Review On different Vehicle Detection Technique

(Harsh Kumar Singh, Akshay Shukla, Abhishek Singh, Krish Chaudhary)

Abstract: methods created so far. The pre-processing,

feature extraction, and classification phases
Vehicle detection is a crucial problem in the
of vehicle detection will all be covered in
field of computer vision and is utilised in a
this review paper. We will also discuss each
number of applications, including
technique's benefits and drawbacks and
surveillance, autonomous driving, and
contrast how well it performs on various
traffic monitoring. Several methods have
datasets. This review paper's main goal is to
been put forth over time to identify
assist researchers and practitioners in better
automobiles in pictures and movies. An
understanding the several vehicle detection
overview of several vehicle detection
strategies that are now available, which will
methods and their applications is given in
eventually help them gravitate towards the
this review study. We begin by talking about
best methodology for their particular
techniques like background removal, the
application. Over the several years, there are
Haar Cascade Classifier, HOG-based
different techniques that have been
algorithms, etc. Then, we go to more
proposed ranging from computer vision-
modern deep learning-based techniques
based techniques to deep learning-based
like SSD, YOLO, and Faster R-CNN. On
techniques. The most common or the
diverse datasets, we examine how well
traditional technique are the computer
these approaches work and assess their
vision-based technique which include
advantages and disadvantages. We
techniques like the Haar Cascade Classifier,
conclude with some recommendations for
HOG-based algorithm. Another one is the
future study in the area of vehicle
Deep Learning based Techniques which
detection. Overall, this review article offers
includes technique like YOLO (You Only Look
a thorough examination of contemporary
Once) which is the most popular deep
methods for vehicle identification and may
learning-based technique. The other
be an invaluable tool for academics and
technique is based on region based deep
professionals involved in this field.
learning which include faster R-CNN
1 Introduction: technique which attains the most accurate
one which we will be discussing further. The
Significant progress has been made in the
next technique on the list is the Motion
field of computer vision recently, especially
based techniques like Background
in the area of object detection. One of the
subtraction which is mostly preferred Video
most important uses of object detection is in
based vehicle detection. The last one on the
vehicle detection, which is used often in
list is the stereo based techniques that
traffic surveillance, autonomous driving, and
includes SSD (Single Shot Detector), these
intelligent transportation systems. In the
methods are less affected by occlusion,
literature, a number of vehicle identification
weather conditions. Overall, the goal of this
strategies have been put out, each having its
review article is to progress the
own advantages and disadvantages. This
development of more precise and effective
review study aims to give a thorough
vehicle detection systems by contributing to
overview of the various vehicle detection
existing research in the field of computer
vision. We use the data stats to compare
these different techniques at different
grounds and providing the graphical
representation for the same. Also, we will be
discussing the problems that are faced by
each technique.
2 Literature Review:

Table 1 Author identification features and techniques

Authors Features Technique

Dalal N This research paper provides a brief on Histogram Oriented

et al. [2] computer vision-based techniques which Gradient (HOG),
include Histogram oriented gradient Computer Vision Based
technique, they provide us the working and
the algorithm for HOG
Viola et al. A set of features are used by the machine Haar Cascade Classifier
[1] learning-based object identification
technique referred to as Haar cascade to
identify objects in an image.
Ren S et al. The deep learning-based object Faster Regions with
[4] identification technique known as Faster convolutional neural
R-CNN was put forward by Ren’s in their networks (faster R-CNN)
paper titled "Faster R-CNN: Towards Real-
Time Object Detection with Region
Proposal Networks." In order to create
object proposals, the authors introduced a
region proposal network (RPN), that is
used in conjunction with a region-based
convolutional neural network (R-CNN) to
recognise objects.
Zang Y The authors introduced the importance of Support vector machines
et al. [10] pedestrian detection and its uses in (SVM), AdaBoost, and
numerous domains. They draw attention to deep learning-based
the difficulties in accurately detecting methods (HOG)
pedestrians, such as changes in appearance,
occlusion, scale, and complicated
backgrounds.
Liu W et A single feed-forward convolutional neural Deep Learning Based
al. [6] network is used by SSD, a deep learning- Technique, Single Shot
based object detection approach, to detect Detector (SSD)
objects. Liu et al. presented the method in
their publication titled "SSD: Single Shot
Multibox Detector."
Redmon et One convolutional neural network is used Deep Learning based
al. [8] for object detection in the YOLO deep techniques, You Only
learning-based object detection method. Look Once (YOLO)
Redmon et al. made the suggestion in their
study titled "You Only Look Once:
Unified, Real-Time Object Detection."
Authors Features Techniques

