Research On The Detection And Segmentation Method Of Highway Guide Signs Based On Smart Car Platform

The detection and recognition of traffic signs have attracted interest from academia and industry in recent years.Some traffic signs in simple format,such as speed limit signs,have been put into use,and the researches on traffic signs in complex format are at an early stage.Highway guide signs contain a lot of road information,and they have a positive significance for vehicle auxiliary positioning as well as analysis of fork and ramp.The layouts of highway guide signs are complex,and there may be a number of sub-regions in a panel,including Chinese characters,letters,numbers,finger symbols,and various graphics.Based on the intelligent vehicle platform,we focus on the detection method of highway guide signs and the segmentation method of information in the panels.The colors of highway guide signs are mainly green and white,and the panels are generally large.In view of these features,we propose a panels detection method based on color and geometric features.The candidate regions are segmented by combining two color spaces with RGB and HSV,and then the noises are filtered through the morphology and the geometric features of the connected components.By the method,we can detect the panels of the highway guide signs under the condition of good real-time performance.The layouts of highway guide signs are complex,and there are many kinds of sub-regions in a panel.We first divide the sub-regions in a panel,and then use different segmentation methods of characters for different types of sub-regions.We propose a segmentation method of sub-regions based on multi-level projection algorithm.Firstly,the image of panel is binarized by block-OTSU and color features;secondly,the border of panel is detected by Hough transform and the tilt correction is carried out;then the multi-level projection on the vertical and horizontal directions is used on the panel to get the image blocks containing characters and symbols;Finally,the image blocks are merged into sub-regions according to the panel rules.In order to filter out pseudo targets as far as possible,we design a SVM classifier based on HOG and LBP features.Through classification,we remove the pseudo target sub-regions,and retain four sub-regions of interest:Highway number,place name of Chinese character,direction indication and distance information.The difference between the four sub-regions in the panels is larger,and the unified characters segmentation method is more complex.Therefore,we propose an adaptive characters segmentation method of sub-regions.The pre-classification of sub-regions is achieved by convolution neural network,and the types of each sub-region are obtained.Then,characters or symbols in different sub-regions are segmented by adaptive projection segmentation algorithm.Finally,we carry out the experiment for the detection and segmentation of highway guide signs based on the intelligent vehicle,and give the results and analysis.The experimental results of the real scene show that we achieve good results.The accuracy of the detection of sign panels is 97.24%,the recall is 98,27%;the accuracy of segmentation of sub-regions in the panels is 97.11%,the recall is 90.65%;the accuracy rate of segmentation of characters and symbols is 95.83%.