A Traffic Sign Classifier Model using Sage Maker

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 5 Issue 4, May-June 2021 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
@ IJTSRD | Unique Paper ID – IJTSRD42411 | Volume – 5 | Issue – 4 | May-June 2021 Page 855
A Traffic Sign Classifier Model using Sage Maker
Arpit Seth1, Vijayakumar A2
1Student, 2Professor,
1,2Department of MCA School of CS & IT, Jain Deemed-to-be University, Bengaluru, Karnataka, India
ABSTRACT
Driver assistance technologies that relieve the driver's task, as well as
intelligent autonomous vehicles, rely on traffic sign recognition. Normallythe
classification of traffic signs is a critical challenge for self-driving cars. For the
classification of traffic sign images, a Deep Network known as LeNet will be
used in this study. There are forty-three different categories of images in the
dataset. There are two aspects to this structure: Traffic sign identificationand
Traffic sign classification. ADASs are designed to perform a variety of tasks,
including communications, detection of road markings, recognition of road
signs, and detection of pedestrians. There are two aspects to this structure:
Traffic sign identification and Traffic sign classification. In the methodologies
for detecting and recognizing traffic signals varioustechniques,suchascolour
segmentation and the RGB to HSI model area unit, were appliedfortrafficsign
detection and recognition. Different elements contribute to recognition of
HOG.
How to cite this paper: Arpit Seth |
Vijayakumar A "A Traffic Sign Classifier
Model using Sage
Maker" Published in
International Journal
of Trend in Scientific
Research and
Development(ijtsrd),
ISSN: 2456-6470,
Volume-5 | Issue-4,
June 2021, pp.855-858, URL:
www.ijtsrd.com/papers/ijtsrd42411.pdf
Copyright © 2021 by author (s) and
International Journal ofTrendinScientific
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
Traffic sign recognition has a lot of economic promise in
clever autonomous vehicles and driver assistance systems.
Without correctly deploying and maintaining road traffic
signs, traffic signals, and road markings, it is impossible to
improve traffic quality and safety. AmazonSageMakercould
be a machine-learning platform in the cloud that allows
developers to create, test, and deploy machine-learning
models.
We frequently use Amazon Sage Maker to create models.
Sage Maker has a number of advantages, including a large
amount of processing capacity (various CPUs and GPUswith
parallel computing) and hence the possibility to deploy the
model as a whole. The entire analysis was completedinSage
Maker's notebook, while training and prediction were
handled by a large number of strong GPU instances. This
paragraph can encompass the entire pipeline, from pre-
processing to final model, for the sake of completeness. A
variety of classification techniques have been used to
recognize traffic signals, ranging from simple model
matching (e.g. pooling, flattening) to complex Machine
learning techniques (e.g. support vector machines,boosting,
dense neural network etc.). Furthermore, moving from
isolated frames to feature maps for traffic sign analysis can
help to reduce the number of false alarms while also
improving the precision and accuracy of the detection and
recognition process.
The training dataset, testing dataset, and validation dataset
are the three elements of the dataset in this work. Accidents
can occur for a variety of causes, including drivers failing to
notice a proof in a timely manner or neglecting to listen at
critical times. In unfavourable weather circumstances such
as severe rain showers, fog, or falls, drivers pay little
attention to traffic signals and focus instead on driving.
Control, in addition to guiding and advising drivers,
necessitates automatic traffic sign recognition. In general,
traffic signs provide the driving force with critical
information for safe and efficient navigation. In order to
better parse and recognize the highway, self-driving cars
need traffic sign recognition. Similarly, in order to assistand
protect drivers, “driver alert” systems in cars must consider
the roadway around them. Computer vision and deep
learning can help with a variety of issues, including traffic
sign recognition. Our goal is to build a multi classifier model
based on deep learning to classify various traffic signs.
II. LITERATURE REVIEW
The three most popular types of traffic sign detection
systems are color-based, shape-based, and learning-based.
Meanwhile, Liu introduces two newmodes:colorandshape-
based methods and LIDAR-based methods. Since it achieves
a detection rate of more than 90%, the learning-based
detection method is the most powerful of these approaches.
On the other hand, the most recent algorithms arestill a long
way from being a real-time classification system.Some
researchers have obtained good results by using the HSI
color space instead of RGB. For traffic sign detection, a new
color space called Eigen color was proposed based on
Karhunen-Loeve (KL).[1]This paper proposes the “German
Traffic Sign Recognition Benchmark” dataset and
competition, as well as their design and analysis. Theresults
of the competition show that cutting-edge machine learning
algorithms excel at the challenging task of traffic sign
recognition.[2] The baseline algorithms considered in this
IJTSRD42411

