Research On Traffic Sign Detection And Classification Algorithm

With the rapid growth of the number of cars in the society,traffic congestion has become a common phenomenon in many cities,and the traffic safety problems have become increasingly serious and caused huge losses to he society.In this context,the the Intelligent Transportation system(Traffic Sign Recognition,for short TSR)came into being.As an important part of intelligent transportation system,traffic sign recognition system has been paid more and more attention.The traffic sign recognition system is mainly used in the auxiliary driving system and the driverless system.It can identify the traffic sign on the road and assist the driver or the autonomous driving vehicle.Due to the complexity of natural scenes,and recognition technology is not yet mature.How to recognize traffic signs in natural scene quickly and accurately still needs deep discussion and research.Because the traffic signs themselves have distinctive color and shape characteristics,they can be quickly positioned to identify the location of the traffic sign and then identify it.This can greatly improve the recognition efficiency of the system.However,the natural scene and its lighting conditions are complex and changeable.The traffic signs themselves may have problems such as occlusion,deformation,fading,etc.,and the color and shape characteristics of the traffic signs collected in the images also change.Aiming at the above problems,this paper proposes a detection and recognition system for traffic signs based on color-shape segmentation.Firstly,in the aspect of image restoration,this thesis mainly studies the image de-fogging algorithm.The innovation of this algorithm is mainly to improve the image de-fogging algorithm based on the prior theory of dark channel.Aiming at the shortcomings of this method in some detail processing,the outline of the object is clearer and the image is closer to the original image than the original dark channel method.According to the high requirement of TSR system,the running speed of the algorithm is improved to a large extent,and the real-time operation can be achieved.Secondly,in color clustering,because the camera shooting effect is greatly affected by illumination and background color interference,the algorithm innovations proposed in this thesis mainly focus on the robustness to different degrees of illumination and complex color background interference.There are many models of color space,and the most adaptable to the robustness of illumination is the HSV color model.Based on the HSV color model,this paper proposes an adaptive color clustering algorithm based on the similarity of local similar illumination intensity,and uses the red degree of the red point and the surrounding crimson point.After experimental comparison,the color clustering of this paper The algorithm has certain advantages in terms of effect and timeliness.Thirdly,in the aspect of contour shape detection,the innovation of the algorithm in this thesis mainly has certain adaptability to the local occlusion of the mark as well as the deformation and the rotation of the mark.This thesis analyzes the commonly used detection algorithms for circles,triangles and rectangles.Aiming at the problem of occlusion,fouling and other influences on the shape of traffic signs,this thesis proposes a contour approximation algorithm based on sag removal.Aiming at the deformation and occlusion of circular traffic signs,this thesis proposes a fixed-leave least squares ellipse fitting method to improve the detection rate and detection speed based on the stochastic least squares ellipse fitting method.For the characteristics of triangle and rectangular contours,this thesis proposes a contour merging algorithm.Experiments show that the detection algorithm of this thesis improves the detection rate and detection speed of deformation and occlusion circular markers,and improves the detection rate of triangles and rectangles,and reduces the false detection rate.Fourthly,in the aspect of feature extraction of traffic signs,this thesis adopts the method of extracting the HOG features of the sample images first,and then applying the techniques to reduce the dimensionality of the image features.Compared with the feature classification alone,the recognition rate and recognition speed of the algorithm are used in this thesis.Increased.In terms of classification and recognition algorithms,this thesis selects the commonly used SVM support vector machine algorithm.The experiment proves that the feature recognition and classification recognition algorithm selected in this thesis has a comprehensive recognition rate of 97.53%,and has achieved good recognition results.Finally,this thesis summarizes the work of this paper,the shortcomings of the work,and the next work focus.