Helmet Detection Using Yolo V3 Technique

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 6 Issue 2, January-February 2022 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
@ IJTSRD | Unique Paper ID – IJTSRD49341 | Volume – 6 | Issue – 2 | Jan-Feb 2022 Page 986
Helmet Detection Using Yolo -V3 Technique
Vijay Khare, Pratiroop Puri, Kratik Chaurasia
Jaypee Institute of Information Technology, Noida, Uttar Pradesh, India
ABSTRACT
Bike mishaps have been quickly enlarged over time in many
countries. A helmet or a protecting cap is the main safety equipment
of motorcycle riders, and two wheelers, however many driver’s
abandonment wearing helmets. The main outcome of wearing a
helmet is to protect the head of a person travelling on two-wheelers
just in case of a major or minor accident or fall from a running bike.
Because of different social and monetary elements, this sort of
vehicle is turning out to be progressively famous. The head protector
is the fundamental security gear of motorcyclists; however, anyway
numerous drivers don't utilize it. This paper proposes a framework
for identification of motorcyclists without helmets. For this, we have
applied the round Hough change and the Histogram of Oriented
Gradients descriptor to remove the picture credits. Then the YOLO
v3 was utilized and acquired outcomes. The system has given an
average recognition accuracy of 75% that is satisfactory.
KEYWORDS: YOLO v3, CNN, CHT, Darknet-53
How to cite this paper: Vijay Khare |
Pratiroop Puri | Kratik Chaurasia
"Helmet Detection Using Yolo -V3
Technique" Published in International
Journal of Trend in
Scientific Research
and Development
(ijtsrd), ISSN:
2456-6470,
Volume-6 | Issue-2,
February 2022,
pp.986-989, URL:
www.ijtsrd.com/papers/ijtsrd49341.pdf
Copyright © 2022 by author (s) and
International Journal of Trend in
Scientific Research and Development
Journal. This is an
Open Access article
distributed under the
terms of the Creative Commons
Attribution License (CC BY 4.0)
(http://creativecommons.org/licenses/by/4.0)
I. INTRODUCTION
Recently, because of the constant advancement of
profound learning research, Girshick [1] and Ren [2]
have separately proposed Fast Regional Convolution
Neural Network (Fast R-CNN) which have worked
on the exactness and realtime identification speed, yet
there is a sure hole between them. In 2015, Redmon J
[3] proposed YOLO object identification calculation,
which adjusted the precision and recognition speed.
In 2016, Redmon made YOLO-V2 [4] and YOLO-V3
[5] discovery calculations through progress. YOLO-
V2 zeroed in on little article identification, expanded
the mean exactness of mAP (mean normal accuracy)
by 2%. The most recent YOLO-V3 further fortified
multi-mark characterization and organization design,
considering both exactness and identification speed,
which has great discovery impact in development
furthermore, different fields. In any case, there are
still a few lacks in discovery exactness in the current
circumstance. In this paper, in light of the YOLO-V3
discovery calculation, bunch calculation is utilized to
anticipate the objective casing of the cap, and
afterward the precision is advanced through the blend
of profundity lingering organization and multi-scale
location preparing in the preparation interaction[6].
II. Methodology
Three phases are merged in the proposed vision-based
framework. The first step is to create a suitable
dataset for training our model, as there is none
available off the shelf. Then does the data pre-
processing, which is divided into three parts: data
acquisition, data enhancement, and data annotation.
The photographs acquired had high resolution,
various angles, and different backgrounds to create a
more realistic scenario. The collected images are
expended through augmentation techniques in terms
of scaling, dropping, and changing brightness to
increase the diversity and richness of the
experimental dataset. The next step after image
augmentation is image annotation, which involves
creating a boundary box surrounding the objects and
its label that is helmet or no helmet. Following the
augmentation and annotation, a dataset of 2480
images was generated, with 80% of the images being
randomly selected for the training dataset and the
remaining 20% for the testing collection. During the
training process, the image size will be reset, and the
batch size will be fixed based on the memory
constraints of the GPU. We will use the optimizer in
the training, with the learning rate set to 0.001 and the
other parameters remaining the same as they were in
IJTSRD49341

International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
the YOLO model. The testing process will follow,
during which a wide variety of images will be run
through the proposed solution and results will be
registered.
A. Data Set
The dataset consists of 2480 images of two classes
1. with helmet
2. without helmet
B. Feature Extraction
For training our custom object detection model, we
will need a lot of images of objects which we’re
going to train nearly a few thousand. Number of
images is directly proportional to accurate precision.
We first perform feature extraction to determine the
distribution and mathematical characteristics of the
dataset; then we build YOLOv3 on our pre-processed
data for training to build our model to detect helmets
on the camera.
Based on the features of the dataset, we can obtain
relevant information that will provide better support
in building neural network training. For feature
extraction, Calculation of the proportion of each
target in the original image, calculate the average
length of the target, calculate the average width,
calculate the average area, calculate the average
percentage of the target.Figure 1 shows the control
flow diagram of Helmet Detection while capturing
live feed. First of all, there will be a background
subtraction from the extracted frame. The next stage
would be that whether the output of first stage
consists of a bike or not. If not then the process would
be ended otherwise it will go to third stage i.e. the
“Helmet Detection Module”, Output of which would
obviously our main concern.
Fig 1: The proposed end to end Helmet
Detection System
Fig. 2 and Fig.3 respectively show the structure of
Optimization and workflow of the approach.
Fig 2 The structure of optimization method
Fig 3: Workflow Diagram
III. Algorithms
In YOLO v3, the new association darknet-53 is used
for incorporate extraction. There are 53 convolutional
layers and 5 most limit pooling layers in the
association structure. To keep away from over-fitting,
a bundle normalization framework and a dropout
movement are introduced after each convolutional
layer. YOLO v3 improves target distinguishing proof
accuracy by using a multi-scale incorporate blend
computation to appraise the position and
characterization on a multi-scale feature map. As far
as possible encases YOLO v3, estimation bunches are

