Research On The Algorithm Traffic Signs Recognition And Detection

With the rapid growth of the global economy,the number of cars is increasing year by year,and the frequency of traffic congestion and traffic accidents also increases.In order to solve traffic problems,scholars from all over the world are actively researching intelligent transportation systems.The identification and detection of traffic signs is one of the important research contents of intelligent transportation systems.The technology in this field mainly faces the problems of low recognition accuracy and high cost.Convolutional neural network technology in deep learning can provide technical support for accuracy improvement.Therefore,how to reduce the cost of equipment use while maintaining a high recognition rate is the focus of this research.Through research on improving accuracy and network lightweight,this research provides new ideas for the development of unmanned driving in the direction of traffic sign recognition and detection.Traffic signs play a vital role in regulating traffic and promoting cautious driving.Automatic traffic sign recognition is one of the key technologies of autonomous driving.The accuracy of traffic sign classification is a very important indicator in the performance evaluation of vehicle navigation systems.This paper studies the methods of improving the accuracy of traffic sign detection and reducing the false detection rate,and improves the existing recognition network framework in terms of lightweight.The main tasks include:1.An improved deep mutual learning network is proposed.Aiming at the problem of large network size and difficult deployment on the mobile terminal,the lightweight feature extraction network and simple classifier are used to reduce the weight of the network;By using loss hyperparameters,the convergence speed of the model is improved.A lightweight network with a recognition accuracy of 98.98% was realized on the German traffic sign recognition data set.Compared with the previous improvement,the accuracy has increased by 2.25%.2.An improved VGG network traffic sign recognition method is proposed.Aiming at the problem of low accuracy of traffic sign recognition,the channel attention module is used to extract important features in the image in a targeted manner to improve the recognition performance of traffic signs.The recognition accuracy of 98.46% was achieved on the German traffic sign recognition data set,which is an increase of 1.67% compared to before the improvement.3.Propose an improved Shuffle Net network traffic sign recognition method.Aiming at the high computational cost of the existing model,the model is reduced by using channel pruning and the memory usage at runtime is reduced.The recognition results on the German traffic sign recognition data set show that after pruning,FLOPs are reduced by 6 times and params are reduced by 8 times.At the same time,a recognition performance of 96.68% is achieved,and the accuracy is improved by 0.81% compared to before pruning.4.An improved YOLOv3 traffic sign detection network is proposed.Aiming at the problem of high false detection rate of traffic sign detection,the feature extraction network is improved to improve the feature extraction performance;the new detection frame positioning loss is introduced to reduce the false detection of the detection model rate.The detection results on the German traffic sign detection data set show that the precision value of the improved model increased by 21.1%,Recall increased by 6.2%,m AP [0.5] increased by 6.5%,and m AP [0.5:0.95] increased by 8.4%.