Traffic Sign Detection And Recognition Based On Deep Learning

Traffic signs can enable drivers to quickly perceive road traffic information and effectively reduce the probability of traffic accidents.Traffic signs are usually designed with conspicuous colors,but the differences in weather,lighting and other conditions make it difficult for drivers to concentrate on distinguishing the category of each traffic sign in more complex natural scenes.In addition,the driving camera on the car needs to capture the image of the traffic sign when driving at high speed.In most cases,the traffic sign image recorded by the driving camera is a very small target for the entire image,and its positioning and recognition face with great difficulties.Therefore,how to assist drivers in accurately detecting and recognizing the types of traffic signs in natural scenes has important practical value.This paper studies the performance of the deep learning model on the traffic sign data set TT100 K,and tries to use different deep learning models to assist the driver in recognizing traffic signs,so that the driver can focus more on the driving process during driving.This article first studies the performance of the deep learning algorithm YOLOv4 in the detection and recognition of traffic signs.In the driver’s perspective,traffic signs in natural scenes are relatively small targets.In training process,we use the method based on global K-means clustering to perform anchor regression,and finally the boundary box prediction for traffic sign recognition in natural scene is obtained.Secondly,in order to solve the phenomenon that the convolutional neural network produces internal covariate shifts during the training,it is different from the traditional convolutional layer followed by batch normalization layer separation.We use the BN fusion technology to fuse the BN layer into In convolution,the speed of network operation is increased,a large number of calculation processes are reduced,and the efficiency of network recognition is improved.Affected by the automobile hardware environment,the current deep convolution model is difficult to apply to mobile devices directly.In YOLOv4,in order to increase the recognition accuracy of the model,a multi-scale bounding box is used to select a large number of candidate regions.This operation makes the speed of the model is reduced.We designed a lightweight deep convolution model that uses ghost net as the backbone feature extraction network.This model uses linear mapping instead of a large number of convolution operations in the network,greatly reduced the amount of calculation of network;The spatial attention module is designed to strengthen the main feature extraction network to capture sensitive information;The path aggregation network is used in the feature fusion part,and all convolution operations in the path aggregation network structure are omitted,and both the up-sampling and the down-sampling are completed by interpolation,and the multi-scale feature maps are directly added,so that the calculation amount of the entire feature fusion operation becomes very small.In addition,channel attention module is added to each fusion scale network to increase the ability of feature fusion.The results show that the improvement of the model by the attention mechanism and the lightening of the network model can improve the overall recognition speed of the network,and the accuracy is also better than that of the network model without the attention mechanism.