International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 09 Issue: 05 | May 2022
www.irjet.net
p-ISSN: 2395-0072
Traffic Sign Recognition System Saloni Pathak#, Rutuja Rane#, Geeta Chavan#, Sumedh Pundkar# #Department
of Computer Science and Technology, SNDT Woman’s University, Mumbai, India
---------------------------------------------------------------***--------------------------------------------------------------
Abstract— Recognition of traffic signs is an important factor in applications such as self-driving cars, traffic mapping and traffic surveillance. Deep Learning models help in automated driving for Traffic Sign Recognition. In this paper, the model is trained with pre-processed RGB images and the saved model is used to classify the image provided.
A. CNNs Designed for Classification The authors in [3] propose a model consisting of basic convolution features with supervised learning technique. It consists of multiple convolution layers and subsample layers, followed by average pooling and fully connected layers. The model is trained with gray-scale images and the Adam optimizer is used to boost the accuracy. Batch normalization is used in [1] after each convoluted layer to prevent vanishing of gradient during back propagation across various layers.
Keywords—Convolution Neural Network, Dropout Layer, Traffic Sign Recognition, Classification, Batch Normalization.
The model [2] uses a modified version of the generalized Hough transform to localize the pre processed images. Point-like noise in the image that occurs during the preprocessing was removed with the help of noise removal algorithm.
I. INTRODUCTION The advancement in technology has led to many evolutions in every field and aspect of life. One such is Convolution Neural Network (CNN or ConvNet) which is ideal and most accurate for image: processing, detection, classification. The same can be done using various architectures available namely AlexNet, VGGNet, GoogLeNet, and ResNet. The images are captured from the front-cameras of vehicles and are processed to give out instructions for the driver or assist in automated cars. Similarly, the system can scan and compare the speed of car with the traffic sign displayed, further informing driver to slow down to avoid over speeding.
One of the most interesting results was seen when the multiple network blocks were used. In [4], the model uses a six-layers neural network that consists of two convolution layers with three subsampling layers and two branched subnets, all followed by a subsampling layer in the end. The subnets layers are trained with different parameters. The weighted cumulative sum of classifiers and alternating data augmentation during model training increases the accuracy of the model. The authors in [5] proposed a classic LeNet-5 architecture [8] and scikit-learn pipeline framework to classify the images. HOG (Histogram of Oriented Gradient) was used to preprocess the images which allow the identification of dominant gradients in the image. The LeNet-5 architecture consists of two sets of convolution and average pooling layers, followed by flattening convoluted layers, a fully connected layer and a softmax classifier.
Hence, we proposed a method using convoluted layers along with other layers to detect and identify the traffic signs.
II. RELATED WORK Traffic Sign Recognition has been researched upon for the last decade. Various techniques have been employed to classify the images. Recently, deep learning techniques involving Convoluted Neural Networks have been used for image detection, classification and localization. Traditional approaches haven’t proved the accuracy comparable to what humans require. Improvements have been made to increase the classification accuracy using deep learning techniques and exceed the average human accuracy.
Taking into account all the problems faced by the above models, we propose a model based on the classic LeNet architecture. The model is trained with Adam’s optimizer and once the required accuracy is reached, a confusion matrix is drawn to help retrain the model.
III.DATASET
Next we study the prior work done on CNN designed to recognise traffic signs.
© 2022, IRJET
|
Impact Factor value: 7.529
The dataset used for this project is the German Traffic Sign Detection benchmark. The dataset consists of 43 traffic
|
ISO 9001:2008 Certified Journal
|
Page 234