2022 International Conference on Computer Communication and Informatics (ICCCI ), Jan.

25 – 27, 2022, Coimbatore, INDIA

Artificial Intelligence (AI) Enabled Vehicle

Detection and counting Using Deep Learning
Mohana, Rahul Kejriwal, Ritika H J, Arpit Arora
Electronics and Telecommunication Engineering, RV College of Engineering®
Bengaluru, Karnataka, India.

Abstract- Urban traffic management is the system for assessing part center around the steady recordings procured by the fixed
and controlling density of vehicles on road, to avoid traffic camera under straightforward conditions. By and large, they
congestion. Vehicle counting is an important module it aids in remove the moving piece of the scene dependent on foundation
traffic prediction and urban infrastructure planning. In this deduction. Consequently, mass examination and following of
2022 International Conference on Computer Communication and Informatics (ICCCI) | 978-1-6654-8035-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICCCI54379.2022.9740929

paper, various Artificial Intelligence tools and their applications

the bouncing box is conveyed to test the vehicles. These
for vehicle detection and counting have been analysed and
elaborated. For vehicle detection, used deep learning based strategies regularly neglect to deal with the impediment,
YOLOv3 algorithm a highly efficient and accurate algorithm for shadows, insecure (shaking) camera, off base situation of the
detecting different classes of vehicles. The role of edge-devices, camera and confounded foundation. For a number of traffic
which locally process the data that they are fed with. Application related issues, vehicle counting can provide a meaningful
of Big Data and different AI tools like GNN, LSTM, and QCNN solution. The number of vehicles on a particular road can
was described for vehicle detection and counting. AI is used to provide insights on the traffic flow. Government guidelines can
predict and detect traffic flow. Big data helps in maintaining an be issued, pertaining to this traffic flow, for the construction or
efficient traffic management system. Both combined, contribute widening of roads. Also, steps like one-way roads can be taken
in developing a systematic traffic management system. The
for some parts of the day.
objective behind using LSTM networks is to prevent the long
term dependency problem. QCNN is used for efficient model
training. Organizations implementing vehicle counting and II. LITERATURE SURVEY
detection focus more on the concise sensors and in house
development of efficient algorithms in order to have extensive Asha C S et al. [1] implemented vehicle counting using YOLO
research and development. and correlation filter algorithms, which can detect, track and
count vehicles. Their work can be applied to any kind of road
Keywords— YOLO, Deep SORT, Traffic Management
video and can be used to detect vehicle classes. Performance
System, Edge Computing, QCNN, LSTM.
parameters of precision, recall and accuracy were analysed.
I. INTRODUCTION However, the processing time is high and only a single lane has
been considered.
Vishal Mandal et al. [2] analysed various object detection
In urban areas, where the quantity of vehicles ceaselessly algorithms, to obtain the most accurate vehicle count. The
increments quicker than the available traffic foundation to help analysis was done for different weather conditions and low
them, traffic is a troublesome issue to manage. This issue has light settings. However, issues such as identity switches and
various impacts, including monetary factors, car crashes, multiple detections of the same vehicle were faced.
extension in nursery emanations, amount of time taken, as well Marek Landowski et al. [3] deployed mathematical models for
as wellbeing harms. Traffic authorities play an important role knowledge collection from traffic density, depending upon
in curbing traffic and its ill effects. To conquer such issues, location and time of the day. This can be applied to reduce
traffic executives' framework collect information via congestion. However, the analysis could not be done for the
heterogeneous sources, adventure this information to classify entire city traffic, but only some parts of it.
risks that may conceivably corrupt the roadway system and Yingjie Xia et al. [4] focused on using the virtual loop
then discuss effective control measures. Constant recordings algorithm for traffic management. In addition, expectation-
represent a few difficulties for an automated traffic maximization and Gaussian mixture models have also been
investigation. The troubles looked by the automated used, to improve the accuracy. The method is efficient in
framework incorporate the presence of shadows, impediment videos with occlusion, and works with an accuracy of 98%.
of vehicles, ecological varieties like downpour, mist, dust, and However, future enhancements in operations are needed.
so forth, which typically corrupt the exhibition. Daming Li et al. [5] analysed perception of images for traffic
Notwithstanding many committed endeavors, a precise management. By the proposed method, a huge amount of
technique for vehicle tallying under complex climate is still a traffic data can be collected and computed. This can help in
long way from being accomplished. There are a few techniques more efficient development and construction. However, this is
to perform expressway traffic checking dependent on just a traffic simulation system.
recordings. Still, they are confined to a specific street situation
and can't be summed up. The current strategies for the most

