
be given as [24]
Pijk =PT
ij
Ar(m+1)
2πdijk
cosm(φij)T(ψijk )G(ψijk ) cosM(ψijk )
(1)
13876 VOLUME 8, 2020

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
FIGURE 2. VLP using fingerprinting: (a) traditional model for fingerprint
positioning, (b) geometric relation between LED and PD, (c) the training
process (initial phase) and the positioning process (positioning phase) of
fingerprint positioning, (d) VLP kernel: Simplified fingerprint positioning
using coordinate transformation.
where PT
ij is the emitted optical power of the LED located
at Cij,Aris the effective area of the PD at the receiver
located at the point ck,m= − ln (2)/ln(cos(φ1/2)) and
M= − ln (2)/ln(cos(ψ1/2)), where φ1/2and ψ1/2are the
half-power angles of the LED and PD. dijk is the distance
between point Cij and point ck.φij is the irradiation angle of
the LED located at Cij and ψijk the incidence angle of the PD
located at ck.
Geometric relation between LEDij and PDkis shown
in Fig. 2 (b). The distance dijk between the LEDij located at
Cij and the PDklocated at ckcan be calculated as
dijk =q(xk−Xij)2+(yk−Yij )2+(zk−Zij)2(2)
The irradiant angle φij can be represented as
cos(φij)=zk−Zij
q(xk−Xij)2+(yk−Yij )2+(zk−Zij)2
(3)
Using Equations (1), (2), and (3), we can calculate the ckthe-
oretically. However, the equation sets for solving (xk,yk,zk)
are nonlinear and it is difficult to obtain the precise received
optical power due to the interference of ambient light in
practical VLP system [11], [14]. This makes the current
three-dimensional positioning algorithm less suitable than the
fingerprint algorithm in the positioning scene using VLC.
B. TADITIONAL VLC FING ERPRI NTING ALDORITHM
As shown in Fig. 2 (a), (n+1) ·(m+1) LEDs are mounted
on the ceiling in the position area in traditional VLC finger-
printing positioning scene. In the initial phase, the vector of
received optical power Pk0can be given as
Pk0=P11k0P12k0... Pijk0... P(n+1)(m+1)k0T
(4)
where Pijk0is the received optical power from LEDij at ck0=
[xk0,yk0,zk0]T,k0=1,2,...,K0.
In the positioning phase, the vector of received optical
power Pkcan be expressed as
Pk=P11k... Pijk ... P(n+1)(m+1)kT(5)
where Pijk is the received optical power from LEDij at ck.
When n≥1 and m≥1, we can use the sampling database
to estimate the coordinates of the measured point and control
the error to a lower level. However, when the values of n and
m increase, the volume of the fingerprint database rapidly
increases, and many useless data are generated. For example,
there are many elements having a value of 0 in the vector Pk0
and Pk. These data do not help fingerprint positioning and
greatly increase the positioning time.
C. MODIFIRED ELM FOR VLP USING FING ERPRI NTING
Considering that the existing fingerprint location algorithm
is not suitable for scenes with large-scale LEDs, we propose
a high-speed, high-accuracy three-dimensional positioning
system based on ELM. We creatively propose the concept of
a VLP kernel for fingerprinting.
In this positioning system, the positioning concept is
divided into absolute positioning and relative positioning
according to the coordinate transformation theory. As shown
in Fig. 2 (d), the relative positioning refers to using ELM
and fingerprint to achieve positioning of the PD, calculating
the relative coordinate (x,y,z) of the PDkwith respect to the
projection (absolute coordinate: Xij,Yij ,0) of LEDij. Abso-
lute positioning refers to calculating the absolute coordinate
of PDkwith respect to the absolute coordinate (0,0,0). Their
relationship can be expressed by
ck=[x,y,z]T+[Xij,Yij ,0]T(6)
In Equation (6), Xij and Yij can be obtained through VLC
because they are coded and transmitted as LED’s position
coordinates and the receiving PD can receive them and obtain
their values. We can divide the entire positioning space into
m×nsubspaces. Each of the four corners of each subspace
has one LED installed. Visible fingerprint positioning only
needs to be performed in these subspaces. The positioning
algorithm in each subspace is the same, so it can be described
abstractly as a visible light fingerprint positioning kernel.
