Content uploaded by Fahimul Hoque Shubho
Author content
All content in this area was uploaded by Fahimul Hoque Shubho on Feb 19, 2022
Content may be subject to copyright.
Real-time traffic monitoring and traffic offense
detection using YOLOv4 and OpenCV DNN
Fahimul Hoque Shubho, Fahim Iftekhar, Ekhfa Hossain, Shahnewaz Siddique
Department of Electrical and Computer Engineering, North South University, Dhaka-1229, Bangladesh
fhshubho@gmail.com, fahimiftekhar55@gmail.com, ekhfahossain@gmail.com, shahnewaz.siddique@northsouth.edu
Abstract— This paper presents a computer vision-based
system for traffic offense detection. The system detects traffic
offenses such as speed limit violations, unauthorized vehicles,
traffic signal violations, unauthorized parking, wrong-way
driving, and motorbike riders without helmets. The traffic
offense detection system consists of a pipeline of four different
modules. These are a vehicle detection module, a vehicle
classification module, a vehicle tracking module, and a traffic
offense detection module. Vehicles on the roads are detected in
the vehicle detection module using visual data such as live
camera feed. Next, after the vehicles are detected, they are
classified into different classes using a vehicle classification
module. A vehicle tracking module is developed to track the
vehicle as it moves through the traffic. Lastly, we have
implemented a traffic offense detection module that analyzes
traffic patterns and detects different types of traffic violations in
real-time. The entire system is implemented using OpenCV
Deep Neural Network (DNN) module. We have used YOLOv4
to detect vehicles on the roads with high accuracy. For
motorbike riders without helmets, we have used a fast YOLOv4-
tiny model. The DeepSORT algorithm is used to track vehicles
in real-time. Obtained accuracies are 86% in YOLOv4 for
Vehicle and 92% in YOLOv4-tiny for Helmet Detection.
Keywords— Computer Vision, Object Detection, Vehicle
Detection, Vehicle Tracking, Traffic Offense Detection, YOLOv4,
YOLOv4-tiny, OpenCV, Deep Neural Network.
I. I
NTRODUCTION
Accidents on the road are a big concern in the modern
world. Many countries spend millions of dollars on reducing
road accidents every year. Various road safety methods and
measures are in place to prevent road users from being
seriously injured or killed. Still, countries are facing a
considerable number of accidents on the roads every day.
Most of these accidents are linked to people not willing to
follow the traffic rules on the road and lack of monitoring.
Research has shown that improvements to road infrastructure,
particularly design standards that consider all road users’
safety, are critical to making roads safe. [1] shows that 112
countries have national design standards for the management
of speed. But infrastructure alone cannot make the roads safer
without monitoring and adherence to traffic rules. Today
artificial intelligence techniques in computer vision and
machine learning offer fascinating and promising solutions to
improving traffic monitoring and increasing road safety.
In this research, we use modern computer vision and
machine learning techniques to identify and classify vehicles
and track them while continually assessing traffic parameters
for potential traffic offenses. Our suggested approach is more
efficient since it avoids the clutter of having separate models
for various objectives. Instead, just one model detects,
classifies, tracks, and assigns traffic violations, if any, in real-
time, while a second model keeps an eye out for special cases.
This method ensures that the system operates in a timely,
clutter-free, efficient, and practical manner. Our system starts
with obtaining a video feed from a traffic camera, then
detecting vehicles on the road and tracking them to analyze
the scenario to detect possible traffic offenses. The video feed
from the video camera is first analyzed with an object
detection model we trained with YOLOv4 (You Only Look
Once) [2] to detect and classify the vehicles on the road. Our
proposed, implemented model has a mean average precision
of 87.19% in detection and vehicle classification. The
YOLOv4-tiny model that we trained from scratch has shown
an incredible 98.1% mean average precision and is developed
to detect riders without helmets.
II. L
ITERATURE
R
EVIEW
Traffic management must rely on a system for estimating
traffic parameters in real-time. Currently, there is a lot of work
going on to detect traffic parameters in the research field. Ou
et al. [3] proposed parallel process techniques to detect traffic
offenses. Their work implemented real-time traffic violation
detection in a monitoring stream using simultaneous video
streams from multiple cameras. The system was designed to
detect three types of vehicles- speeding vehicles, blacklisted
vehicles, and plate-cloned vehicles. The system was
implemented and evaluated using both real and synthetic
traffic data.
