Research On Traffic Sign Recognition Based On Deep Learning

With the rapid development of science and technology,assisted driving technology continues to advance,as an important part of assisted driving technology,the traffic sign recognition system can identify the traffic sign information in front of the driver and inform the driver,remind the driver of the current road condition information,in order to make a correct judgment on the information obtained,which can reduce a large number of traffic accidents caused by ignoring traffic signs.Therefore,traffic sign recognition has become the core research direction in the field of assisted driving.At present,in the research of traffic signs,researchers at home and abroad have achieved certain results,but because the traffic sign recognition system is often on the road to identify small targets,but also affected by weather,complex background,lighting and many other factors,the difficulty of recognition is further improved.In view of the above problems,this paper designs a traffic sign recognition method based on the improved YOLOv5 s algorithm,which can better realize traffic sign recognition in the natural environment,and the main work includes the following points:(1)This analysis and comparison of some open source traffic sign datasets at this stage,select the TT100 K dataset as the basis of the dataset trained in this paper,and in view of the uneven distribution of the number of categories in the TT100 K dataset,the number of traffic sign categories greater than 100 is extracted by script,and a dataset containing 45 types of traffic signs is made,and then the categories with less than 200 are supplemented by two methods of realistic acquisition images and data augmentation,and finally 13138 pictures are obtained.Re-TT100 K was made as the dataset for the experiment in this paper.(2)The network structure and loss function of YOLOv5 s algorithm are elaborated in detail,and the YOLOv5 s network model is improved according to the problems existing in the traffic sign recognition method based on YOLOv5 s algorithm at this stage.Firstly,the CBAM attention mechanism module is added to the YOLOv5 s network,so that the network can pay more attention to the important features in the picture.Then,according to the problems of the original loss function CIo U,the EIo U loss function is selected to replace it,which improves the convergence speed and detection accuracy of the model.Finally,aiming at the problem of small target detection,the K-means++ algorithm is used to recluster the dataset to obtain a prior box size more suitable for traffic sign recognition,and the Bi-FPN network is introduced to better extract small target features and extract more features while reducing network parameters.The experimental results show that the improved B-YOLOv5 s model has good effect in recognizing traffic signs in a variety of natural scenes,and has certain generalization ability,which is 3.737% higher than YOLOv5 s in m AP＿0.5,reaching 90.348%,and the recognition speed reaches 101 f/s,which meets the requirements of traffic sign recognition.(3)Based on the B-YOLOv5 model,a traffic sign recognition system is designed,which is written using Pyqt5 framework and Python language,and the completed system function is actually tested,and the results show that the system can stably and accurately identify traffic signs.