Content uploaded by Mohana ..
Author content
All content in this area was uploaded by Mohana .. on Nov 26, 2022
Content may be subject to copyright.
Anomaly Detection in Videos for Video
Surveillance Applications using Neural Networks
Ruben J Franklin, Mohana, Vidyashree Dabbagol
Department of Electronics & Telecommunication Engineering,
RV College of Engineering® Bengaluru- 560059, Karnataka, India.
Abstract— Security is always a main concern in every domain,
due to a rise in crime rate in the crowded event or suspicious
lonely areas. Abnormal detection and monitoring have major
applications of computer vision to tackle various problems.
Due to growing demand in the protection of safety, security
and personal properties, the needs and deployment of video
surveillance systems can recognize and interpret the scene and
anomaly events play a vital role in intelligence monitoring.
Anomaly detection is a technique used to distinguish various
patterns and identify unusual patterns with a minimal period,
this pattern is called outliers. Surveillance videos can capture a
variety of realistic anomalies. Anomaly detection in video
surveillance involves breaking down the whole process into
three layers, which are video labelers, image processing, and
activity detection. Hence, anomaly detection in videos for video
surveillance application gives assured results in regards to
real-time scenarios. In this paper, we anomaly was detected in
images and videos with an accuracy of 98.5 %.
Keywords – Convolutional neural network, Common objects in
context, Feature pyramids networks, Masked region
convolutional neural network, Visual object classes.
I. INTRODUCTION
Ano maly d etection is the identification of irregular,
unexpected, unpredictable, unusual events or items which
are not considered as a normally occurring event or a regular
item in a pattern or items present in a dataset and thus
different from existing patterns. An anomaly is a pattern that
occurs differently from a set of standard patterns. Therefore,
anomalies depend on the phenomenon of interest. Nowadays
protection for personal and personal property becoming very
important. Video surveillance gives a good role in real-time.
Because of these needs deployment of cameras take place at
every corner, video surveillance system understand the
scene and it automatically detects abnormal activities [1].
The main aspect understands the action then reports the
operator or users automatically when an unexpected event
happens. Video surveillance performs efficiently improve in
the application of safety and security for the management of
personal life and public area [2]. It also develops an
automatic surveillance system to replace human observed
oriented services with a reduction in the workload of an
observer. In the process of anomaly detection, cameras are
used to collect the data of varies events representing the
behavior of anomaly in an environment under surveillance.
The system performance performs feature extraction on the
data collected, process it and after that the resulting features
become varies inputs to the specified algorithm. Mainly
there are three anomaly detection techniques such as
supervised, semi-supervised and unsupervised anomaly
detection.
Fig.1. Block diagram of methodology.
Figure 1 depicts the block diagram of the methodology.
Anamoly objects are detected for images of COCO dataset
and as well as real-time video [3] [4]. Frames of a video
extracted and objects are detected using CNN by extracting
features in the required region of interest [5] [16]. Usually,
the features are measured and compared based on
considered patches such as motion and appearance. The
obtained output in this step is feature representation, which
is a very important aspect of anomaly detection. By using
M-RCNN, three features are obtained such as a class of the
object, bounded box and semantic segmentation [12] [13].
II. PERFORMANCE METRICS USED IN ANOMALY
DETECTION
This section describes some of the performance metrics used
in anomaly detection.
Eyeball evaluation: This the easiest way to judge the
performance of a detector is by considering visual aspects of
results, or anomaly scores. For example, first, make use of
histogram graphs to plot the different distribution of scores.
Then compare two graphs by overlaying histograms in the
same plot, one representing "normal" points and another one
as the "anomaly". In this respect, eyeballing are fantastic
because they are simple to create and interpret, and contain
Proceedings of the Fourth International Conference on Inventive Systems and Control (ICISC 2020)
IEEE Xplore Part Number: CFP20J06-ART; ISBN: 978-1-7281-2813-9
978-1-7281-2813-9/20/$31.00 ©2020 IEEE 632
Authorized licensed use limited to: R V College of Engineering. Downloaded on October 11,2022 at 06:42:07 UTC from IEEE Xplore. Restrictions apply.
all of this information. However, if the data structure is
complex, a lot of data points lie on the same point
interpreting the graphs becomes much harder. Moreover, if
the inspector misses the data and interpretation is different
from the result will be completely different from the actual
results, performance goes below an acceptable level. The
efficiency of performance is very low when compared to the
rest of the different methods [14] [15].