978-1-6654-8035-2/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on June 05,2023 at 17:48:02 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Computer Communication and Informatics (ICCCI ), Jan. 25 – 27, 2022, Coimbatore, INDIA

Antonito Galletta et al. [6] investigated the usage of edge- delay in data processing. This project also focuses on giving
devices to accomplish different tasks at a time. The paradigm priority to emergency vehicles like ambulances. The paper
used is Function as a Service (FaaS). The outcome of the model analyses the green traffic light duration in a dynamic manner.
deploys versatile, adaptable and scalable systems to respond to Also, it tries to effectively service emergency vehicles.
dynamic and changing requirements of cities. The precision of Mun Chon Ho et al. [15] discuss a Proactive Travel time based
the model is 95%, but it is limited to analysis using four Pheromone Routing system. Using the duration of traffic and
cameras only. duration of travel, the system predicts the future traffic levels.
Ahmed Gomaa et al. [7] explained an effective algorithm for Each automobile in this system lays out a pheromone on
traffic management. The algorithm uses CNN as the backbone various road segments. Another algorithm is proposed to
of the algorithm. As per the proposed model, tracking and decrease the network stress. This algorithm effectively helps in
detection are used to generate feature points, for every reducing the average time of travelling and decreases
particular frame interval. It provides a dynamic real-time consumption of fuel.
analysis. Nejdet Dogru et al. [16] studied about detecting and preventing
Johan Barthelemy et al. [8] described the usage of a smart road car accidents. Majority of which includes recognizing the
visual sensor for the system. Computer vision and deep neural objects which lead to car accidents or an analysis on the same.
network algorithms have been used. The proposed work In this, a system that helps in detection of accidents is studied
suggests a framework which is designed to take and store data .This system will gather all the information necessary from the
from various sensors. The privacy of people is a consideration neighbouring automobiles and will process that given
here. However, for the present implementation, simplicity of information using the tools for machine learning to be able to
prototyping is considered. detect any possible accidents. These machine learning
Erik Sonnleitner et al. [9] described an approach towards algorithms have successfully been able to distinguish between
developing a general system for classifying and counting the normal and abnormal behaviours.
vehicles on webcam pictures. The system uses ML and CV to Nyothiri Aunga et al. [17] simulated the incorporation of
count and differentiate various types of vehicles. The potency Internet of things (IoT) in modern cities would resolve many
in the proposed system under real-world situations could not traffic related issues like decreasing the congestion and
be proven. The proposed system is capable of processing preventing accidents. Various charging policies for electrical
simple roads only, a more general approach is not found yet. vehicles and others have been initiated in recent times. In this
Kapileswar Nellore et al. [10] suggested an algorithm that paper, the authors propose a modern and effective pricing of
considers the space between the emergency truck and the traffic congestion and vehicle charging systems for the IoT in
hospital. The system uses count of vehicles and time related a modern city.
alerts in the processing. The distance is measured by various Muhammad Shafiq et al. [18] curated this paper which
techniques, to increase the accuracy of the computation. discusses the concepts of smart cities, data mining and machine
Allan M de Souza et al. [11] gave an extensive study of Traffic leaning, and other aspects required for implementation.
Management Systems (TMS), focussing on the obstacles and Various techniques for the classification of Sustainable smart
the future of the current systems available today and propose cities network were given out with more information.
some suggestions that would improve the efficiency of the Nourdine Aliane et al. [19] introduced an in-vehicle framework
traffic management. A complete overview of the modern TMSs for criminal traffic offense alarms and the executives. The
were described including the three key phases: information structure gear involves a mix of an on-board PC vision system
processing, gathering of information, and service delivery. In for traffic sign ID and affirmation with a data recorder
addition, they detected and discussed different methods to contraption for trivial criminal offense alert and the chiefs. As
solve it. of now, criminal traffic offenses are simply revolved around
Vipin Jain et al. [12] discussed ways for detection and restraint three signs: explicitly in speed limit, stop sign, and no-no
congestion in a local setting. In spite of the fact that the turning. The structure is expected to alert drivers about likely
proposed solution is practical to affect the local congestion but negligible criminal offenses by creating acoustical messages
not ideal to stop the traffic that extends to miles. We can have through vehicle enhancers.
an improved analysis by using an array of cameras along the Ninad Lanke et al. [20] contemplated that Traffic blockage is a
highway, with facilities for local processing of data collected. significant issue in numerous urban areas of India along with
With this, the system can make informed analysis, keeping the different nations. Disappointments of signs, powerless law
global context of the frames into consideration. execution and dreadful traffic the leaders have provoked
Mahesh Lakshminarasimhan et al. [13] studied the well- gridlock. One of the significant issues with Indian metropolitan
founded growth of the urban population throughout the world regions is that the current structure can't be broadened more,
and the effect it has on traffic control across various big cities. and thus the singular decision available is better organization
The flaws in the current traffic management systems are of the traffic.
presented along with a traffic control mechanism. An effective Mandhare et al. [21] saw that a significant estimation for the
system is presented with big data analytics and Radio financially savvy development of any country is quickly
Frequency Identification which is backed by IoT. This expanding vehicle check. Rise in vehicle count and wastage of
proposed system, after implementation, would extensively help energy, time and biological defilement is huge issue, that most
in reducing the traffic congestion in various cities and help of the metropolitan networks are taking a gander at
improve the security system. notwithstanding measures being taken to hide and diminish it.
Partha Sarathi Chakrabortya et al. [14] explains an algorithm Recently gridlock has become clear as one of the huge troubles
for traffic management with better response time and lower for trained professionals, coordinators, and policymakers, not

