Traffic Sign Classification by Using Learning Methods: Deep Learning and SIFT Based Learning Algorithm

Abstract

Automatic detection of road signs is of significance to automated driver assistance systems contributing to the safety of the drivers. With the development of different traffic sign recognition systems, several detection methods appeared: color-based, shape-based, and learning-based methods. In this paper, we consider learning-based methods and compare the performance of deep learning methods and other learning methods based on hand-crafted features (Scale-Invariant Feature Transform(SIFT)) in Landmark recognition. We also compare several algorithms based on deep learning, including a convolutional neural network with 2 and 3 layers, and LeNet. [1] pointed that SIFT is outperformed other learning methods based on hand-crafted features like SURF and BRISK, therefore we choose the SIFT with the best road sign recognition rate to compare with the two basic deep learning methods CNN and Lenet in order to know which is more suitable for road sign recognition. In the SIFT -based classification algorithm, we extract features from images, use K-means, Bag of Words Model to generate feature vectors, and the SVM model is then trained by feature vectors and their labels. The same operation will be done in the test images to feature vectors that are put into the well-trained SVM model and a traffic sign is detected, while in deep learning methods, we use two different convolutional neural network architects. All the results show that all deep learning algorithms perform well and LeNet has higher accuracy than CNN while the SIFT -based classification algorithm has the lowest accuracy even though it has the highest speed.