Percentiles: Percentiles describe the distribution of scores
between data points labeled "normal" and "anomaly". This
parameter gives us a well-defined value by painting the
whole picture in a single value but insight aspect or in detail,
results are not well defined. This is similar to “Yes” or
“NO” with what percentage. Percentiles are easily
comprehensible metrics, with the advantage of being able to
indicate high false positive or false negative rates. There are
indeed interesting conclusions to draw if percentile
calculations are applied to the scores of data points from
both “normal” and “abnormal” classes.
Usual Companion (AUC): AUC is a popular metric that is
designed to represent the ranking capability of scorers. To
what extent can we say that "unusual" data points get a
higher score than "usual" ones?" It is the well-known Area
under the ROC Curve, or ROC-AUC — simply AUC fro m
now on. It practically means that if we randomly pick a
"usual" and "unusual" user activity from our test set, the
"unusual" one will have a higher score with a probability
that equals to AUC. This is essential since we want to make
sure that abnormal events are highlighted with an anomaly
score larger than that of normal events. The AUC value of
an anomaly scorer's performance ranges from 0 to 1. An
AUC of 1 indicates a flawless anomaly scorer that perfectly
separates the two classes ("usual" and “unusual” events). If
the AUC is below 1, that means that some "usual" events
have larger scores than "unusual" ones do. Although not
being perfect, an AUC of 0.7–0.9 is considered acceptable,
on the other hand, if the AUC is 0.5, the algorithm is as
effective a classified as random guessing AUC fails to
encapsulate this aspect of performance of an anomaly
scorer.
An extension to AUC, AUC (probabilistic AUC) method aims
to address the above issue. AUC stands for probabilistic AUC.
Consider a score of 100 as the best possible anomaly score that
an "abnormal" event can obtain and of score 0 as the optimal
anomaly score for every "normal" event. AUC gives
probabilistic scores. This is because AUC is much stricter in its
assessment of those test cases that yielded too many highly
false negatives or false positives.
RP distance: Percentiles in performance evaluation, the
simplicity in interpreting and the amount of power it provides
to express in terms of percentiles inspired our performance
metric that we call reverse percentile distance. This concept is
based on a reinterpretation of this quality as of how far the
scores obtained by the “normal” and “abnormal” are from each
other. With the new metric, we attempt to capture this distance
numerically. This main assumption is that the greater this
distance is, the accurate is the scores of each to its respective
scores The RP distance at a given percentile p, or RP@p, is
defined as:
RP@p = perc (d abnormal, 100 — p) — perc (d normal, p)
Where: d abnormal refers to the score distribution of data
points labeled “unusual” d normal (d, p) refers to the pth
percentile of a score distribution d [6] [7] [8].
III. DESIGN AND IMPLEMENTATION OF
ANOMALY DETECTION
Anomaly detection is used to distinguish anomaly
patterns that do not conform to a set of expected
behavior. The anomaly detection algorithm is designed
by breaking into three separate yet correlated processes
they are the convolutional neural network, mask
recurrent convolutional neural network and spatial
awareness using semantic segmentation. The design
process is implemented in the anaconda environment
A.Convolution Neural Network (CNN)
Fig.2.Layers of convo6lutional neural network
In anomaly detection, a Convolutional Neural Network
(CNN or ConvNet) in figure 2, Layers of CNN is a class
of Deep Neural Networks (DNN) to analyze images or
videos. It is a class of multi-layer perceptron, refers to a
Fully Connected Network (FCN) in which all neurons of
one layer is connected to other neurons in the next layer
[21] [26] [27][28].
B. Mask RCNN
Figure 3 shows mask RCNN structure, it is designed and
built to solve instant segmentation problems in various
commuter vision applications. It is capable to separate
various objects in video or image.
Fig.3. Mask RCNN Structure
Proceedings of the Fourth International Conference on Inventive Systems and Control (ICISC 2020)
IEEE Xplore Part Number: CFP20J06-ART; ISBN: 978-1-7281-2813-9
978-1-7281-2813-9/20/$31.00 ©2020 IEEE 633
Authorized licensed use limited to: R V College of Engineering. Downloaded on October 11,2022 at 06:42:07 UTC from IEEE Xplore. Restrictions apply.