We only need to use the data of one subspace during the ini-
tial phase of fingerprint positioning. Therefore, Equation (4)
and (5) can be re-described as
Pk=Pijk Pi(j+1)kP(i+1)jk P(i+1)(j+1)kT(7)
The complete fingerprint database can be defined as
P1... Pk... PK⇔c1... ck... cK(8)
Assuming that a set of K0training samples in the visible
light fingerprint library are independent of each other, the
e
K(e
KK) hidden neurons and the standard SLFNs are
shown in Fig. 3, whose activation function is h(x). If the
VOLUME 8, 2020 13877

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
FIGURE 3. The standard single layer feedforward networks.
input data Pkwas fed into the neuron network, the estimate
positioning coordinate ˆckcan be described as [20]
β·h(α·Pk+b)=bck,k=1,···,K(9)
where βis the weight matrix of size 3 ×e
Kconnecting
the hidden neurons and output neurons and αis the weight
matrix of size e
K×4 connecting the input neurons and hidden
neurons. b=b1,b2,...,be
KTis the threshold of the hidden
neurons.
Consider the K sampled data together, the corresponding
positions can be expressed as [21]
βH=ˆ
C(10)
where ˆ
C=[ˆc1,..., ˆck,..., ˆcK] is the output matrix of size
3×K.H(α,b,P1,···,PK)=[h(α·P1+b),...,h(α·PK+b)].
The traditional backward propagation (BP) algorithm
learning algorithm requires a lot of time for learning in most
cases because it adopts a gradient learning method. Differ-
ently, The input weight matrix αand hidden layer threshold b
of the network based on the extreme learning machine do not
need to be adjusted during the learning process. Therefore,
the least squares solution of (10) can be written as [20]


ˆ
βH(α,b)−C

=min
β
kβH(α,b)−Ck(11)
where C=c1... ck... cKis the ideal output matrix.
The least squares solution of Equation (11) can be
expressed as [22]
ˆ
β=CH †(12)
where H†is the Moore-Penrose generalized inverse matrix
of H. Using orthogonal projection, H†can be described as
H†=((HTH)−1HT,HTH:Non-singular matrix
HT(HHT)−1,HH T:Non-singular matrix (13)
In the actual VLC-based positioning process, we need to
consider that collecting samples is very difficult. In order to
reduce the number of samples, we can let the positioning
system continue learning after initialization [25]. This will
reduce the difficulty of training and the difficulty of data
FIGURE 4. The algorithm of VLP system based on ELM.
collection, and it will not affect the positioning accuracy. The
VLP system based on ELM is divided into 3 steps, as shown
in Algorithm 1 in Fig. 4.
III. SIMULATION AND ANALYSIS
A. INDOOR 3-D POSITIONING SIMUL ATION MODEL
Simulation experiments on the established ELM for VLP
using fingerprinting model is conducted to test the stability
and accuracy of the algorithm. As shown in Fig. 2(d), four
LEDs located on the ceiling in the simulation experiment.
The 3D visible light fingerprinting positioning database is
generated by the VLP kernel. The parameters used for the
3-D indoor VLP system using ELM and fingerprinting are
shown in Table 1. The ELM model needs to be trained before
positioning estimation. It is easy to obtain a large amount of
data through simulation, but in the actual process, data sam-
pling is very difficult, especially when the data size reaches
13878 VOLUME 8, 2020

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
TABLE 1. Simulation parameters of the visible light fingerprint
positioning kernel.
10,000 or even 100,000. Therefore, we chose 10000, 20000,
30000, 40000, and 50000 as training data sizes respectively
to compare their effect on training time and positioning accu-
racy. In our experiment, the number of ELM Hidden Layer
Nodes e
Kis set to 21.
Before carrying out three-dimensional positioning, the fin-
gerprint database needs to be trained. We need the com-
puter’s computational load to be as small as possible. One
of the important points is the size of the fingerprint database.
As shown in Fig. 5 (a) and (b), when the number of sampling
dots of fingerprint database increase from 10000 to 50000, the
training time increases rapidly while the average positioning
error does not drop significantly. Although the larger the fin-
gerprint database, the higher the positioning accuracy, but the
excessive fingerprint database will only increase the difficulty
of data collection and waste computing resources without
significantly improving the positioning accuracy. Therefore,
in the following experiments, we set the fingerprint database
density to 10000/12 m3.
