Research On Traffic Sign Detection And Recognition Method Based On Deep Learning

Road traffic in China has been developing rapidly in recent years and traffic conditions are becoming more and more complex.Achieving efficient detection of road traffic signs is not only an important component in the field of autonomous driving,but also a focus of intelligent transportation research.Efficient and accurate detection of traffic signs can not only give instructions for manual driving,but also lay the foundation for realizing driverless driving,and the subject has a relatively wide range of application scenarios.With the progress of digital era,Convolutional Neural Network has gradually replaced Machine Learning as the mainstream method in traffic sign recognition and detection due to its excellent detection effect.In this paper,the YOLO v3 algorithm is optimized for the problem of traffic sign detection and recognition,so that it can detect and recognize traffic signs in complex scenarios.The main work and innovation points of this paper are as follows:(1)For the problem that the classification of traffic sign dataset is not specific enough,an improvement method is proposed in two aspects.First,the existing dataset is processed with geometric distortion,optical distortion,bad weather data enhancement,and motion blur data enhancement according to the situations occurring in real driving scenarios,which expands the richness of the dataset;second,a more detailed classification of traffic signs is made,which is originally only briefly divided into warning,indication,and prohibition classes,and not more specific.In this paper,the common traffic signs are divided into 39 specific categories,so that each category has a specific representative meaning.(2)Optimization is made for the problem of high arithmetic power consumption and long computation time of backbone.After optimizing the residual structure in Darknet53 using the cross-stage partial network CSPNet,the network is lightened using Mobile Net v2.CSPNet alleviates the contradiction between the long computation time in the residual structure and the need for higher real-time traffic sign detection,and Mobile Net v2 reduces the computation of standard convolution and lightens the network.(3)Optimized the neck part of YOLO v3 neural network.The 3-level SPPNet is added to the FPN and the attention mechanism CBAM is used to enhance the feature representation to focus on important features.By using 3-level SPPNet,the feature extraction network can perform multiple feature fusion,which solves the problem of losing target detail information due to too many convolutions;the attention mechanism CBAM enhances the feature expression and focuses on important features,which improves the detection accuracy.(4)The experimental results show that the optimized algorithm in this paper improves detection accuracy by 5.67% and detection speed by 14.29frame/s compared with YOLO v3.Compared with the mainstream target detection algorithms SSD,Retina Net,Center Net,M2 Det and YOLO v4,the improved algorithm in this paper has a better performance in terms of model capacity and number of parameters.The improved algorithm has better detection speed and detection accuracy with relatively small number of models and parameters.