Traffic monitoring using object detection

Naga Rithesh N
B.Tech Computer Science with
AI and Robotics
Vellore Institute of Technology
Chennai, India
Naga Rithesh Sukesh S Roopesh Mathav
[email protected] [email protected]
[email protected]

Abstract-Determining the optimal object detection algo- tection there is a need to find which of the algorithms can prove to be
rithm for traffic monitoring requires balancing speed and ac- faster, has more accuracy and efficiency and lesser error for which Ross
curacy. YOLOv3 excels in this area, outperforming Faster Girshick [4] divided the solution into three modules: 1) Region propos-
R-CNN and Haar Cascade in both speed and accuracy based als were generated using a selective search algorithm, 2) Each region
on tests with custom image sets and the COCO database. proposal was
While Faster R-CNN offers higher precision for complex
scenes, YOLOv3's speed makes it more suitable for real-
time traffic management. The choice depends on the specific
environment; high-traffic areas may benefit from Faster R-
CNN's accuracy, while YOLOv3's speed is advantageous in
less congested settings. Further research involving real-time
simulations and diverse traffic conditions is recommended.
Keywords: Object Detection, Traffic Monitoring, Algo-
rithm Selection, Speed vs. Accuracy, YOLOv3, Faster R-
CNN, Haar Cascade, Real-time performance, COCO
database, SURF algorithm

