See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/370233434

A Research Survey on Real-Time Intelligent Traffic System

Article · April 2023

DOI: 10.46338/ijetae0423_05

CITATIONS READS
0 91

6 authors, including:

Deepak T. Mane Prashant Babarao Kumbharkar

Vishwakarma Institute of Technology Rajarshi Shahu College of Engineering
49 PUBLICATIONS   247 CITATIONS    10 PUBLICATIONS   23 CITATIONS   

SEE PROFILE SEE PROFILE

Nilesh J Uke
Trinity Academy of Engineering
46 PUBLICATIONS   381 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Development of Innovative Digital Technologies for construction & Architecture View project

Ensemble Learning for Plant Classification View project

All content following this page was uploaded by Deepak T. Mane on 10 May 2023.

The user has requested enhancement of the downloaded file.

 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
Manuscript Received: 20 February 2023, Received in Revised form: 27 March 2023, Accepted: 02 April 2023 DOI: 10.46338/ijetae0423_05

A Research Survey on Real-Time Intelligent Traffic System

Deepak T. Mane1, Prashant Kumbharkar2, Nirupama Earan3, Komal Patil4, Sakshi Bonde5, Nilesh Uke6
1
Vishwakarma Institute of Technology, Pune-411037, Maharashtra, India
2,3,4,5
JSPM’s Rajarshi Shahu College of Engineering, Pune-411033, Maharashtra, India
6
Trinity Academy of Engineering, Kondhwa, Pune-411048, Maharashtra, India
Abstract– Nowadays, it is crucial to continuously monitor The extraction of crucial traffic information such as
and provide real-time analysis to reduce traffic-related congestion level and lane occupancy is made possible by an
accidents and practices such as vehicle overloading on the accurate vehicle count. In our present day and age, the
roads daily. As a result, we reviewed the literature's numerous usage of information to give insights and meaning to the
methods and techniques for vehicle detection, recognition,
data has increased significantly. One of these ways is
identification, speed estimation, and license plate recognition.
In this analysis, we examined 42 articles published in the last machine learning, where the data is generalized from
ten years, from 2012 to 2022. Based on our research, we found numerous sets of instances. Artificial Neural Networks
that the Deep CNN is the optimum method for vehicle (ANNs), which attempt to replicate their biological
categorization. The motivation of this review is that none of counterparts, are one method employed. The CNN
the aforementioned models are combined into a single model, (Convolutional Neural Network) is one of the methods
so we present a comprehensive list of all these models that used in it, and it mimics computer vision to recognize,
may be helpful to anyone conducting the study in this area. detect, and categorize the objects or entities in the given
Therefore, after reviewing the chosen research publications, field of view. By applying N layer processing to the data,
we propose 20+ datasets that might be used in the field for
the techniques land with the generalization of samples in
more research. We also discovered 15+ different ML models
used to detect and identify vehicles. Finally, we observed that this case. CNN also offers a variety of ways to consider,
combining machine learning and AI (Artificial Intelligence) to one of which is YOLO. This strategy is ideal for a variety
create intelligent traffic control systems is a promising of different applications, including real-time vehicle
research area. recognition, because of its high accuracy and capacity to
outperform other approaches.
Keywords—Intelligent Traffic System, Vehicle Detection, The most underrated issue in the modern world is traffic.
Recognition and Identification, Automatic License Plate The more automobiles on the road, the more congestion
Recognition, Traffic density, Deep Learning, Convolutional there is, which results in increase of fuel consumption,
Neural Network, Machine Learning,.
unnecessary energy and time waste, air pollution, chronic
stress, and other physiological issues. As a result, we learn
I. INTRODUCTION
that increased traffic offenses are the main reason for
The total number of registered automobiles in India was accidents on the road. Road, flyover, bridge, and underpass
over 295 million, in the year 2019. With an upward construction as well as the development of alternative mass
increase in the number of vehicles on the road, with that the public transportation systems are insufficient to ease traffic
traffic difficulties have increased as well. Accidents congestion, which is getting harder to control every day.
involving roads and traffic frequently go unnoticed and Therefore, it is imperative to implement a cutting-edge and
unreported. Illegal traffic operations, such as the transit of intelligent system to address these problems. A tool for
overloaded vehicles carrying commodities, the jumping of managing traffic is provided by the Intelligent Traffic
traffic signals, vehicle over speeding, and so on, must be Management System, an AI-based application combined
curtailed. Calculating vehicle density in traffic images is a with cameras mounted at traffic intersections is used to
difficult study topic because it involves dealing directly detect: Accidents, illegal activities related to traffic law
with hostile yet realistic road conditions such as violation, vehicle density to manage green light timing to
uncontrolled illuminations, cast shadows, and optical save wasted time, and all of this with good accuracy under
occlusion. Nonetheless, being able to count and resolve different weather conditions.
vehicles effectively in such settings has a significant impact
on traffic surveillance.