FIGURE 5. The time and error of ELM training fingerprint database.
FIGURE 6. Mean square error of the training epochs.
The Mean Square Error (MSE) [26] can serve as the loss
function, that is defined as
MSE =1
K
K
X
k=1
[(xk− ˆxk)2+(yk−ˆyk)2+(zk− ˆzk)2] (14)
The behavior of MSE in different training epochs is displayed
in Fig. 6. When the epoch reach 160, the MSE is basically
close to 0.
B. SIMULATION RESULTS AND ANALYSIS
In order to verify the superiority of ELM for 3D vis-
ible light fingerprinting positioning, a simulated exper-
iment is conducted about the positioning accuracy of
ELM, KNN, and SVM algorithms, respectively. As shown
in Fig. 7 (a), (b), and (c), We can intuitively see the effect of
VOLUME 8, 2020 13879

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
FIGURE 7. The distribution of three-dimensional positioning. Orange
points are the exact positions and the blue point are the estimated 3-D
position. (a) Represents the 3-D position with SVM, (b) represents the 3-D
position with KNN, (c) represents the 3-D position with ELM.
three algorithms for 3D visible light fingerprinting position.
Among them, ELM has the best positioning effect. No matter
where it is located, positioning using ELM is closer to the real
point.
We can get this conclusion more accurately from the 3-D
positioning error cumulative distribution function (CDF) [27]
FIGURE 8. Mean square error of the training epochs.
TABLE 2. Computing resources and positioning error comparison of ELM,
KNN and SVM.
curves in Fig. 8. VLP system based on ELM performs better
than KNN and SVM. If 90% of the test points are con-
sidered to be acceptable, the 3-D position error CDF curve
in Fig. 8 shows that the test points have a maximum 3-D
position error of 0.055 m, which proves the VLP system
based on ELM performs very well in simulated environment.
As shown in TABLE 2, ELM has the shortest positioning time
and the lowest positioning error. Though VLP system based
on ELM training time is not the shortest, its excellent posi-
tioning time and positioning accuracy make it more suitable
for VLP system than KNN and SVM.
The whole simulation result of 3-D VLP system based
on ELM and fingerprinting is discussed in the Fig. 9. The
six figures show that the modified ELM for VLP using
fingerprinting works accurately, and their position error are
all smaller than 0.07 m. The resolution of the test positions
is 0.3 m with height from 0.3 m to 1.8 m in the room
and 216 positions are included. All of them are respectively
shown in Fig. 9 (a) to (f) for readers, where the solid green
points represent the estimated position and the solid red
points represent the exact position of the tested points. The
proposed positioning system works very good in the whole
room in the simulation experiment.
As shown in Fig. 10, the orange line represents the random
trajectory and the blue point represent the estimated position
13880 VOLUME 8, 2020

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
FIGURE 9. Estimated position and 3-D positioning error of the simulated experiment with different height.
of the trajectory using the VLP system based on ELM. From
Fig. 10, we can intuitively see that the proposed system has
good trajectory tracking effect. In order to further present the
positioning effect of the system, the average positioning error
is plotted as a function of height in Fig. 11. It can be seen
that the average positioning error will increase as the height
increases higher than 1.2 m. This is because as the height
increases, the angle ψijk between PD and LED increases, and
the accuracy of the measurement decreases.
IV. EXPERIMENTS IN REAL ENVIRONMENT AND RESULT
A. EXPERIMENT DESIGN
The data acquisition circuit structure of the three-dimensional
VLP system based on ELM is shown in Fig. 12. The computer
VOLUME 8, 2020 13881

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
FIGURE 10. 3-D estimated positioning in trajectory tracking using VLP
system based on ELM.