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on June 05,2023 at 17:48:02 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Computer Communication and Informatics (ICCCI ), Jan. 25 – 27, 2022, Coimbatore, INDIA

in each and every metropolitan setting, but instead around the related to the condition of weather, traffic flow and data related
world. to navigation
C. Edge Computing
Inference or findings from literature survey Edge computing facilitates the local computation of data. It
After extensively glancing through each research paper, we involves the processing and storage of data close to the devices,
deduced that a recognition following structure is carried out to which create or consume these data. This reduces the
precisely match the quantity of vehicles on streets. The dependency on cloud server based computation, thereby
identifier tracker model blends are also clarified to accomplish reducing the delay in processing. Edge computing is preferred
critical upgrades in vehicle checking results in spite of the fact for applications that need quick and real time responses from
that there are as yet numerous deficiencies which should be the system. For vehicle detection and counting, edge devices
addressed. In spite of the fact that conditions like sub-par can be connected to cameras used for capturing videos of
camera quality, impediment and low light conditions could vehicles on roads. These devices can thus be programmed to
demonstrate interest in precisely recognizing objects. Profound store the number of vehicles of different classes [25]. The edge
adapting fundamentally based article recognition models devices can be further configured to perform analysis of the
combined with both on the web and disconnected multi-object data that it collects [26].
global positioning frameworks might incorporate continuous D. Quantum Convolutional Neural Network (QCNN)
item location related to following vehicle development Classical convolutional neural networks are the structured
directions. By and large, for checking all vehicles on the street, neural network comprising multiple layers of image
test results from the study demonstrate that YOLO and Deep processing. Here, the input image goes through multiple layers
SORT will be the best mixes. The upside of the proposed of image processing before it gets classified (car, truck, person
strategy is that it very well may be summed up to any sort of etc). A Quantum Convolutional Neural Network (QCNN)
street video caught utilizing handheld versatile cameras. would have the same types of layers but it has a quantum circuit
model, so in this case the input state will be a generic pure or
mixed state and this could be a classical or quantum state, the
III. AI - ENABLED TRAFFIC MANAGEMENT qubits are in a one-dimensional array but more generally this
SYSTEM could be easily generalized to two or higher dimensions [27].
A. Artificial Intelligence The convolution layers are now going to be translationally
Artificial intelligence is the science of creating machines or invariant quasi local unitary, this means to take one unitary
systems that do things that may need intelligence if done by acting on a small number of qubits versus this two qubit
men. For their implementation it is necessary to induce unitary, apply it in the same fashion to all the nearest neighbour
information and knowledge and use information and pairs of qubits. Now for the pooling layer, we want to reduce
knowledge to realize the goal or solution. AI in transportation the system size by measuring a fraction of the qubits and apply
systems would require: unitaries to the remaining qubits depending on the outcome.
• A big evolution in the IT sector and technology. Again the convolution and pooling layers can be interleaved
• An expansion of computer based components like fast and repeated until you get to a small enough number of
processors and machines with a high capacity Qubits that you perform a fully connected operation which is
memory. in general a non-local measurement and so the weights to be
• Digitization of image and videos learned in the quantum CNN would just be the parameters to
Artificial intelligence is used to predict and detect traffic flow each of these unit areas.
and conditions. It is used in resolving control and optimization E. Popular online simulators
of problems [22]. • Google Colaboratory
B. Big data Technology • Brain simulator
In the earlier days very little data was processed victimising the • Arnold simulator
traditional and humble RDBMS. As the information • Unity Simulation
magnified, the traditional systems began to fail dramatically. • Artificial Intelligence SimBot
To process such enormous data we needed some special
storage units and a powerful processing capability. Big Data is IV. IMPLEMENTATION
basically a term used which means a huge volume of
unstructured and structured data which is so vast that it makes A. Object detection
it complicated to comprehend using the available database and YOLOv3 is a highly efficient object detection algorithm. It is
techniques. The role it plays here is by providing the highly preferred for real-time detections [24]. The YOLO
subsequent benefits: algorithm applies a single neural network to the entire image
• An efficient traffic management can scale back which means that only a single layer of input nodes sends data
congestion, enhance overall performance measures to subsequent receiving nodes. The architecture of YOLO v3
for a smooth flow of traffic, and efficiently manage is based on Darknet-53 as the backbone, which has 53
the roadway related assets. convolutional layers. For detection, three scales are used,
• The creation of intelligent vehicles will help reduce which helps in detecting various objects of different scales in
congestion throughout various cities in the world. an image [23]. Also, fine grained features of the image are
Telematics is comprehensively applied in managing traffic and saved, by performing the upsampling and downsampling of
other traffic systems, it is used to produce figures and specifics feature layers.The v3 version faces less localization errors and
has higher precision in detection of smaller objects.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on June 05,2023 at 17:48:02 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Computer Communication and Informatics (ICCCI ), Jan. 25 – 27, 2022, Coimbatore, INDIA

B. Vehicle counting
Vehicle counting is a method wherein the estimation of road
vehicle density and assessment of road traffic is done. This can
be done with the substantial utilization of various surveillance
videos and camera footage, which has become a huge digital
data resource.
Count of vehicles is carried out by separating the moving
vehicles from the background part in each video frame. It is
performed by the below mentioned algorithms which can
classify the objects into various types like a car, truck, bus,
motorbike. The proposed methodology would require:
• A trained dataset (prepare a model, train it).
• Images are sent into the YOLO framework to detect
and classify the vehicles.
• Multi vehicular tracking will be done using Deep
SORT with a Deep association metric.
• Vehicle counting is done using the integrated
algorithms followed by an analysis and determination
of the performance parameters as shown in Figure 1.