36
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
For reducing traffic congestion in urban areas, Automatic vehicle identification and detection is a key
controlling traffic lights at road intersections and junctions component of ITS, and it remains a primary priority area
can be implemented as the key strategy. Increasing traffic for many applications; it also includes law and traffic
throughput, improving traffic network capacity, and management enforcement. License plates are very
reducing road deadlock are all benefits of optimizing signal important when it comes to visible labels for an
timing plans at intersections. As a result, traffic signal automobile's identification. ALPR is a technology that
control has become a popular area of study in intelligent extracts the necessary license plate information
transportation and advanced traffic management, with the automatically. It helps with custom checkpoints, nonstop
goal of planning signal timing at intersections in a scientific vehicular toll collecting, regulating access to restricted
and reasonable manner. By allowing a huge amount of areas, ticket-free parking charge collection, and rule
vehicle data to be processed, AI technology provides enforcement by various law enforcement organizations
support to manage and control the traffic[23]. Figure 1, [27].
shows the impact and influence of the ITS on various A comprehensive summary of all the referenced papers
factors. is provided in the related work part of this publication. The
literature review sheds light on the paper's main idea, the
various strategies and procedures employed the accuracy
and performance measures, and the research's future scope
and problems. In this article, the architecture of CNN,
YOLO, and Deep learning algorithms is briefly explored.
The pros and cons of various algorithms used in various
articles are also discussed. This document informs the
reader about current traffic systems and the improvements
and alterations that can be made to make them more
efficient. This work may be valuable in future research in
this domain. The requirement to maintain, document, and
assess the state of the transportation system is growing
along with the number of vehicles. In order to increase
traffic efficiency, maintain the flow, and protect people
from non-rule followers, it is important to analyze the cars
and the road conditions. This analysis may provide several
essential areas for making modifications. This paper
Figure 1: The Influence of ITS
provides a summary of the many methods used by the
Applications of Intelligent Traffic System (ITS): The vehicle detecting system. We shall examine several
pressures of an expanding urban population are causing strategies employed by various authors and their outcomes
city roadways to deteriorate. Smart traffic management is in this study[4],
one of those applications of intelligent transportation
systems that are very important. The following are some of II. LITERATURE REVIEW
the other applications: 2.1 Methodology
● Real-time parking management Figure 2 depicts the methodology we used to design and
● Electronic toll collection carry out the survey. First, we discovered the answers to
● Emergency vehicle notification systems our questions, such as why this survey is necessary, how
● Automated road speed enforcement beneficial it is, and what the best course of action should
● Speed alerts be. We downloaded research and survey articles on
● RFID in freight transportation pertinent subjects covering the years 2012 through 2022.
● Variable speed limits
● Dynamic traffic light sequence
● Collision avoidance systems

37
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
2.1.1 Selection Criteria For Survey Papers: OBJ-2: To discuss the various methods and approaches
We collected a total of 65 papers and articles and employed in the publications.
shortlisted 42 papers after screening them through and OBJ-3: To illustrate the degree of accuracy provided for
considering only those papers that are relevant to our each method presented in the paper.
subject. We followed the below measures to configure and
2.2.2 Observation Table:
filter the selected 42 papers:
Table 1, displays the papers that we have reviewed. The
1. We chose the papers that were most pertinent to our
table includes the paper's publication year, the datasets
subject and gave explanations of the expected
discussed, the methodology used, performance metrics, and
approaches and algorithms.
future scopes or challenges faced. Some of the often
2. Simple-to-comprehend papers were given higher
discussed real-time datasets include the KITTI dataset,
priority.
CDnet2014, VEDAI dataset, COCO, and BDD100K. Many
3. We looked at the articles that provided reliable
publications have also used their own real-time feeds. The
performance measures.
most common techniques and methodologies discussed are
2.2 Discussion of Survey: CNN, YOLO, and other DL and DRL techniques. The
2.2.1 Objectives (OBJ) of the Review: performance metrics discussed in the paper focus on
accuracy, precision, recall, F-score, etc.
OBJ-1: To provide a variety of datasets that were used in
The following table reviews a total of fourty papers in an
this domain.
orderly and understandable fashion.

Figure 2: Methodology Used for Survey

38
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
TABLE I:
LITERATURE ON DIFFERENT METHODOLOGIES, PERFORMANCE AND FUTURE SCOPE / CHALLENGES

Sr Year, Dataset Method Used Performance Metrics Challenges/Future Scope

No. Used

1. 2021 Hybrid model: Recognition rate using ALPR: In the future the problems of illumination due to
Dataset: 1) Optical Character 92.2% overexposed and underexposed regions could be
UoJ dataset Recognition solved with help of
2) R-CNN HDR images.
3) YOLOv5

2. 2021 1) RBFNN Accuracy : 95% They might start by using the fusion process to
Dataset: 2) ANN enhance the quality of the vehicle photographs.
Real-time Finally, to identify the characters on the license
dataset plate, they might employ an artificial neural
network.

3. 2021 1) CNN Accuracy : 95.31%. They plan to work the model in low light and
Dataset: adverse conditions, i.e. at night as sometimes it
Real-time fails to detect the blue and black color of the
dataset vehicle and also to make the other attributes of
the vehicle accurately classified.

4. 2021 1) YOLOv4 Recall: 0.25 A new dataset can be used with many videos,
Dataset: Precision: 0.53 and also different kinds of vehicles can be used
Real-time in various weather conditions. They can work on
dataset vehicle collision systems.

5. 2021 1) YOLOv2 Average absolute error (speed) The system can be lightened by integrating it
Dataset: 2) OCR : 3.2 km/hr into an embedded device, reducing construction
Real-time Average relative error: 9.94% costs.
input

6. 2020 1) YOLOv3 Estimated error (speed The subject of vehicle speed estimation using
Dataset: resolutions) : 9.2%. Estimated moving UAVs will be investigated more in the
Real-time error (speed UAV Height) : future.
input 8.8%.

7. 2020 1) Lucy Richardson The majority of the assessment In the future, they plan to improve the fine
Dataset: algorithm index algorithms arrive at tuning to get high level efficiency and high level
Real-time findings that are similar to accuracy.
input those reached by the naked
sight. The greater the value of
the assessment index, the
clearer the image.

8. 2014 1) Haar-like feature Accuracy : 81.4%. Their algorithm should be able to solve the
to detect vehicles majority of the problems and thus they can use it
2) AdaBoost to detect the back of a moving camera's road
3) Cascade classifier vehicle. They are also planning to develop and
improve their algorithms.

39
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)

9. 2016 1) CNN Mean Average Precision: The system struggles when dealing with near and
Dataset: 2) Faster-RCNN 80.5%. small objects in groupings.
PASCAL
VUC 2007

10. 2014 OCR The results of the systems are Instead of restricting to developing only two
Dataset: analyzed in real-time with data systems for the detection of suspicious vehicles,
Real-time acquired from the Department they plan to do more research on this.
input of Land Transportation in
order to identify suspicious
cars and generate reports on
them.