I. INTRODUCTION

The root causes of traffic are different depending on many scenarios

such as dysfunctionality of traffic lights,accidents at the middle of
the road or even due to road raging, thus Traffic management can be
very tough at many such situations, police and other officials can at
times help to ease the flow of traffic but this isn't possible for all
cases, therefore there arises a necessity to monitor the traffic at all
times and help find the root cause of such traffic and any possibilities
on how that traffic can be cleared in the most efficient way so that it
doesn't disturb or waste time of the public. This can be achieved by
the use of technologies like cameras simply by placing them on traf-
fic signals or also on drones in remotely accessible areas. Inorder to
achieve, first a live video or image of the current ongoing traffic has
to be recorded in the camera Souhail Guennouni, Et al. [1] used the
FEI face database, which contains numerous face images captured
from various angles and positions. For other objects, we established a
database of images extracted from video captures, encompassing ob-
jects subjected to training from various angles and positions.. Then
the taken video is filtered or made clearer by using various image
processing methods so that all the objects will be identified easier
and classified during object detection Sander Soo[2] employed a
technique known as MOG (Mixture of Gaussians) to model the back-
ground pixels of an image, representing them as a mixture of Gaus-
sians with different weights that signify pixel intensity. After pro-
cessing the images, it is through an object detection algorithm that
will analyse if all vehicles are above a particular speed limit, or if any
collision has occurred between any vehicles, Mohana Et al.[3]
utilised the YOLOv2 algorithm for object detection, implementing
Microsoft Visual C++ 2017 to build an .exe file. They also integrated
pre-trained YOLO9000 weights and its configuration. For Object de-
tect smaller or partially occluded objects is
passed through an architecture with five convolutional layers fol- paramount. Algorithms like Faster R-CNN,
lowed by two dense layers, generating a feature vector of size 4096, known for their precision, excel in such sce-
and 3) The third module had independent linear classifiers pre-
trained for each class. For this reason, this research paper deals
narios, even though their processing speed
with the comparison of 3 such algorithms namely, Haar might be comparatively lower. The trade-off
Cascade(Haar cascade is a well known machine learning based al- between speed and accuracy becomes a
gorithm used for object detection and identification. These algo- critical consideration, and in urban con-
rithms work based on machine learning approach in which lots of
texts, the enhanced accuracy offered by
images are used,whether positive or negative, to train the classi-
fier.),Faster RCNN(The "Faster Region-based Convolutional Neu- Faster R-CNN often outweighs the slightly
ral Network," or "Faster R-CNN," is a ground-breaking object iden- slower processing time. The reliability of de-
tification model that tackles the difficulties of precisely and effec- tections in preventing accidents and ensur-
tively identifying objects in images.The paper "Faster R-CNN: The ing smooth traffic flow justifies the empha-
"Faster Region-based Convolutional Neural Network," or "Faster
R-CNN," is a ground- breaking object identification model that sis on accuracy in these complex urban
tackles the difficulties of precisely and effectively identifying ob- landscapes.
jects in images.The paper "Faster R-CNN: Towards Real-Time Ob-
ject Detection with Region Proposal Networks," presented at NIPS Conversely, rural environments typically ex-
2015, by Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun hibit lower traffic density and simpler
introduces this innovative approach.[5]), SURF(The SURF method
(Speeded Up Robust Features) is a fast and robust algorithm for lo-
scenes. In these settings, the speed of de-
cal, similarity invariant representation, and comparison of images. tection becomes a more prominent factor.
The main interest of the SURF approach lies in its fast computation Algorithms like YOLO, renowned for their
of operators using box filters, enabling real-time applications such speed, are particularly well-suited for such
as tracking and object recognition.[6] This algorithm has been
added for the sole purpose of locating a particular car in traffic by
environments. Their rapid processing capa-
comparing the number plate of different cars with the number plate bilities enable efficient monitoring, even
of a car that the police need to find.) and YOLO((You Only Look with the presence of fewer vehicles or ob-
Once) is a deep learning based algorithm for object detection. Un- stacles. While YOLO's accuracy might be
like the traditional methods which had a two-step process where
marginally lower compared to Faster R-
you first select a bunch of region proposals and then classify each
proposal using CNN, YOLO uses a single neural network to predict CNN, the speed advantage becomes signifi-
bounding boxes and class probabilities directly from full images in cant in less congested areas. The balance
one evaluation, making it incredibly fast.[7]). Sometimes during a between speed and accuracy shifts towards
car chase or a car escape there is a need to track down the exact lo-
speed in rural contexts, as the reduced
cation of that particular car, or to check if the car that has passed a
particular signal's camera has any resemblance to the features of the number of objects to detect minimizes the
car that needs to be tracked down. For this reason there has also impact of slight inaccuracies.
been an inclusion of a fourth algorithm named SURF, which can
help with feature comparison and recognition. This is also an essen- The selection of an appropriate object de-
tial for traffic management as it helps to eliminate any unwanted tection algorithm for traffic monitoring
vehicle from creating further traffic or disruption to the public.
hinges on a comprehensive evaluation of
Object detection, a crucial aspect of computer vision, plays a piv- various factors, including the specific envi-
otal role in various applications, including autonomous driving, ronment, the desired level of accuracy, and
surveillance systems, and robotics. Its significance in traffic moni- the available computational resources.
toring is particularly noteworthy, contributing to improved road
safety, efficient traffic management, and the development of intelli-
High-traffic, densely populated urban areas
gent transportation systems. This research delves into the selection benefit from the precision of algorithms like
of optimal object detection algorithms for real-time traffic monitor- Faster R-CNN, despite their slower process-
ing, focusing on the critical balance between speed and accuracy. ing speed. The enhanced accuracy ensures
The dynamic nature of traffic scenes, with vehicles and pedestrians
constantly in motion, necessitates rapid and precise object identifi-
reliable detection in complex scenes, con-
cation to prevent accidents, optimize traffic flow, and enhance tributing to improved safety and traffic
overall safety measures. management. On the other hand, in less
The complexity of traffic environments congested rural settings, the speed advan-
presents unique challenges for object de- tage offered by YOLO becomes more signifi-
tection algorithms. Urban settings, charac- cant, enabling efficient monitoring with
terized by high vehicle and pedestrian fewer computational demands.
density, demand high accuracy to effec-
tively analyze intricate scenes. In these This research explores the performance of
congested environments, the ability to de- different object detection algorithms, in-
cluding YOLOv3, Faster R-CNN, and Haar rithm, commonly used for Haar Cascade
Cascade, using custom image sets and es- training, efficiently calculates these fea-
tablished databases like the COCO tures using integral images.
dataset. The evaluation considers both
speed and accuracy metrics to determine The training process involves a supervised
the most suitable algorithm for real-time learning approach using a large dataset of
traffic monitoring applications. The find- positive (images containing the object) and
ings will provide valuable insights into the negative (images without the object) sam-
strengths and weaknesses of each algo- ples. Adaboost, a boosting algorithm, is em-
rithm, guiding the selection process based ployed to select the most discriminative
on specific environmental conditions and Haar features and combine them into a
application requirements. Furthermore, strong classifier. The cascade structure is
the research will discuss the potential in- built by sequentially adding stages, each
tegration of additional algorithms like consisting of multiple weak classifiers. This
SURF for enhanced vehicle recognition cascade efficiently filters out negative sam-
and number plate matching, further refin- ples at early stages, reducing computa-
ing the accuracy and effectiveness of tional cost.
traffic monitoring systems.

