Article
Design of an SVM Classifier Assisted Intelligent Receiver for
Reliable Optical Camera Communication
Md. Habibur Rahman 1 , Md. Shahjalal 1 , Moh. Khalid Hasan 2 , Md. Osman Ali 1,3 , and Yeong Min Jang 1, *
Citation: Lastname, F.; Lastname, F.;
Lastname, F. Title. Sensors 2021, 1, 0.
https://doi.org/
Received:
Accepted:
Published:
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional af-
filiations.
Copyright: © 2021 by the authors.
Submitted to Sensors for possible
open access publication under the
terms and conditions of the Cre-
ative Commons Attribution (CC
BY) license (https://creativecom-
mons.org/licenses/by/ 4.0/).
Version May 22, 2021 submitted to Sensors
1 Department of Electronics Engineering, Kookmin University, Seoul 02707, Korea;
rahman.habibur@ieee.org; mdshahjalal26@ieee.org; osman@kookmin.ac.kr; yjang@kookmin.ac.kr
2 Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ, USA;
khalidrahman45@ieee.org
3 Department of Electrical and Electronic Engineering, Noakhali Science and Technology University,
Noakhali, Bangladesh; osman.eee@nstu.edu.bd
* Correspondence: yjang@kookmin.ac.kr; Tel.:+82-02-910-5068
1 Abstract: Embedding optical camera communication (OCC) commercially as a favorable com-
2 plement of radio-frequency technology has led to the desire for an intelligent receiver system
3 that is eligible to communicate with an accurate Light emitting diode (LED) transmitter. To shed
4 light on this issue, a novel scheme for detecting and recognizing data transmitting light-emitting
5 diode (LED) has been elucidated in this paper. Since, the optically modulated signal is captured
6 wirelessly by a camera that plays the role of receiver for the OCC technology, the process to detect
7 LED region and retrieval of exact information from the image sensor is required to be intelligent
8 to achieve a low bit error rate (BER) and high data rate to ensure reliable optical communication
9 within limited computational abilities of the most used commercial cameras such as those in
10 smartphones, vehicles, and mobile robots. In the proposed scheme, we have designed an intelli-
11 gent camera receiver system which is enabled to separate accurate data transmitting LED regions
12 removing other unwanted LED regions employing support vector machine (SVM) classifier along
13 with convolutional neural network (CNN) in the camera receiver. CNN is used to detect every
14 LED region from the image frame and then essential features are extracted to feed into an SVM
15 classifier for further accurate classification. The receiver operating characteristic curve and other
16 key performance parameters of the classifier has been analyzed broadly to evaluate the perfor-
17 mance and justify the assistance of the SVM classifier in recognizing accurate LED region and
18 decode data with low BER. To investigate communication performances, BER analysis, data rate,
19 and inter-symbol interference have been elaborately demonstrated for the proposed intelligent
20 receiver. Besides, BER against distance and BER against data rate have also been exhibited to
21 validate the effectiveness of our proposed scheme comparing with only CNN classifier based
22 receiver, and only SVM classifier based receiver. Experimental results have ensured the robustness
23 of the proposed scheme and the applicability of this scheme even if in the mobile scenario.
24 Keywords: Support vector machine (SVM) classifier, intelligent receiver, optical camera communi-
25 cation (OCC), image sensor.
26 1. Introduction
27 In the recent years, optical camera communication (OCC) has been considered a
28 potential wireless access option owing to the recent advancement in the luminous efficacy
29 of light-emitting diode (LED) technology. At present, a dense formation of LED-based
30 lighting infrastructures is being observed almost everywhere owing to offering low
31 power consumption, low cost, and large optical spectrum immune to electromagnetic
32 interference [1–3]. The large cost-free optical spectrum has created a new venture for
33 resolving data congestion into the radio frequency (RF) spectrum. Thus, LEDs have
34 become a significant option for illumination and communication simultaneously. In
https://www.mdpi.com/journal/sensors
Version May 22, 2021 submitted to Sensors 2 of 17

35 contrast, the successive improvement of complementary metal-oxide-semiconductor