11. 2017 1) Gaussian Mixture Average error rate : 0.3 The results have a low error rate that can be
Dataset: Model (GMM) improved further also it does not capture the
Real-time 2) OCR technique long-distance image.
dataset

12. 2013 1) Regression Model Accuracy: 92% In the future, their techniques can be used to
Dataset: replace radar guns. They are planning to develop
Real-time new techniques for speed detection.
input

13. 2016 1) Thresholding Vehicle recognition rate: 85%. When a vehicle has high speed, the algorithm
Dataset: algorithm fails. So to correct this, dimensions of the exit
Real-time 2) Convex hull zone based on vehicle speed can be chosen.
input algorithm and hole
filling algorithm

14. 2012 1) GMM Detection accuracy: 0.9639, The challenges include:

Dataset: 2) Kalvin Filter Classification accuracy: SVM fails in detecting background and
Real-time Model 0.9469 foreground when both have similar colors.
dataset 3) SVM

15. 2018 1) Transfer Learning Mean Average Precision: 0.74 They intend to look at smoothing approaches for
Dataset: expected vehicle speeds to improve model
Provided by performance.
NVIDIA

16. 2017 1) Iterative A comparison study of low Their method performed better than state-of-the-
Reweighted Least light color image restoration art technologies at reconstructing clear color
Square (IRLS) between their method and images and they plan to improve their method
2) Conjugate regular method was done. furthermore.
Gradient (CG)

17. 2013 1) Image Processing They compared different Future scope includes developing new models
Dataset: Toolbox and Signal deblurring applications and which are reliable and fast.
Real-time Processing Toolbox softwares and the obtained
input 2) Point Spread results were studied.
Function (PSF)

40
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)

18. 2020 1) ANN model- The NN model attained an They didn‘t consider the challenges caused by
Dataset: prediction of results. accuracy of 0.9274. weather conditions. A neural network model can
Real-time 2) PCA- train the further be developed into a multi-agent network
dataset neural network. and test its efficiency.

19. 2019 1) YOLOv3. Detection accuracy: 0.8346. The future scope includes improving vision
Dataset: based vehicle identification and contributing into
KITTI the development of traffic infrastructure.
benchmark
dataset,
Traffic-Sign
Dataset

20. 2019 1) ANN As a result, the suggested It can be used to assist the government in
Dataset: 2) MLP approach correctly classifies developing policies for the building of a flyover
Real-time the traffic condition for or alternate routes to improve traffic
dataset various types of road sections. management.

21. 2021 1) RNN The plots show that in the The system can be expanded in the future to
Dataset: 2) DRL LSTM network, the Deep regulate traffic signals on real-time road
Spatio- 3) Gated Recurrent Reinforcement Learning networks..
temporal data Unit (GRU) (DRL) agent develops a
superior action selection
strategy.

22. 2021 1) YOLOv4 target The technology can effectively The future work includes
detection algorithm tackle increasingly critical 1) Improving detection rate, and accuracy
2) DeepSORT traffic problems, as well as 2) Implement the proposed system in real time
algorithm assist traffic management traffic management system for a variety of
departments in better applications.
managing urban traffic and
promoting smart city
development.

23. 2021 1) Multiagent Reduced average delay, The future scope includes: a reward function
Dataset: algorithms smaller queue size, lower that can be given for coordination of many
Historical 2) Evolutionary waiting time, higher average crossovers in traffic.
traffic flow algorithms speed and higher throughput
data 3) Q - learning were all used to validate the
extended RL models and
algorithms' performance.

24. 2017 1) HOG A confusion matrix is used to It has yet to develop a very accurate and rapid
2) SVM present and describe such procedure that is resistant to all situations.
3) SIFT measurements. We also give Finally, we will discuss some open difficulties
4) Nearest Neighbors speed-related measures that and intriguing future research areas in the field
(K-means) can be used to assess the cost of AVC.
of training, retraining, and
testing a system.

41
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)

25. 2021 1) CNN - Models with high precision and reliability that
Dataset: 2) SVM operate well in real-world situations have yet to
KITTI, DLR- 3) Recurrent be developed. Inter-class similarity and intra-
MVDA, Neural Networks class variation, speed and model size, occlusions
Stanford Cars, (RNNs) and appearance distortion, evaluation under
CompCars, various environmental conditions, and multi-
BoxCars modal sensing were some of the difficulties
presented.

26. 2019 1) AdaBoost Detection rate: 0.925 The detection rate and false positive rate can be
Dataset: classifier Error detection rate: 0.05 improved in future scope.
Real-time 2) Haar-like features
dataset 3) Computing
methodologies

27. 2019 1) CNN Accuracy: 0.983 While detecting the license plate, the model ends
Dataset: 2) ALPR and OCR CPU time: 0.343 up extracting the rotating plate as well as other
Real-time unwanted elements which causes noisy data. So
dataset by reducing the noise in data, the performance of
the proposed model can be enhanced.

28. 2016 1) Deep CNN Processing time per frame: The study may include video-based information
Dataset: Classifier 200 ms. such as optical flow and temporal data in order
Real-time 5-layer CNN model accuracy: to produce better and fewer suitable zones for LP
input 98.7% solutions.

29. 2019 1) Sequential SMO accuracy (LF): 79-82% The dataset only contains a small number of
Dataset: Minimal images.
Real-time Optimization (SMO) SMO accuracy (EHF): 82.36% Future development on more images is possible.
dataset 2) Simple Logistic
(SL) SL accuracy: 81.79%

30. 2016 1) ULEA (Unwanted Success rate: 95% Character segmentation in the license plate must
Line Elimination segment the characters accurately; otherwise, the
Algorithm) character cannot be recognized appropriately.
2) ANN

31. 2015 WSN (Wireless Vehicle detection accuracy: They intend to run extensive field tests on
Sensor Networks) 0.99 congested roadways to gauge the algorithms'
Estimate velocity: 0.94 efficiency in terms of accuracy and energy
usage. The ability to classify cars based on
vehicle size using the sampling rate, breadth, and
height of the detected pulse is also something
they wish to include.