II. METHODOLOGY
Haar cascade method: A machine learning object detection algo-
rithm uses Haar features to detect objects in images. The algorithm
works by training a cascade of classifiers, each designed to identify
specific Haar features. These classifiers are arranged in a cascade,
ensuring that each classifier is applied only to windows that pass
the previous one.[8] This arrangement makes the algorithm very ef-
ficient, allowing it to quickly reject windows that are unlikely to
contain the object of interest. The false positive rate provides an es-
timate of how many objects fail to be identified out of a given num-
ber of datasets. This ratio usually depends on the data model, which
is either trained by the programmer or based on pre-existing pre-
trained data models.
The Haar Cascade algorithm, a machine
learning approach for object detection,
employs Haar-like features to identify ob-
jects within images. These features, repre-
sented as rectangular regions within the
image, capture contrasting pixel intensi-
ties. The algorithm's training process in-
volves creating a cascade of classifiers,
each designed to detect specific Haar fea-
tures. This cascade structure ensures effi-
ciency by quickly discarding image re-
gions unlikely to contain the target object.

Haar-like features are calculated by sum-

ming the pixel intensities within rectangu-
lar sub-regions and taking the difference
between these sums. Various types of
Haar features, such as edge features, line
features, and four-rectangle features, are
used to capture different aspects of the
object's appearance. The Viola-Jones algo-
with a Region Proposal Network (RPN). The algorithm begins by pass-
ing the input image through a backbone CNN to extract features, corre-
sponding to the "Backbone Network" process box. The next step in-
volves using the RPN to generate region proposals based on the ex-
tracted features, as depicted by the "Region Proposal Network (RPN)"
process box. Subsequently, the flowchart correctly indicates the RoI
(Region of Interest) pooling to extract fixed-size feature

Figure 1: flowchart for haar cascade algorithm.[9]

The presented flowchart is for object detection using Haar Cascades.

The algorithm initiates by loading the trained Haar cascades and ini-
tialising the Haar Cascade classifier, both accurately depicted by the
"Load Pre-trained Haar Cascade XML" and "Initialize Cascade Clas-
sifier" process boxes in the flowchart, respectively. Following this,
the algorithm involves loading the input image, converting it to
grayscale, and resizing it to the required dimensions, all of which are
represented in the flowchart. The heart of the algorithm is encapsu-
lated in the "Haar Cascade Detection" process box, where, for each
possible position in the resized image, the algorithm extracts the nec-
essary portion and runs it through the Haar classifiers. The decision
diamond appropriately checks whether objects are detected, leading
to the subsequent steps of drawing rectangles around the detected ob-
jects and displaying the result image, showcased in the "Draw Rec-
tangles Around Objects" and "Display Result Image" process boxes.
The flowchart concludes with the "End" symbol.

Faster R-CNN:An object detection system called Faster R-CNN en-

ables end-to-end training by fusing a single convolutional neural net-
work and a Region Proposal Network (RPN). It carries out object
categorization, bounding box regression, and region proposal genera-
tion efficiently inside a single architecture. Anchor boxes are useful
for managing differences in aspect ratios and object scales. Faster R-
CNN surpasses its predecessors in computing efficiency by a large
margin and achieves competitive accuracy thanks to a streamlined
workflow and shared convolutional layers.[5] It is now the corner-

stone paradigm for applications involving real-time object detection.

Figure 10: flowchart for faster r-cnn.[10]

The provided flowchart presents the step-by-step process of object

detection using a pre-trained Convolutional Neural Network (CNN)
vectors from each region proposal and the subsequent processing Disadvantages of Faster R-CNN for Object Detection
through fully connected layers for object classification and bound-
ing box regression, captured by the "RoI Pooling" and "Fully ● Computational Intensity: The computational intensity of
Connected Layers" process boxes, respectively. The following Faster R-CNN is one of its primary limitations. There are
several phases in the model, starting with the Re-
steps, including object classification, bounding box regression, gion
and the application of Non-Maximum Suppression (NMS) to filter Proposal Network (RPN) and going on to classification and
out redundant bounding boxes, are accurately represented in the regression stages. When compared to faster single-shot de-
flowchart. Finally, the output of detected bounding boxes and cor- tectors like YOLO, this complexity can make inference and
responding object classes is visualised on the original image, as training rather slow.
indicated by the "Output Bounding Boxes and Classes" and "Dis- ● Complexity in Implementation and Tuning: It can be more
difficult to implement Faster R-CNN than it is to use sim-
play Result Image" process boxes. The flowchart concludes with pler techniques for object detection. It entails
the "End" symbol. configuring the RPN, fine-tuning the hyperparameters, and
Similar to its predecessors, Faster R-CNN network design optimization. It can be more difficult for
novices to attain optimal performance since deep learning
utilizes a convolutional neural network as and computer vision skills are frequently needed.
a backbone to extract feature maps from
the input image. This shared convolutional SURF: It works by first detecting interest points in an image. Interest
points are regions of the image that are likely to be invariant to
layer computation across both the RPN changes in illumination, rotation, and scale. SURF detects interest
and the detection network contributes to points by using a Hessian matrix, which is a measure of the curvature
of a surface and can be used to identify regions of the image that have
its efficiency. Commonly used CNN archi- high curvature[6].
tectures include VGG16, ResNet, and In-
ception, each offering different trade-offs
between speed and accuracy. The choice
of CNN architecture depends on the spe-
cific application requirements and compu-
tational resources.