36 (CMOS) technology has integrated a new dimension in OCC development providing
37 high pixel resolution and high-speed built-in camera inherited smart devices [4], [5].
38 Therefore, OCC has become a promising indoor and outdoor communication technology.
39 Recently, with the enrichment of OCC technology, it opened up a new frontier
40 for the future massive connected network. Recently, academia and industry both have
41 a surge of interest in taking advantage of OCC technology. A variety of beneficial
42 applications are expected to be achieved leveraging OCC technology in some prospective
43 areas such as vehicle to everything (V2X) communications, robotics, automation, and
44 e-Health. Furthermore, to deploy 5G communication system effectively and to meet
45 the requirement of the upcoming 6G communication system, OCC provides ample
46 opportunity [6–10]. While this technology is being continuously improved, it is not
47 yet adequate to coronate OCC technology commercially in the initially envisioned
48 areas. Currently, artificial intelligence (AI) based approach has been adopted to enhance
49 different performance features, for example, data rate, SNR, and communicating distance
50 [11].
51 In the traditional scheme, the stream of data is converted into an optical signal
52 being modulated using LEDs. Various modulation techniques have been proposed so
53 far including both single-level intensity modulation schemes such as ON-OFF keying
54 (OOK), under-sampled frequency shift ON-OFF keying (UFSOOK), under-sampled
55 phase shift ON-OFF keying (UPSOOK), and spectral-efficient transmission scheme
56 techniques such as color shift keying (CSK), multi-level intensity modulation. Afterward,
57 the modulated signal is sent through a wireless medium [12–16]. The transmitted data
58 is captured in the image sensor receiver. Deep learning techniques, notably CNN, are
59 employed at the receiver side to detect LED and the data bits are retrieved using several
60 image processing techniques. Recently, it has been reported that stripe distortion can be
61 effectively reduced when CNN is employed to extract features from the bright and dark
62 striated frames during decoding [17,18]. The key issues that can characterize the OCC
63 performance, are broadly investigated in [19]. A new study was conducted to improve
64 decoding methods. They demonstrated differentiating multi-valued data more correctly
65 for intelligent transport systems [20]. Besides, the authors also demonstrated how OCC
66 performance is varied according to the key performance parameters. Machine learning
67 techniques at the receiver side has been also investigated due to the low computational
68 complexity than deep learning algorithm [21]. Authors in [22] used support vector
69 machine (SVM) in 8-SPAM and DCO-OFDM visible light communication (VLC) system
70 in LED detection purpose. SVM utilized optical barcode detection technique based on
71 VLC have been studied in [23]. In [24], the authors have described SVM for constellation
72 classification in two kinds of geometrically-shaped 8 quadrature amplitude modulation
73 seamless integrated fiber and VLC system. However, in all the traditional detection
74 scheme, all the detected LED have been considered as data transmitting LED. But, in
75 practical many unwanted LED regions can be detected which demolishes the robustness
76 of the communication performance. In object detection purpose, CNN outperforms
77 than the other machine learning technique. But, it cannot detect the actual transmitting
78 LED regions deleting other unwanted LED region from the camera frame because it
79 does not consider spatial information of the object. So it fails to differentiate actual data
80 transmitting LED from other unwanted LEDs [25]. Consequently, the data decoding
81 process gets vulnerable in practical scenarios such as V2V and robotics in case of the
82 presence of many LEDs in the field of view (FOV) of the camera. This seems to be a large
83 roadblock towards the potential application of OCC technology.
84 To the best of our knowledge, it is not investigated yet how a receiver can recognize
85 actual data transmitting LED region from the frame among other unwanted LED regions
86 and decode data intelligently from that accurate region in the OCC case. Therefore, in
87 this paper, a novel intelligent camera receiver system employing SVM, a supervised
88 learning model, along with CNN for OCC has been designed. Our designed receiver
Version May 22, 2021 submitted to Sensors 3 of 17

Figure 1. Scenario of the intelligent receiving system for OCC.

89 system is capable of decoding data bit separating accurate data transmitting LED region
90 among various light-emitting sources present in the camera frame. Taking account of the
91 visual flickering, we have modulated the data stream by FSOOK technique, a low cost
92 and low complexity modulation technique, which encodes bits into different frequency
93 [26]. This modulated optical signal was transmitted through a wireless medium to the
94 receiver in the presence of other blinking LEDs. The scenario of how the proposed
95 intelligent camera receiver will work is depicted in Fig. 1. Based on our scheme, we
96 have developed a new data decoding algorithm for FSOOk based OCC Our designed
97 receiver first detects every LED region employing CNN. From each region, features are
98 extracted to train the SVM classifier. Using the trained classifier model, the camera can
99 successfully separate accurate regions for data decoding purposes. As a result, unwanted
100 LED regions will be removed which will result in low computational complexity and
101 mitigation of neighboring interference.
102 The remainder of this paper is structured as follows. Section II describes the SVM
103 classifier. The overall OCC system including modulation technique and algorithm for
104 decoding data intelligently from the recognized region is has been illustrated in Section
105 III. We have demonstrated how the data set has been prepared using extracted features
106 in Section IV. In Section V, we elaborately discussed the experimental results to justify
107 our scheme. Finally, we concluded our article in Section VI.

108 2. Support vector machine (SVM)

109 The Support vector machine (SVM) has been considered one of the most widely
110 used and accurate classification algorithms than others after being introduced in 1990
111 within the machine learning community [27]. Fundamentally, it’s a supervised learning
112 algorithm whose main objective is creating an optimal hyperplane between two classes
113 of data set focusing on the training samples even if for a miniature training set. The
114 hyperplane acts as a decision boundary to categorize the data in different classes. In Fig.
115 2, points nearer the hyperplane, called support vector, are used to determine optimized
Version May 22, 2021 submitted to Sensors 4 of 17

Figure 2. Illustration of SVM classifier.

