Recognition On Road Traffic Signs Based On Feature Fusion

As a necessary tool for people's daily travel,automobiles have been integrated into the lives of millions of households.People's requirements for the safety and convenience of automobiles are becoming higher and higher.Therefore,driver-assisted driving system emerges as the times require.Road Sign Recognition is an important component of environmental perception.Accelerating the pace of popularization and application of related technologies is of great significance for improving the quality of life and reducing traffic accidents.The environment of road signs is complex,and the accuracy of traditional detection and recognition methods is difficult to guarantee.Therefore,two-level detection algorithm and feature fusion algorithm are studied in this paper.Firstly,two main methods of road sign location are analyzed,which are based on the specific color and shape of the road signs.Through the experimental reproducing of the two methods,the advantages and disadvantages are analyzed,and the grading detection method of the road sign in this paper is determined.Two main road sign recognition algorithms are analyzed.They are machine learning algorithm based on Hu moment invariant feature and support vector machine and deep learning algorithm based on convolution neural network.Through the experimental repetition of the two recognition algorithms,the advantages and disadvantages are analyzed,and the machine learning classification algorithm based on fusion feature is determined.Secondly,rough detection of all possible locations of road signs can be carried out through gray level transformation,Otsu color segmentation algorithm,noise removal according to the unique color and shape of road signs compared with the surrounding road environment.Each location of road signs to be detected can be obtained.HOG features of corresponding locations can be obtained,which are put into the SVM classification model to detect the accurate location of road signs.Then,LBP feature,SIFT feature and HAAR feature of the detected road sign accurate location image are extracted separately,and the fusion of three image features is completed by serial fusion method.The fusion feature of the image to be recognized is obtained as the basis of image classification.Finally,the corresponding NBC model is established,and the relevant data sets of road signs are produced by the project in the summer internship of Shanghai Automobile Corporation in 2018 for training and testing of the NBC model.This paper has compared the recognition results of fusion features and each single feature,and has also compared the classification results of the fusion features combined with NBC model,fusion features combined with SVM multi-classification model,CNN and other main image classification algorithms.Through analysis of experimental data of various recognition methods,the feasibility of the recognition method combining fusion features with NBC model is obtained.This paper completes the software architecture of the road sign recognition system through two-level detection algorithm of traffic signs,fusion features and classification algorithm of NBC model.The detection rate of the algorithm reaches 99.3% and the recognition rate reaches 98.7% through experiments on the driving video collected by intelligent car when driving on the track,which are higher than the accuracy of human identification of road signs in the driving process.This method has a certain degree of popularization and application value.