Rectangle Traffic Sign Detection And Character Extraction Algorithm From Natural Scene Images

In recent years,researchs of computer vision make great progress in the image classification,target detection,recognition,and other areas.On the other hand,in order to alleviate the urban traffic congestion and improve traffic efficiency,intelligent transportation system has received great attention.Besides,auxiliary driving,unmanned intelligent transportation system is an important component of intelligent transportation system.Moreover,the detection and recognition of traffic signs in natural scenes is an important part of auxiliary driving and unmanned driving systems.Rectangular traffic signs,which contain text,icons and symbols,provide valuable information such as road names,instructions,warnings and others.For the reason that the traffic signs in natural scenes are easily affected by different weather or light intensity,the traffic signs may be obscured.Furthermore,the characters in the traffic signs containing text are arranged irregularly,and the spacing of characters and the size are not unified.Therefore,the detection of rectangular traffic signs and text extraction has some challenges.The main work of this paper includes the following points:1.A rectangle target proposal detection algorithm based on line segments merging is proposed,which realizes the rapid extraction of rectangular traffic sign candidate regions.As for the rectangle traffic sign,it has the characteristics of wide aspect ratio change and serious debris occlusion,but the rectangular shape is a stable feature that we can use for detection.To this end,we extract the straight-line segments in the image,and then merge the rectilinear segments to achieve the detection of rectangle target,so as to realize the extraction of rectangular traffic sign candidate regions.Experimental results show that the proposed algorithm can not only run fast but also can achieve high recall rate of rectangular traffic signs with a small number of candidate regions.2.In order to filter the candidate regions which are not traffic signs,multi-stage classification and multi-feature fusion are used to detect the rectangular traffic signs.In short,the ex-HOG feature,the CLBP feature and the HOG and BoW fusion feature are combined with the SVM linear classifier to filter the candidate regions of non-traffic signs,HOG + BoW is a coding framework that the paper proposed for traffic sign detection.The experimental results show that the proposed algorithm is fast and accurate.3.For the text in rectangular traffic signs,a method based on contracting extremum region(CER)is used to extract the text components,and a single character detection based on stroke hierarchical clustering is used to solve the problem that Chinese characters can not be detected due to stroke separation;The algorithm of distance metric learning is used to merge the detected single text components to extract the Chinese text line fast,which effectively overcomes the difficulty of ununiform character spacing in the text line.For purpose of testing the performance of the algorithm,this paper collected a total of 1524 pictures with traffic signs,which cover a wide variety of different situations.Every one of the traffic sign and the text lines are annotated manually.In the randomly selected 424 images,543 rectangular traffic signs,the detection algorithm obtains 96%recall rate with average 2 to 3 false positive in one image;the text line extraction algotithm yields a recall rate of 86.2%and an accuracy of 84.7%.Experimental results show the effectiveness of the algorithm.