FIGURE 11. Average positioning error of the experiment with different
height.
at the transmitting part controls the LED through the serial
port, transmits the information and the position ID of the
LED. The STM32F407ZG receives the data from the com-
puter and sends data to the LED driving circuit through
the I/O port. The driving circuit controls the LEDs. The
PD at the receiving part receives the LED signal, and after
passing through the first stage amplifier circuit and the sec-
ond stage amplifier circuit, the STM32F407ZG acquires the
ADC measurement and obtains the RSSI value. At the same
time, the amplified signal passes through the comparator
and enters the STM32F407ZG through the I/O port, per-
forming demodulation and decoding to obtain the LED’s
position ID.
Visible Light Fingerprint Positioning Kernel Platform
Setup: The experiment in real environment is established to
verify the practicality and feasibility of the VLP kernel and
accuracy of the proposed positioning algorithm. The exper-
imental platform of VLP system based on ELM is shown
in Fig. 14 (length is 1.1 m, width is 0.9 m, height is 1.8 m).
Four LEDs (pattern: Kecent, KC-CBF09, white: 6000 K,
power: 9 W) as positioning base station are installed at the
top four corners. The PD circuit board of receiving part
FIGURE 12. The data acquisition circuit structure of the
three-dimensional positioning system.
FIGURE 13. The PD circuit board of receiving part.
can be seen in Fig. 13, which includes a digital output and
an analog output. The receiver samples the voltage through
analog output after the second stage amplifier circuit with
an ADC. At the same time, STM32F407 can get the LED
ID by acquiring digital output signal of the comparator and
decoding by CDMA [13]. The regenerated electrical signal is
uploaded to the PC for further processing.
B. RESULTS AND ANALYSIS
Implementation of the proposed three-dimensional posi-
tioning algorithm using python3.5, the algorithm runs in
Ubuntu16.04. Time spent on each positioning is 0.00177 s,
including the time that the signal flows through the receiv-
ing circuit, the time of serial communication and the time
of algorithm running. We collected 1,200 sets of data for
establishing a fingerprint database and trained the fingerprint
database using ELM algorithm. In the experiment, there are
24 evenly distributed test points at the height of 0.25 m,
0.50 m, 0.75 m, 1.00 m, 1.25 m and 1.50 m. The positioning
results are shown in Fig.15 and Fig. 18. To directly present
the result of 3-D positioning using the VLP system based on
ELM in real world, the CDF curves of horizon positioning
error, vertical positioning error and 3-D positioning error are
shown in Fig. 16. From the CDF curve shown in Fig. 16, 90%
13882 VOLUME 8, 2020

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
FIGURE 14. Experimental platform of VLP system using fingerprinting
and ELM.
FIGURE 15. Three-dimensional positioning effect of ELM in real world.
of the test points realize a positioning accuracy of 0.08 m.
As we can see from Fig.17 and Fig. 18, when the height
is 0.25 m, the average error is smaller than 0.025 m and
the maximum error is 0.04 m; when the height is 0.50 m,
the average error is smaller than 0.020 m and the maximum
error is 0.03 m; when the height is 0.75 m, the average error
is smaller than 0.0225 m and the maximum error is 0.04 m;
when the height is 1.00 m, the average error is smaller than
0.0275 m and the maximum error is 0.07 m; when the height
is 1.25 m, the average error is smaller than 0.0375 m and
the maximum error is 0.08 m; when the height is 1.50 m,
FIGURE 16. The cumulative distribution function (CDF) curves of
positioning error in real world.
FIGURE 17. Average positioning error of the experiment with different
height in real world.
TABLE 3. Accuracy and time performance for different methods.
the average error is smaller than 0.0525 m and the maximum
error is 0.10 m. It can be concluded that the VLP system based
on ELM works well for 3-D positioning.
VLP Kernel Study: We proposed the VLP kernel to reduce
the size of the fingerprint database. In the simulation,
VOLUME 8, 2020 13883

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
FIGURE 18. Estimated 3-D position of the experiment in real world.
we supposed that the VLP kernel =2 m×2 m×3 m.
Compared with four-layer neural network proposed by
Alonso-González et al. [28], the number of LEDs used
in our system is 9/16 of their, when the indoor area is
the same. In actual robot or drone positioning, the differ-
ence between 4mm and 5cm is not large. However, posi-
tioning time is more likely to affect their work. At each
positioning execution, the proposed VLP system based
ELM only need to process 28 nodes (4+21+3), whereas
their four-layer neural network has to process more than
129 nodes (16+80+30+3). Therefore, it can be estimated
that the proposed VLP system is much smaller in positioning
time than four-layer neural network, as can be appreciated
in Table 3.