116 hyperplane. For a given training sample (x1 , y1 ), (x2 , y2 ), . . . .., (xm , ym ) where class
117 labels are represented by

yi = {+1, −1}, ∀ie{1, 2, 3, .., m}. (1)

The optimal separating hyperplane is mathematically formulated as [28]

θ T xi + b = 0, (2)

118 where θ = [θ1 , θ2 ...., θm ] is m-dimensional vector of weights and xi = [ x0 , x1 , ...., xm ]

119 is an m-dimensional input vector where x0 = 1 by convention and b is termed as the
120 biasing unit. The optimization problem associated with finding the hyperplane can be
121 expressed as follows

1 1
kθ k2 = θ T θ, (3)
2 2
which is subjected to
θ T xi + b ≥ +1 for yi = +1. (4)
T
θ xi + b ≤ +1 for yi = −1. (5)
The final nonlinear decision function can be obtained as follows
!
m  
f (x) = sign ∑ αi θ xi + b .
T
(6)
i=1

Table 1: Kernel function for the proposed system.

Type of Kernel Function Kernel Function

Linear x T xi + c
 
 k x − x k2 
RBF exp − 2σ2 
 i

Polynomial ( x T xi + c ) p
Sigmoid tanh(( x T xi + c) p )
Version May 22, 2021 submitted to Sensors 5 of 17

122 The commonly used kernel functions are Linear, Polynomial, sigmoid, the radial basis
123 function (RBF) which are briefly described in the following Table 1.
To generalize SVM classifier about complex data features, SVM uses a technique
called Kernel k(xi , x). The value k(xi , x) corresponds to ϕ(xi ).ϕ(x) which maps linearly
non-separable patterns into a higher dimension feature space. Finally, the decision
function can be modified as follows
!
m
f (x) = sign ∑ αi (ϕ(xi ).ϕ(x)) + b . (7)
i=1

124 3. OCC System Overview

125 3.1. OCC System architecture

Figure 3. Overall OCC architecture with the proposed classifier model.

126 In Fig. 3, the block diagram delineates the overall OCC architecture in addition to
127 the proposed classifier model. The structure mainly consists of two parts: the transmitter
128 part and the receiver part but unlike the traditional receiver system, the receiver part
129 includes some additional features that assist to differentiate the accurate data transmit-
130 ting LED over other LEDs. On the transmitter side, numerous LED regions are present.
131 The data are modulated by FSOOK modulation at a frequency level between 2-4 kHz
132 before sending to the receiver. While the modulation is performed, the LED blinking
133 should not be perceivable to the human eye. Since, the perceivable flickering rate is
134 up to 200 Hz [29]. This modulation frequency used in FSOOK can attenuate flickering
135 issue in a significant margin. With regards to the transmitting LEDs and other sources of
136 noise, they are not modulated in that particular frequency. The projection of all these
137 light-emitting sources is captured by the image sensor. Interference sources are dispelled
138 when a trained CNN model is used. All the possible LED sources are detected and
139 segmented using image processing techniques. Due to the rolling shutter effect of CMOS
140 based image sensor, each image frame gets striated of white and black shade. Afterward,
141 necessary geometrical features are extracted to classify and recognize accurate data
142 transmitting region from the range of all possible regions.
Version May 22, 2021 submitted to Sensors 6 of 17

143 3.2. FSOOK modulation:

Figure 4. Illustration of FSOOK modulation technique and data packet architecture.

144 It is a modulation scheme that encodes bit ‘1’ and bit ‘0’ with two different fre-
145 quencies known as mark and space frequency respectively [26]. The header and footer
146 is appended at the beginning and ending of of data bits respectively to identify the
147 data packets. In order to create distinction among header, footer and data bits, we use
148 synchronization bit pattern ‘111’ as the header and footer. Due to the rolling shutter
149 effect, we get sequence of bit pattern in each data packet depicted in Fig. 4.

150 3.3. Proposed SVM classifier Assisted Intelligent Receiver

Figure 5. Workflow inside proposed intelligent OCC receiver.

151 The workflow of the proposed intelligent receiver system is depicted in Fig. 5.
152 The receiver system consists of CMOS based image sensor that collect optical signals
Version May 22, 2021 submitted to Sensors 7 of 17

153 transmitted through the optical channel. The captured image is now input to the
154 object classifier designed by CNN algorithm. These algorithms are eligible to extract
155 image features automatically and can learn from those features effectively. The CNN
156 recognizes all light source regions by performing the functions like reason proposal,
157 feature extraction, classification, etc. Several layers inside CNN namely convolutional
158 layers, pooling layers, and fully connected layers, perform those functions but the
159 convolutional layer is the main part of CNN. The convolutional layers are composed
160 of convolutional kernel which is used to generate proposed region and feature maps
161 from the original image. Convolutional kernel can be of many sizes. In our cases, we
162 have used 4 × 4 convolutional filter to extract the features. The convoluted output is
163 estimated by the following equation [30]
q −1 p −1
y= ∑ ∑ (w pq xa+1,b+ p + Θ), (0 ≤ a ≤ a f , 0 ≤ b ≤ b f ). (8)
q=0 p=0

