Traffic Sign Recognition with Light Convolutional Networks

Abstract

In this work, we aim to design a light net that can be executed on the embedded system in real time. We modify VGG Net to a small net, called Safe Net, and utilize multi-scale features for traffic sign recognition. Moreover, we convert the dataset into grayscale, which has been proved that has a better performance on GTSRB dataset. In addition, we augment the training data by about 6.6 times more via spinning, distorting and flipping to boost the accuracy. On Nvidia Jetson TX1, Safe Net only takes 4.58ms per image including preprocessing at the testing and Safe Net can even achieve 99.34% accuracy.