32. 2021 1) ViBe algorithm Precision: 0.8534 Although the optical flow technique may more
Dataset: 2) GMM Recall: 0.7451 efficiently extract moving objects, real-time
CDnet2014 detection is difficult due to its high computation
requirements, making it unsuitable for traffic
scenarios.

42
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)

33. 2018 1) Faster RCNN Precision: 0.907 The accuracy of vehicle recognition in aerial
Dataset: 2) Feature Pyramid F1 score: 0.878 photographs will be improved in their future
VEDAI Network (FPN) Recall: 0.864 work, and they believe that altering the
annotation process and taking orientation into
account may be a good idea.

34. 2020 1) YOLOv3 mAP score of 0.703 at 41.2 Because the vehicles in aerial imagery are often
Dataset: 2) K-means FPS. tiny, their characteristics blend seamlessly with
VEDAI clustering algorithm the background. The surrounding environment is
also quite complex. As a result, it is difficult to
distinguish vehicles in aerial views.

35. 2021 RoIAlign and License plate detection The performance of LP detection, identification,
Dataset: MaskOps accuracy: 95.1% and its practical application deployment skills
Chinese City License plate recognition must still be improved under tough
Parking accuracy: 96.1%. circumstances.
Dataset

36. 2022 1) Neural Networks Frame loss incidence: 0.24 Since the cameras are frequently fitted with
Dataset: 2) Deep Hold time: 0.155 heterogeneous and turbulent networks, it is
Mahimahi Reinforcement extremely difficult to play back traffic
Learning monitoring films in real time smoothly and with
good quality.

37. 2022 1) Faster R-CNN mAP: 0.728 Unbalanced data utilization. Even though a
Dataset: 2) YOLOv3 particular category has a good detection effect,
BDD100K the network tends to focus more on its traits,
which lessens the detection effect of other
categories.

38. 2022 YOLO models mAP: 74-80% They intend to assess the effects of Internet
Dataset: Average accuracy: 94.4% network speed, from 4G to 5G, on the
COCO-2017, application of DL models for road safety in
V-TLD, future work. To improve the accuracy they also
KITTI advise testing various backbones and necks
using the YOLOv5 design.

39. 2022 1) R-CNN Average IoU: 0.9325 We plan to use generative neural networks in the
Dataset: 2) YOLO models mAP_0.5: 78.5% future to enhance the data, create more accurate
DAWN artificial scenes that represent challenging
driving situations, and then make the data set
accessible.

40. 2022 1) YOLO Average Accuracy : 86.8% The YOLO DL framework is investigated for the
Dataset: 2) RES-YOLO Training time: 229s challenges of vehicle target recognition,
BDD100K statistical errors in huge data sets, and the
sophisticated automatic judgment of vehicle
target self-position.

43
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
2.2.3 Major Techniques Used: This presents both a statistical difficulty and a
A. Deep Learning (DL): computational challenge, which are two separate
challenges. The statistical difficulty with generalization
DL is a subdivision of machine learning that allows
stems from the fact that the number of configurations we
computers to learn by doing, much like humans learn and
might want to distinguish might increase exponentially
unlearn. DL is an important part of self-driving or
with the increase of dimensions of interest, and this fast
automatic automobiles because it allows them to identify
exceeds the number of instances one can conceivably have.
stop signs and distinguish between pedestrians, traffic
Because many learning or using trained models algorithms
signals or lampposts. It plays an important role in consumer
involve intractable computations that become exponentially
electronics such as hands-free speakers, phones,
more difficult as the number of dimensions increases, high-
televisions, tablets as it enables voice control. DL has
dimensional distributions pose a computational challenge.
received a lot of attention recently, and rightfully so. It's
Figure 3, shows the architecture of the DL model.
about achieving goals that were previously unreachable.
Multilayer neural networks and a huge amount of B. Convolutional Neural Network (CNN):
labeled data are used to train models. DL achieves higher A ConvNet/CNN is a DL method that takes an image as
recognition accuracy than ever before. It has progressed to input and then applies learnable weights to the various
the point where it now outperforms humans in some tasks, features or objects in the image to enable it to recognize
such as categorizing objects in photographs. Large volumes different groups of images. The amount of pre-processing
of labeled data are required for deep learning. For example, required in CNN is comparatively less than in other
the creation of self-driving cars involves millions of photos classification methods. With enough training, CNN can
and thousands of hours of video. DL takes a lot of learn the various filters and properties. The convolution
computational power. The parallel design of high- function extracts dissimilar information from a picture
performance GPUs is ideal for deep learning. When used in input. The first convolutional layer retrieves tiny features
conjunction with clusters or cloud computing, this enables such as lines, edges, and corners. Other deep layers extract
development teams to cut the time needed for DL network the image's most prominent and apparent features.
training from weeks to just a few hours or less. The large Convolutional operations begin from the top-left corner of
dimension of the random variables being modeled is one of the matrix and work your way down to the bottom-right
the main reasons why unsupervised learning is difficult. corner.

Figure 3: Deep Learning Architecture

44
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
The following are the components of a traditional Take pedestrian detection as an example. An illustration
Convolutional Neural Network: of a moving impediment is a pedestrian. To assess if a
1. Convolutional Layer. collision is likely, a CNN must be able to recognize the
2. After each convolutional layer, perform an activation pedestrian and extrapolate their current speed. Given that it
operation. must classify an object in addition to delivering the four
3. Pooling layer, particularly the Max Pooling layer, as coordinates of its bounding box, a CNN for object
well as others as needed. identification is a little more complicated than a
4. Finally, the layer is fully connected. classification model.
When a person logs in to a website, they are asked to
CNN for Autonomous Vehicles classify photos like traffic lights, autos, fire hydrants, etc.
Many businesses, use CNNs as the computer vision part using Google's Captcha system. This is a practical method
of a self-driving car. A self-driving car's computer vision for gathering tagged training photos for applications like
system must be able to localize objects, avoid obstacles, Google StreetView and self-driving cars.
and plan routes. Figure 4, shows the architecture of the CNN model.