International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
paper are the Viola-Jones detector based on Haar features
and a linear classifier based on HOG descriptors, which are
two of the most common detection approaches.[3] This
paper focuses on real-time traffic sign recognition, which
involves deciding in a short period of time which type of
traffic sign appears in which region of an input image. To
accomplish this, a two-module architecture (detection and
classification modules) is proposed. The color probability
model in the detection module converts the input color
image into probability maps. The road sign
recommendations are then derived by identifying the most
secure extreme areas on the charts.[4] In this research, the
HIS color model is used to detect the traffic sign, which is
accompanied by circle detection. [5]The precise sign area
cannot be determined from the color-detected regions. This
approach traces the edge of interested regions to obtain
their contours after morphologic operations.
III. PROPOSED METHODOLOGY
In this purposed system we use the LeNet CNN architecture
to develop a multi classifier model based on deeplearningto
categorize multiple traffic indicators in this purposed
system. Image classifiers work by guessing the types of
objects in a photo. We employ a convolutional neural
network as a classification model (CNN). In general, when it
comes to classifying images on large datasets, CNN beats
classic neural networks and other machine learningmodels.
For the classification of traffic sign photos, a Deep Network
known as LeNet will be used. Below are the steps we
followed to option our goals and to achieve the final Le Net
multi classifier model for traffic sign prediction
Fig 3.1 - LeNet Architecture
We collect many of these neurons together in a systematic
format, in a layered fashion. After the processing we have
bunch of neurons here in the hidden layer and then will
connect all the neurons in one layer and will be fully
connected to all the neurons in the subsequent layer this is
called dense artificial neural network or fully connected
artificial neural network. So let's assume that we have a
bunch of images here. And these are the images from the
traffic sign data set. Here we have all the different classes. A
dense, fully connected artificial neural network is created.
However, in order to interact with photos, we must first do
extra processing procedures. Following that, we perform
convolution. Then we'll employ non-linearity. Using the
RELU activation function, or rectified linear unit activation
function, and subsequently pooling, we hope to add non-
linearity to the CNN after the convolution layer. In the
pooling layer, we aim to compress all of these feature maps,
all of the outputs that we extract or collect from the first
layer. Pooling layers produce same findings, but in a
compressed form of that information. We notice that when
we use the pooling function, the sizehereshrinks.Wesimply
apply flattening after we've createdthesepoolingfilters. Asa
result, we bring all of the pixels together.Weflattenthemup.
After final step model will be ready to take all these inputs
and feed them to dense or fully connected artificial
networks.
Fig 3.2- Proposed FunctionalArchitecture
Below are the process done to train the model.
1. Convolution Layer
First step for identification is convolutional layer. These
convolutional are Kernelsor featuredetectors.Thesekernels
or feature detectors just filters the image to extract features
out of the original image. These kernels with the help of
some 2X2 matrix filters the image by performing some
algorithm functions on original image. Convolutions simply
use the kernel matrix to scan a given image and apply a filter
to obtain a certain effect. Where you take an image, you
apply kernel or a feature detector to extract features out of
my image. So what we do is after we apply these kernels or
feature detectors, we actually generate feature maps.
Feature maps are kind of different variations of the original
image, as we are getting the image and we can perform
sharpening to the image or blurring to the image. We will
creating all these different feature maps to extract features
out of original image. With the help of image kernel we have
here is a matrix that is used to apply effects such as blurring
and sharpening. And kernels are used in machine learning
for feature extraction to select the most important pixels in
the image.
2. RELU
The function map incorporates non-linearity using RELU
Layers. It also improves the sparseness, or dispersion, of the
function map. The RELU's gradient does not vanish as x
grows, unlike the sigmoid function.
3. Pooling
Pooling or down sampling layers are placed after
convolutional layers to reduce feature map dimensionality.
This improves computing efficiency while keeping features
intact. Pooling helps the model generalize by preventing
over fitting. If one of the pixels is moved, the pooled function
map will remain unaltered. Max pooling in a feature map
maintains the maximum feature answer within a given
sample size. The max pooling layer reduces the size of the
previous layer's performance, making training easier. To
avoid over fitting, a drop out layer is also used.
The sagemaker.tensorflow.TensorFlowestimator handles
locating the script mode container, uploading script to a S3
location and creating a Sage Maker training job. File is
created and saved in S3 bucket in .npz files so the model
could access the data.
Fig 3.3- upload data to S3 bucket

