Performance Comparison of Road Sign Recognition Methods based on Machine Learning

Abstract

Road sign recognition technology refers to the use of computer vision technology to identify, detect and analyze road signs in images, and is one of the research hotspots in the computer vision community and the field of autonomous driving. Most of the early road sign recognition methods rely on manual features, and their recognition accuracy cannot meet the actual application requirements. Thanks to the rapid development of convolutional neural networks, the accuracy and speed of road sign recognition have made breakthroughs in recent years. However, due to the complexity of actual traffic scenes, such as illumination changes, motion blur, partial occlusion, etc., there are significant differences in the application boundaries of different methods. Most of the existing works are devoted to further improving the recognition performance. In this paper, we explore the accuracy of three traditional methods: linear support vector machine, logistic regression and multi-layer perceptron (MLP), and compare them with the current mainstream methods. The study also explores the influence of illumination and definition factors on accuracy. Experimental results show that the accuracy of the three traditional methods is lower than that of CNN. For the accuracy of the three methods, where the MLP had the highest accuracy, followed by logistic regression, and Linear SVM had the lowest accuracy. For the testing sets with different features, the three methods all have the highest accuracy of good illumination and definition, the second accuracy of good definition but poor illumination, and the lowest accuracy of good illumination but poor definition. In terms of accuracy on both training set and three testing set, MLP model has the highest comprehensive accuracy, followed by SVM, and logistic regression has the lowest comprehensive accuracy.