Video Analysis Of Subway Tunnel Inspection Based On Deep Learning

Subway tunnel safety currently relies mainly on subway track inspectors to manually inspect rails when there is no train on the track.This method has slow speed,long line occupation time,and low working efficiency.The inspection effect is completely dependent on the experience and status of the track inspector.Aiming at this problem,this paper proposes a subway tunnel inspection video anomaly alarm system based on the improved GoogLeNet network model.The system can classify the tunnel conditions in the video in real time and sound an alarm when an abnormality is detected.The analysis result is accurate and efficient.Firstly,the collected subway tunnel inspection video is processed into a Subway-5 image data set suitable for convolutional neural network classification.Secondly,using the depth separable convolution in MobileNet_V1 and the skip connection in MobileNet_v2 to improve the GoogLeNet inception_v2 structure.Two building blocks including Subway_inception_v1 and Subway_inception_v2 are proposed.The proposed building blocks are used to build two networks,which are SubwayNet v1 and SubwayNet_v2.The network model and parameters are saved for subsequent work.At the same time,in order to facilitate human-computer interaction and make the detection effect display more intuitive,a graphical user interface is created.Alarm function is also added to the system.The alarm sound changes with the detection result.And then,the program files are packaged into an executable file for portability,and the application scenario of the analysis system is also designed.Finally,the accuracy and parameter quantities of the lightweight network such as SubwayNet_v1,SubwayNet_v2,MobileNet_v1 and MobileNet_v2 are compared on the Subway-5 dataset.The processing speed of the single inspection image of different networks is also compared.The network's feature map and class activation map are visualized to analyze the features and the basis for category determination extracted by the network.Furthermore,in order to test the versatility of the proposed network,the accuracy and parameters of different networks are compared on the CIFAR-10 data set.The accuracy of the SubwayNet_v1 network model on the Subway-5 dataset reached 94%,and the image processing speed on a single Nvidia TESLA-K80 GPU is 52 frames per second,enabling real-time and accurate analysis of the video.And the SubwayNet_v2 network model improves the accuracy to 96%on the basis of unchanged processing speed.The results show that the video analysis system based on the SubwayNet_v2 network model can detect abnormal conditions in the subway tunnel inspection video in real time and automatic alarm when abnormality is found.Manual inspection is prone to miss detection due to the subway tunnel is in low light environment and there are many power cables,water pipelines and other facilities The analysis system has strong anti-interference and high accuracy,and also has strong portability.