The RPN is a fully convolutional network

that takes the shared feature maps as in-
put and generates region proposals. It
slides a small network over the feature
maps, predicting objectness scores and
bounding box regressions at each spatial
location. The objectness score represents Figure 8: flowchart for surf.[11]
the probability of an object being present
at that location, while the bounding box The outlined flowchart is for comparing objects using the SURF
(Speeded-Up Robust Features) The algorithm initiates by defining
regression refines the coordinates of the
parameters, such as the Hessian threshold, for keypoint detection, and
proposed region. Anchor boxes, pre-de- creating a SURF detector object. This initialization step corresponds
fined boxes of various scales and aspect to the "SURF Feature Detection (Image 1)" process box in the flow-
ratios, are used to handle different object chart. The subsequent steps involve applying the SURF detector to
sizes and shapes. The RPN outputs a set find key points in the input images, computing SURF descriptors for
each keypoint, and utilising a matching algorithm for finding corre-
of region proposals, each with an associ- spondences between keypoints in different images. These operations
ated objectness score and bounding box. are accurately depicted by the "SURF Feature Detection (Image 2),"
"Feature Matching," and "Filter Matches" process boxes

Advantages of Faster R-CNN for Object Detection

● End-to-End Training: Because of the unified design, end-

to- end training is possible, which simplifies and expe-
dites the training procedure.
● Improved Speed: Faster R-CNN achieves competitive ac-
curacy with reduced computation time compared to its
predecessors.
● Flexibility: The use of anchor boxes provides flexibility
in handling objects of different scales and aspect ratios.
Figure 4:Loss function formula[7]

Figure 5: flowchart for YOLO.[14]

The flowchart outlines the algorithm for object detection, specifically

using a Convolutional Neural Network (CNN) and incorporating steps
for bounding box prediction, class probabilities, and confidence scores.
in the flowchart. computing of the transformation matrix and evaluat- The algorithm begins by initialising the model with a pre-
ing match quality, through the "Compute Transformation" and "Eval-
uate Match Quality" process boxes. The final visualisation of
matched features is represented by the "Draw Matched Features"
process box, and the algorithm concludes with the "End" symbol.
SURF begins by calculating the integral
image of the input image. The integral im-
age allows for fast computation of Haar-
like features at different scales, which are
essential for detecting interest points.
SURF uses a Hessian matrix-based detec-
tor to identify interest points. The Hessian
matrix, a measure of image curvature, is
approximated using box filters for compu-
tational efficiency. Interest points are lo-
cated where the determinant of the Hes-
sian matrix is maximal. This approach al-
lows for scale-invariant detection by ap-
plying the Hessian detector at different
scales using a scale space representation.

Advantages of SURF for Object Detection

● SURF is very fast. It can detect and describe feature points

in an image in real time.
● SURF is very robust to changes in illumination, rotation,
and scale.
● SURF is accurate. It can accurately match feature points
between different images.
● SURF is easy to implement. There are many open source
SURF implementations available.

Disadvantages of SURF for Object Detection

● SURF is not as accurate as some newer object detection

algorithms, such as deep learning-based algorithms.
● SURF can be sensitive to noise and clutter.