Yiqin Cao In the research paper authors discussed Image extraction,

et al. [18] about the simplest technique that is Background Subtraction
background subtraction. Using a
comparison between the current frame and
a backdrop model, the traditional computer
 vision technique known as background
subtraction detects objects.
Chen et al. The authors investigate the idea of using Training deep learning
[3] massive amounts of web data to train models and Convolutional
convolutional neural networks (CNNs) in a neutral network (CNN)
web-supervised fashion. The authors
discuss the difficulty of getting a
substantial amount of labelled training
data, which is essential for successfully
training deep learning models.
Xia Y et al. The authors propose a method for Colour Feature
[5] accurately detecting cars in difficult Extraction, Texture
backdrops and lighting situations by Feature Extraction
utilising colour and texture signalling. The
primary goal of this research is to raise the
precision and dependability of vehicle
detection systems in challenging
environments.
Yingfeng The authors present a novel method for Deep Learning and
et al. [7] increasing the precision of vehicle Feature Extraction
recognition in low-light or night time Classification
conditions that blends visual saliency and
deep learning approaches. The programme
shows good results in detecting
automobiles under difficult illumination
situations by leveraging the distinctive
properties of vehicles and utilising deep
learning.
Mark The author highlighted the significance of Computer Vision Based,
Bugeja vehicle detection methods for IP camera Image Capturing and
et al. [11] video feeds in the context of intelligent Monitoring
transport systems. The authors evaluate
many methods, from deep learning-based
algorithms to conventional computer vision
techniques.
Huansheng The author of this research paper Deep Learning and
et al. [12] investigates the creation of a highly Computer Vision Based
advanced computer vision system for the Techniques.
detection and counting of vehicles in
highway scenarios. The research focuses
on using deep learning methods.
Authors Features Techniques