Fig. 3. Simulation results

Figure 3 shows the results of simulation. Various objects like
cars, motorbike and person being detected with confidence
scores.

V. KEY PLAYERS IN RESEARCH AND DEVELOPMENT

Fig.1.Block diagram of proposd work Some of the popular key industry people working in this
research domain are -
• Intel Corporation
• tuSimple
• Autonomous solutions
• Intelligent Automation, Inc
• Vayyar

VI. RESEARCH AND REAL TIME IMPLEMENTATION

CHALLENGES
The dataset is available mainly for foreign roads and not Indian
roads. It is therefore, tough to train the model for vehicles
pertaining specifically to India, for example three-wheeler
auto-rickshaw. Also, model training for YOLOv3 requires an
Fig.2. simulation code snipet
Nvidia graphics card to be present in the system. In the absence
Figure 2 shows the simulation code snipet. The time taken to of this GPU, open source online platforms like Google
process each frame is 0.872s and the total time to finish entire Colaboratory have to be used for model training. The online
video is 5.429 minutes. GPU availability is limited and poses a challenge when a
higher number of iterations is required during training. The
vehicle counting model counts vehicles that cross a line. The
position of this line has to be modified for different roads,
depending upon the orientation and direction of the particular
road. Also, for capturing videos on roads, high resolution
cameras will be needed, which increases the cost of
implementation.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on June 05,2023 at 17:48:02 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Computer Communication and Informatics (ICCCI ), Jan. 25 – 27, 2022, Coimbatore, INDIA