It generates region proposals based on objects in an image
and capable to predict the various class of objects by
drawing a bounding box around an object to generate masks
in pixel level. This stage acts as a backbone for Feature
Pyramids Networks (FPN) to detect objects at different
scales.
C. Semantic Segmentation
Fig.4. Mask R-CNN Semantic Segmentation
Figure 4. Shows semantic segmentation of Mask R-CNN. It
is one of the high-level tasks that helps to paint a complete
picture of understanding. The importance of understanding
frame by frame is a core computer vision constrains which
is highlighted by the fact that with a high increasing number
of applications. Some of those applications are self-driving
cars, human and computer interaction with each other,
virtual-reality and so on. With many semantic segmentation
changes are being solved using deep learning architectures,
mostly convolutional neural networks, which yields much
better results surpass other approaches by far better
efficiency and accuracy [20].
D. Anomaly detection
Fig.5. Block Diagram of Anomaly Detection
Fig.6. Flow diagram of M-RCNN
Figure 5 shows the anomaly detection block diagram.
Features are extracted from a video. The frame per second is
set to 20 frames for better results in a minimum amount of
time. The features are extracted frame by frame, each video
frame is divided into foreground and background region.
Most of the time's background region information is
unchanged and the foreground region contains moving
objects [17] [22] [23] [25]. Features of both the region are
extracted. Moving objects are detected and given as input to
mask – RCNN. Mask R-CNN used there is nothing but
faster R-CNN. Faster R-CNN has two outputs produced for
each different object, one is a class label and another a
bounding-box. Additionally, a third output is added which is
an object mask. This additional mask is different from the
other two outputs; extraction of information is much finer to
construct a spatial layout of the object. Figure 6 shows the
flow diagram of M-RCNN. The Pattern detection phase
used in this model to identify patterns in the video sequence.
Finally, the desired operations for the abnormal activities in
the university campus and traffic flow monitoring is
performed. The system can also be programmed to take
some decisions upon pattern detection.
Some of the assumptions and constraints made during
implementation are:
x The observed object brightness at any given time
should be constant.
x The video is captured from a single stationary
source
x The captured video should contain RGB frame
structure
x Nearby objects in the image should move in a
similar manner meaning smooth change in velocity.
x This camera unit should be connected to the
computer and the video capture should be
available.
IV. SIMULATION RESULTS AND ANALYSIS
Anomaly detection is the identification of various data
points, events, and observations that deviate from the
dataset's normal behavioral patterns. Anomaly detection also
referred to as outlier detection, is used to find critical
incidents, such as a technical glitch, fraud, or logistical
obstacle, or potential opportunities.
Proceedings of the Fourth International Conference on Inventive Systems and Control (ICISC 2020)
IEEE Xplore Part Number: CFP20J06-ART; ISBN: 978-1-7281-2813-9
978-1-7281-2813-9/20/$31.00 ©2020 IEEE 634
Authorized licensed use limited to: R V College of Engineering. Downloaded on October 11,2022 at 06:42:07 UTC from IEEE Xplore. Restrictions apply.
A. Detection of Anomaly Object in the image
Fig.7. Detection of anomaly objects in an image (a) input image (b) output
image.
Fig.8. Output parameters
Figure 7 shows the detection of an anamoly object (knife) in
an image. Fig. 7(a) is an input image, Fig. (b) Shows
detection of knife with an accuracy of 98.5 % and detection
of a person in the masked area with an accuracy of 96.5%.
Also, the corresponding output parameters as shown in
figure 8. Feature Pyramids Networks (FPN) are a basic
component in recognition systems for detecting objects at
different scales, which is used in this part of the algorithm.
It describes the frame format, size, type, and pixel
dimension.
B. Detection of Anomaly Object in video
Case -1: For home security related video
Fig.9. Detection of a person in a video. (a) Input image (b) Output image
Fig.10. Detection of a person in a video plotted frame by frame
Figure 9 shows the detection of humans in a video, and the
corresponding frame of video as shown in figure 10. Since
this model is capable of detecting the presence of a knife in
an image it is has been enhanced to detect the presence of a
human in a video. The corresponding frame of the video is
shown in fig.9a, which is the input frame of the video, the
shown frame number in this video is 57 and the
corresponding fig.9b. This is the output frame of the video,
the shown frame number in this video is 56 where a person
is detected and is masked with a color to locate the person in
that particular frame. This is a real video surveillance
camera footage model that can detect any presence of
human activity around the ho me. In this scenario, if owners
of the house were not present in the home. If a motion
detector were installed to detect any presence, it would
generate alarms even if wild birds and animals enter the
area. This algorithm eliminates all the drawbacks mentioned
and can detect selective anomalies. Figure 10, detection of a
person in a video plotted frame by frame.