164 Max pooling with a 2 × 2 filter further used to reduce the spatial size of the warped
165 images and computation in the network resulting in less prone to overfitting preserving
166 all features. This pooled feature map was flattened into one column while passing
167 through the FC layers. Here, FC layers tries to determine most correlated features of
168 a particular class. The non linear softmax function is used to classify LED from other
169 objects in Fc layer.
170 In Fig. 5, the detected LED regions with bounding boxes are the output of the CNN.
171 The bounding box objects are then finally processed to precisely determine the accurate
172 data transmitting LED and decode data as well. To make the receiver an intelligent
173 and robust retrieval system, the system needs to respond by differentiating accurate
174 LEDs as quickly as possible. Therefore, the selection of features is very important.
175 Extraction of appropriate features is very challenging since the objects are almost the
176 same. LED objects of the same type can have a different shape, stripe pattern based on
177 the communication distance, camera frame rate. mark and space frequency. To analyze
178 the different geometrical shapes, computation of the features from the contour line has
179 been considered significant distinguishable features [31]. Contours are the outline that
180 is designed using the edges of the object to represent the shape. They contain some
181 geometrical attributes that are effective to recognize and segment objects.
182 Based on the extracted input features, a trained SVM classifier detects accurate data
183 transmitting the LED region. Finally, from that region data bits are decoded accordingly.
184 The overall demodulation scheme of our proposed intelligent receiver is shown in
185 Algorithm 1.

186 4. Data Set Preparation

The prime objective of our proposal is to enable the camera to intelligently deter-
mine data transmitting LED and communicate with this LED. Hence, the camera needs
a trained AI-based classifier by which it can perform the required task. To train the
classifier, a data set is needed which contains some necessary features from the bounding
boxes of detected LED regions. In order to build the data set, 1000 bounding box images
of the LED region are used. While extracting features, parameters from Table 2 are used.
Input feature vectors are constructed considering different situations such as varying
distance from LED to the camera, moving LED from left to right. But before performing
feature extraction, a closing operation is applied on the selected regions to combine all
the neighboring stripes, and contour is drawn on the shape by combined stripes. It is a
morphological filtering technique that is composed of gray level dilation followed by
gray level erosion taking input of the selected regions and a structuring element. The
closing operation can be expressed as [32]

f pn∗q · bm∗n = ( f pn∗q ⊕bm∗n ) · bm∗n , (9)

Version May 22, 2021 submitted to Sensors 8 of 17

Algorithm 1 Demodulation scheme at the receiver.

Input: Captured image of all LEDs state.
Output: x = "11101011..." from accurate data transmitting LED region.
1: read each RGB image Iu∗v∗3 ;
2: detection of each LED region by CNN and separate regions
Iu1∗v∗3 , Iu2∗v∗3 , Iu3∗v∗3 , ..., Iun∗v∗3 ;
3: convert to grayscale image Iu∗v∗3 → f p∗q (gray);
4: draw contour on each f pn∗q ;
5: for i = 0 : no. of separated LED region do
6: extract features [ST, M00 , PR , NL ];
7: pass features to trained SVM classifier;
8: recognize accurate data transmitting LED region;
9: end for
p
10: normalize each row intensities values ∑i=1 Zi ;
11: set median of normalized values as threshold zth ;
12: calculate average Zi (avg) of each ten Zi successive values;
p
13: for i = 0 : 10 do
14: if Zi (avg) > zth then
15: set xi = "1";
16: set xi−1 = "0";
17: else
18: pass;
19: end if
20: end for

187 where f pn∗q is nth region and bm∗n is the structuring element of which m, n  a, b.

Table 2: Parameters used in the implementation of the proposed model.

Parameter Value
LED 10 mm (diameter)
Camera exposure time 15.6 ms, 3.90 ms, 976.6 µs
Camera frame rate 30 fps
Camera image resolution 1280 ∗ 720
Mark and Space frequency 4000 kHz and 2000 kHz
Learning rate parameter 0.001
No. of epoch 30
Kernels RBF, linear, polynomial, sigmoid

188 In Fig. 6, we have presented a single frame where the shape of combined stripes
189 contour is shown. For each projected LED either data transmitter or unwanted LED, we
190 found this type of shape of the contour formed based on the combined stripe. Analyzing
191 the geometrical pattern of the contour, there is considerable change in the pattern when
192 the communication distance is varied or mark and space frequency has been changed.
193 So based on the contour pattern of the data transmitting LED region and other unwanted
194 LED regions, we have selected the following features and the features are elaborately
195 described in the next section.
Version May 22, 2021 submitted to Sensors 9 of 17

Figure 6. Example of a single frame where all the generated stripes are combined together and a
contour is formed.

