Research On Railroad Intrusion Target Identification Method Based On Edge Computing

With the rapid construction and development of various railroads in China,the role played by railroads in transportation has become more and more important,however,due to the inevitable loopholes in railroad security protection facilities,foreign object intrusion events may occur,thus posing a serious threat to the normal operation of trains.Existing detection methods are broadly divided into two types: one is the traditional integrated railroad video surveillance method,which relies on human eyes to check afterwards and thus has low detection efficiency and poor accuracy;the other is the use of deep learning-based target detection algorithms,but this method needs to rely on servers deployed in the back end to achieve,which is affected by network bandwidth and cannot guarantee the real-time detection.For this reason,it is necessary to design and establish a railroad foreign object intrusion recognition detection system based on edge computing and deep learning to ensure railroad transportation safety.In this paper,I have mainly completed three aspects of work: the design of intrusion target detection algorithm based on CBAM-DIo U＿nms-YOLOv5 model,the design of concise network based on BN layer pruning,and the optimization of detection system based on edge computing.Firstly,I collected the image data along the railroad line to establish the railroad dataset,and used the public dataset for data type leveling to provide data support for subsequent model training and testing;then,I used the YOLOv5 x network as the base model,and introduced the CBAM detection module to suppress useless information on the basis of the original network to help the model perceive the target to be detected quickly from the complex railroad scene;based on the improved model,the original weighted nms strategy was replaced by the DIo U＿nms strategy to improve the recognition results of the detection model for the obscured target.Secondly,the obtained CBAM-DIo U＿nms-YOLOv5 model is subjected to a new model trimming operation based on BN layer pruning to obtain a lightweight model,which can not only achieve the effect of fast pruning to save computational resources and experimental time costs,but also achieve the purpose of reducing the model size to enhance the detection speed with almost no loss of detection accuracy.Finally,to truly detect the target at the railroad site,the recognition algorithm was migrated and deployed to NVIDIA AGX jetson Xavier through the processor environment configuration,and the algorithm model was accelerated at the inference level using the Tensor RT model inference acceleration library.The railroad intrusion detection algorithm based on the CBAM-DIo U＿nmsYOLOv5 model designed in this paper achieved a detection accuracy of 95% for pedestrian detection and 86.2% for combined pedestrian,train and animal detection in tests on the railroad dataset.The edge computing-based detection system implemented in this paper achieves a detection speed of 15 fps in experiments in a real railroad environment and maintains normal operation for a long time.The algorithm experiments and the overall system experiments in the real railroad environment show that the edge computing-based railroad foreign object intrusion detection system designed in this paper has good accuracy and real-time performance,which is important to ensure the safe operation of trains.