Figure 4: Convolutional Neural Network Architecture

C. YOLO Algorithm: This algorithm is used to find and identify various

YOLO uses a neural network to identify objects in real- objects in a picture (in real-time). The YOLO approach
time. This algorithm finds popularity because of its speed uses CNN to recognize objects in real-time. As the name
and accuracy. YOLO has a wide variety of applications implies, we learn that this technique is crucial for object
such as identifying traffic signals, parking meters, animals, detection since it uses a method that only requires one
and pedestrians. The image is divided into N grids via the forward propagation via a neural network. This suggests
YOLO method, with an identical-sized SxS dimensional that the complete image is anticipated using a single
sector in each grid. ‗You Only Look Once‘ is referred to as algorithm run. CNN is used to forecast multiple bounding
YOLO. boxes and class probabilities at the same time. The real-
In YOLO, object detection is carried out as a regression time data analysis capabilities of YOLO make it suited for
problem, and the class probabilities are provided by the several applications, including real-time vehicle
detected photographs. identification.

45
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
It is very accurate and outperforms many approaches commonly used versions of YOLO like YOLOv3,
[42]. Figure 5, shows the architecture of the YOLO mode YOLOv4, YOLOv5
and Table 2, depicts a comparison study of the most

Figure 5: YOLO Architecture

TABLE II:
COMPARISON STUDY OF DIFFERENT VERSIONS OF YOLO

Parameter YOLOv3 YOLOv4 YOLOv5

Backbone Feature Darknet-53 CSP Darknet-53 CSP Darknet-53

Extractor

Neural Network Type Fully convolution Fully convolution Fully convolution

Loss Function Binary cross entropy Binary cross entropy Binary cross entropy and
Logits loss function

Neck FPN SSP and PANet PANet

III. OBSERVATIONS AND OUTCOMES OF SURVEY The new algorithm YOLOv4, which is designed to
The important takeaways from this paper are: For recognize different kinds of vehicles on the road with the
alphanumeric recognition, the clustering-based multi- help of real-time CCTV footage, has produced some
language license plates repudiate the CNN model. When promising results. Using the YOLOv3 algorithm, we
the license plate number is present in bulk and there is learned about different developing methods for vehicle
more than one different character on the license plate, it speed and track prediction based on aerial recordings.
becomes a little difficult to identify the vehicle. Following Following that, a survey of vehicle attribute recognition
that, we learn about vehicle color classification in video algorithms is conducted for both coarse and fine-grained
and image. attributes.

46
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
We performed an entire survey in the ANPR system, The observed images with bare eyes are distinct from the
with better results than both the real-time and simulated acquired image system, which is what we have learned in
algorithms. The images of various environments are used to this work. The focus of the following research phase is on
test the quality of the image assessment indication on the creating a flexible, density-based traffic control system that
paper. We got to know that Haar-like technology is will offer real-time analysis. Then a vehicle object dataset
employed as it has a significant edge over most other having high definition was provided from the viewpoint of
features in terms of calculation speed. To catch the unseen a surveillance camera, and it was applied to an object
object, the multi-frame filter is used. The 9 convolutional detection and tracking algorithm for scenes from highway
layers in the network architecture are used for real-time surveillance footage. Figure 6, shows the different varieties
vehicle recognition and tracking. Two subsystems are of the datasets discussed in the papers.
included in a suggested framework for the automatic real-
time detection of suspicious vehicles: an ALPR system and
a vehicle model recognition system. Vehicle speed
detection is a valuable technique in vehicle tracking and
processing applications. Similarly, different methodologies
and algorithms are discussed which demonstrates the
components of each technique and gain a better
understanding of their similarities, differences, and
obstacles. Then there's the study, which is a mix of several
fields in the realm of vehicle identification and
classification, such as background subtraction and
classification. Then, to mitigate the detrimental impact of
camera vibration, a hybrid algorithm technique is
developed, which includes a background subtraction
module and foreground extraction. Based on the detect-
then-track paradigm, the model is introduced for vehicle
tracking in traffic video.
The state-of-the-art technique in vehicle speed Figure 6: Names of the Datasets Use
estimation, vehicle detection, and object tracking is
presented in the studies.
3.1. Strengths and Weaknesses of different techniques:
TABLE 3:
STRENGTHS AND WEAKNESSES OF DIFFERENT TECHNIQUES STUDIED

NAME OF STRENGTH WEAKNESS

TECHNIQUE

YOLOv5 It swiftly trains the model for object detection and It is almost as accurate as YOLOv4 on the same task.
recognition. Only Pytorch and a few lightweight python
modules are required for YOLOv5.

Artificial Neural We can train ANNs since they learn from their The ANNs are hardware-dependent since they require
Network experiences and make decisions. It is capable of processors that support parallel processing.
completing multiple tasks at the same time.

Conventional It automatically detects any relevant traits without the The most difficult part is classifying images in various
Neural Network need of human intervention. positions. When CNN is shown an image with noise, it
recognizes it as an entirely distinct image.

47
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)

Optical Character OCR information may be read with a great degree of OCR text works efficiently only with printed text; it
Recognition precision. It is also processed quickly. does not work well with handwritten text.

Gaussian Filter It is very smooth for removing Gaussian noise. It takes time and reduces details.

Regression Regression techniques result in an algebraic equation that If the input data contains errors, regression models will
Algorithm is simple to implement and use as a forecasting tool. not work effectively (that is poor-quality data). If data
Regression models are easy to understand since they are preprocessing is improperly done, the regression model's
founded on fundamental statistical concepts like output could be erroneous, necessitating the removal of
correlation and least-square error. missing values, duplicate data, outliers.

YOLOv3 Because of its simplified architecture, YOLOv3 operates The Yolov3 technique isn't perfect for using specialized
faster than R-CNN. models because huge dataset models are difficult to
come by.

Radial Basis They are not very complex and have a design that is Training is quick, but classification is very slow as
Function Neural understandable and has good generalization, a high compared to the multimedia perceptron.
Network tolerance for input noise, and the capacity to learn online.
(RBFNN) These are very ideal for designing flexible control
systems.