196 4.1. Zero order moments or area of LED region (M_0_0)

The moment is a unique feature that defines the relationship between pixels of the
corresponding objects with their surroundings. For many years, it has been using to
recognize objects from image frames. Fundamentally, it computes the weighted average
intensity of the pixels and therefore it cannot be applied on a single pixel except a definite
area of the projected object [? ]. For an image, f (p ∗ q) the moment of order (x, y) can be
defined as
m n
Mx y = ∑ ∑ σx p yq f (p ∗ q); x, y = 0, 1, 2, ..., (10)
p=0 q=0

197 where, σx p yq denotes moment weighting kernel. Different image properties lies in the
198 raw moments. In our case, we have denoted moment of order (0, 0) as area of the
199 combined stripes for data transmitting LED.

200 4.2. No. of Stripes Per LED region (S_t)

201 Since LED modulated data results in dark and bright stripes in the captured image,
202 the number of stripes and the width plays an important role to recover the data stream
203 from the image frame. The number of stripes is varied proportionally with the size of
204 LED, mark, and space-frequency of LED and also a distance between the camera and LED
205 [16]. It is noticed that high-speed cameras with high read-out capability outperformed
206 in terms of communication speed. The width of the bright and dark stripes can be
207 theoretically determined as [19]

1
ST = , (11)
f tr
208 where, f represents mark or space frequency and tr is the time needed to read-out a
209 single pixel of the image.

210 4.3. Perimeter of combined stripes contour (P_r)

211 A contour is drawn considering all the stripes generated by the LED blinking and
212 rolling shutter effect of the camera. Hence, the arc length of the enclosed contour varies
213 according to the number of stripes. So, the perimeter is also considered a significant
214 feature for our data set.

215 4.4. No. of line segment of combined stripes contour (N_L)

216 The Douglas-Peucker algorithm developed by D. Douglas and T. Peucker in 1973
217 splits a curve into line segments depending upon the specified precision (e). It determines
218 the dominant point on the curve by digitizing the original curve and generates a similar
Version May 22, 2021 submitted to Sensors 10 of 17

219 curve of fewer points [34]. We have used this algorithm to relate the number of line
220 segments with the no. of stripes It is noticeable that with the variation of LED on-
221 off frequency and distance, getting accurate contour of the object is difficult. So, in
222 the counting of stripes error may present. Combining all stripes, we segmented the
223 combined contour using the Douglas-Peucker algorithm based on the corner point of
224 each stripe.
225 The scatter plot of the selected features has been shown in Fig. 7. In the figures,
226 the distribution of features has shown clear distinction among data transmitting LED
227 and unwanted LED which affirms our motive to introduce an SVM classifier in the
228 conventional receiver system.

Figure 7. Scatter plot of selected features (a) area vs. no. of stripes per LED region, (b) area vs.
perimeter of the combined stripes contour, and (c) area vs. no. of the line segment of the combined
stripes contour.

229 5. Experimental results

In this section, the performance of our proposed scheme for OCC receiver has been
demonstrated explicitly. The overall performance has been measured experimentally
and the experiments were conducted considering different conditions. We have taken
three LED for our experiment. The whole decoding procedure of receiver side has been
performed and analyzed in Python 3.7 environment. Firstly, we examined classifier
performance to classify and recognize accurate data transmitting LED varying kernel
function of the SVM classifier. During the experiment, we have used three LEDs among
which we varied no. of actual data transmitting LED’s and the rest were considered as
interfering sources and we also varied the communication distance (10 cm to 60 cm) to
justify the effectiveness of our scheme. However, to examine the efficiency of separating
accurate LED region intelligently, the SVM classifier performance has been investigated
Version May 22, 2021 submitted to Sensors 11 of 17

demonstrating receiver operating characteristic (ROC) curve and other key performance
parameters, such as precision, recall, F1 Score. These parameters are defined as

True positive True positive

Precision = = ,
True positive + False positive Total predicted positive
True positive True positive
Recall = = , (12)
True positive + False negetive Total actual positive
Recall ∗ Precision
F1 Score = 2 ∗ .
Recall + Precision
230 We have also evaluated the performances from the communication perspective. It
231 shows satisfactory performance on the BER, data rate, and inter-symbol interference
232 (ISI).

233 5.1. Classifier performance

Figure 8. ROC curve of SVM classifier for (a) RBF kernel, (b) linear kernel, (c) polynomial kernel,
and (d) sigmoid kernel.