Beymer et al. [4] proposed a model that deals with traffic
and lighting conditions while using segmentation,
classification, and tracking methodologies. The model focuses
on vehicle segmentation and tracking and the computation of
traffic parameters from tracking data. They have used a
feature-based traffic tracking approach to track vehicles under
congestion. The system tracks vehicle sub-features, which
makes it less susceptible to partial occlusion. A network of
C40 DSP (Digital Signal Processor) chips linked to a host PC
has been used to construct a real-time version of the system.
Wang et al. [5] proposed an approach to use an enhanced
background-updating algorithm and feature-based tracking
method. In their paper, they have used video-based traffic
detection through an improved background-updating
algorithm, then tracking the moving vehicles by a feature-
based tracking method.
Zhu et al. [6] proposed a vehicle detection and tracking
volume statistics algorithm based on an improved single
Gaussian model. The algorithm has three sections: moving
target detection, shadows suppression, and traffic volume
count. The single Gaussian model detects moving vehicles
while shadow suppression is performed in the RGB feature
space. Traffic volume was calculated in the virtual lanes.
Deng et al. [7] proposed a model that can be used for object
segmentation, classification, and tracking methodologies to
know the real-time measurements in urban roads. The model
used the background subtraction method to detect stationary
or slow-motion objects. Experimental results showed a
diminishing 20% degradation of infrastructure capacities.
Wen-juan et al. [8] proposed a model that can locate and
track the vehicles in a video and calculate the traffic flow,
queue length, queue waiting time, the average speed, and other
vital parameters. The model used online learning mechanisms
and optical flow to track objects. Experimental results showed
the accuracy of daytime detection was 98%, and nighttime
detection was 92%. The average speed detection of daytime
was 95%, while nighttime was 90%.
Kim et al. [9] compared different Artificial Neural
Network (ANN) models for real-time vehicle type
recognition. They compared deep learning-based object
detection models R-CNN (Region Based Convolutional
Neural Network), Fast R-CNN, Faster R-CNN, YOLO, and
SSD (Single Shot Detector) in processing speed and accuracy
for best performance. Faster R-CNN had less FPS (frames per
second) with better accuracy, while SSD had less accuracy
with better FPS. The YOLO model was a middle ground.
Zinchenko et al. [10] proposed an Artificial Neural
Network (ANN) model based on Warren McCulloch and
Walter Pitts’ work on the simulation of nervous activity. It
takes a similar approach to Kim et al. [9] in using a faster and
effective model. The proposed system's primary aim is to
deliver better outcomes in decreased pedestrian and vehicle
waiting times, shorter travel times, and higher average vehicle
velocity. The best result of vehicle detection was 94.65%, and
the worst result was 88.71%.
III. M
ETHODOLOGY
Our system monitors and detects traffic offenses by
processing the video feed from traffic cameras in several
stages. An object detection model first processes the video
feed from the traffic camera to detect and classify the vehicles
present in the video. Detected vehicles are tracked between
frames, and all the available information from each frame of
the video is stored temporarily and analyzed to detect any
possible traffic offense. An additional object detection model
processes every detected motorbike to monitor if the riders are
wearing helmets. All detected offenses are stored as a visual
snap for further use. Detected traffic offenses are- Overspeed
Detection, Unauthorized Vehicles Detection, Traffic Signal
Violation, Unauthorized Parking, Wrong-Way Driving,
Motorbike Riders without Helmet. Fig. 1 represents the
overview of our proposed system.
Fig. 1. Overview of the system.
A. Vehicle Detection, Classification and Tracking
Detecting vehicles from the video feed of traffic cameras
and classifying them is a crucial part of this system. A good
amount of data and an efficient object detection model are
required for this part. We decided to use multiple datasets and
train YOLOv4 object detection models for detection and
classification on our custom dataset.
1) Dataset: A combined dataset consisting of vehicles
images of Bangladesh from different sources were used.