Wang X The challenges and improvements in Image Capturing,

et al. [13] background subtraction for image and Computer Vision and
video processing are the main topics of this Background Subtraction
paper. In computer vision and image
processing, background subtraction is a
fundamental operation that is frequently
 applied to surveillance systems, object
tracking, and activity recognition.
Sriashika The research's sole goal was to examine Traditional method with
[14] every method and technique for vehicle CCTV Detection like
detection currently in use, with the sole Haar cascade, HOG and
goal of implementing them at the Background Subtraction
prestigious Lovely Professional University
in Punjab, India. The author only chose the
optimal way to build the vehicle detection
and recognition system on a busy campus
like LPU after carefully examining all the
already-used techniques.
Sourajit The goal of the study is to present a Deep Learning Based,
et al. [15] thorough overview of the improvements in Support Vector Machines,
vehicle detection and classification Single Shot Detector
approaches over the past ten years. It (SSD), You only Look
provides a summary of the many Once (YOLO)
approaches and algorithms used for vehicle
identification and classification in a range
of situations.
Shikha PredictV, a predictive strategy put forth by Occlusion Handling,
et al. [16] the authors, seeks to address the difficulties Motion Modelling,
provided by occlusions and boost the Machine Learning
precision of vehicle prediction systems.
When things, especially automobiles, are
partially or entirely obscured from vision,
this is known as an occlusion, which makes
it impossible to precisely forecast how they
will behave in the future.
Joseph R The authors propose an innovative Deep Learning, You Only
et al. [21] technique for real-time object detection. By Look Once (Yolo)
combining object detection and
classification into a single network, the
YOLO framework described in the study
provides a unified, effective, and accurate
approach.
3 Vehicle Detection Techniques:
Vehicle detection is a critical problem in principle of background subtraction is to
contemporary traffic monitoring and detect the pixels that are noticeably
control systems. Although it is employed different from the surrounding area by
for these purposes, it is also used for the e- comparing each frame of the video with a
chalan system, toll booths, and even object reference or background picture. These
detection in automobiles. Automated pixels are then categorised as being in the
driving, intelligent transportation systems, forefront, or the scene's moving objects.
and traffic monitoring all require accurate
Background subtraction may be done
and trustworthy vehicle detection. There
using a variety of methods, from
are several vehicle detection techniques
straightforward thresholding approaches to
that are used in our daily life. Let’s find
more intricate statistical and machine
about all the techniques there working,
learning-based approaches. Among the
there application and compare at different
often-employed methods are:
parameters.
Thresholding: For this approach, the
3.1 Background Subtraction:
difference between the current frame and
the background picture is the threshold.
Foreground pixels are defined as those
having a difference beyond the threshold.
Running average: To build a background
model, this technique computes a running
average of the previous frames in the video
stream. The pixels with a difference larger
than a threshold are categorised as
foreground after computing the difference
between the current frame and the
background model.
Gaussian mixture model: This technique
uses a combination of Gaussian
distributions to describe the pixel
intensities in the foreground and
Fig1 (Background subtraction Gaussian background. The model is then updated as
mixture model sourced from [13]) a result of the algorithm's estimation of the
Gaussian distributions' parameters using
With this method, every frame of a video
the values of the pixels in the current
is compared to a reference frame that
frame.
depicts the background without any
moving parts. Any pixels that vary Deep learning-based techniques: Several
between two frames are seen as being a deep learning-based techniques for
component of a moving object, such as a background removal have been presented
moving automobile. Although this method as a result of recent advances in deep
is quick and easy, changes in lighting, learning. These techniques categorise the
shadows, and camera movement may pixels as being in the foreground or
create an impact. The fundamental background by learning characteristics
directly from the video frames using determine if an item is present or absent in
convolutional neural networks (CNNs). a picture. The Haar cascade classifier
trains a classifier with a collection of
Although background removal can be a
positive and negative samples that can
quick and efficient method for spotting
subsequently be used to identify objects in
moving items in a video stream, it can
a picture. The Haar cascade classifier,
encounter difficulties due to lighting
however, is renowned for its low accuracy
variations, shadows, and occlusion. The
and sensitivity to changes in light
effectiveness of the procedure is also
conditions. With this method, items in a
influenced by the standard of the
frame of an image or video are identified
background model and the selection of the
using a set of attributes. The difference
threshold or model parameters.