13884 VOLUME 8, 2020

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
TABLE 4. Algorithm running environment.
FIGURE 19. The LED driving circuit of transmitting part.
FIGURE 20. The photodiode circuit of receiving part.
V. CONCLUSION AND FUTURE WORKS
In order to make indoor visible light fingerprint positioning
can be applied to scenes with large-scale LEDs, this paper
proposed a VLP system that firstly split the large indoor
positioning environment into regular VLP kernels and then
used fingerprinting and ELM algorithm for real-time posi-
tioning and less positioning error. In the simulation exper-
iment, the average positioning time is 0.000469 s and the
average 3-D positioning error is 0.0211 m. In the experiment
in real world, the average positioning time is 0.00177 s and
the average 3-D positioning error is 0.0365 m. Both the simu-
lation and the experiment results show that the proposed VLP
system achieves real-time 3-D positioning and much less
positioning error, which demonstrates the system proposed
is suitable for indoor robot positioning, UAV positioning and
other indoor scenes that require real-time positioning.
The three-dimensional VLP system supposed the receiver
is parallel to the ceiling. However, in the actual working
environment, the VLP system will tilt the receiver due to
movement, rotation, etc. Further work is needed to measure
device rotation or tilt while tracking and then calibrate the
fingerprinting database.
APPENDIX
A. HARDWARE AND SOFTWARE TO RUN ELM
See Table 4.
B. CIRCUIT DESIGN
See Figs. 19 and 20.
ACKNOWLEDGMENT
(Yirong Chen and Weipeng Guan are co-first authors.) Their
contributions to the paper are consistent and their ranking is
based solely on the first letter of the last name.
REFERENCES
[1] P. H. Pathak, X. Feng, P. Hu, and P. Mohapatra, ‘‘Visible light communica-
tion, networking, and sensing: A survey, potential and challenges,’’ IEEE
Commun. Surveys Tuts., vol. 17, no. 4, pp. 2047–2077, 4th. Quart., 2015,
doi: 10.1109/COMST.2015.2476474.
[2] A. Yassin, Y. Nasser, M. Awad, A. Al-Dubai, R. Liu, C. Yuen,
R. Raulefs, and E. Aboutanios, ‘‘Recent advances in indoor localization:
A survey on theoretical approaches and applications,’’ IEEE Commun.
Surveys Tuts., vol. 19, no. 2, pp. 1327–1346, 2nd Quart., 2016, doi:
10.1109/COMST.2016.2632427.
[3] N. Ul Hassan, A. Naeem, M. A. Pasha, T. Jadoon, and C. Yuen, ‘‘Indoor
positioning using visible LED lights: A survey,’’ ACM Comput. Surv.,
vol. 48, no. 2, Nov. 2015, Art. no. 20, doi: 10.1145/2835376.
[4] S. Hann, J.-H. Kim, S.-Y. Jung, and C.-S. Park, ‘‘White LED ceiling lights
positioning systems for optical wireless indoor applications,’’ in Proc.
36th Eur. Conf. Exhib. Opt. Commun. (ECOC), Nov. 2010, pp. 1–3, doi:
10.1109/ECOC.2010.5621490.
[5] A. Naz, H. M. Asif, T. Umer, and B.-S. Kim, ‘‘PDOA based indoor
positioning using visible light communication,’’ IEEE Access, vol. 6,
pp. 7557–7564, 2018.
[6] L.-C. Chen and J.-K. Lain, ‘‘Two-dimensional indoor visible light posi-
tioning using smartphone image sensor,’’ in Proc. IEEE Int. Conf. Con-
sum. Electron.-Taiwan (ICCE-TW), May 2018, pp. 2575–8284, doi:
10.1109/ICCE-China.2018.8448675.
[7] M. F. Keskin, S. Gezici, and O. Arikan, ‘‘Direct and two-step positioning in
visible light systems,’’ IEEE Trans. Commun., vol. 66, no. 1, pp. 239–254,
Jan. 2018, doi: 10.1109/TCOMM.2017.2757936.
