Research On Foreign Object Detection Method Of Railway Drainage Structure Based On UAV Inspection

Railroads are developing rapidly in China,and with them,railroad safety hazards are increasing.Among them,water damage is one of the important factors causing railroad safety hazards,so it is necessary to regularly inspect and clean up the railroad drainage structure.The current inspection method on the railroad,mostly manual inspection,is difficult to meet today’s inspection efficiency needs,especially the railroad tunnel area,its complex terrain,drainage system built on the mountain,greatly increasing the difficulty of manual inspection.Therefore,how to efficiently realize the railroad drainage structure foreign body detection has become an urgent problem.This paper relies on the portability and efficiency of UAV inspection to obtain the image data of the railroad site,and classify,detect and segment the images through deep learning methods to assess the danger of foreign objects in the drainage structure to achieve efficient and intelligent detection of railroad drainage structure,and the method can provide a new way of thinking for the inspection of railroad drainage structure.This paper focuses on the following aspects of research:1.Aiming at the problem of inconvenient data acquisition at railroad sites,the UAV image acquisition scheme was designed and the inspection route of railroad tunnel cavity elevation slope was improved.Through the field test,the image data of railroad sites were successfully acquired for the construction of data sets in the subsequent network model.2.To address the problem that the UAV inspection images do not contain only one detection target of railroad drainage structure,an automatic classification method of UAV images based on Efficient Net V2 is proposed.The image data acquired by UAV inspection is divided into four categories,and the ECA attention mechanism is introduced in the MBConv module and the size of the convolution kernel of the depth-separable convolution is adjusted,while the overall module stacking structure of the network is changed.While the network is lightweight,the classification efficiency of UAV images is improved,and the automatic classification of railroad drainage structures in railroad UAV images is realized.3.To address the problem that the foreign object features in railroad drainage structures are consistent with the characteristics of the drainage structure surroundings,a YOLOX-based foreign object detection method for railroad drainage structures is proposed.The classified drainage structure images are used as the data set,and the network is divided into five categories of detection targets according to the foreign object features,and the triage of the network is increased to obtain deeper feature information.The residual module and spatial feature pooling pyramid module in the backbone network are replaced,while the EIo U loss function is chosen to improve the problem of unbalanced positive and negative samples to a certain extent,and the accuracy of foreign object detection of railway drainage structures is improved.4.To address the problem of low efficiency of foreign object hazard assessment of traditional railroad drainage structure,an intelligent assessment method of foreign object hazard of railroad drainage structure based on U2 Net is proposed.The semantic segmentation dataset is constructed based on the target detection dataset,and considering the model size of the segmentation network,the depth separable convolution is used to replace the ordinary convolution involved in sampling in the network,and the ECA attention mechanism is added to the feature fusion module to obtain the foreign object segmentation results in the form of a mask.Finally,a hazard assessment module is designed to calculate the pixel occupancy of the segmented foreign objects in the drainage structure,assess their hazard level to the railroad drainage structure,and propose remediation suggestions.In this paper,based on the UAV image data collected from the railroad site,three networks of image classification,target detection and semantic segmentation are tested in turn.Aiming at the target of foreign objects in railway drainage structure,the improved network is compared with some mainstream networks,which proves the advantages of the research method in this paper in terms of detection efficiency and accuracy,and the method can realize efficient and intelligent detection of railroad drainage structure,which has certain practical application value.