Research On Traffic Sign Small Target Detection And Recognition Based On Deep Learning

In recent years,thanks to the rapid development of artificial intelligence technology,as one of ITS main applications,Intelligent Transportation System(ITS)has become increasingly mature and perfect.Traffic Sign Recognition(TSR)is a key subsystem of ITS,and ITS performance has been improved by leaps and bounds under the promotion of computer vision technology in artificial intelligence.However,when traffic signs enter the driver’s field of vision,it is usually a small target at first.At present,the detection effect of TSR still has some room for improvement.The real-time and accurate detection of small targets in traffic signs can increase the reaction time of drivers and reduce traffic accidents caused by inadequate response.Therefore,this study mainly improves the detection and recognition technology of small targets in traffic signs.The specific work is as follows:(1)Expand the small target data set of Traffic signs in China based on the TT-100 K data set compiled and published by Tsinghua university and Tencent,in order to make up for the lack of high quality and large scale small target data set of traffic signs in China.In view of the unbalanced distribution of the tt-100 K dataset,Mosaic and Cutmix data were used to supplement the data and enhance the equalization by combining the characteristics of traffic signs,so as to truly improve the generalization ability of the small target detection model of traffic signs.Finally,the small target data set of traffic signs is 21,702 sample images,which contains more than 58,000 traffic signs.Experiments are designed and the effectiveness of equalization processing is compared and analyzed.The average accuracy of 35 types of traffic signs increases from 86.49% to 88.12%,which proves that the equalization processing of data set is effective.(2)Improve the detection algorithm model of small targets in traffic signs based on YOLOv4 target detection and recognition algorithm.Considering that the detection accuracy of YOLOv4 algorithm for small targets and minimal targets of traffic signs still has room to improve,the multi-scale prediction structure of YOLOv4 algorithm is modified,and the detection scale number is expanded from 3 to 4,so that more details of image location can be obtained by the network.In order to reduce detail information loss and enhance semantic feature information,ASPP feature extraction module based on empty convolution was used to replace SPP module.The feature fusion strategy is improved and the adaptive spatial feature fusion structure ASFF is added to improve the effectiveness of feature fusion.For the size of anchor frame,k-means++ clustering optimization is adopted to further improve the detection accuracy.The design experiment compares and analyzes the detection effect before and after the improvement.The average accuracy of 35 types of traffic signs increases from 88.12%to 92.93%.Finally,network lightweight was carried out to improve the detection rate.CSPDarknet53 was replaced with Moblie Netv3,and experiments were designed to verify the effectiveness.At a small cost of 0.79% average accuracy loss,the detection frames per second increased from 18.8 to 27.1.The test results prove that the improved algorithm is reasonable and effective.