between the pixel totals in the white and
Therefore, in order to improve the black rectangles of an image is used to
fundamental technique's efficacy in compute these characteristics. In
addressing these difficulties, researchers complicated settings, recognising
have suggested a number of tweaks and automobiles may not be as successful
improvements. using Haar cascades, which are good at
detecting items with distinctive traits like
faces. The main principle of the Haar
Cascade Classifier is to classify sub-
regions of an image as either belonging to
the item of interest or not by using a series
of straightforward Haar-like
characteristics. Simple rectangular
characteristics called Haar-like features
may be calculated rapidly and effectively.
The difference in the sum of the pixel
brightness in two rectangular areas of an
image constitutes what these
characteristics are known as. The picture is
scanned at various sizes and locations by
the Haar Cascade Classifier, which then
determines whether or not each sub-region
contains the item of interest. AdaBoost is a
machine learning technique that is used to
choose the most discriminating features
3.2 Haar Cascade Classifier:
and train the Haar Cascade Classifier. By
Fig 2(Haar Cascade Classifier: object giving the incorrectly classified samples
detection. Sourced from [1]) greater weights and retraining the classifier
A machine learning-based object using these examples, AdaBoost combines
recognition system called the Haar cascade numerous weak classifiers into a powerful
classifier employs a collection of classifier. The final classifier consists of a
predefined features to find things. It makes series of these boosted classifiers, where
use of Haar-like properties to find items in each step applies a subset of the Haar-like
a picture. Rectangular patterns known as features to the picture and advances the
Haar-like features are employed to sub-regions that successfully classify to
the following stage. Face detection is one cells, calculating the gradient magnitude
of the most popular uses for the Haar and orientation in each cell, and then
Cascade Classifier, where it has been averaging the gradient information over a
demonstrated to attain excellent accuracy block of cells, the HOG features are
and real-time performance. To learn the computed. HOG is a well-liked feature
distinguishing characteristics for face extraction technique for detecting vehicles
identification, the classifier is trained using and pedestrians. It is renowned for its
an extensive database of both positive and precision and resistance to variations in
negative face samples. The Haar Cascade lighting conditions. Using the distribution
Classifier has undergone a number of of gradient orientations as a basis, HOG's
extensions and modifications since it was primary concept is to extract features from
first introduced in order to enhance its a picture. The technique splits the picture
functionality and overcome its drawbacks. into tiny blocks by computing the gradient
Use of more sophisticated features like magnitude and orientation for each pixel in
LBP and HOG as well as deep learning- the image. The gradient orientations inside
based techniques like CNNs are a few of each block are quantized into a collection
the most well-liked adaptations. of discrete directions, and an orientation
histogram is generated. The histograms
3.3 Histogram Oriented Gradients:
from all the blocks in the image are then
concatenated to provide the final feature
descriptor. Compared to other feature
descriptors like SIFT and SURF, HOG
offers a number of benefits. First off, it is
computationally effective, making real-
time applications possible. Second, it is
trustworthy in a variety of lighting
circumstances as it is resistant to changes
in illumination and contrast. Third, it is
highly suited for jobs involving object
identification and recognition since it can
record shape and edge information about
Fig 3 (Histogram of Oriented objects. In pedestrian detection, one of
Gradients feature) sourced from [2] HOG's most popular uses, it has been
demonstrated to deliver cutting-edge
performance. To determine whether the
A feature extraction technique called image contains a pedestrian or not, a
Histogram of Oriented Gradients (HOG) classifier like a support vector machine
counts the instances of gradient orientation (SVM) uses the HOG feature descriptor as
in certain localised areas of a picture. In input. HOG has undergone several
order to detect objects in a frame of an extensions and modifications since it was
image or video, this method computes first introduced in order to enhance its
histograms of gradient orientations. It is functionality and fix its drawbacks. Some
good at identifying items with of the prominent developments include
recognisable edges, such as cars, but may multi-scale HOG, which computes the
struggle to identify objects with less clear HOG features at several scales to
edges. By segmenting the picture into tiny accommodate scale variance, and HOG
with spatial pyramids, which captures the target items. Although deep learning-
spatial arrangement of the features in the based approaches need a lot of training
picture. HOG is a strong and useful feature data and computer power, they can be
descriptor for tasks involving the detection useful at identifying a variety of objects,
and identification of objects. It is a popular
such as cars, people, and animals. The
option in many applications due to its
objectness score and offsets for each
tolerance to changes in illumination and