YOLO: The YOLO algorithm takes an image as input and then uses
a simple deep convolutional neural network to detect objects in the
image. The architecture of the CNN model that forms the backbone
of YOLO is shown below.
YOLO divides an input image into an S × S grid. If the centre of an
object falls into a grid cell, that grid cell is responsible for detecting
that object. Each grid cell predicts B bounding boxes and confidence
scores for those boxes. These confidence scores reflect how confi-
dent the model is that the box contains an object and how accurate it
thinks the predicted box is.[7,12]
trained architecture, such as from the COCO dataset. The flowchart
accurately captures this initialization step with the "Initialize Con-
volutional Neural Network (CNN)" process box. Subsequently, the
input image is resized and normalised, aligning with the flowchart's
"Pass Image through CNN" process box, indicating the processing
of the image through the neural network. The forward pass through
the network, as mentioned in the algorithm's third step, is reflected
in the "Objectness Score and Class Prediction" process box. Fur-
thermore, the steps involving filtering overlapping boxes, calculat-
ing Intersection over Union (IoU), and obtaining the final set of
bounding boxes and class labels are captured by the "Non- Maxi-
mum Suppression (NMS)" and "Output Bounding Boxes and
Classes" process boxes.
Each grid cell predicts B bounding boxes,
each with five parameters: x, y, w, h, and
confidence. (x, y) represent the center of
the bounding box relative to the grid cell,
(w, h) represent the width and height of
the bounding box relative to the whole im- Figure 2: Image While Being Processed
age, and confidence represents the Inter-
section over Union between the predicted
Faster R-CNN:
box and the ground truth box.

After obtaining the bounding boxes and

their corresponding class probabilities,
NMS is applied to filter out overlapping
boxes and retain only the most confident
detections for each object. This step helps
to refine the results and avoid redundant
detections.

Advantages:
It is suitable for real-time applications mainly because the image is
looked at only once.

The training of the custom model for the YOLO algorithm is com-
paratively easier, where the modules can be trained end-to-end.

YOLO tends to be less prone to errors caused by background clut-

ter or overlapping objects. This is because it learns to directly pre-
dict bounding boxes from the full image, leveraging context from
the entire scene.

It has multiple versions like YOLOv2 to YOLOv8, with each suc-

ceeding version proving to be more reliable than the previous one.

.
III.
IV.RESULTS
Haar cascade method:
SURF:

on the specific application and environment. In high-traffic,

densely populated areas, the accuracy of algorithms like Faster
R-CNN may be more advantageous, while in less congested, ru-
ral settings, YOLO's speed advantage can significantly improve
the efficiency of traffic monitoring systems.

YOLO:

IV.DISCUSSION
Real-time processing plays a critical role in traffic monitoring
applications, where timely detection and response are crucial.
Algorithms such as YOLO (You Only Look Once) offer sig-
nificant advantages in terms of speed. YOLO processes im-
ages faster by detecting objects in a single forward pass
through the network, allowing it to achieve real-time perfor-
mance. This speed advantage is particularly beneficial for
traffic monitoring, as vehicles and pedestrians are constantly
moving, requiring prompt identification to prevent accidents,
optimize traffic flow, and enhance safety measures.However,
the trade-off between accuracy and speed is a key considera-
tion when selecting an algorithm for traffic monitoring. In ur-
ban environments, where the density of vehicles and pedestri-
ans is higher, accuracy becomes more critical. Algorithms
like Faster R-CNN, which may process images slower than
YOLO, offer higher precision in detecting smaller or partially
occluded objects, which are common in crowded cityscapes.
The slower processing speed is often an acceptable compro-
mise in exchange for more reliable detections, especially
when accuracy is prioritized over real- time performance.In
contrast, rural environments tend to have lower traffic density
and fewer complex scenes. In such scenarios, the speed of
YOLO may outweigh its slightly lower accuracy compared to
other algorithms. The faster detection allows for more effi-
cient monitoring, with fewer vehicles or obstacles to compli-
cate the detection process. The balance of speed and accuracy
in rural areas leans more towards speed, as there are fewer
objects to detect, and slight inaccuracies are less likely to lead
to critical errors.Ultimately, the choice of algorithm depends
V.CONCLUSION
Concluding which algorithm for object detection might be the best
is not that simple; it involves comparing a set of images and
running each algorithm mentioned above. Additionally, sim-
ulating them in real-time traffic environments of all kinds is
necessary. When comparing the time complexity of all the
mentioned algorithms, YOLO proves to be faster.

In this research paper, all the algorithms have been run on a

custom set of images and various databases, including the
COCO database. Upon doing so, the YOLOv3 Algorithm
appears to be the most fruitful, even though there are newer
versions available. This algorithm was primarily designed to
balance speed and accuracy. When run on different sets of
images, it achieved an accuracy of 58%, whereas Faster
RCNN only provided an accuracy of 40%, and Haar Cas-
cade fell behind with an accuracy of 30%, as it is an older
algorithm for object detection.

