Regularized CNN for Traffic Sign Recognition

With the increasing urban population, city expan- sions come along with the problem of traffic management. In today’s era of the computational world, where there is a smart AI-based solution for almost every problem, AI-based Traffic Sign Recognition (TSR) comes to our rescue. This work develops a customized convolutional neural network for TSR. This work explores the impact of different regularization techniques namely L1, L2, Dropout, and batch normalization on the performance of a developed TSR system. To conduct the experiments, this work utilizes the benchmark GTSRB dataset (German Traffic Sign Recognition Benchmark). Experimental results show that our proposed classifier outperforms the existing approaches for TSR. This paper tries to cover one such aspect and tries to train the system to respond in accordance with a rich dataset containing 39,209 color images spanned over 43 different categories of traffic signs, whereas the system is tested with 12,630 examples, achieving an accuracy of 99.56%.