234 We have classified accurate data transmitting region from all the possible cap-
235 tured LED regions detected by CNN using the trained SVM classifier. Based on the
236 features that we have selected, the SVM classifier provides an accurate LED region for
237 demodulation. So, the other communication performances are greatly dependent on
238 the classification and resulted separated LED region. Therefore, to evaluate classifier
239 performance, we have shown the ROC curve for different Kernel functions, while the
240 communication distance was set to 30 cm and there were two LEDs among which one
241 LED was transmitting data actually, in Fig. 8. Beside, the exposure time Te was kept 976.6
242 µs. The area measured under the ROC curve known as ’AUC’ has been measured and the
243 measured ’AUC’ value is larger for RBF kernel than others and hence the classification
244 accuracy was also
Version May 22, 2021 submitted to Sensors 12 of 17

Figure 9. Comparison among precision, recall, F1 score varying exposure time (Te ).

245 Larger for the RBF kernel. In Table 3, we have presented overall accuracy and ’AUC’
246 measurement considering the different conditions. In Table 3, the recorded maximum
247 accuracy for the RBF kernel is 94.92% and the lowest accuracy for the sigmoid kernel is
248 79.87%. Moreover, we have also given a bar plot to show the comparison among key
249 performance parameters varying exposure time Te while the number of data transmitting
250 LED was one in Fig. 9. For the exposure time 976.6 µs, all the performance indexes are
251 high. However, the achieved performance has validated our scheme being capable of
252 recognizing accurate LED regions.

Table 3: Classifier performance (Accuracy and AUC) analysis with different no. of LED

No. of
Linear RBF Polynomial Sigmoid
LED
Accuracy
92.09 94.92 89.45 80.93
One (%)
AUC .89 .94 .87 .79
Accuracy
90.05 93.10 87.91 81.36
Two (%)
AUC .85 .91 .83 .81
Accuracy
87.76 89.91 88.13 79.87
Three (%)
AUC .84 .88 .85 .77
Version May 22, 2021 submitted to Sensors 13 of 17

253 5.2. BER, data rate, and ISI analysis

Figure 10. Comparison of bit error rate (BER) among receiver using only SVM classifier, receiver
using only CNN classifier, and our proposed scheme.

254 After classifying accurate data transmitting LED, data bit has been decoded from
255 the cropped region using decoding algorithm described in the Algorithm 1. We have
256 decoded data and evaluated the communication performance of our proposed intelligent
257 receiver. According to our scheme, the receiver intelligently selects accurate LED regions
258 so the probability of interference from other sources or reflection from the surrounding
259 surface has been reduced. Fig. 10 depicts the measured BER performance against
260 communication distance for the receiver employing only SVM classifier, only CNN and
261 our proposed scheme. Significant reduction of BER has been observed when the SVM
262 classifier along with CNN is used. At the maximum distance of 60 cm, the BER increases
263 up to 7.9× 10−4 , which is below the FEC limit of 3.8 × 10−3 . At this point, only SVM
264 based receiver shows BER above than FEC limit whereas CNN based receiver is very
265 close to the FEC limit. It signifies the effectiveness of our proposed scheme.

Figure 11. BER analysis by (a) varying exposure time (Te ), and (b) varying no. of data transmitting
LED.

266 The BER performance varies based on exposure time. The effect of changing
267 exposure time Te on BER for a single data transmitting LED keeping two unwanted
268 LED regions has also appeared in Fig. 11(a). With the increasing of exposure time BER
269 shows enhancement in BER. This considerable enhancement in BER has appeared due
270 to the effect of interfering noise from other sources. Although with the increasing of the
Version May 22, 2021 submitted to Sensors 14 of 17

271 exposure time Te BER has increased, the maximum BER at the maximum distance is also
272 below the FEC limit in our scheme. We have also investigated BER performance varying
273 number of LED and it is shown in Fig. 11(b). When the no. of actual data transmitting
274 LED increases, efficient separation of LED regions have enormous influences on BER.
275 Therefore, a significant change in BER has been found when no. of data transmitting
276 LED has been increased from one to three. When the no. of data transmitting LED is
277 three, the BER has been achieved around 3.5 × 10−3 for exposure time Te = 976.6 µs.

Figure 12. Eye diagram of SVM assisted intelligent OCC receiver while exposure time (Te ) is 977.6
µs.

278 Furthermore, the eye diagram depicted in Fig. 12 at exposure time Te = 976.6 µs
279 has ensured the low ISI at the receiver side. During multiple data transmitting LED
280 region cases, if the receiver fails to differentiate distinct regions accurately, recovering
281 distinct pulse for a distinct bit can be difficult for the receiver. The large eye opening
282 affirms the low ISI presence at separated regions classified by the SVM classifier.

Figure 13. Data rate comparison varying no. of LED while exposure time (Te ) is 977.6 µs.

283 Since the SVM classifier separates LED regions and data is decoded from the
284 separated regions, it has a significant effect on data rate too. With the change in no. of
285 LED, data rate variation against communication distance has been shown in Fig. 12. For
286 a single LED, the maximum estimated data rate is 960 bps at distance 10 cm and 830 bps
Version May 22, 2021 submitted to Sensors 15 of 17

287 is the lowest data rate which is achieved at 60 cm for the three data transmitting LED.
288 The data rate can be augmented significantly using LED array with BER lower than FEC
289 limit.