The training script is a python file that can be run on itsown.
Sage Maker will run this script on a giveninstanceduring the
training job.
A. CORPORATE IMPLEMENTATION OF MACHINE
LEARNING AND CLOUD COMPUTING
In every industry, machinelearningandcloudcomputing are
required for data storage, processing, and management. We
offer AWS data storage and administration services as part
of our cloud computing services. The data in the CSV file will
be saved in the same register. The AWS files are used to
collect and update data for a model that we utilize
frequently. The AWS EC2 and S3 architecture is shown here,
which allows us to store data and map it to the IAM service
model. AWS Sage maker assistsinthetrainingofa prediction
model by using built-in algorithms. You can start using AWS
sage creation right immediately. Boto 3 and Sage Maker are
required to be installed in the notebook. The Amazon Web
Services (AWS) Software Development Kit (Boto3) is a
Python version of the Amazon WebServices(AWS)Software
Development Kit(SDK).PythonprogrammersmayuseBoto3
to construct applications that use Amazon S3 and Amazon
EC2. Then it's time to schedule a Sage maker session. Sage
Maker facilitates the storing of data and training datasets in
S3, as well as the training of models on S3 for later use. Sage
Maker has the advantage of being able to train an unlimited
number of models from which the best or most dependable
model can be chosen. The Linear Learner is a supervised
learning algorithm for training data by fitting a line to it.
Fig 3.4 – Architectural working of sage maker with S3
Bucket
IV. RESULT and DISCUSSIONS
This is the expected outcome, and this section will explain
the results as well as the example codes. In this design, we
employed a multi classifier model that could accurately
predict the class of pre-processed images existing in a given
image. We can describe and train the model's training job.
An instance of ml.p2xlarge was used. After two epochs, the
validation accuracy is 0.8523. The assignment took the
instance 76 seconds to complete, with a 0.923 accuracy on
the testing set.
Fig 4.1- Trained Models
A. DATASET DESCRIPTION
GTSRB dataset to train our own custom trafficsignclassifier.
The Description of the data:
The sizes of each picture are slightly different.
There are 39209 training examples in total.
The total number of test examples is 12630.
There are 43 classes total.
The images range in scale from 20x20 to 70x70 pixels and
have three channels: RGB. AWS S3 storage will be used to
store all of the data. And all of the research and training is
done in Sage Maker instances on the dataset. The first thing
we have to do is to resize all the images to 32x32x3 and read
them into a numpy array as training features.
All of the classes go through the same procedure. Data is
stored in pickle form to avoid having to do pre-processing
again if the notebook was disconnected. We can easily load
the processed data from pickle the next time we need it.
Fig 4.2 - Input Dataset
B. Final Output
The below is the sign recognition of 5 samplesfrom the
testing dataset out of 43 classes.
Fig 4.3 – Predicted output from testing dataset
Fig 4.4 – Dataset classes

V. CONCLUTION
We used the new Tensor Flow framework to train a model
for traffic sign classification. The pipeline includes pre-
processing, model construction, training, estimate, and
endpoint deployment. The accuracy of testingis0.923,while
the accuracy of validation is 0.8523. The accuracyofthebulk
of the classes is better than 90%.
The proposed model, according to the findings, is capable of
resolving some faults.
A. FUTURE WORK
The following are some suggestions for improvement:
1. This project lacks a user interface.
2. Other variables may be used to improve prediction
accuracy.
References
[1] M. S. J. S. a. I. J. Stallkamp, “The German traffic sign
recognition benchmark: a multi-class classification
competition,” IEEE, 2011.
[2] J. S. J. S. S. a. C. I. S. Houben, “Detection of traffic signs
in real-world images: The German traffic sign
detection benchmark,” IEEE, pp. 1-8, 2013.
[3] H. L. H. X. a. F. W. Yi Yang, “Towards Real-Time Traffic
Sign Detection and Classification,” IEEE, 2014.
[4] W. L. Kai Li, “Traffic indication symbols recognition
with shape contex,” IEEE, 2011.
[5] J. K. K. J. S. Bao Trung Nguyen, “Fast traffic sign
Detection under challenging conditions,” IEEE, 2014.

A Traffic Sign Classifier Model using Sage Maker

More Related Content

What's hot (20)

Similar to A Traffic Sign Classifier Model using Sage Maker (20)

More from ijtsrd (20)

Recently uploaded (20)

A Traffic Sign Classifier Model using Sage Maker