Automatic Road Sign Recognition And Localization From In-vehical Images

Automatic road sign recognition and localization are crucial in the field of Intelligent Transportation System(ITS).Accurate and fast road sign recognition and localization are of great important not only for Intelligent Road Asset Inventory(IRAI)but also for Advanced Driver Assistance System(ADAS).But the accuracy and time efficiency of the existing methods can't meet the demand of ITS.Vehicle GPS(Global Positioning System)can locate the location of vehicle,but can't locate road signs.To achieve accurate and fast road sign recognition and localization,the following research are carried out,which are based on in-vehicle images and computer vision:(1)To solve the problems of high computational complexity and poor time efficiency existing in local feature extraction method for road signs,the holistic features method were proposed,such as ORB(Oriented FAST and Rotated BRIEF)holistic features and SURF(Speeded-Up Robust Features)holistic features.Based on holistic features proposed and affine transformation,Affine-ORB(A-ORB)and Affine-SURF(A-SURF)holistic features were proposed.When computing holistic features,images are resized to a standard patch image firstly,then the center of the resized image is set as the feature point position to compute the ORB and SURF local features.The local features can be regarded as the holistic features of sign images.The results demonstrate that holistic features method enhance computation speed by skipping the time-consuming step of feature point detection,which lays a good foundation for accurate and fast road sign recognition.(2)Based on K-Nearest Neighbors(KNN),Weighted Hybrid K-Nearest Neighbors(WH-KNN)were proposed.The WH-KNN can solve recognition ambiguity problem in KNN by computing the Gauss weights of the input fused features and hold the advantage of recognizing multiple targets simply and efficiently in KNN.Based on WH-KNN,the proposed method has been validated with the public German Traffic Sign Recognition Benchmark(GTSRB)dataset,the collected dataset,and the public LISA traffic sign dataset.The results validate that the proposed WH-KNN method can solve the recognition ambiguity problem and achieve the multiple targets recognition simply and efficiently.(3)Finally,this paper proposed fusing GPS and vision measurement for road sign localization.Vehicle GPS can locate the location of vehicle in real time,but can't locate road signs.The vision local localization and GPS are fused to achieve the localization of the road signs in this paper.The proposed method has been validated with simulation experiment and field experiment,the results demonstrate the proposed method can achieve the accurate localization of road signs.