VII. CONCLUSION [21] Mandhare et al., “Intelligent Road Traffic Control System for Traffic
Congestion A Perspective”, International Journal of Computer Sciences
and Engineering, 2018.
Deep learning based vehicle detection model YOLOv3 is used [22] C. V. Krishna et al., “A Review of Artificial Intelligence Methods for
for vehicle detection and counting, which is a highly efficient Data Science and Data Analytics: Applications and Research
and accurate algorithm. Various AI tools like QCNN, LSTM, Challenges,” International Conference on I-SMAC (IoT in Social,
Mobile, Analytics and Cloud) (I-SMAC) I-SMAC (IoT in Social, Mobile,
GNN have been analysed and elaborated for the design and Analytics and Cloud) (I-SMAC), 2018, pp. 591-594.
development of AI enabled vehicle detection and counting [23] R. K. Meghana et al., “Background-modelling techniques for foreground
system for urban traffic management system using deep detection and Tracking using Gaussian Mixture Model,” 3rd
learning. Usage of AI instead of IoT, reduces the cost of International Conference on Computing Methodologies and
Communication (ICCMC), 2019, pp. 1129-1134.
implementation, as sensors are not needed for detection and [24] C. Kumar B et al., “Performance Analysis of Object Detection Algorithm
counting. Usage of edge devices, along with high resolution for Intelligent Traffic Surveillance System,” Second International
cameras, which facilitates local data processing and does not Conference on Inventive Research in Computing Applications (ICIRCA),
2020, pp. 573-579.
rely totally on cloud computation. The constraints of the model [25] P. Pradhyumna et al.,“Graph Neural Network (GNN) in Image and Video
is that it requires a dedicated GPU for vehicle detection and Understanding Using Deep Learning for Computer Vision Applications”
even in the absence of it, online simulators can be used for Second International Conference on Electronics and Sustainable
simulation. Communication Systems (ICESC),2021,pp. 1183-1189.
[26] M. Rohith et al., “Comparative Analysis of Edge Computing and Edge
Devices: Key Technology in IoT and Computer Vision Applications,”
REFERENCES International Conference on Recent Trends on Electronics, Information,
[1] C. S. Asha et al., “Vehicle Counting for Traffic Management System Communication & Technology (RTEICT), 2021, pp. 722-727.
using YOLO and Correlation Filter”, IEEE International Conference on [27] V. Rajesh et al., “Quantum Convolutional Neural Networks (QCNN)
Electronics, Computing and Communication Technologies (CONECCT), Using Deep Learning for Computer Vision Applications,” International
Bangalore, India, pp. 1-6, 2018. Conference on Recent Trends on Electronics, Information,
[2] Mandal Vishal et al., “Object Detection and Tracking Algorithms for Communication & Technology (RTEICT), 2021, pp. 728-734.
Vehicle Counting: A Comparative Analysis”, Conference of Civil and
Environmental Engineering, University of Missouri-Columbia, 2020.
