An Identification Method Of Intelligent Rail Transit Road Sign Based On Convolutional Neural Network

Rail transit plays an important role in China’s comprehensive transportation system.Intelligent perception of environmental information around rail traffic is also becoming more and more attractive.In the rail transit system,power transmission is the basis of the train operation,and timely processing of power transmission faults requires inquiry of the transmission frame number.By using the deep learning method,the traveling video of the train can be utilized,so that the computer can automatically detect the sign number of the front transmission frame,improve maintenance efficiency.And it plays an important role in timely troubleshooting and ensuring the normal operation of the railway system.The key point in the identification of the road sign is the identification and recognition of the road sign.The trouble is that there are many numbers in the road sign;the road sign is small and difficult to detect;the video image quality cannot be guaranteed because there are too many noise.All of these make the design of the algorithm very difficult,and there are still great challenges in balancing the accuracy and speed of the algorithm.In view of the above problems,after analyzing,the road sign detection module,semantic segmentation module and digital recognition module are used to identify the numbers in the road signs:1)In the road sign detection algorithm,a target detection algorithm based on convolutional neural network is proposed.The algorithm fully considers the problem of small road sign target,and uses multi-scale feature map detection to solve the scale change problem caused by the signs from far and near.Secondly,the algorithm uses the idea of the anchor box and sets the a default box to make the algorithm more suitable to detect the target.Finally,the feature fusion method is introduced in the algorithm,so that the spatial information of the shallow features in the convolutional network can be fully utilized to improve the detection effect.2)In the semantic segmentation algorithm,the different situations in the day and night are discussed separately,and the full convolutional neural network is used for semantic segmentation under daytime conditions to solve the problem of the noise in the image;In the dark we use the condition of high digital and background contrast in the road sign,the method of adaptive threshold segmentation and morphological operation is used for semantic segmentation.It balances segmentation accuracy and algorithm speed.3.In the digital recognition algorithm,we use a small convolutional neural network,by reducing the model parameters,the operation speed of the algorithm is guaranteed under the condition of ensuring the accuracy of the algorithm identification,and the experimental results in accordance with the expected results are achieved.Using the proposed road sign recognition algorithm based on convolutional neural network,the experiment is carried out under the test dataset,and the experimental results are analyzed.The mAP(mean Average Precision)of the final road sign detection algorithm reached 88.32%(in the day)and(in the dark);the semantic segmentation algorithm mIOU(mean Intersection over Union)reached 83.67%(in the day)and 85.39%(in the dark);the correct rate of digital recognition algorithm reached 93.44%.Finally,the detection accuracy of the overall algorithm in the video is 87.98%(in the day)and 72.92%(in the dark).