Road Traffic Identification Based On Deep Learning

Road traffic signs,as an important way of road traffic information,have become an indispensable part of driverless technology,which including lane speed limit reminder,lane direction indicator,lane warning reminder and other road traffic information.Research around this question has long been popular.In this paper,the traditional machine learning and image processing technology are combined with the deep learning model.The specific work details are as follows:1)Preprocess the data set.The completeness of training set has an important influence on the model of deep learning network in the deep learning training.A relatively complete data set is obtained by reducing the noise of road traffic sign data set image and the influence of complex background and other factors.By introducing fuzzy theory,fuzzy clustering algorithm alleviates the overly strict boundary division in the original clustering algorithm.On the basis of the classical FCM algorithm,PCM algorithm reduces the requirement for membership degree,thus reducing the influence of noise on the algorithm.On the basis of the kernel possibilitic c-means algorithm,the image segmentation is more accurate by introduces a maximum penalty term,overlapping of the centers between classes and reducing the fuzzy data misclassification problem at the boundary.A complete data set is obtained by rotating data enhancement operation,multi-color space image enhancement operation and normalized image size operation.Experiments show that the pre-processed data set has better recognition accuracy than the raw data set on the same deep learning network model.2)A new network model MRESE model is proposed.In the deep learning network model,deepening the depth of the network model is a method to improve the recognition rate.With the increase of computer performance,the depth of network model becomes deeper and deeper,and the training time consumed gradually increases.In order to solve the time consuming problem of deep learning model training.The MRESE model combines the Residual Neural Network(ResNet)model structure and Squeeze-and-Excitation(SENet)Network model structure.The ResNet model solves the problem of gradient disappearance or explosion to some extent.SENet model starts from the relationship between feature channels to improve the learning of useful features,and on the basis of reducing the depth of network model so as to reduce the training time,ensures the high accuracy of network model.The validity of MRESE model was verified by experiments with two different databases,the small database with BelgiumTSC and the large database with GTSRB.