Support Vector SVM provides good results when the classes are already It is not very suitable for larger data sets.
Machine (SVM) classified, and it is more successful in high-dimensional
environments.

YOLOv4 YOLOv4 is a refinement of the YOLOv3 algorithm. It YOLOv5 is about 180% faster than YOLOv4.
has a 10 %t increase in mean average precision (mAP)
and a 12 % increase in the number of frames per second.

Histograms of Because HoG characteristics are strong at characterizing Image rotation has a big impact on HOG. As a result,
Oriented Gradient object shapes, they're useful for detecting pedestrians. HOG is a poor choice for classifying textures or objects
(HOG) that are frequently identified as rotating images.

Recurrent Neural RNN has an added advantage over ANN, a collection of Problems with gradients exploding and vanishing, it is
Network (RNN) records can be modeled (i.e. time collection) presuming quite difficult to train an RNN.
that each pattern is dependent on the previous one.

Scale Invariant The SIFT descriptor is a classic technique, it is more It is both mathematically and computationally
Feature precise than other adjectives and it is scale and rotation challenging. SIFT is based on the Gradient Histogram,
Transform invariant. that is, each Pixel in the patch's gradients must be
(SIFT) computed, and these computations take time.

AdaBoost Adaboost is less prone to overfitting since the input The problem with Adaboost is that it requires a high-
Classifier parameters are not jointly optimized. Adaboost can help quality dataset and it is necessary to eliminate noisy data
enhance the accuracy of weak classifiers. and outliers.

Haar-like A Haar-like feature has a significant edge over most other The drawbacks of Haar cascades include their propensity
Features features in terms of calculation speed. A Haar-like feature for false-positive detections, the need for parameter
of any size can be estimated in constant time thanks to the changing when used for inference and the fact that they
usage of integral pictures. aren't as accurate as more "modern" techniques used
today.

48
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)

Figure 7: Usage Of Different Techniques And Algorithms

From the Figure 7, we come to know that CNN, OCR, There has been a lot of work done in intelligent traffic
Gaussian filter algorithms are most popularly used in this systems using AI and ML, but most work is discrete and
domain. contains only some of the features mentioned above. The
review done in this paper will help provide all the relevant
IV. LIMITATIONS O F SURVEY information to someone who wants to work in the field of
Our survey report can save the new researcher time and intelligent traffic systems. Our study focuses on articles
effort by eliminating the need to read pointless study that were published between 2012 and 2022. We reviewed
publications. As a result, the researchers will be able to cut over 42 papers from reputable journals and conferences for
out errors from the beginning of their research and will this study. Our assessment considers all of the research
have a clear path for carrying out a solid systematic survey from the following viewpoints: Abstract, Literature
on the subject of ITS. However, we only included a limited Review, Major Techniques Used, Observations, Results
number of studies that we felt were crucial in our analysis. and Accuracy, and Future Scope/Challenges faced. This
This review has limitations in terms of the time period survey report provides insight into machine learning for
taken into account and the number of research papers intelligent traffic management systems. After reviewing all
examined. But we think that relying on more sources might these research papers, we plan to establish an intelligent
have improved this assessment. traffic management system. Our suggested system would
utilize live video feeds from CCTV cameras at traffic
intersections for real-time traffic management. For the
V. CONCLUSION
future, we aim to establish a traffic signal controller based
Our comprehensive literature review covered various on computer vision that can adapt to the current traffic
approaches and techniques for vehicle attribute recognition, scenario.
vehicle makes and model recognition, license plate
recognition, and vehicle type recognition. We also REFERENCES
discussed various techniques for vehicle tracking and speed [1] Syed Talha Abid Ali, Abdul Hakeem Usama, Ishtiaq Rasool Khan,
estimates. This survey taught us how to measure traffic Muhammad Murtaza Khan, Asif Siddiq, ―Mobile Registration
Number Plate Recognition using Artificial Intelligence‖,
density over a defined region, use picture de-blurring International Conference on Image Processing (ICIP) 2021, DOI:
algorithms to obtain an exact image, and deploy 10.1109/ICIP42928.2021.9506699
technologies like computer vision for real-time analysis and [2] Farheen Ali, Himanshu Rathor, Wasim Akram, ―License Plate
monitoring of traffic intersection scenes. Recognition System‖, 2021 International Conference on Advance
Computing and Innovative Technologies in Engineering (ICACITE),
DOI: 10.1109/ICACITE51222.2021.9404706