Case -2: Detection for surveillance video
Fig.11. Detection of a person in an input video frame
Fig.12. Detection of a person in an output video frame
Proceedings of the Fourth International Conference on Inventive Systems and Control (ICISC 2020)
IEEE Xplore Part Number: CFP20J06-ART; ISBN: 978-1-7281-2813-9
978-1-7281-2813-9/20/$31.00 ©2020 IEEE 635
Authorized licensed use limited to: R V College of Engineering. Downloaded on October 11,2022 at 06:42:07 UTC from IEEE Xplore. Restrictions apply.
Fig.13. Detection of anomaly event in video frames
A particular anomaly is been set to detect if any anomaly is
detected in and around a fence to monitor if any person is
"jumping" the fence. This is achieved by extracting the
masked binary value of the person detected in the video
frame by frame and using the mathematical model. Which
will generate a summed location values of the person in
each frame will be stored to find the initial and current
position of the person. This way a person trying to jump or
jumping the fence location in that frame for the normal
position is, higher this will be the direct indicator for
anomaly detection. Figure 11 shows a person climbing
across the gate and figure 12 shows the detection of the
person with the probability values, finally figure 13 shows
the anomaly value plotted for every frame from start until
the end of the video. In fig.13, the percentage of anomaly
detected for frames, we can conclude that an anomaly event
occurs from frame 150 to 225, which is a person-climbing
gate. Also from frame 0 to 88, we see the anomaly value is
a constant 0.4% this is due to the person is on a flight of
stairs which is relatively higher position the normal.
V. CHALLENGES IN ANOMALY DETECTION
Evolving “anomaly” – Different trade of, no universal
model for an anomaly, timestamp difference, and change
over time research in anomaly detection has been carried out
for over several years in a wide area of disciplines. Even
then still open research, the key reason is that the definition
of anomaly is contextual, this is because what may be an
anomaly in E-commerce may be different in the context of
networking or computer vision. Furthermore, a different
application needs different trade-offs to detect an anomaly.
Moreover, what anomaly is today may not be the same
tomorrow and the existing algorithm may not be useful later
because the norm might change over time.
Lack of Data: Due to the fact that data is necessary for the
model to be trained or to be tested. Lack of labeled data in
"class imbalance" adds more time factor to implement a
well-defined working anomaly detector. Increasing in WSN
and IOT without doughty generates large amounts of data
which makes it even more difficult to label them and
differentiates between "normal" and "abnormal" data set.
Due to all the facts stated above makes the development of a
new anomaly detection algorithm is challenging.
Continuous learning and Training: Dynamic nature of the
data streams calls for anomaly detection based on
continuous learning. This way the constantly changing
anomaly definition can be kept track and system updates can
be performed from time to time.
Ve r a c i t y : To identify objects labels and their corresponding
location in video frame and image requires a computational
power in both hardware and software and a very large
amount of computing time. The resolution of the interested
detection has to improve localization accuracy on small
objects under partia l occlusio ns. Rea l-time analysis of data,
centralized anomaly detection system is not realistic and
feasible.
VI. ANOMALY DETECTION ISSUES
Number of Attributes that are to be used to define
anomalies: Consider a scenario when there is one attribute
for an object, for some data values are anomalous but other
values of other attributes are normal. Whether we have to
decide based on those values of attributes or not. For
example, consider people who are 6 feet tall and people who
weigh 90kgs. But it is uncommon to have people who are
6feet tall and 90kgs weight. So, there must be a definition of
anomaly specifying how multiple attributes are used to
identify whether an object is an anomaly or not.
Limits to consider when to detect anomaly: To decide
whether an object is anomalous or not is based on a divisive
decision in a binary format for some techniques. But it is
important to know that, there are some extreme anomalies
and very less in degree. This is possible by giving a score to
each anomaly based on the degree of it being anomalous.