In terms of speed, YOLOv3 again proves to be the fastest,

followed by Haar Cascade, and then Faster RCNN. Faster
RCNN is relatively slow, especially on complex images
with various objects, involving multiple stages, including re-
gion proposal generation, which can impact its overall
speed, especially in real-time applications. In contrast, Haar
Cascade is faster than Faster RCNN, being simpler and able
to run easily on simple and moderate images.

Hence, for traffic management, it is wiser to incorporate the

use of the YOLO algorithm, preferably the newer versions
when compared to Faster RCNN and Haar Cascade. When
the need arises to run the algorithm with complex images,
the use of Faster RCNN is recommended, as even though it
might be slow, it provides better results in those cases. It is
also advisable to include the SURF algorithm to recognize
certain vehicles or to compare the picture of a number plate
with the number plate of a vehicle.

VI. REFERENCES
1. h t t p s : / / w w w . r e s e a r c h g a t e . n e t / p u b l i c a t i o n /
283556334_Multiple_Object_Detection_using_OpenCV_
on_a n_Embedded_Platform
2. https://ieeexplore.ieee.org/document/10098596

3.https://www.ijitee.org/wp-content/uploads/papers/v8i8/
H6362068819.pdf
4. Girshick R, Donahue J, Darrell T, Malik J (2014) Rich fea- [18].Ross Girshick1 Jeff Donahue1,2 Trevor Darrell1,2
ture hierarchies for accurate object detection and semantic Jitendra Malik1 1UC Berkeley and 2 ICSI. Rich feature hierar-
segmentation. In: Proceedings of the IEEE conference on chies for accurate object detection and semantic segmentation
computer vision and pattern recognition, pp 580–587

5. https://proceedings.neurips.cc/paper/2015/file/
14bfa6bb14875e45bba028a21ed38046-Paper.pdf

6. https://medium.com/@deepanshut041/introduction-to-surf-
speeded-up-robust-features-c7396d6e7c4e

7. Joseph Redmon* , Santosh Divvala*†, Ross Girshick¶ , Ali

Farhadi*†.

You Only Look Once: Unified, Real-Time Object

Detection.

[ 8 ] https://www .academia.edu/38877608/
Object_detection_using_Haar_cascade_Classifier

1. https://www.researchgate.net/figure/Car-detection-flow-
diagram_fig3_347959851

2. https://www.researchgate.net/figure/Faster-RCNN-
algorithm-flow-chart_fig2_341652820

3. https://www.researchgate.net/figure/Figure-1-Flow-
chart-of-SURF-algorithm_fig1_353367737

4. https://www.researchgate.net/figure/Pseudo-Code-of-
Yolo-v3-Algorithm-Model-Evaluation-This-involves-the-
testing-and_fig3_362062548

5. https://www.researchgate.net/figure/The-Flowchart-for-
training-of-YOLO_fig2_326535574

6. P. Viola, M. Jones, “Robust Real-Time Face Detection”,

International
Journal of Computer Vision 57(2), 137154, 2004

7. Y. Freund and R. E. Schapire, “A decision-theoretic

generalization of
on-line learning and an application to boosting” Journal of
computer and
system sciences, no. 55, pp. 110 – 140, 1997

8. F. Fleuret and D. Geman. Coarse-to-fine face detection.

Int.
J. Computer Vision, 2001 [17].github.com/wfs123456/CC-

Trans
19. github.com/bertsky/ocrd_detectron2

20. https://pyimagesearch.com/2021/04/12/opencv-haar-
cascades/

[21] https://chowdera.co
m / 2022/206/202207251843390865.html

22. https://www.researchgate.net/figure/T-raining-of-
Faster-
R-CNN-Pseudo-code-of-faster-RCNN_fig6_343438688
23. https://www.researchgate.net/figure/Faster-RCNN-
algorithm-flow-chart_fig2_341652820

24. https://www.researchgate.net/figure/Figure-1-Flow-
chart- of-SURF-algorithm_fig1_353367737
25. https://www.researchgate.net/figure/Speeded-Up-
Robust-
Feature-SURF-algorithm_fig1_369092276
26. https://ieeexplore-ieee-org.egateway.chennai.vit.ac.in/
stamp/stamp.jsp?tp=&arnumber=9993862