49
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
[3] Abdullah Tariq, Muhammad Zeeshan Khan, Muhammad Usman [16] Hiraki Yamashita, Daisuke Sugimura and Takayuki Hamamoto,
Ghani Khan, ―Real Time Vehicle Detection and Colour Recognition ―Rgb-Nir Imaging With Exposure Bracketing For Joint Denoising
using tuned Features of Faster-RCNN‖, 2021 1st International And Deblurring Of Low-Light Color Images‖, 2017 IEEE
Conference on Artificial Intelligence and Data Analytics (CAIDA), International Conference on Acoustics, Speech and Signal
DOI: 10.1109/CAIDA51941.2021.9425106 Processing (ICASSP), DOI: 10.1109/ ICASSP. 2017. 7953319
[4] Vibhanshu Singh Sindhu, ―Vehicle Identification from Traffic Video [17] Al-Ameen, Zohair & Sulong, Ghazali & Md Johar, Md Gapar.
Surveillance Using YOLOv4‖, 2021 5th International Conference on (2013). Computer Forensics and Image Deblurring: An Inclusive
Intelligent Computing and Control Systems (ICICCS), DOI: Investigation. International Journal of Modern Education and
10.1109/ICICCS51141.2021.9432144 Computer Science, vol. 5, pp 42-48. 10.5815/ijmecs.2013.11.06.
[5] Ying-Cheng Wu, Chih-Chun Hsu, Jenq-Shiou Leu, Senior Member, [18] W.A.C.J.K. Chandrasekara, R.M.K.T Rathnayaka, L.L.G
IEEE, Jui-Tang Wang, ―Design and Implementation of Vehicle Chathuranga, ―A Real-Time Density-Based Traffic Signal Control
Speed Estimation Using Road Marking-based Perspective System‖, 2020 5th International Conference on Information
Transformation‖, 2021 IEEE 93rd Vehicular Technology Technology Research (ICITR), DOI: 10.1109/ICITR51448.
Conference (VTC2021-Spring), DOI: 10.1109/VTC2021- 2020.9310906
Spring51267.2021.9448813 [19] Song, H., Liang, H., Li, H. et al. Vision-based vehicle detection and
[6] DONGYANG ZHAO, YUQING CHEN, SHUANGHE YU, counting system using deep learning in highway scenes. Eur.
―Tracking and Speed Estimation of Ground Vehicles Using Aerial- Transp. Res. Rev., vol. 11, issue 51 (2019), DOI:
view Videos‖, 2020 5th International Conference on Automation, https://doi.org/10.1186/s12544-019-0390-4
Control and Robotics Engineering (CACRE), DOI: [20] Md Ashifuddin Mondal and Zeenat Rehena. 2019 ―Intelligent
10.1109/CACRE50138.2020.9230274 Traffic Congestion Classification System using Artificial Neural
[7] Boxin Zhao, Xiaolong Chen, Guoqiang Feng, Xiaolin Zhao, Jun Network‖, In Companion Proceedings of the 2019 World Wide Web
Jiang, ―An Improved LR Algorithm for Image Deblurring‖, 2020 Conference (WWW‘19 Companion), May 13–17, 2019, San
Chinese Control And Decision Conference (CCDC), DOI: Francisco, CA, USA. ACM, New York, NY, USA, 7 pages, DOI:
10.1109/CCDC49329.2020.9164762 https://doi.org/10.1145/3308560.3317053
[8] Qing Xu, Feng Gao, Ke Shi, Guoyan Xu, ―Real-time rear of vehicle [21] Ananya Paul and Sulata Mitra. 2021. Management of Traffic Signals
detection from a moving camera‖, The 26th Chinese Control and using Deep Reinforcement Learning in Bidirectional Recurrent
Decision Conference (2014 CCDC), DOI: Neural Network in ITS. In 2021 5th International Conference on
10.1109/CCDC.2014.6852989 Intelligent Systems, Metaheuristics & Swarm Intelligence (ISMSI
[9] Xiao-Feng Gu, Zi-Wei Chen, Ting-Song Ma, Fan Li, Long Yan, 2021), DOI: https://doi.org/10.1145/ 3461598.3461608
―Real-Time vehicle detection and tracking using deep neural [22] Dingjie Zhang et al. 2021. Application of Computer Vision in
networks‖, 2016 13th International Computer Conference on Intelligent Transportation Management System. In 2021 10th
Wavelet Active Media Technology and Information Processing International Conference on Internet Computing for Science and
(ICCWA TIP), DOI: 10.1109/ICCWAMTIP.2016.8079830 Engineering (ICICSE 2021), DOI:
[10] Anucha Tungkastan, Nipat Jongsawat, Wichian Premchaiswadi, ―A https://doi.org/10.1145/3485314.3485325
proposed framework for automated real-time detection of suspicious [23] Li Kuang, Jianbo Zheng, Kemu Li, and Honghao Gao. 2021.
vehicles‖, 2014 Twelfth International Conference on ICT and ―Intelligent Traffic Signal Control Based on Reinforcement Learning
Knowledge Engineering, DOI: 10.1109/ICTKE.2014.7001541 with State Reduction for Smart Cities‖, ACM Trans. Internet
[11] Praveen Dhulavvagol, Abhilash Desai , ―Vehicle Tracking and Technology, vol. 21, issue 4, Article 102 (July 2021), 24 pages.
Speed Estimation of Moving Vehicles for Traffic Surveillance DOI: https://doi.org/10.1145/3418682
Applications‖, 2017 International Conference on Current Trends in [24] Azzedine Boukerche et. al. 2017 ―Automated Vehicle Detection and
Computer, Electrical, Electronics and Communication (CTCEEC), Classification: Models, Methods, and Techniques‖, ACM Comput.
DOI: 10.1109/CTCEEC.2017.8455043 Surv. vol. 50, issue 5, Article 62 (October 2017), 39 pages. DOI:
[12] Mayank Garg & Sahil Goel, ―Real-Time License Plate Recognition https://doi.org/10.1145/3107614
and Speed Estimation from Video Sequences ‖, ITSI Transactions on [25] Azzedine Boukerche and Xiren Ma. 2021 ―Vision-based
Electrical and Electronics Engineering (ITSI-TEEE) 2013 Autonomous Vehicle Recognition: A New Challenge for Deep
[13] Shashank Ramesh, Vikas Upadhyaya, ―Vehicle classification and Learning-based Systems‖, ACM Comput. Surv., Vol.54, 4, Article
lane categorization‖, 2016 International Conference on Computer, 84 (April 2021), 37 pages. DOI: https://doi.org/10.1145/3447866
Control, Informatics and its Applications (IC3INA), DOI: [26] Zhou Yan, Yuan Deming, and Zhou Jun. 2019. Research on Vehicle
10.1109/IC3INA.2016.7863016 Identification Method Based on Computer Vision. In Proceedings of
[14] Zezhi Chen, Tim Ellis, Sergio A Velastin SMIEEE, ―Vehicle the 2019 International Conference on Data Mining and Machine
Detection, Tracking and Classification in Urban Traffic‖, 2012 15th Learning (ICDMML 2019), Association for Computing Machinery,
International IEEE Conference on Intelligent Transportation New York, NY, USA, 140–145.
Systems, DOI: 10.1109/ITSC.2012.6338852 https://doi.org/10.1145/3335656.3335700
[15] Shuai Hua, Manika Kapoor, David C. Anastasiu, ―Vehicle Tracking [27] Imran Shafiq, Ahmad and Boubakeur Boufama. 2020 ―Automatic
and Speed Estimation from Traffic Videos‖, 2018 IEEE/CVF Vehicle Identification Through Visual Features‖, In Proceedings of
Conference on Computer Vision and Pattern Recognition the 17th International Conference on Advances in Mobile
Workshops, DOI: 10.1109/CVPRW.2018.00028 Computing & Multimedia (MoMM2019). Association for
Computing Machinery, New York, NY, USA, pp. 185–194.
https://doi.org/10.1145/3365921.3365938