[8] Y. Li, Z. Ghassemlooy, X. Tang, B. Lin, and Y. Zhang, ‘‘A VLC
smartphone camera based indoor positioning system,’’ IEEE Photon.
Technol. Lett., vol. 30, no. 13, pp. 1171–1174, Jul. 1, 2018, doi:
10.1109/LPT.2018.2834930.
[9] M. Yoshino, S. Haruyama, and M. Nakagawa, ‘‘High-accuracy positioning
system using visible LED lights and image sensor,’’ in Proc. IEEE Radio
WirelessSymp., Jan. 2008, pp. 439–442, doi: 10.1109/RWS.2008.4463523.
[10] S. H. Yang, H. S. Kim, Y. H. Son, and S. K. Han, ‘‘Three-dimensional
visible light indoor localization using AOA and RSS with multiple opti-
cal receivers,’’ J. Lightw. Technol., vol. 32, no. 14, pp. 2480–2485,
Jul. 15, 2014.
[11] Y. Cai, W. Guan, Y. Wu, C. Xie, Y. Chen, and L. Fang, ‘‘Indoor high
precision three-dimensional positioning system based on visible light com-
munication using particle swarm optimization,’’ IEEE Photon. J., vol. 9,
no. 6, pp. 1–20, Dec. 2017, doi: 10.1109/JPHOT.2017.2771828.
[12] H. Steendam, ‘‘A 3-D positioning algorithm for AOA-based VLP with
an aperture-based receiver,’’ IEEE J. Sel. Areas Commun., vol. 36, no. 1,
pp. 23–33, Jan. 2018, doi: 10.1109/JSAC.2017.2774478.
[13] H. Chen, W. Guan, S. Li, and Y. Wu, ‘‘Indoor high precision three-
dimensional positioning system based on visible light communication
using modified genetic algorithm,’’ Opt. Commun., vol. 413, pp. 103–120,
Apr. 2018, doi: 10.1016/j.optcom.2017.12.045.
VOLUME 8, 2020 13885

Y. Chen et al.: Indoor Real-Time 3-D VLP System Using Fingerprinting and ELM
[14] D. Konings, B. Parr, F. Alam, and E. M.-K. Lai, ‘‘Falcon: Fused
application of light based positioning coupled with onboard net-
work localization,’’ IEEE Access, vol. 6, pp. 36155–36167, 2018, doi:
10.1109/ACCESS.2018.2847314.
[15] B. Zhu, J. Cheng, Y. Wang, J. Yan, and J. Wang, ‘‘Three-dimensional
VLC positioning based on angle difference of arrival with arbitrary tilting
angle of receiver,’’ IEEE J. Sel. Areas Commun., vol. 36, no. 1, pp. 8–22,
Jan. 2018, doi: 10.1109/JSAC.2017.2774435.
[16] M. Xu, W. Xia, Z. Jia, Y. Zhu, and L. Shen, ‘‘A VLC-based 3-D indoor
positioning system using fingerprinting and K-nearest neighbor,’’ in Proc.
Veh. Technol. Conf. (VTC Spring), Nov. 2017, pp. 1–5, doi: 10.1109/VTC-
Spring.2017.8108345.
[17] S. Berchtold, C. Böhm, D. A. Keim, and H.-P. Kriegel, ‘‘A cost model
for nearest neighbor search in high-dimensional data space,’’ in Proc.
16th ACM SIGACT-SIGMOD-SIGART Symp. Princ. Database Syst., 1997,
pp. 78–86, doi: 10.1145/263661.263671.
[18] K. Kaemarungsi and P. Krishnamurthy, ‘‘Modeling of indoor positioning
systems based on location fingerprinting,’’ in Proc. 23rd Annu. Joint Conf.
IEEE Comput. Commun. Soc. (INFOCOM), Mar. 2004, pp. 1012–1022,
doi: 10.1109/INFCOM.2004.1356988.
[19] T. Wigren, ‘‘Adaptive enhanced cell-ID fingerprinting localization by
clustering of precise position measurements,’’ IEEE Trans. Veh. Technol.,
vol. 56, no. 5, pp. 3199–3209, Sep. 2007, doi: 10.1109/TVT.2007.900400.