Labeled images from the PoribohonBD dataset [11] and the
Dhaka-AI dataset [12] were used. Additionally, publicly
available images from online were labeled and added to the
dataset. The combined dataset contained 11,210 images with
46,662 labeled objects of 17 different types of vehicles. The
types of vehicles are bus, rickshaw, motorbike, bike rider
without helmet, car, three-wheelers, pickup, microbus, van,
truck, bicycle, police car, ambulance, human hauler (leguna),
wheelbarrow, auto-rickshaw, army vehicle, and horse cart.
This dataset was used to train the object detection model to
detect and classify the vehicles from the video feed at the first
stage of the system.
Another dataset, the helmet dataset from Kaggle [13],
containing 764 labeled images, was combined with an
additional 502 images collected and labeled by us into a
dataset of 1,266 labeled images of different helmet types.
This dataset was used to train an additional object detection
model to detect helmets on bike riders. Figure 2 shows the
vehicle count of each class in the combined dataset images.
Fig. 2. Vehicle counts in the combined dataset images.
2) Data pre-processing: To prepare the dataset for object
detection model training, the image labels were unified in the
combined dataset. The ProibohonBD dataset and Dhaka-AI
dataset had different image labels for the same class of
vehicles. We prepared a list of labels for 17 types of vehicles
and converted the image labels from both datasets according
to it to unify them. We used a python script to unify the labels.
We also labeled the images collected online according to the
list. Table I shows the data difference of data labels in the
datasets and the unified label for our combined dataset.
Some vehicle types such as SUV and taxi were separately
labeled in the Dhaka-AI but were labeled as cars in the
PoribohonBD dataset. As there were not enough images of
SUVs and taxis to treat them as two separate classes, we
decided to merge them into car images and modified their
labels as cars in the combined dataset. Some vehicles were of
YOLOv4
Model
[Vehicle detection]
Detected &
Classified Vehicles
Vehicle Tracking
with DeepSORT
YOLOv4-tiny
Model
[Helmet detection]
Detection of
Traffic Offense Store & Analyze
frame data
Visuals with
Detected Offense
Video Frame
Detected Motorbikes
the same type but labeled differently in the two datasets. As
the same vehicle had multiple names locally and in datasets,
we kept one name as label and merged them. We did the same
for a few other labels, and it is shown in Table I.
TABLE I. I
MAGE LABELS BETWEEN THE DATASETS
Class labels
PoribohonBD
Dataset
Class labels Dhaka-
AI Dataset
Class labels
Combined Dataset
Bicycle bicycle bicycle
Bus bus bus
- minibus bus
Car car car
- SUV car
- taxi car
- minivan microbus
- ambulance ambulance
Van van
CNG three-wheelers(CNG) three-
wheelers(CNG)
Easy-bike auto-rickshaw auto-rickshaw
Horse-cart - horsecart
Leguna human hauler human
hauler(leguna)
Motorbike motorbike motorbike
- scooter motorbike
Rickshaw rickshaw rickshaw
Truck truck truck
Wheelbarrow wheelbarrow wheelbarrow
- police car police car
- pickup pickup
- army vehicle Army vehicle
We didn’t include vehicle images of the boat and launch
from the PoribohonBD dataset as these are not seen on the
road. The PoribohonBD dataset and Dhaka-AI dataset had
image labels in ‘.xml’ format. To use the images from
datasets labels were converted to YOLO supported ‘.txt’
format using python script. Our collected images from online
were labeled in YOLO-supported ‘.txt’ format. We followed
the same procedure for our helmet detection dataset to
convert the labels into YOLO supported format.
3) Object Detection Model: For such a monitoring
system, it is critical to have an object detection model that can
perform efficiently while providing enough speed to monitor
the traffic in real-time and accuracy to classify the vehicles
with acceptable precision. YOLOv4 provides acceptable
speed and accuracy in detecting objects and classifying them.
A YOLOv4 model and a YOLOv4-tiny model were trained
with the combined dataset of vehicle images of Bangladesh
to detect and classify the vehicles from the video feed of
traffic cameras. Another YOLOv4-tiny model was trained
with the second dataset to detect helmets from the motorbike
riders that are detected by the previous model. YOLOv4-tiny
was chosen to detect helmets on bike riders for its speed, this
sustained the speed and efficiency of the whole system while
using two object detection models at the same time.
