Design Of Traffic Identification Intelligent Identification System Based On TensorFlow

As an important part of intelligent driving system,traffic sign recognition system plays an important role in many aspects,such as advanced driving assistance,automatic traffic sign maintenance,autonomous driving of unmanned vehicle,and so on.It is the key and foundation of intelligent driving system research.However,the real road traffic environment is much more complex and changeable,and many problems such as light conditions,weather conditions,partial occlusion,angle of view tilt and background color similarity interference make the research of traffic sign recognition system face many difficulties.Based on summarizing the research status of traffic sign recognition technology,the problem of the existing methods are analyzed in this dissertation.A traffic sign recognition method based on TensorFlow framework and convolutional neural network is proposed.Traffic sign recognition system is designed and implemented.The main work is as follows:1.Aiming at traffic sign recognition problem,a multi-scale feature convolutional neural network based on TensorFlow framework is achieved.The convolutional kernel slip filtering is used to extract features,and the maximum value pooling technique is used to reduce the dimension.Optimize network performance by adjusting parameters in the network structure and using rectified linear unit activation functions.Dropout,L2 regularization and early stopping are used to prevent overfitting.The network model is tested on the GTSRB dataset,and the recognition accuracy is 99.26%.2.To solve various types of imbalance in data samples of GTSRB dataset,horizontal and vertical mirroring,random rotation,and projective transformations are used to increase the dataset's balance.Aiming at the problem that the traffic sign will interfere with the feature extraction under different lighting conditions,the improved histogram equalization method is used to preprocess the data samples in the GTSRB traffic sign dataset.In order to enhance the security,reliability,flexibility and get faster speed,this dissertation configures the data preprocessing and training process to run on the cloud server.3.The software of the traffic sign recognition system is designed and implemented.In the software,the convolution neural network model is imported,and the test of the live pictures and some video samples is completed.The experimental result is relatively good.In summary,the proposed method can achieve high detection rate and classification accuracy,while ensuring recognition efficiency.The software of the traffic sign recognition system which is developed in this dissertation can realize the recognition of the traffic sign in the real picture and the video stream,finally obtains the test accuracy rate of 80.36%.