Research On Real-time Detection Method Of Railway Perimeter Disaster Based On Video

It is estimated that by the end of 2020,the total mileage of my country’s railway business will reach 146,000 kilometers,the environment along the railway line is complex,and the perimeter of the railway is vulnerable to natural disasters.The occurrence of natural disasters along the route will affect the safety of railway operations.Failure to detect them in time may lead to accidents and cause significant loss of personnel and property.Intelligent detection based on video surveillance has been gradually applied in perimeter security detection.Combining with the current development direction of railway perimeter safety detection technology,this paper mainly studies the problem of railway perimeter disaster detection,and proposes a video-based railway perimeter disaster realtime detection method.First,according to the perimeter environment of the railway,the types of disaster data to be detected are determined,and the disaster data set is constructed through operations such as data collection,data labeling,and data augmentation.Through the research on the target detection algorithm based on deep learning,three currently mainstream deep learning target detection algorithms are selected: YOLOv3,Faster R-CNN and SSD,and these three algorithm models are built and deployed on the server.By using the disaster data set to train the three algorithms separately,observe the training process and compare the detection effect of the trained model,and use the algorithm with the best performance as the algorithm framework of natural disaster detection in this paper.Finally,this paper proposes a video-based real-time detection method of railway perimeter natural disasters.According to the characteristics of railway perimeter video images,the background difference method of mixed Gaussian background modeling is used for video preprocessing,which reduces the method to a certain extent.Consumption of computer resources;parameter adjustment,model optimization,and comparative training of the deep learning disaster monitoring part of the algorithm,further improve the detection effect on the disaster data set.