Research On Traffic Sign Detection Method Based On Lightweight YOLO

Traffic signs contain important road condition information,but traditional traffic sign detection methods cannot solve the problems efficiently,which brought by complex road environment in practical applications.In recent years,with the rapid development of Artificial Intelligence,high-performance computing technology and the automotive industry,traffic sign detection based on deep learning methods can make full use of the complex feature information in traffic sign images,which have gradually become an important research direction in this field of study.Aiming at the vehicle-mounted mobile devices with limited performance,this paper improves the YOLOv4 target detection framework to achieve end-to-end and real-time traffic sign detection.The main tasks accomplished in this paper are as follows:(1)Analyze the existing public dataset about the categories and data volumes,select a traffic sign data set which has large sample size and high quality,and expand and complete the labeling of the missing categories.Analyze the data enhancement method based on pix2 pix and the data enhancement method based on basic graph transformation.Use the basic graphics transformation methods to complete the data enhancement of the CCTSDB.Finished training the model.The test results show that the model trained with data augmentation data set has better performance.(2)Training and testing of YOLOv4 preliminary,the results show that the accuracy is97.8%,and the test time of a single picture is 0.178 s,both of them are at a high level.Analyze the YOLOv4-tiny target detection framework.Aiming at the problem of the missed detection of small target,this paper uses the k-means++ clustering algorithm to complete the redesign of the size of prior frame,and combined with the size distribution of traffic signs in the data set,matches the input size of images with the output size of the YOLO feature layer before model training.The improved YOLOv4-tiny traffic sign detection model has been tested with an 46 FPS and a model volume of 22 MB.The results show that it can meet the requirements of high efficiency and light weight.(3)Analyze the lightweight feature extraction network Mobile Net,and use its effective feature layer and depth separable convolution to replace the original feature layer in YOLOv4 and the ordinary convolution block of PANet,so as to reduce the amount of parameters and the scale of the model.The improved lightweight YOLOv4 model has trained and tested,and the results show that it has high average accuracy and real-time performance for all classes,which can meet the requirements of relevant performance-constrained devices to achieve end-to-end traffic sign target detection.This paper has accomplished a comprehensive analysis of six traffic sign detection models,and the results show that the two models proposed in this paper can meet performance-constrained devices to achieve end-to-end traffic sign target detection.