All the models were trained using open source Neural
Network framework Darknet [14] in Google Colab.
Parameters modified for training YOLOv4, and YOLOv4-
tiny object detection models on our dataset are shown in
Table II. Batches refer to the number of images loaded for
each iteration during training and split into the number of
subdivisions to process in the GPU. Max batches are the
number of iterations during training for each model. It is
calculated as classes multiplied by 2000. There are 17 classes
of vehicles in our combined dataset. The number of iterations
for YOLOv4 and YOLOv4-tiny models are 17*2000 or
34000, respectively. For the helmet detection model, we have
set the number of iterations to 6000 as the minimum number
of iterations in YOLOv4 models is required to be more or
equal to 6000. Steps are the number of iterations at which
learning rates are changed. It is set to 80% and 90% of the
max batches for each model. Network size refers to the
resized input image for the model. The number of filters is
calculated as follows-
filters = (classes + 5) × 3 (1)
For our vehicle classification models, the number of filters is
(17+5)*3 or 66. For the helmet detection model, it is (2+5)*3
or 21.
TABLE II. M
ODIFIED PARAMETERS FOR TRAINING
YOLO
V
4
AND
YOLO
V
4-
TINY
O
BJECT
D
ETECTION
M
ODELS
Parameters
YOLOv4-
tiny Vehicle
Detection
Model
YOLOv4
Vehicle
Detection
Model
YOLOv4-tiny
Helmet
Detection
Model
classes 17 17 2
batch 64 64 64
subdivisions 16 32 16
max_batches 34000 34000 6000
steps 27200,
30600 27200, 30600 4800, 5400
Network size
(width,
height)
416, 416 608, 608 608, 608
filters 66 66 21
4) Vehicle Tracking: Vehicles that are detected by
YOLOv4 object detection models are tracked throughout the
video frames. When a vehicle enters into the video frame for
the first time, it is given a unique id and tracked between
frames using DeepSORT [15]. Each vehicle's id and position
in each frame are stored temporarily and analyzed for
possible traffic offenses. We have implemented the
DeepSORT algorithm in our system following the procedure
of [16].
B. Traffic Offense Detection
1) Overspeed Detection from Speed Estimation: The speed
of each detected vehicle is estimated as shown in Fig. 3 by
the change in their position.
Fig. 3. Speed estimation of vehicles on the road.
Speed is estimated when a vehicle stays at least five frames
in the video. The change of vehicle’s position in the x-axis
and y-axis is considered for five consecutive frames, and
speed is estimated in terms of pixel per second. When the
total area covered by the traffic camera is available to the
system, it can also calculate the estimated speed in terms of
meter per second. When a vehicle’s estimated speed goes
over the allowed speed, it is visually marked on the video
frame and logged for overspeeding.
2) Unauthorized Vehicles Detection: Unauthorized
vehicles are detected in the first stage when the video feed
goes through the YOLOv4 model. Any detected vehicle is
classified as rickshaw, bicycle, van, auto-rickshaw,
wheelbarrow, or horse cart; it is immediately marked and
logged as an unauthorized vehicle that is not permitted to be
on the main roads. Fig. 4 shows a bicycle and several
rickshaws being marked on the video frame as unauthorized
on the main street of Dhaka, Bangladesh.
Fig. 4. Vehicles 1, 2, 3 & 4 are detected as unauthorized vehicles on street.
3) Traffic Signal Violation: The traffic signal is
synchronized with our system, and it monitors the road
intersection points for a possible signal violation when the
traffic light indicates red. A visual line on the intersection
point of the road works as a reference for our system, and
when any vehicle crosses the line while the traffic light
indicates red, it is marked and logged for traffic signal
violation. The vehicle remains marked even if it reverses back
after violating the traffic signal. Fig. 5 shows an example of
a traffic light violation detected by the system on a video
collected from YouTube.
Fig. 5. Traffic Signal Violation detection.
4) Unauthorized Parking: Vehicles that stand still on the
main road and cause roadblocks are marked and logged for
unauthorized parking. All the vehicle’s IDs and positions
from the previous frames are stored in the system. The current
position of each detected vehicle is compared with the
positions they were on previous frames to detect how long the
vehicles stood still in the same places. Depending on the time
the vehicle is standing idle on the street, the vehicle is marked
for unauthorized parking. Figure 6 shows a bus standing idle
on the main street is being marked on the video frame for
unauthorized parking.