anchor box in relation to a set of default
contrast, capacity to record object shape
boxes are then predicted by the RPN. The
and edge information, and computing
Fast R-CNN network is then provided with
the RPN's suggestions for object
categorization and bounding box
regression. The suggested areas are fed
into the Fast R-CNN network, which
produces a probability distribution across
the object classes and enhanced bounding
boxes for each suggestion. In comparison
to earlier object detection algorithms like
R-CNN and Fast R-CNN, Faster R-CNN
provides a number of advantages. The
RPN and Fast R-CNN networks may be
trained together in a single training session
since it is end-to-end trainable. Second, it
performs state-of-the-art object detection
efficiency.
on a number of benchmark datasets,
3.4 Faster R-CNN: including COCO and PASCAL VOC.
Thirdly, it is computationally effective and
Fig 3(Faster R-CNN sourced from [10] capable of real-time object detection.
3.5 YOLO (You Only Look Once):
Faster R-CNN is a deep learning-based
object identification system that generates
object suggestions using a region proposal
network (RPN) and classifies them using a
convolutional neural network (CNN). The
CNN categorises each proposal into one of
the object classes after the RPN creates a
collection of object proposals that are
likely to include an object. Faster RCNN is
recognised for its accuracy, but it takes a
lot of computing time and resources to
process. These methods identify and
categorise items in an image or video
frame using deep neural networks. A vast
collection of photos is used to train the
network, which then learns to recognise Fig 4 (You Only Look Once YOLO
traits that are crucial for recognising the image processing sourced from [21]
 though. For instance, it could struggle to
A single neural network is used by the You
Only Look Once (YOLO) deep learning-
based object recognition technique to
forecast the bounding boxes and class
probabilities of objects in an image. In
order to estimate the bounding box and
class probabilities for each grid cell,
YOLO splits the picture into a grid. Yolo is
a popular real-time item identification tool
because of its speed and accuracy. Deep
learning-based object identification
algorithms are frequently used in
techniques for finding cars in aerial
photography. To increase accuracy, they
could also make advantage of other
features like object tracking or motion
detection. These methods have the identify small or low-contrast objects, and
potential to be useful for finding it might produce false positives or false
automobiles across wide areas, but they negatives in complicated scenarios.
could be constrained by the resolution of
the images. In order to forecast the
bounding boxes and class probabilities of 3.6 SSD (Single Shot Detector):
the objects included in each cell, the
YOLO method splits an input picture into
a grid of cells and applies a convolutional
neural network to each cell. The x and y
coordinates of the box's centre, the box's A well-liked deep learning-based object
width and height, and a confidence score recognition system called Single Shot
that indicates how likely it is that the box Detector (SSD) is created to quickly and
will include an item make up each accurately identify objects in photos. Wei
bounding box prediction. The likelihood Liu and colleagues initially described the
that the item in the box belongs to a certain technique in their 2016 publication "SSD:
class is represented by the class Single Shot MultiBox Detector." The SSD
probabilities. The YOLO method is approach is based on a single neural
superior to conventional object detection network that receives an image as input
algorithms in a number of ways. As and produces a collection of bounding
opposed to other algorithms like Faster boxes and associated class scores for each
RCNN, it doesn't need a separate area object that is recognised in the picture. A
proposal phase, hence it is very quick. multi-task loss function that
Additionally, it is quite precise in detecting simultaneously optimises the localization
Fig 5 (Singletasks
and classification ShotisDetector) SSD [6]
used to train the
objects of all shapes and sizes. Thirdly, it
is more effective than other algorithms that network. The use of several convolutional
require numerous runs to identify items feature maps at various scales to identify
since it can find multiple things in a single objects of various sizes is one of the
pass. YOLO has certain restrictions, essential components of SSD. This
enhances the algorithm's effectiveness in
managing scale variation and enables it to
recognise objects of various sizes and
aspect ratios. Anchor boxes are pre-defined
bounding boxes that are used with SSD
and are positioned at various sizes and
places on the picture. In order to precisely
localise the objects in the picture, the
network learns to anticipate the offset and
size of these anchor boxes. It is also Occlusions: Vehicles may be completely or
renowned for having a quick time to partially blocked from view by other
inference, which makes it appropriate for automobiles or objects, making it
real-time applications. However, SSD can challenging to identify them. In these
still run into issues like occlusion, situations, occlusion-tolerant techniques
appearance variation, and object distortion, like CNNs are helpful. When a part of the
just like any other object identification car is blocked by another item in the scene,
method like a tree or a building, this is known as
partial occlusion.

