The New Advanced Society
by Sandeep Kumar Panda, Ramesh Kumar Mohapatra, Subhrakanta Panda, S. Balamurugan
4Automated TSR Using DNN Approach for Intelligent Vehicles
Banhi Sanyal*, Piyush R. Biswal, R.K. Mohapatra, Ratnakar Dash and Ankush Agarwalla
Department of Computer Science, NIT, Rourkela, India
Abstract
Traffic Sign Recognition (TSR) system has become an indispensable component for intelligent vehicles. The primary focus is to develop an efficient DNN with a reduced number of parameters to make it real-time implementable. The architecture was implemented on GTSRB. Four variations of Neural network architectures: Feed Forward Neural Network (FFNN), Radial Basis Function NN (RBN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) are designed. The various hyper-parameters: Batch Size, Number of epochs, Momentum, Initial Learning Rate of the architectures, are tuned to achieve the best results. Extensive experiments are performed to study and improve the effects on efficiency. The effects of other techniques such as validation split (0.1 and 0.2) and data augmentation are also investigated. All results are tabulated to learn the effects of different techniques. The best performing model was selected as the real-time implementable architecture of our research. Four pre-trained models, namely LeNet, GoogleNet, ResNet, and AlexNet were also implemented on the same database for comparative studies. Various other schemes of other researchers have also been provided. The comparative studies prove the supremacy of our proposed architecture.
Keywords: Traffic sign ...