Automatic Traffic Signs Recognition Using Convolutional Neural Network For Self-driving Cars

Traffic sign classification is an actuality which involves a lot of constraints and complications.A petite misclassification of the traffic sign by a self-driving car can lead to disastrous repercussions and worse,loss of life.Through the application of Artificial Intelligence(AI)in the automotive industry,numerous algorithms are being refined to increase the performance of Traffic Sign Recognition systems.However,there is an ambiguity in selecting the most efficient algorithm for designing such systems due to variations in experimental platforms and lack of common datasets.The standpoint on this project is using deep learning for traffic sign identification for selfdriving cars.A deep learning model,specifically Convolutional Neural Network(CNN)will be trained on the images of traffic signs to for automatic recognition.Since the Image Net competition in 2012,important advances have been made in the deep learning field supported by the advancement of computing technologies like the invention of Graphical Processing Unit(GPU).Nonetheless,there is limited information on how.Deep learning for traffic signs has major overlap concerning control as well as perception.Our main focus shall be on the perception part.This research aims to fill the gap in the literature by reviewing deep learning approaches by analysis.This notion primarily presents contributions to computer vision plus deep learning techniques for traffic prospects on object recognition.Deep learning models including CNN are known by the need of large training data set,and has been successfully applied in domains which are easy to collect large amount of data like general object and voice recognition.The scarcity of enough training data set in domain like traffic signs is addressed by using transfer learning and data augmentation techniques into the training process.Transfer learning is a technique of using weights of a model trained by dataset from the domain where it is relative easier to find enough training dataset,then layers of the model will be fine-tuned with data set from the target domain.Data augmentations is method to artificially increase the training data size by performing various operations like zooming,flipping and rotations on the available data set.VGG16,a 16 layer network pre trained on Image Net will be fused with our proposed model through transfer learning.Image Net is a popular dataset of general objects containing 14 million images of more than 20,000 categories general object.The VGG16 uses its weights to understand the images that are passing through,however main purpose for the VGG16 is not for classification but only feature extraction.This greatly scales down the amount of labeled real data and it is strenuous to collect adequate training data for deep learning models,will compare training our developed model with transfer learning techniques.Before training of our proposed model,several preprocessing steps were followed.The first approach will be data importation and visualization through explanatory analysis shown in graphs and lastly fine-tuning of the model was conducted.Recent techniques for fine-tuning includes an ad-hoc tuning which is considered the best way of selecting hyper parameters.For the implementation of this project on self-driving cars,we created a user interface that saves as an information portal for the facilitation of vehicle communication.This interface is a visual interpretation of what our model predicts.Streamlit is a latest library framework for Artificial Intelligence that will facilitate the creation of our user interface.To implement streamlit we include text editor python and a virtual environment set up.In this thesis,will use the Chinese traffic TSR dataset which is publicly available.