Fig. 6. Vehicle 1 is detected for unauthorized parking on the street.
5) Wrong Way Driving: Vehicles that drive in the wrong
direction of the road are detected by our system from their
movement. Fig. 7 demonstrates such a situation where the
system treats all the lanes as one-directional and all the
vehicles moving into opposite directions are marked for
wrong-way driving. The direction of driving on the road is
predefined in our system. It can be changed depending on the
road that is being monitored through the system. Our system
checks the change of position of every vehicle on the x-axis
and y-axis. Change of positions along the x-axis of the video
frame indicates in which direction the vehicles are taking
turns. And the changes in position along the y-axis of the
video frame indicates in which direction the vehicles are
moving forward. From the change of position along both axis,
our system detects when vehicles move in the wrong
direction.
Fig. 7. Wrong way driving detection.
6) Motorbike Riders without Helmet: Our system detects
motorbike riders without a helmet in four steps, as illustrated
in Fig. 8.
Fig. 8. Steps of detecting helmet on motorbike riders.
In the first two steps, every motorbike is detected
and classified along with the riders as a motorbike. In step 3,
the portion of the video frame that contains the detected area
of motorbikes are separated as individual images, and they go
through the YOLOv4-tiny model in step 4. The YOLOv4-
tiny model detects helmets on the riders of the motorbikes
images passed to it. If no helmet is detected, our system labels
it as a bike rider without helmet and marks it in the video
frame for riding a motorbike without helmet. Running two
object detection models, one for detecting the motorbike and
another one to detect the helmet, increases time complexity.
To tackle this complexity, YOLOv4-tiny is used for helmet
detection. YOLOv4-tiny model is faster than a regular
YOLOv4 model, but the trade-off is accuracy.
IV. R
ESULT AND
D
ISCUSSION
Maintaining reasonable accuracy and speed was crucial
while implementing the system as computer vision-based
techniques are computationally expensive. Our system can run
on both CPU and GPU. But the GPU is required to get
minimum speed to monitor the traffic in real-time. We have
implemented and tested the system with OpenCV[17] DNN
(Deep Neural Network) module on an Nvidia 1050Ti 4GB
GPU. The OpenCV library does not support GPU by default,
and it was compiled from source code to enable GPU support.
The first stage of our system is crucial for detecting traffic
offenses. The accuracy in this stage in detecting and
classifying vehicles affects how accurately the vehicles'
positions and classes are being logged and analyzed for the
possible traffic offenses. We have trained and tested one
YOLOv4-tiny model and one YOLOv4 model to detect and
classify vehicles in the first stage of our system. The
YOLOv4-tiny model produced 24~30 FPS on our setup and
showed 72.02% mean average precision. While using the
YOLOv4 model in the first stage of the system, our system
was able to process the video feed from the traffic camera at
15~20 FPS speed with our setup. It was slower than the
YOLOv4-tiny model but showed better accuracy. The
YOLOv4 object detection model showed 86% average
precision in detecting and classifying the vehicles. Table III
shows the comparison of class-wise average precision of the
YOLOv4-tiny model and YOLOv4 model in the detection and
classification of the vehicles.
The accuracy in classifying vehicles differed with the
number of samples of each class present in the dataset. The
lower number of samples of police cars and ambulances in the
dataset reduced the average precision of the YOLOv4-tiny and
YOLOv4 models in detecting and classifying them. Table III
shows that though the YOLOv4 model produces less FPS, it
outperforms YOLOv4-tiny in average precision in most
vehicle classes. So, we have used the YOLOv4 model in the
first stage of our system for its better accuracy.
TABLE III. C
LASS
-
WISE AVERAGE PRECISION OF
YOLO
V
4-
TINY
MODEL AND
YOLO
V
4
MODEL
.