used as before boxes. The k-infers approach is used to
perform dimensional bundling on the goal encloses
the dataset, achieving 9 priori boxes of various sizes
that are reliably passed on among incorporate outlines
of various scales. More unobtrusive concluded boxes
are utilized for incorporate charts with a more
noteworthy scale. Finally, the pack local area will be
used to do security defensive cap wear recognizable.
A. Architecture of YOLO v3
YOLO v3uses a variant of Darknet CNN
architecture of Darknet has 53 layer network
trained on Imagenet.
In YOLO v3, the detection is done by applying
3x3 and 1 x 1 detection kernels on feature maps
of three different sizes at three different places in
the network.
The output is a list of bounding boxes along with
the recognized classes. Each bounding box is
represented by 6 numbers (pc, bx, by, bh, bw, c).
Finally, we do the IoU (Intersection over Union)
and Non-Max Suppression to avoid selecting
overlapping boxes.
YOLO v3usesbinary cross-entropy for
calculating the classification loss for each label
while object confidence and class predictions
are predicted through logistic regression.
B. Hyper-parameters used
Class_threshold- Defines probability threshold
for the predicted object.
Non-Max suppression Threshold - It helps
overcome the problem of detecting an object
multiple times in an image. It does this by taking
boxes with maximum probability and suppressing
the close-by boxes with non-max probabilities
(less than the predefined threshold).
input_height&input_shape - Image size to
input.
C. Training and Optimization
The preparation information is parted into 8:2 with 8
sections for training and 2 sections for testing. As the
camera situations are mind boggling and diverse
camcorders have various goals, the full association
layer is eliminated in YOLO v3, so the prepared
model can be taken care of pictures of various scales.
So, we focus harder on the most proficient method to
distinguish the far off, little and unclear targets better.
In the exploratory interaction, we tracked down that
the YOLOv3 model has a decent reaction to the ID of
"individual". Hence, right off the bat, the specialists
in the video are distinguished and caught by utilizing
the YOLO v3 model, and afterward certain and
negative examples are made. Because of the low goal
of the video, which is fluffy and difficult to
recognize. Part of the information from the positive
examples are haphazardly separated and fluffy
handling is to recreate the little impact.
IV. Result and Discussion
The preparation information is parted into 8:2 with 8
sections for preparing and 2 sections for testing. As
the camera situations are perplexing and diverse
camcorders have various goals, the full association
layer is taken out in YOLO v3, with the goal that the
prepared model can be feeded pictures of various
scales. So we focus harder on the best way to
recognize the inaccessible, little and ambiguous
targets better. In the exploratory interaction, we
tracked down that the YOLOv3 model has a decent
reaction to the distinguishing proof of "individual".
Hence, first and foremost, the laborers in the video
are distinguished and captured by utilizing the YOLO
v3 model, and afterward certain and negative
examples are made. Because of the low goal of the
video, which is fluffy and difficult to distinguish. Part
of the information from the positive examples are
arbitrarily removed and fluffy preparing is to recreate
the little and fluffy examples, to improve the location
precision.
In this paper helmet detection using YOLOv3 has
been implemented and Figures 4 and 5 show the
implemented results as helmet detection with a
accuracy of 75%and non- helmet detection with 47%
accuracy.
Fig. 4 Helmet is detected with a probability of
75%

Fig. 5No Helmet with a probability of 47%
V. Conclusions
In this paper, we have used the YOLO v3 for
identification of real time person with and without
helmets. YOLO is suitable to detect the single object
from the image, YOLO has a limitation that if there
are multiple object in a single cell then YOLO is not
suitable to all objects. Therefore if you know that
your data set consist many small object in group then
YOLO will unable to detect all the objects.
VI. Acknowledgment
We would like to place on record our deep sense of
gratitude to Dept. of Electronics and Communication
Engineering, Jaypee Institute of Information
Technology, Noida, for their guidance, support,
continuous encouragement and supervision
throughout the course of the work.
References
[1] Girshick, R. (2015) Fast R-CNN[C]//IEEE
International Conference on Computer Vision.
IEEE, 2015:1440-1448.
[2] Ren, S., He, K., Girshick, R., et al. (2015)
Faster R-CNN: towards real-time object
detection with region proposal networks[C]
//International Conference on Neural
Information Processing Systems. MIT Press,
2015: 91-99.
[3] Redmon, J., Divvala, S., Girshick, R., et al.
(2015) You Only Look Once: Unified, Real-
Time Object Detection. J.2015: 779-788.
[4] Redmon, J., Farhadi. A., (2017) YOLO9000:
Better, Faster, Stronger[C]//IEEE Conference
on Computer Vision and Pattern Recognition.
IEEE, 2017:6517-6525.
[5] Sanjana, S., et al. "A review on various
methodologies used for vehicle classification,
helmet detection and number plate
recognition." Evolutionary Intelligence (2020):
1-9.
[6] VinayakMudhomath | Prof. S. A. Angadi
"Automatic Detection of Bike-Riders Without
Helmet" Published in International Journal of
Trend in Scientific Research and Development
(ijtsrd), ISSN: 2456-6470, Volume-2 | Issue-5,
August 2018, pp.1972-1975, URL:
https://www.ijtsrd.com/papers/ijtsrd18217.pdf

Helmet Detection Using Yolo V3 Technique

More Related Content

What's hot (20)

Similar to Helmet Detection Using Yolo V3 Technique (20)

More from ijtsrd (20)

Recently uploaded (20)

Helmet Detection Using Yolo V3 Technique