[20] G.-B. Huang, H. Zhou, X. Ding, and R. Zhang, ‘‘Extreme learning
machine for regression and multiclass classification,’’ IEEE Trans. Syst.,
Man, Cybern. B, Cybern., vol. 42, no. 2, pp. 513–529, Apr. 2012, doi:
10.1109/TSMCB.2011.2168604.
[21] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, ‘‘Extreme learning machine:
Theory and applications,’’ Neurocomputing, vol. 70, nos. 1–3,
pp. 489–501, 2006, doi: 10.1016/j.neucom.2005.12.126.
[22] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, ‘‘Extreme learning machine:
A new learning scheme of feedforward neural networks,’’ in Proc.
IEEE Int. Joint Conf. Neural Netw., Jan. 2004, pp. 985–990, doi:
10.1109/IJCNN.2004.1380068.
[23] Y. Miche, A. Sorjamaa, P. Bas, O. Simula, C. Jutten, and A. Lendasse,
‘‘OP-ELM: Optimally pruned extreme learning machine,’’ IEEE Trans.
Neural Netw., vol. 21, no. 1, pp. 158–162, Jan. 2010, doi: 10.1109/
TNN.2009.2036259.
[24] T. Komine and M. Nakagawa, ‘‘Fundamental analysis for visible-
light communication system using LED lights,’’ IEEE Trans. Consum.
Electron., vol. 50, no. 1, pp. 100–107, Feb. 2004, doi: 10.1109/
TCE.2004.1277847.
[25] Y. Sun, Y. Yuan, and G. Wang, ‘‘An OS-ELM based distributed ensemble
classification framework in P2P networks,’’ Neurocomputing, vol. 74,
no. 16, pp. 2438–2443, Sep. 2011, doi: 10.1016/j.neucom.2010.12.040.
[26] C. M. Theobald, ‘‘Generalizations of mean square error applied to ridge
regression,’’ J. Roy. Stat. Soc. B (Methodol.), vol. 36, no. 1, pp. 103–106,
1974, doi: 10.2307/2984775.
[27] M.-H. Chuna, S.-J. Han, and N.-I. Tak, ‘‘An uncertainty importance mea-
sure using a distance metric for the change in a cumulative distribution
function,’’ Rel. Eng. Syst. Saf., vol. 70, no. 3, pp. 313–321, Dec. 2000, doi:
10.1016/S0951-8320(00)00068-5.
[28] I. Alonso-González, D. Sánchez-Rodríguez, C. Ley-Bosch, and
M. A. Quintana-Suárez, ‘‘Discrete indoor three-dimensional localization
system based on neural networks using visible light communication,’’
Sensors, vol. 18, no. 4, p. 1040, Mar. 2018, doi: 10.3390/s18041040.
YIRONG CHEN received the B.S. degree in elec-
tronic science and technology from the South
China University of Technology, China, in 2019.
He is currently pursuing the Ph.D. degree with
the School of Electronic and Information Technol-
ogy, South China University of Technology. His
research interests include visible light communica-
tions, natural language processing, and emotional
calculation. He is also a Journal Reviewer of the
IEEE PHOTONICS JOURNAL and IEEE ACCESS.
WEIPENG GUAN received the B.E. degree
from the Electronic Science and Technology
Department (Electronic Materials and Compo-
nents Department) and the M.S. degree from
the Control Theory and Control Engineering
Department, South China University of Technol-
ogy, Guangzhou, China. He is currently pursuing
the Ph.D. degree with the Department of Infor-
mation Engineering, The Chinese University of
Hong Kong, Hong Kong. His research is focused
on visible light wireless communication technology and visible light posi-
tioning technology.
JINGYI LI was born in Gansu, China, in 1998.
He is currently pursuing the B.E. degree with the
School of Automation Science and Technology,
South China University of Technology (SCUT),
Guangzhou, China. He joined R&C Academic
Studio, SCUT, when he was a Sophomore. His
research interests are visible light communica-
tions, machine vision, and optimization and edge
computing.
HONGZHAN SONG was born in Jiangxi, China,
in 2001. She is currently pursuing the B.E. degree
with the School of Automation Science and Engi-
neering, South China University of Technology.
She is also focused on the research of mobile
robots and visible light positioning.
13886 VOLUME 8, 2020