4 Problems and Challenges: Changes in Vehicle Appearance and

Orientation: Vehicles can appear in a
Finding and locating vehicles in an image variety of shapes, sizes, and colours, as
or video sequence is the problem definition well as be oriented in a variety of
for the vehicle detection technique. This directions, which can make it challenging
involves developing algorithms and to notice them. These changes can impact
methods that can recognise vehicles based feature-based methods like HOG and Haar
on their outside appearance, such as their cascades. Vehicles, for instance, can have a
size, colour, and shape. For a range of variety of colours, and these colours might
applications, including traffic monitoring, change depending on things like lighting,
surveillance, and autonomous driving, it is weather, and time of day.
critical to precisely and quickly identify
the presence and location of vehicles. The False positives: False findings are
issue with vehicle detection technology is produced when non-vehicle items are
that it must be able to detect and track mistakenly identified as vehicles. By
vehicles in a range of locations and making the feature selection and
circumstances, including on roadways, in classification of the algorithm better, this
parking lots, and at intersections. Different issue can be solved. When the algorithm
vehicle detection techniques face several wrongly classifies an object as a target
common problems that impact their object when it is not, false positives
performance and accuracy, as well as happen. False positives in the instance of
several unique challenges that are specific vehicle detection can happen when the
to each technique. Some of the common algorithm incorrectly classifies non-
problems for different vehicle detection vehicle items as cars, such as shadows,
techniques are: trees, or pedestrians.

Changes in weather and lighting: In Speed and computing complexity:

outdoor settings, specifically, changes in Techniques for real-time vehicle
weather and lighting can make it identification applications must be quick,
challenging to identify cars. Techniques effective, and computationally simple.
resistant to changes in illumination and Techniques appropriate for real-time
contrast, including HOG and CNNs, can applications include background removal
be used to solve this issue. and Haar cascades. In addition to these
widespread issues, several vehicle
detection algorithms also confront unique
difficulties.
5Design and Comparison:  Computational Complexity: The

To compare different vehicle detection complexity of computation refers to how

techniques, it involves various terminology much computing power the vehicle

which includes speed, accuracy, versatility detection method needs. It can be

and real time processing and on basis of measured by taking into account elements

these we compared them and represented like the quantity of operations, memory

each technique in graphical format. So, lets use, and required computing power. Faster

understand all of the terminology and their detection is correlated with lower

criteria for each technique. computational complexity.

