Research On Traffic Sign Recognition Algorithm Based On Deep Learning

Aided by big data and powerful computing capability,artificial intelligence technology has developed rapidly and its application has enriched people’s lives.In the face of complex and changeable traffic conditions,in order to ensure driving safety and improve driving experience,intelligent driving technology has entered people’s field of vision.As an indispensable part of intelligent driving technology,it is important to carry out research on road traffic sign recognition technology.Road traffic signs are small in size and diverse in types.At the same time,the road traffic signs are often polluted or blocked by foreign objects in real road conditions.Therefore,it is difficult to recognize road traffic signs through traditional recognition algorithms.In view of the above problems,this thesis carries out researches based on deep learning in terms of network structure design and related algorithm optimization,etc.The main work is as follows:1.Construction of road traffic sign data set: Aiming at the problems existing in the public data set of road traffic signs,this thesis constructs a data set of road traffic signs containing 25 categories,with a total of 9888 data images.Meanwhile,detailed annotations of object categories and location information are completed,with a total of 58560 ground truth boxes.By comparing and analyzing with the public data set of road traffic signs and the object detection data set,the reasonableness and superiority of the data set constructed in this thesis are verified.2.Design of road traffic sign recognition algorithm: This thesis proposes to replace K-means clustering with gaussian mixture model to improve the accuracy of prediction box regression.Design a multi-scale recognition network,in which the depth of the network is increased by multilevel residual structure,the feature pyramid network is used to share semantic information,and the five-scale prediction is used to improve the recognition effect of small targets.A new loss function is designed to solve the problem of poor recognition effect caused by unbalanced number of the training samples.Several basic training strategies are adopted to improve network training ability.3.The overall performance test of the algorithm: In this thesis,a series of experiments are carried out on our road traffic sign data set.The experimental results show that the m AP of the proposed algorithm is 95.0%,22.3%,9.6%,5.7%,4.9%,4.3% and 1.9% higher than the m AP of SSD,Refine Det,YOLO v3,SPP-YOLO v3,YOLO v4 and YOLO v5-x,respectively,and the recognition speed reaches 23 frames per second,283.3%,91.7%,130.0%,15.0%,4.5% and 27.8%faster than Faster RCNN,SSD,Refine Det,YOLO v3,SPP-YOLO v3 and YOLO v5-x,respectively.Through experiments conducted on the server,the superiority of the algorithm proposed in this thesis is verified.By deploying the proposed algorithm to the on-board hardware platform,a fast and accurate recognition of 0.04 seconds per frame is realized,and real-time requirement is met.Outdoor experiments further show the feasibility of the algorithm proposed in this thesis at the application level.