Hierarchical Traffic Sign Recognition By Super-resolution Transfer

Traffic Sign Recognition(TSR)is a significant part of Intelligent Transportation System,having already played an important role in manless driving system,driver assistance system,and road resource control system.The research of traffic sign recognition contributes immeasurably to both practical application and scientific study.However,the traffic signs in the natural scenes are greatly influenced by factors such as lighting,viewpoints and occlusion.Thus,the multi-class TSR,especially the application-oriented TSR,is still a challenging problem,for its updated algorithms are lacking in enough real-time performance and accuracy.To overcome these challenges,the current thesis focuses on the following two problems:(1)the unbalanced training data of the traffic sign classification datasets;(2)the small-size images in traffic sign classification datasets.The main contributions of this thesis are as follows:(1)The thesis proposes a hierarchical recognition method for traffic sign recognition.Due to the large number of the traffic sign classes and the unbalanced training data,the horizontal classification structure has high compute complexity and low classification accuracy.Aiming at this problem,this paper adopts the hierarchical recognition method.Firstly,a visual tree is constructed,which consists of two layers.The non-leaf node is constructed based on shape classification.According to aggregated channel features and the trained shape classifier by Adaboost,the traffic signs can be divided into three shape classes:circle,triangle,and square.Each leaf node contains traffic sign identification,which is constructed based on the random forest classifier and histogram of oriented gradient for multi-classifying traffic signs.Extensive experiments have been done on three famous traffic sign datasets:the German Traffic Sign Recognition Benchmark(GTSRB),Swedish Traffic Signs Dataset(STSD),and the 2015 Traffic Sign Recognition Competition Dataset,and experimental results validate the efficiency and effectiveness of the aforementioned method.(2)The thesis also proposes an algorithm for traffic sign recognition based upon super-resolution transfer.The small-size images in traffic sign classification datasets seriously affect the image recognition performance.In this paper,the application of the super resolution reconstruction helps to assign the small traffic signs to the specified size,to restore the high frequency details of the original image,to enhance the visual effect of traffic signs,and to improve the recognition rate of small-size images in traffic signs,so as to avoid losing detailed information during upsampling process.Therefore,we can get abundant information of texture and edge features in order to train and test the classifier,and finally we are capable of improving the recognition accuracy.The thesis applies the deep convolutional network SR method and builds a traffic sign super-resolution dataset,in which four objective standards are used to evaluate the image quality and compare the results in traffic sign recognition test set.The experimental results verify the effectiveness of this algorithm.