YOLO- You Only Look Once

5.1 Speed
SSD- Single Shot Detector
By taking into account a variety of factors, HOG- Histogram Oriented Gradient
speed can be defined in various vehicle BG SUB- Background Subtraction
detection algorithms. Here are some typical HAAR- Haar Cascade Classifier
standards for determining speed in vehicle FasterR-CNN- Region convolutional neural network
detection methods:

 Processing Time: The processing time is

the amount of time the algorithm needs to
analyse a given input image or video frame
and find any moving vehicles. Typically, it
is expressed as a number of frames per
second (fps) or milliseconds (ms). Faster
detection is indicated by a lower
processing time.
 Real-Time Performance: The capability
of a vehicle detection technique to process
image or video frames in real-time without
Fig 6 Graph comparing speed for
substantial lags is referred to as real-time
different vehicle detection
performance. In order to match the speed techniques
of the incoming data, real-time
performance is often described as
accomplishing a specific number of frames
per second (for example, 30 fps).
 5.2 Accuracy:  False Positive Rate: The frequency of
false detections, or times when the
For defining accuracy for different vehicle
algorithm incorrectly classifies a region
detection algorithms, multiple
as a vehicle when it is not, is measured
considerations must be taken into mind.
by the false positive rate. Lower false
Here are some typical standards for
positive rates represent the technique's
defining vehicle detection methods'
capacity to prevent pointless detections or
accuracy:
false alarms, which indicates improved
 Detection accuracy: describes the accuracy.
precision with which a vehicle detection
YOLO- You Only Look Once
technique can locate and identify vehicles
SSD- Single Shot Detector
inside a frame of an image or video. HOG- Histogram Oriented Gradient
Metrics like recall, F1-score, and BG SUB- Background Subtraction
precision are frequently used to gauge it. HAAR- Haar Cascade Classifier
The proportion of accurately detected FasterR-CNN- Region convolutional neural network

vehicles out of all the detected vehicles is

called precision, and the proportion out of
all the ground truth vehicles is called
recall. The harmonic mean of recall and
precision, known as the F1-score,
provides a balanced indicator of
accuracy.
 Localization Accuracy: The ability of a
vehicle detection technique to precisely
find the bounding box or region of
interest surrounding the identified cars is
referred to as localization accuracy.
Metrics like intersection over union (IoU)
or mean average precision (mAP) are Fig 7 Graph comparing accuracy
for different vehicle detection
frequently used to measure it. IoU
quantifies how closely the predicted and
actual bounding boxes overlap, whereas
mAP takes the precision-recall curve for
various IoU levels into account.
5.3 Versatility: well at various scales, resolutions, and
processing capacities. Without sacrificing
Numerous factors that evaluate a technique's
precision or speed, a flexible approach
capacity to perform well in multiple
should be usable on a variety of hardware
contexts and adapt to different situations are
platforms, from resource-constrained
taken into account while defining versatility
embedded systems to high-performance
in relation to various vehicle detection
computer sets.
techniques. The following are some typical YOLO- You Only Look Once

standards for defining the adaptability of SSD- Single Shot Detector

vehicle detection techniques: HOG- Histogram Oriented Gradient

BG SUB- Background Subtraction

 Generalisation Across Datasets: Versatile
HAAR- Haar Cascade Classifier
vehicle detection methods need to
FasterR-CNN- Region convolutional neural network
function well when applied to various
datasets. They ought to be able to adapt
to different vehicle shapes, sizes,
orientations, lighting, weather, and
backdrops. It is essential to evaluate the
technique's precision and consistency
across various datasets to determine how
versatile it is.
 Adaptability to Changing Environments:
Versatile approaches must be able to
change with the environment while still
maintaining accurate detections. This
includes dealing with changes in weather,
such as day-to-night transitions, rain, and
fog, as well as dynamic scenes with
moving objects or camera motion.
Versatility is demonstrated by the
capacity to manage various changes
without noticeably degrading
performance.
 Scalability: A vehicle detection technique
is said to be scalable if it can function
 5.4 Real Time Processing: adjust and maintain constant detection
capabilities even when the frame rate or
While defining real-time performance for
processing resources fluctuate can be
various vehicle recognition systems, it is
used to define real-time performance.
important to take into account a number of
factors that evaluate how quickly the
YOLO- You Only Look Once
methodology can process image or video SSD- Single Shot Detector
frames. Here are a few typical standards HOG- Histogram Oriented Gradient
for measuring real-time effectiveness in BG SUB- Background Subtraction
vehicle detection technique HAAR- Haar Cascade Classifier

FasterR-CNN- Region convolutional neural network

 Frame Rate: The term "frame rate"
describes how many frames per second
(fps) a technique can handle in real time.
Real-time performance often calls for
reaching a frame rate that is equal to or
higher than the input data's frame rate. A
real-time approach should process at least
30 frames per second, for instance, if the
input video was recorded at 30 frames per
second.
 Processing Time per Frame: The time
required by a vehicle detection system to
analyse and detect vehicles inside a
single frame is known as processing time
per frame. Usually, it is expressed in
milliseconds (ms). To ensure prompt
Fig 9 Graph comparing Real Time for
detection and action, real-time different
vehicle detection techniques
performance demands that the processing
time per frame be within a predetermined
range.
 Adaptive Framerate: In some cases, the
input data may be subjected to shifting
computational resources or varying frame
rates. The capability of the method to
6Result and Discussion: Real-time applications or complex

Table 2 Comparison scenarios are not recommended.

