Research On Traffic Sign Detection,Tracking And Recognition Algorithm In Natural Scene

Traffic Sign Recognition(TSR)algorithm based on vision is usually composed of Traffic Sign detection and Traffic Sign Recognition.Although traffic sign recognition methods have made good progress in the field of Advanced Driving Assistant System(ADAS),there are still many challenges in designing a real-time detection and accurate identification method of traffic signs in natural scenes.On the one hand,traffic signs are in a natural scene for a long time,which is likely to cause the distortion of the characteristics of traffic signs themselves,including the effects of fading,deformation,shielding,tilt,light interference and unbalanced distribution of the categories of traffic signs.On the other hand,self-driving cars not only require TSR algorithm to have high detection and recognition accuracy,but also require TSR algorithm to still show good performance in the context of limited computing resources.In the face of the above challenges,this paper attempts to use a Cascade traffic sign detection algorithm based on significant features and a traffic sign detection and tracking algorithm based on nuclear correlation filtering to make up for some problems existing in the current target detection algorithm.At the same time,we combine the characteristics of traffic signs and propose a traffic sign recognition method based on data enhancement to solve the problem of the decrease of recognition accuracy caused by the unbalanced distribution of traffic signs in the natural scene.This paper will carry out relevant research work from the following aspects:1)Considering that the current algorithm for traffic sign detection cannot achieve a good balance between computational cost and detection performance,we proposed a Cascade traffic sign detection algorithm based on significant characteristics.The algorithm makes full use of the prior knowledge of the color,shape and spatial position of traffic signs to extract the candidate regions,and USES the Cascade classifier of two stages to verify the candidate regions step by step.The proposed method accelerates the extraction of traffic sign candidate areas and improves the detection performance of traffic signs.2)For the traditional target detection method,it is required to search the target in the image globally frame by frame,and the detection algorithm is prone to environmental interference and cannot continuously detect the target,which leads toproblems such as high time consumption of the detection algorithm and unrobust detection effect.This paper proposes a traffic Detection and Tracking framework based on the idea of detection-by-tracking,which integrates the target Tracking algorithm into the traffic sign Detection algorithm,replaces the target Detection with target Detection and Tracking,improves the efficiency of traffic sign Detection,and at the same time makes up for the shortcomings of the Detection results of the target Detection algorithm.3)In view of the unbalanced distribution of traffic sign categories in natural scenes,most categories will block the "right of speech" of a few categories,thus affecting the accuracy of identification.In this paper,a method of traffic sign data enhancement is proposed.By expanding the sample data of minority categories,the quantity gap between minority categories and majority categories is narrowed,so as to improve the accuracy of traffic sign identification.