Class Labels
YOLOv4-tiny
Model’s Average
Precision
YOLOv4 Model’s
Average Precision
bus 75.82% 95.40%
rickshaw 69.17% 92.66%
motorbike 69.13% 92.61%
car 74.48% 94.11%
three wheelers(CNG) 79.52% 96.38%
pickup 61.27% 91.27%
microbus 57.28% 89.48%
van 88.95% 96.97%
truck 84.62% 96.55%
bicycle 77.73% 94.32%
police car 21.30% 17.83%
ambulance 21.17% 51.75%
human hauler (leguna) 89.55% 97.58%
wheelbarrow 88.80% 94.50%
auto rickshaw 86.67% 97.47%
army vehicle 67.46% 75.27%
horsecart 92.35% 97.04%
Our second object detection model used in the system is a
YOLOv4-tiny model for detecting helmets from motorbikes
that are detected in the first stage of the system by the previous
model. Fig. 9 shows the mean average precision and
validation loss over iteration during training YOLOv4-tiny
model.
Fig. 9. mAP vs loss during training YOLOv4-tiny model for helmet
detection.
This model achieved 98.1% mean average precision in
detecting helmets with 6000 iterations in training. The
YOLOv4-tiny model ran faster than the YOLOv4 model and
maintained the speed of the system.
We have tested the system with several video footage we
captured from the streets of Dhaka and collected from
YouTube. In our observations, the efficiency of the system
depended on the accuracy of the object detection models.
Table IV shows performance metrics for all the three models
we have tested in our system. Precision shows the accuracy of
the models, recall refers to the actual positives identified by
the models, f1-score refers to the balance between precision
and recall, and the last metric, mAP@0.50 refers to the mean
average precision with a threshold of 0.50 intersection-over-
union. Replacing the YOLOv4-tiny model with the YOLOv4
model for detecting and classifying vehicles increased the
detection efficiency of the traffic offense. With the YOLOv4
model for detecting vehicles and the second YOLOv4-tiny
model for detecting helmets, the system analyzed the
information from the object detection models and detected
most of the traffic offenses it was intended to. An increased
number of vehicles reduced the speed of the system and
resulted in an FPS drop, which might be improved by running
the system on better hardware.
TABLE IV. PERFORMANCE
METRICS
OF
THE
MODELS
Class Labels Precision Recall F-1
score
mAP
@0.50
YOLOv4-tiny for
Vehicle Detection 69% 72% 71% 72.07%
YOLOv4 for
Vehicle Detection 86% 92% 89% 87.19%
YOLOv4-tiny for
Helmet Detection 92% 98% 95% 98.1%
V. C
ONCLUSION
In this paper, a computer vision-based road safety
system has been presented. We performed real-time
monitoring to get information of vehicles along with their
various traffic offenses (Over speeding, unauthorized
vehicles, signal violation, unauthorized parking, wrong-way
driving, riders without helmets) if any. We investigated the
vehicle detection, classification, and tracking methodologies
as a guiding point of a road safety system. Existing traffic
CCTV cameras can be used to capture live footages and then
fed into a central server for further analysis with our system.
In future work, better hardware and models with higher
accuracy level can be integrated with the proposed system. It
can be helpful for both traffic monitoring and automatic
traffic offense detection.
R
EFERENCES
[1] Global status report on road safety 2018. Geneva: World Health
Organization; 2018. Licence: CC BY-NC-SA 3.0 IGO.
Retrieved from
https://www.who.int/publications/i/item/9789241565684
[2] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: Optimal
speed and accuracy of object detection,” 2020. [Online]. Available:
arXiv:2004.10934
[3] G. Ou, Y. Gao and Y. Liu, "Real-Time Vehicular Traffic Violation
Detection in Traffic Monitoring Stream," 2012 IEEE/WIC/ACM
International Conferences on Web Intelligence and Intelligent Agent
Technology, 2012, Macau, China, December 4-7, 2012. pp. 15-19, doi:
10.1109/WI-IAT.2012.91.
Available: https://ieeexplore.ieee.org/document/6511640
[4] D. Beymer, P. McLauchlan, B. Coifman and J. Malik, "A real-time
computer vision system for measuring traffic parameters," Proceedings
of IEEE Computer Society Conference on Computer Vision and
Pattern Recognition, 1997, San Juan, PR, USA, June 17-19, 1997. pp.