Figure 14. BER against data rate comparison among our scheme based classifier, receiver using
only CNN classifier, and receiver using only SVM classifier.

290 Fig. 14 illustrates BER against data rate comparison while communication distance
291 was 50 cm for three different cases. At each case, two LED was in data transmitting
292 mood and one was unwanted region. For achieving 960 bps. BER recorded for SVM
293 based receiver is beyond the FEC limit while CNN based receiver shows very closeness
294 of FEC limit. Our scheme based receiver can decode data at 960 bps in BER of 3.5 × 10−3 .
295 This is due to the decoding of data from separated exact data transmitting LED region.

296 6. CONCLUSION
297 In this article, we proposed accurate data transmitting LED detection scheme for
298 reliable OCC communication. In the early receiving mechanism, only CNN has been
299 used to detect LED regions without considering unwanted LED regions or multiple
300 LED regions from multiple users. But, we have designed a scheme by which the
301 camera receiver can differentiate accurate data transmitting LED among unwanted
302 LEDs leveraging SVM classifier additionally for the first time. FSOOK modulation
303 scheme has been used to modulate data and send blinking of LED accordingly. We have
304 evaluated the overall performance elaborately demonstrating the ROC, and other key
305 performance parameters of the classifier, as well as other communication performance
306 features, such as BER, data rate, and ISI. The ROC of the SVM classifier affirms its efficacy
307 to accurately classify accurate LED. The BER analysis exhibits 10 times lower BER at
308 the maximum communication distance (60 cm). Besides, by eye diagram, we have
309 also investigated ISI of the decoded data from the recognized LED region. Moreover,
310 data rate against distance varying No. of LED have been observed for our proposed
311 scheme. In addition, BER against data rate is also discussed to ensure the necessity of our
312 proposed scheme. However, it is perceptible from the overall analysis that our scheme is
313 a very novel approach for reliable OCC communication. Based on our scheme, a new
314 frontier is expected to be outstretched for the OCC to deploy in future communication
315 technology.
316 Author Contributions: The contribution of the authors are as follows: conceptualization, M.H.R.;
317 methodology, M.H.R. and M.S.; software, M.H.R.; validation, M.S. and M.O.A.; formal analysis,
318 M.S. and M.K.H.; writing—original draft preparation, M.H.R.; writing—review and editing,
Version May 22, 2021 submitted to Sensors 16 of 17

319 M.K.H., M.S., and M.O.A.; supervision, Y.M.J.. All authors have read and agreed to the published
320 version of the manuscript.
321 Funding: “This work was supported by Institute for Information & communications Technology
322 Promotion (IITP) grant funded by the Korea government (MSIT) (No.2017-0-00824, Development of
323 Intelligent and Hybrid OCC-LiFi Systems for Next Generation Optical Wireless Communications)”.