50
 International Journal of Emerging Technology and Advanced Engineering
Website: www.ijetae.com (E-ISSN 2250-2459, Scopus Indexed, ISO 9001:2008 Certified Journal, Volume 13, Issue 04, April 2023)
[28] Vishal Jain et. al. 2016 ―Deep automatic license plate recognition [35] Haijie Wang, Yanjie Ke, and Ge Yang. 2021 ―RT-LPDRnet: A Real-
system‖, In Proceedings of the Tenth Indian Conference on Time License Plate Detection and Recognition Network‖, In 2021
Computer Vision, Graphics and Image Processing (ICVGIP '16), the 5th International Conference on Innovation in Artificial
Association for Computing Machinery, New York, NY, USA, Intelligence (ICIAI 2021), March 05–08, 2021, Xia men, China.
Article 6, pp. 1–8, https://doi.org/10.1145/ 3009977 .3010052 ACM, New York, NY, USA, 6 pages. DOI:
[29] Agnes S. Suguitan and Lucille N. Dacaymat. 2019 ―Vehicle Image https://doi.org/10.1145/3461353.3461391
Classification Using Data Mining Techniques‖, In Proceedings of [36] A Zhang, X., Zhao, Y., Min, G., Miao, W., Huang, H., & Ma, Z.
the 2nd International Conference on Computer Science and Software (2022) ―Intelligent Video Ingestion for Real-time Traffic
Engineering (CSSE '19), Association for Computing Machinery, Monitoring‖, ACM Transactions on Sensor Networks (TOSN), vol.
New York, NY, USA, pp. 13–17. 18, pp. 1 - 13.
https://doi.org/10.1145/3339363.3339366 [37] Yidan Chen, Zhenjin Li, ―An Effective Approach of Vehicle
[30] Sangita Kumari, D. K. Gupta, and Raj Mohan Singh. 201 ― A Novel Detection Using Deep Learning‖, Computational Intelligence and
Methodology for Vehicle Number Plate Recognition using Artificial Neuroscience, vol. 2022, Article ID 2019257, 9 pages, 2022.
Neural Network‖, Proceedings of the Third International https://doi.org/10.1155/2022/2019257
Symposium on Computer Vision and the Internet (VisionNet'16)., [38] Neupane, B.; Horanont, T.; Aryal, J. Real-Time Vehicle
Association for Computing Machinery, New York, NY, USA, pp. Classification and Tracking Using a Transfer Learning-Improved
110–114. https://doi.org/10.1145/2983402.2983432 Deep Learning Network. Sensors 2022, 22, 3813. DOI:
[31] Fawad Ahmad and Sahibzada Ali Mahmud. 2015 ―Poster: Accurate https://doi.org/10.3390/s22103813
Vehicle Detection in Intelligent Transportation Systems (ITS) using [39] Humayun, M.; Ashfaq, F.; Jhanjhi, N.Z.; Alsadun, M.K. Traffic
Wireless Magnetic Sensors‖, In Proceedings of the 13th ACM Management: Multi-Scale Vehicle Detection in Varying Weather
Conference on Embedded Networked Sensor Systems (SenSys '15), Conditions Using YOLOv4 and Spatial Pyramid Pooling Network,
Association for Computing Machinery, New York, NY, USA, pp. Electronics 2022, 11, 2748. DOI: https://doi.org/10.3390/
385–386. https://doi.org/10.1145/2809695.2817876 electronics11172748
[32] Dongming Xu and Guogang Han. 2021 ―Application of Improved [40] Li, Y.; Wang, J.; Huang, J.; Li, Y. Research on Deep Learning
ViBe Algorithm in Vehicle Detection‖, In 2021 4th International Automatic Vehicle Recognition Algorithm Based on RES-YOLO
Conference on Artificial Intelligence and Pattern Recognition (AIPR Model. Sensors 2022, vol. 22, 3783. DOI: https://doi.org/10.3390
2021), September 24–26, 2021, Xiamen, China, ACM, New York, /s22103783
NY, USA, 6 pages. DOI: https://doi.org/10.1145/3488933.3489019
[41] Avinash Bhujbal, Deepak Mane 2019. ―A Survey On Deep Learning
[33] Yinjuan Gu, Bin Wang, and Bin Xu. 2018 ―A FPN-Based Approaches For Vehicle And Number Plate Detection‖, International
Framework for Vehicle Detection in Aerial Images‖, In Proceedings Journal of Scientific & Technology Research, Volume 8, Issue 12,
of the 2018 the 2nd International Conference on Video and Image pp. 1378–1383.
Processing (ICVIP 2018), Association for Computing Machinery,
[42] Deepak T. Mane et. al. 2022. ―Detection of Anomaly using Machine
New York, NY, USA, pp. 60–64.
Learning: A Comprehensive Survey‖, International Journal of
https://doi.org/10.1145/3301506.3301531
Emerging Technology and Advanced Engineering. Volume 12, Issue
[34] Moran Ju, Haibo Luo, and Zhongbo Wang. 2020 ―An improved 11, pp. 134-152, DOI: 10.46338/ijetae1122_15.
YOLO V3 for small vehicles detection in aerial images‖, In 2020 3rd
International Conference on Algorithms, Computing and Artificial
Intelligence (ACAI 2020), December 24–26, 2020, Sanya, China,
ACM, New York, NY, USA, 5 pages. DOI:
https://doi.org/10.1145/3446132.3446188

51

View publication stats