Vehicle
Accura Object
Detection Speed Training Usage
cy Detection
Technique
You Only Look Real-time
Fastest High High Complex
Once (YOLO) applications
Less
Single Shot Slower Slightly Object
complex
Detector than High lower than recognition
than
(SSD) YOLO YOLO in images
YOLO
Histogram of
Object
Oriented
Slow Medium Low Simple recognition
Gradients
in images
(HOG)
Static
Background
Fast Low Low Simple camera
Subtraction
applications
Face
Haar Cascade
Fast Low Low Simple detection in
Classifier
images
Object
Faster R-CNN Medium High High Complex detection in
images

Comparison involves evaluating the

benefits, drawbacks, and effectiveness of
various vehicle detection techniques. Here
is a summary comparison of a few well - Histogram-Oriented Gradient (HOG):
known vehicle detection methods: Strengths: capable of identifying cars based
HaarCascade: on their shape and edge characteristics,
Strengths: quick processing, effective decent accuracy, and acceptable processing
frontal vehicle detection, ideal for areas speed.
with limited resources. Limitations: susceptible to changes in
Limitations: in accuracy and resistance to illumination and partial occlusions; may
changes in scale, position, and occlusions. have trouble with intricate backgrounds
and crowded situations. vehicles; real-time performance capability;
Faster R-CNN (region-based suitable for scenarios with static
convolutional neural network) processing backgrounds.
speed: Limitations: sensitive to variations in
lighting conditions ; may have
difficulties with complex scenes and
occlusions;
Strengths: high localization accuracy,
resilience, and handling of difficult limited accuracy compared to more
situations and occlusions sophisticated techniques.
Limitation: relatively slower processing It is crucial to remember that the
speed as compared to certain other effectiveness and applicability of these
techniques; high computer resource strategies might change depending on a
requirements. variety of elements, including the
SSD (Single Shot MultiBox Detector): implementation's details, the dataset being
Strengths : good balance between accuracy utilised, and the application's needs. The
and processing speed ; effective at exact requirements of the vehicle detection
detecting vehicles of different scales; real- situation, such as the need for accuracy,
time performance capability. real-time performance, processing
Limitation : can struggle with small resources, and environmental complexity,
objects and occlusions ; may have lower will determine which technique is most
localization accuracy compared to some appropriate. The best method for a given
other techniques. application can be chosen by evaluating
YOLO (You Only Look Once): and comparing techniques based on these
Strengths : very fast processing speed , criteria.
real-time performance capability , good
accuracy, effective at handling complex 7 Conclusion:
scenes and occlusions. In conclusion, the analysis of various
Limitation : can have lower localization vehicle detection methods reveals a
accuracy compared to some other complex environment with a range of
techniques ; may struggle with small advantages and disadvantages. Although
objects and crowded scenarios. Haar Cascade and HOG algorithms
Background Subtraction: handle data quickly, they might not be
Strengths: Effective at detecting moving accurate or robust enough. Higher
accuracy and robustness are demonstrated Vision and Pattern Recognition,
by faster R-CNN and SSD, but at the CVPR 2005, 1, 886-893
expense of greater processing demands. [3] Chen, X., & Gupta, A. (2015).
Yolo is notable for its effectiveness in Webly supervised learning of
complex situations and real-time convolutional networks.
performance. Although background Proceedings of the 2015 IEEE
subtraction is excellent at spotting International Conference on
moving vehicles, it is inaccurate and has Computer Vision, ICCV 2015,
trouble managing occlusions. The best 1431-1439.
strategy to use will depend on factors [4] Ren, S., He, K., Girshick, R.,
including accuracy, real-time & Sun, J. (2015). Faster R-CNN:
performance, computer power, and Towards real-time object
environmental complexity. Future detection with region proposal
developments ought to concentrate on networks. Advances in Neural
enhancing the compromise between Information Processing Systems,
precision, speed, and flexibility for 28, 91-99.
trustworthy vehicle detection. [5] Xia, Y., Zhang, W., Wang,
X., & Yang, J. (2015). Vehicle
8 References: detection in complex scenes using
Here are some references for research color and texture features. IEEE
papers on different vehicle detection Transactions on Intelligent
techniques: Transportation Systems,
[1] Viola, P., & Jones, M. (2001). 16(3), 1321-1333.
Rapid object detection using a [6] Liu, W., Anguelov, D., Erhan,
boosted cascade of simple D., Szegedy, C., Reed, S., Fu, C.
features. Proceedings of the IEEE Y., & Berg, A. C. (2016). SSD:
conference on computer vision Single shot multibox detector.
and pattern recognition, 1, I-511 Proceedings of the 2016 European
[2] Dalal, N., & Triggs, B. (2005). Conference on Computer Vision,
Histograms of oriented gradients ECCV 2016, 21-37.
for human detection. Proceedings [7] Yingfeng Cai,1 Xiaoqiang
of the 2005 IEEE Computer Sun,2 Hai Wang,2 Long Chen,1
Society Conference on Computer and Haobin Jiang2(2016). Night-
Time Vehicle Detection
Algorithm Based on Visual Transport Infrastructure and
Saliency and Deep Learning Systems
[8] Redmon, J., Divvala, S., [12] Huansheng Song, Haoxiang
Girshick, R., & Farhadi, A. Liang* , Huaiyu Li, Zhe Dai and
(2016). You only look once: Xu Yun.(2019). Vision-based
Unified, real-time object vehicle detection and counting
detection. Proceedings of the system using deep learning in
2016 IEEE Conference on highway scenes
Computer Vision and Pattern [13] Wang, X., & Li, Q. (2019). A
Recognition, CVPR 2016, 779- comprehensive study of
788. background subtraction
[9] Zhang, Y., & Yang, Y. (2016). algorithms based on deep
A survey on recent advances in learning. Journal of Real-Time
pedestrian detection. Proceedings Image Processing, 16(5), 1673-
of the 2016 23rd International 1689.
Conference on Pattern [14] Sriashika Addala(2020).
Recognition, ICPR 2016, 2613- Vehicle Detection and
2620. Recognition
[10] Zhang, L., Lin, L., Liang, X., [15] Sourajit Maity,
& He, K. (2018). Is faster R-CNN Avirup Bhattacharyya,
doing well for pedestrian Pawan Kumar Singh,
detection? Proceedings of the Munish Kumar,
2018 IEEE Conference on Ram Sarkar(2022). Last Decade
Computer Vision and Pattern in Vehicle Detection
Recognition, CVPR 2018, 5225- and Classifcation:
5233. A Comprehensive Survey(2022).
[11] Mark Bugejaa, , Alexiei Proceedings of the 2022 to
Dinglib, Maria Attarda , Dylan International Center for
Seychel (2019). Comparison of Numerical Methods in
Vehicle Detection Techniques Engineering (CIMNE) 2022.
applied to IP Camera Video Feeds [16] Shikha Tuteja, Shashi
for use in Intelligent Transport Poddar, Divya Aggarwal,& Vinod
Systems. Proceedings of the 2019 Aggarwal(2022). PredictV: A
International Congress on Vehicle Prediction Scheme to
Circumvent Occluded Frame. [21] Joseph Redmon∗ , Santosh
Proceedings of the 2022 IEEE Divvala∗†, Ross Girshick¶ , Ali
Computer Society Conference on Farhadi. You Only Look Once:
Computer Vision. Unified, Real-Time Object
[17] Shaik Hafeez Shaheed, Detection
Rajanala Sudheer, Kavuri Rohit,
Deepika Tinnavalli, and Shahana These references cover a range of
Bano.(2022) Vehicle Spotting in vehicle detection techniques,
Nighttime Using Gamma including Haar cascades, HOG,
Correction deep learning-based methods such
[18] Yiqin Cao* , Xiao Yun, Tao as YOLO, SSD, and Faster R-
Zhong and Xiaosheng Huang. CNN, as well as pedestrian and
Research on Vehicle Detection aerial vehicle detection.
and Tracking Algorithm Based on
the Methods of Frame Difference
and Adaptive Background
Subtraction Difference
[19] Ke Li,1,2 Rongchun Deng,3
Yongkang Cheng,3 Rongqun
Hu,4 and Keyong Shen(2022).
Research on Vehicle Detection
and Recognition Based on
Infrared Image and Feature
Extraction
[20] Samuel Olugbade 1 , Stephen
Ojo 2 , Agbotiname Lucky Imoize
3,4,* , Joseph Isabona 5 and
Mathew O. Alaba 6(2022). A
Review of Artificial Intelligence
and Machine Learning for
Incident Detectors in Road
Transport Systems.