495-501, doi: 10.1109/CVPR.1997.609371. Available:
https://ieeexplore.ieee.org/document/609371
[5] Xiaoling Wang, Li-Min Meng, Biaobiao Zhang, Junjie Lu and K. -. Du,
"A video-based traffic violation detection system," Proceedings 2013
International Conference on Mechatronic Sciences, Electric
Engineering and Computer (MEC), 2013, Shenyang, China, December
20-22, 2013. pp. 1191-1194, doi: 10.1109/MEC.2013.6885246.
Available: https://ieeexplore.ieee.org/document/6885246
[6] H. Zhu, C. Xu and F. Li, "The Traffic Volume Count Algorithm Based
on Computer Vision," 2013 5th International Conference on Intelligent
Human-Machine Systems and Cybernetics, 2013, Hangzhou, China,
August 26-27, 2013. pp. 130-133, doi: 10.1109/IHMSC.2013.38.
Available: https://ieeexplore.ieee.org/document/6643850
[7] L. Y. Deng, N. C. Tang, Dong-liang Lee, Chin Thin Wang and Ming
Chih Lu, "Vision based adaptive traffic signal control system
development," 19th International Conference on Advanced
Information Networking and Applications (AINA'05) Volume 1
(AINA papers), 2005, Taipei, Taiwan, March 28-30, 2005. pp. 385-388
vol.2, doi: 10.1109/AINA.2005.342. Available:
https://ieeexplore.ieee.org/document/1423717
[8] X. Wen-juan and L. Jian-feng, "Application of vision sensing
technology in urban intelligent traffic control system," 2018 4th
International Conference on Computer and Technology Applications
(ICCTA), 2018, Istanbul, Turkey, May 3-5, 2018. pp. 74-77, doi:
10.1109/CATA.2018.8398659.
Available: https://ieeexplore.ieee.org/document/8398659
[9] J. -a. Kim, J. -Y. Sung and S. -h. Park, "Comparison of Faster-RCNN,
YOLO, and SSD for Real-Time Vehicle Type Recognition," 2020
IEEE International Conference on Consumer Electronics - Asia (ICCE-
Asia), 2020, Seoul, South Korea, November 1-3, 2020. pp. 1-4, doi:
10.1109/ICCE-Asia49877.2020.9277040.
Available: https://ieeexplore.ieee.org/document/9277040
[10] V. Zinchenko, G. Kondratenko, I. Sidenko and Y. Kondratenko,
"Computer Vision in Control and Optimization of Road Traffic," 2020
IEEE Third International Conference on Data Stream Mining &
Processing (DSMP), 2020, Lviv, Ukraine, August 21-25, 2020. pp.
249-254, doi: 10.1109/DSMP47368.2020.9204329. Available:
https://ieeexplore.ieee.org/document/9204329
[11] S. Tabassum, S. Ullah, N. H. Al-Nur, and S. Shatabda, “Poribohon-
BD. V2”, October 1, 2020, Distributed by Mendeley Data. Available:
https://data.mendeley.com/datasets/pwyyg8zmk5/2
[12] ASM Shihavuddin, M. R. A. Rashid, “DhakaAI. V1”, September 19,
2020, Distributed by Harvard Dataverse. Available:
https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/
DVN/POREXF.
[13] Larxel, “Helmet Detection. V1”, June 1, 2020, Distributed by Kaggle.
Available: https://www.kaggle.com/andrewmvd/helmet-detection
[14] J. Redmon. “Darknet: Open Source Neural Networks in C.”
pjreddie.com. https://pjreddie.com/darknet/ (accessed May 6, 2020).
[15] N. Wojke, A. Bewley and D. Paulus, "Simple online and realtime
tracking with a deep association metric," 2017 IEEE International
Conference on Image Processing (ICIP), 2017, Beijing, China,
September 17-20, 2017. pp. 3645-3649, doi:
10.1109/ICIP.2017.8296962.
Available: https://ieeexplore.ieee.org/document/8296962
[16] OpenCV. 2020. opencv (Version 4.5.0) [Source code].
https://github.com/opencv/opencv
[17] The AI Guy. 2020. yolov4-deepsort [Source code].
https://github.com/theAIGuysCode/yolov4-deepsort