This assessment is called anomaly or outlier score.
Masking and Swamping: In some anomaly detection
techniques anomalies are identified one at a time. In such
techniques, there is a chance of missing out anomalies
because of masking. Masking means the presence of several
anomalies masks all other objects. Whereas identifying
multiple anomalies at once can the involvement of
swamping. Swamping is where normal objects are
considered as anomalies.
Precision, recall false positive rate: If class labels are
available to identify anomalies and normal data, then it is
usually normal that the class of anomalies is smaller than
that of a normal class. In such cases measures like precision,
recall and false-positive rate are more important than the
accuracy [9][10][11].
Efficiency: There are some significant differences between
anomaly detection techniques. It is very important to choose
the right technique. Based on the time complexities
choosing an anomaly detection technique is important as
finding an anomaly, it is also important to reduce the cost
too.
VII. CONCLUSIONS
Detections of an anomaly in video surveillance are a
challenging aspect due to the fact of different factors
affecting results like video noise, outliers, and resolution. In
this paper segmentation and classification model using
thresholds, classification model and graph-based
segmentation algorithms are used to obtain results with an
accuracy of 98.5%. By utilizing the features from normal
and anomalous surveillance videos as well, to avoid long
training time; a general model of anomaly detection using
deep learning yields best results with very minimal time. To
Proceedings of the Fourth International Conference on Inventive Systems and Control (ICISC 2020)
IEEE Xplore Part Number: CFP20J06-ART; ISBN: 978-1-7281-2813-9
978-1-7281-2813-9/20/$31.00 ©2020 IEEE 636
Authorized licensed use limited to: R V College of Engineering. Downloaded on October 11,2022 at 06:42:07 UTC from IEEE Xplore. Restrictions apply.
explore the importance of locality in anomaly detection,
experimental results show that, locating anomaly in the
frame helps to achieve good results over a long surveillance
video and this method is very robust. Further, it can be
implemented for larger datasets by training using GPUs and
high-end FPGA kits [18] [19] [24].
REFERENCES
[1] Henry et al., "Anomaly Detection in Videos Recorded b y Dron es in a
Surveillance Context", International Carnahan Conference on
Security Technology (ICCST), Madrid, 2017.
[2] Mohammad Farhadi et al., “AAD: Adaptive Anomaly Detection
through traffic Surveillance videos”, arXiv:1808.10044, 2018.
[3] Abhiraj Biswas et.al.,“Classification of Objects in Video Records
using Neural Network Framework,” International conference on
Smart Systems and Inventive Technology (ICSSIT-2018).
[4] Arka Prava Jana et.al.,“YOLO based Detection and Classification of
Objects in video records,” International Conference On Recent
Trends In Electronics Information Communication
Technology,(RTEICT) 2018.
[5] Nehashree M R et.al., “Simulation and Performance Analysis of
Feature Extraction and Matching Algorithms for Image Processing
Applications” International Conference on Intelligent Sustainable
Systems (ICISS-2019).
[6] Y. Wang, et al., “Traffic Camera Anomaly Detection”,2014 22nd
International Conference on Pattern Recognition, Stockholm, 2014,
pp. 4642-4647.
[7] Jefferson Ryan Medel et al.,“Anomaly Detection in Video Using
Predictive Convolutional Long Short-Term Memory Networks”,
arXiv:1612.00390, 2016.
[8] Biao Yang et al., “Anomaly Detection in Moving Crowds through
Spatiotemporal Autoencoding and Additional Attention” Advances in
Multimedia,2018.
[9] G. Chandan et.al., “Real Time Object Detection and Tracking Using
Deep Learning and OpenCV”, International Conference on Inventive
Research in Computing Applications (ICIRCA), 2018.
[10] Abhiraj Biswas et.al.,“Survey on Edge Computing - Key Technology
in Retail Industry” International Conference on Intelligent Computing
and Control Systems (ICICCS 2019).
[11] Krishna C V et.al., “A review of Artificial Intelligence methods for
data science and data analytics: Applications and Research
Challenges” International Conference on I-SMAC (IoT in social,
mobile, Analytics and cloud), 2018.
[12] A. Raghunandan et. al., "Object Detection Algorithms for Video
Surveillance Applications," International Conference on
Communication and Signal Processing (ICCSP), 2018.
