Intrusion Detection System For Key Railway Areas Based On Deep Learning

In the current railway system,key areas(including bridges,tunnels,and throat areas)are not allowed to enter the road.However,the objects such as people often illegal enter the key areas of the railway without permission,leading to major accidents.With the rise of artificial intelligence technology,deep learning has become a more common technology to replace repeated manual operations,which can greatly improve efficiency with the same accuracy or even higher than human recognition accuracy.This paper proposes a new type of intrusion detection system for railway critical areas based on deep learning,which can effectively identify foreign objects in key areas,ignore foreign object of non-critical areas,and identify and detect key areas of interest automatically.When the system detects the foreign objects in key areas,it will actively send out alarm information to avoid the occurrence of intrusion accidents.Data analysis and statistics in the web terminal are provided to the staff with comprehensive intrusion information.The system uses the deeplab semantic segmentation model to ensure that the railway area can still be identified when the background of the image changes,and the key and non-key areas are segmented.The part of object detection uses the YOLO model.The system compares whether personnel intrusion has occurred acoording to the foreign object and the key area of the railway are overlap.In order to improve the detection efficiency,this paper proposes optimization and improvement methods for the object detection model which is sparsely trained,compressed and cut.The final detection data is analyzed and calculated through the terminal data platform based on finereport.The terminal data platform reports detection information in real time and issues an early warning of intrusion events.The main research content and research work of this paper are:(1)In this paper,the system uses the combination of deeplabv3 and YOLOv3 to realize the detection of foreign objects in key areas of the railway,distinguishing between foreign objects in the rail area and foreign objects in the non-rail area.The system can intelligently re-identify key railway areas after the background of the image is changed when the camera is rotated or zoomed,without human correction.(2)This paper optimizes and improves the YOLOv3 object detection model to make the model more suitable for object recognition in railway scenarios.By compressing and cutting the model parameter volume,the model parameter volume is made smaller,which not only reduces the memory consumption during model calculation,but also improves the detection performance of the model,increases the calculation speed,and improves the detection efficiency.(3)This paper builds a terminal platform webpage based on intrusion detection results.The intrusion detection data is reported to the terminal platform of the intrusion detection system,and the intrusion detection results can be viewed on the terminal platform in real time,and the statistical analysis of the data is completed to provide more comprehensive information.The detection accuracy of the railway key area of intrusion detection system can reach more than 95%.Optimized detection model is more suitable for object detection in railway scene.The m AP is increased from 0.77 to 0.95.After the compression and cutting of the model,the loss does not exceed 0.01.The memory usage of the cut model is reduced to one-fifth of the original model.The number of simultaneously detected video streams increases from 6 streams to 10 streams.The efficiency has increased by 8 times under the same computing resources.