Research On Traffic Sign Detection And Recognition Algorithm Based On Convolutional Neural Network

In the field of vehicle unmanned driving and assisted driving,traffic sign detection and recognition related algorithms are important research topics.The performance of these algorithms directly affects the driving safety of motor vehicles in road traffic.Especially today,under the wave of intelligence,how to achieve the accuracy,real-time,stability and anti-interference of traffic sign detection and identification has always been the focus of research.The subject of this paper is to use the "CNN convolutional neural network" method to do research on traffic sign detection and recognition related algorithms.The key research is as follows:A traffic sign detection algorithm based on improved SSD network model is proposed.The original SSD model has a problem of performance degradation in detecting small targets such as traffic signs.Firstly,the related reasons are analyzed through the “receptive field” in CNN.It is believed that the relevant network layer at the top of the original model lacks the ability to extract local and detailed features,which leads to this problem.For related reasons,an improved method is proposed: the original model Adding the "Feature Fusion Module" significantly improves the detection performance of the original model for small targets such as traffic signs;replacing some of the network layers in the original model with the "improved Inception network layer" can minimize the computational cost of the model and keep the original The advantages of the model "quick detection" and the increase in the number of multi-scale convolution kernels further improve the performance of traffic sign detection;it also optimizes the structure of the extra layer in the original model and continues to reduce the amount of parameters calculated by the model.Experiments were carried out on the GTSDB public dataset.The mAP value of the improved method reached 0.894,which solved the problem of the performance degradation of the previous model.A traffic sign recognition method based on ResNet residual network model is proposed.Firstly,the data set is expanded by the "angle rotation" method,which avoids the occurrence of over-fitting problems in the model training process.Secondly,the data set image is converted into "YUV color space",and the YUV compares the RGB color space.The features extracted by the network model are more effective,and the classification accuracy of the model is further improved.A “six-block ResNet residual network model” is designed for the identification of traffic signs,which can extract deeper features and prevent them from being over-exposed.In the end,the data set enters the design model to train and test,and the related experiments are carried out on the GTSRB public data set.The test accuracy of the proposed method reaches 99.72%,which is better than many current methods.