[13] A. Mangawati et al., “Object Tracking Algorithms for Video
Surveillance Applications,” 2018 International Conference on
Communication and Signal Processing (ICCSP), 2018.
[14] X. Ma, T. Lu, F. Xu and F. Su, “Anomaly detection with spatio-
temporal context using depth images”, International Conference on
Pattern Recognition (ICPR2012), Tsukuba, 2012, pp. 2590-2593.
[15] Samir Bouindour et al., “An On-Line and Adaptive Method for
Detecting Abnormal Events in Videos Using Spatio-Temporal
ConvNet” Applied Science, 2019.
[16] Manjunath Jogin et.al., “Feature extraction using Convolution Neural
Networks (CNN) and Deep Learning” International Conference On
Recent Trends In Electronics Information Communication
Technology, 2018.
[17] Meghana R K et.al., “Background Modelling techniques for
foreground detection and Tracking using Gaussian Mixture model”
International Conference on Computing Methodologies and
Communication (ICCMC 2019).
[18] S. K. Mankani et al., “Real-time implementation of object detection
and tracking on DSP for video surveillance
applications,” International Conference on Recent Trends in
Electronics, Information & Communication Technology (RTEICT),
2016.
[19] S. Sajjanar et. al., “Implementation of real time moving object
detection and tracking on FPGA for video surveillance applications,"
2016 IEEE Distributed Computing, VLSI, Electrical Circuits and
Robotics (DISCOVER) 2016.
[20] W. Sultani et al., “Real-World Anomaly Detection in Surveillance
Videos”, 2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition, Salt Lake City, UT, 2018, pp. 6479-6488.
[21] Federico Landi et al., “Anomaly Locality in Video Surveillance”,
arXiv:1901.10364, 2019.
[22] Mohana et.al., “Elegant and efficient algorithms for real time object
detection, counting and classification for video surveillance
applications from single fixed camera” International Conference on
Circuits, Controls, Communications and Computing (I4C),2016,
[23] Mohana et.al., “Simulation of Object Detection Algorithms for Video
Survillance Applications”, 2nd International Conference on I-SMAC
(IoT in Social, Mobile, Analytics and Cloud),2018.
[24] V. P. Korakoppa et.al., “An ar ea efficient FP GA imp lementation of
moving object detection and face detection using adaptive threshold
method,” International Conference on Recent Trends in Electronics,
Information & Communication Technology (RTEICT), 2017.
[25] H. V. Ravish Aradhya et.al., “Real time objects detection and
positioning in multiple regions using single fixed camera view for
video surveillance applications," International Conference on
Electrical, Electronics, Signals, Communication and Optimization
(EESCO), 2015, pp. 1-6.
[26] Bashar, Abul. "SURVEY ON EVOLVING DEEP LEARNING
NEURAL NETWORK ARCHITECTURES." Journal of Artificial
Intelligence 1, no. 02 (2019): 73-82.
[27] Kong, Lingchao, Ademola Ikusan, Rui Dai, and Jingyi Zhu. “Blind
Image Quality Predi ction for Object Detection”, In 2019 IEEE
Conference on Multimedia Information Processing and Retrieval
(MIPR), pp. 216-221. IEEE, 2019.
[28] Moosbauer, Sebastian, Daniel Konig, Jens Jakel, and Michael
Teutsch. “A Benchmark for Deep Learning Based Object Detection in
Maritime Environments”, In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition Workshops, 2019.
Proceedings of the Fourth International Conference on Inventive Systems and Control (ICISC 2020)
IEEE Xplore Part Number: CFP20J06-ART; ISBN: 978-1-7281-2813-9
978-1-7281-2813-9/20/$31.00 ©2020 IEEE 637
Authorized licensed use limited to: R V College of Engineering. Downloaded on October 11,2022 at 06:42:07 UTC from IEEE Xplore. Restrictions apply.
Publication Link:
https://ieeexplore.ieee.org/document/9171212
Cite This:
R. J. Franklin, Mohana and V. Dabbagol, “Anomaly Detection in Videos for Video
Surveillance Applications using Neural Networks,” 2020 Fourth International Conference
on Inventive Systems and Control (ICISC), 2020, pp. 632-637, Doi:
10.1109/ICISC47916.2020.9171212.