References
1. Chow, C. W.; Yeh, C. H.; Liu, Y.; Liu, Y. F. Digital signal processing for light emitting diode based visible light communication.
IEEE Photon. Soc. Newslett. 2012, 26, 9–13.
2. C.-L. Liao,; Ho, C.-L.; Chang, Y.-F.; Wu, C.-H.; Wu, M.-C. High-speed light-emitting diodes emitting at 500 nm with 463-MHz
modulation bandwidth. IEEE Electron Device Lett. 2014, 35, 563–565.
3. Luo, P. et al. Experimental demonstration of RGB LED-based optical camera communications. IEEE Photonics J. 2015, 7, 1–12.
4. Le, N. T.; Hossain, M. A.; Jang, Y. M. A survey of design and implementation for optical camera communication. Signal Process.
Image Commun. 2017, 53, 95–109.
5. Yamazato T. et al. Image-sensor-based visible light communication for automotive applications. IEEE Commun. Mag. 2014, 52,
88–97.
6. Chowdhury, M. Z.; Hasan, M. K.; Shahjalal, M.; Hossan, M. T.; Jang, Y. M. Optical wireless hybrid networks: Trends, opportunities,
challenges, and research directions. IEEE Commun. Surv. Tutor. 2020, 22, 930–966.
7. Takai, I.; Harada, T.; Andoh, M.; Yasutomi, K.; Kagawa, K.; Kawahito, S. Optical vehicle-to-vehicle communication system using
LED transmitter and camera receiver. IEEE Photonics J. 2014, 6, 1–14.
8. Chowdhury, M. Z.; Hossan, M. T.; Shahjalal, M.; Hasan, M. K.; Jang, Y. M. A new 5G eHealth architecture based on optical camera
communication: An Overview, Prospects, and Applications. IEEE Consum. Electron. Mag. 2020, 9, 23–33.
9. Katz, M.; Ahmed, I. Opportunities and challenges for visible light communications in 6G. Proceedings of 2nd 6G Wireless Summit
(6G SUMMIT), Levi, Finland, 2020.
10. Marabissi, D.; Mucchi, L.; Caputo, S.; Nizzi, F.; Pecorella, T.; Fantacci, R.; Nawaz, T.; Seminara, M.; Catani, J. Experimental
Measurements of a Joint 5G-VLC Communication for Future Vehicular Networks. J. Sens. Actuator Netw. 2020, 9,32.
11. Pham, T. L.; Shahjalal, M.; Bui, V.; Jang, Y. M. Deep Learning for Optical Vehicular Communication. IEEE Access 2020, 8 ,
102691–102708.
12. Chen, Q. et al. Spaced color shift keying modulation for camera-based visible light communication system using rolling shutter
effect. Opt. Commun. 2019, 449, 19–23.
13. Roberts, R. D. Undersampled frequency shift ON-OFF keying (UFSOOK) for camera communications (CamCom). Proceedings of
22nd Wireless and Optical Communication Conference, Chongqing, China, 2013.
14. Liu, N.; Cheng, J.; Holzman, J. F. Undersampled differential phase shift on-off keying for optical camera communications. J.
Commun. Netw. 2017, 2, 47–56.
15. Monteiro, E.; Hranilovic, S. Design and implementation of color-shift keying for visible light communications. J. Lightwave Technol.
2014, 32, 2053–2060.
16. Rachim, V. P.; Chung, W.-Y. Multilevel intensity-modulation for rolling shutter-based optical camera communication. IEEE
Photonics Technol. Lett. 2018, 30, 903–906.
17. Islam, A.; Hossan, M. T.; Jang, Y. M. Convolutional neural networkscheme–based optical camera communication system for
intelligent Internet of vehicles. Int. J. Distrib. Sens. Netw. 2018, 14, 1550147718770153.
18. Yu, K.; He, J.; Huang, Z. Decoding scheme based on CNN for mobile optical camera communication. Appl. Optics 2020, 59,
7109–7113.
19. Hasan, M. K.; Chowdhury, M. Z.; Shahjalal, M.; Nguyen, V. T.; Jang, Y. M. Performance analysis and improvement of optical
camera communication. Applied Sciences 2018, 8, 2527.
20. Nagura, T.; Yamazato, T.; Katayama, M.; Yendo, T.; Fujii, T.; Okada, H. Improved decoding methods of visible light communication
system for ITS using LED array and high-speed camera. Proceedings of IEEE 71st Vehicular Technology Conference, Taipei,
Taiwan, 2010.
21. Chi, N.; Jia, J.; Hu, F.; Zhao, Y.; Zou, P. Challenges and prospects of machine learning in visible light communication. Journal of
Communications and Information Networks 2020, 5, 302–309.
22. Yuan, Y. et al. SVM-based detection in visible light communications. Optik 2017, 151, 55–64.
23. Li, J.; Guan, W. The optical barcode detection and recognition method based on visible light communication using machine
learning. Applied Sciences 2018, 8, 2425.
24. Niu, W., Ha, Y.; Chi, N. Support vector machine based machine learning method for GS 8QAM constellation classification in
seamless integrated fiber and visible light communication system. Sci. China Inf. Sci. 2020, 63, 202306.
25. Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of International
Conference on Learning Representation, virtual conference, 2021.
26. Yamga, G. M. M.; Ndjiongue, A. R.; Ouahada, K. Low complexity clipped Frequency Shift Keying (FSK) for Visible Light
Communications. Proceedings of IEEE 7th International Conference on Adaptive Science Technology (ICAST), Accra, Ghana,
2018.
Version May 22, 2021 submitted to Sensors 17 of 17

27. Bazi, Y.; Melgani, F. Toward an optimal SVM classification system for hyperspectral remote sensing images. IEEE Trans. Geosci.
Remote Sens. 2006, 44, 3374–3385.
28. Brereton, R. G.; Lloyd, G. R. Support vector machines for classification and regression. Analyst 2010, 135, 230–267.
29. Chavez-Burbano, P.; Rabadan, J.; Guerra, V.; Perez-Jimenez, R. Flickering-Free Distance-Independent Modulation Scheme for
OCC. Electronics 2021, 10, 1103.
30. Li, P.; Zhao, W. Image fire detection algorithms based on convolutional neural networks. Case Stud. Therm. Eng. 2020, 19, 100625.
31. Pedronette, D. C. G.; Torres, R. Shape retrieval using contour features and distance optimization. Proceedings of the International
Conference on Computer Vision Theory and Applications, Angers, France, 2010.
32. Vincent, L. Morphological transformations of binary images with arbitrary structuring elements. Signal Process. 1991, 22, 3–23.
33. Ashtputre, S.; Sharma, M.; Ramaiya, M. Image segmentation based on moments and color map. Proceedings of International
Conference on Communication Systems and Network Technologies, Gwalior, India, 2013.
34. Douglas, D. H.; Peucker, T. K. Algorithms for the reduction of the number of points required to represent a digitized line or its
caricature. Cartogr. Int. J. Geogr. Inf. Geovisualization 1973, 10, 112–122.