[3] Marek Landowski et al., “Knowledge extraction from the experimental
data of the vehicle traffic volume”, Transportation Research Procedia,
Volume 39, pp. 270-279, 2019.
[4] Yingjie Xia, et al., “Towards improving quality of video-based vehicle
counting method for traffic flow estimation”, Signal Processing, Volume
120, pp. 672-681, 2016.
[5] Daming et al., “Intelligent vehicle network system and smart city
management based on genetic algorithms and image perception”,
Mechanical Systems and Signal Processing, 2020.
[6] Galletta, Antonino et al., “MeSmart-Pro: Advanced Processing at the
Edge for Smart Urban Monitoring and Reconfigurable Services”,
Journal of Sensor and Actuator Networks, 2020.
[7] A. Gomaa, et al., “Robust Vehicle Detection and Counting Algorithm
Employing a Convolution Neural Network and Optical Flow,” Sensors,
vol. 19, no. 20, 2019.
[8] J. Barthélemy et al., “Edge-Computing Video Analytics for Real-Time
Traffic Monitoring in a Smart City,” Sensors, vol. 19, no. 9, p. 2048,
2019.
[9] E. Sonnleitner, et al., “Traffic Measurement and Congestion Detection
Based on Real-Time Highway Video Data,” Applied Sciences, vol. 10,
no. 18, p. 6270, 2020.
[10] K. Nellore et al., “Traffic Management for Emergency Vehicle Priority
Based on Visual Sensing,” Sensors, vol. 16, no. 11, p. 1892, 2016.
[11] Souza, A. et al., “Traffic management systems: A classification, review,
challenges, and future perspectives 13, 2017.
[12] Jain, Vipin et al., “Road traffic congestion in the developing world”,
International Journal of Distributed Sensor Networks,2012.
[13] Mohammed Sarrab et al., “Development of an IoT based real-time traffic
monitoring system for city governance”, Global Transitions, Volume 2,
pp 230-245, 2020.
[14] Chakraborty et al., “Adaptive and Optimized Emergency Vehicle
Dispatching Algorithm for Intelligent Traffic Management System”,
Procedia Computer Science, 2015.
[15] Chun Yong et al., “An improved pheromone-based vehicle rerouting
system to reduce traffic congestion”, Applied Soft Computing, 2019.
[16] Dogru et al., “Traffic accident detection by using machine learning
methods”, International Symposium on Sustainable Development, 2012.
[17] Nyothiri Aung et al., “Dynamic traffic congestion pricing and electric
vehicle charging management system for the internet of vehicles in smart
cities”, Digital Communications and Networks, 2021.
[18] Mengting Bai et al., “PrePCT: Traffic congestion prediction in smart
cities with relative position congestion tensor”, Neuro computing, 2020.
[19] N. Aliane et al., “Traffic violation alert and management,” 14th
International IEEE Conference on Intelligent Transportation Systems
(ITSC), Washington, DC, USA, pp. 1716-1720, 2011.
[20] Lanke et al., “Smart Traffic Management System”, International Journal
of Computer Applications, pp. 19-22, 2013.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on June 05,2023 at 17:48:02 UTC from IEEE Xplore. Restrictions apply.