Traffic Sign Recognition Method Based On Multi Scale Convolutional Neural Network

Traffic sign recognition is one of the key technologies for self-driving cars and driver assistance systems.High quality traffic sign recognition methods can provide traffic information,traffic rules and some other such information for self-driving cars and drivers in real-time and accurately so as to help driver make decisions,therefore finally to improve the traffic safety factor,reduce even avoid traffic accidents.Since traffic sign recognition is usually faced with the recognitions within natural scenes,the recognition performance is easily affected by the factors such as illumination,weather,motion blur,rotation,tilt,artificial damage,etc.Therefore,the traffic sign recognition is a worth studying and challenging subject.An improved traffic sign recognition method is proposed and implemented based on a multi-scale convolutional neural network(CNN),the simulation experimental results show that the method obtains improved performances.To improve the accuracy and real-time performance,first of all,the quality of the images to be recognized is improved via optimized option on the image preprocessing methods;and then a multi-scale CNN is designed to extract multi-scale features;finally,a SoftMax classifier is used to classify the traffic signs.This method obtains 98.82% recognition accuracy and 0.1 ms per image real-time performance on the German Traffic Sign Recognition Benchmark(GTSRB).The main work is as follow:1.In view of the fact that the image quality of traffic signs collected in natural scenes is uneven,the comparison experiments on image preprocessing methods are carried out at first.ROI extraction,size normalization and contrast limited adaptive histogram equalization based on color image are chosen as the preprocessing method this work used,the images’ quality is improved.2.In view of the limitations of single scale CNN in traffic sign recognition,another multi-scale CNN model is proposed and implements,experiments show that the combination of local and global features can effectively improve the final recognition performance.3.During CNN implements,it is difficult to set up parameters and their assignments,comparison experiments are carried out on weights initialization,activation function,optimizer,dropout,etc.The parameter selection scheme is optimized in order to improve the recognition accuracy.