Research On Key Technologies Of Railway Intrusion Prevention Based On Edge Computing And Deep Learning

Railway security is related to the national economy and people’s livelihood,and the prevention of railway line intrusion is an indispensable part of railway work.The development of chip technology has gradually made intelligent monitoring based on deep learning into the mainstream.However,due to the weak computing power of edge computing platforms compared with server clusters,developers often need to seek a compromise between algorithm detection speed and effect.This paper studies the effect optimization method in the process of lightweight model design and the application technology of the model on the edge computing platform.The main contents are as follows:1)In view of the lack of current railway intrusion datasets,a dedicated dataset containing 1823 pedestrian targets and a total of 1066 images was produced through multiperiod fixed-point shooting,and it was applied to the research of the algorithm in this paper.2)A new type of stackable cross-network segment connection module is designed for the YOLO-Fastest network,aiming at the problem of accuracy loss that may be caused by the convolutional neural network in the process of feature extraction and downsampling.This module can provide a simple and rich feature reference for the previous-level network for the latter-level network,while requiring only a small increase in the amount of parameters and computation.The experimental results of the comparative test show that the introduction of this module can stably improve the m AP of the network,and the improvement effect can reach up to 2.4% in different combinations and stacking schemes.The results of the supplementary test experiments on the VOC2007-2012 joint dataset show that the improvement of the network detection effect of this module is independent of the dataset.3)By designing and comparing a variety of different transfer learning methods,it is verified that the improvement of the network detection effect based on the improvement of the inter-segment connection module is stable and positive.Finally,an intrusion detection model that can be used in specific railway scenarios is obtained,and the m AP performance of the improved model on the test set can reach 95.61%.4)An intelligent IPC software framework is designed for general intrusion detection tasks,and the improved network proposed in this paper is deployed on Hi Silicon Hi3516DV300 and Hi3519AV100 IPC So C embedded platforms based on the framework.The results show that the new network proposed in this paper In the case of satisfying the detection effect,the detection speed of 33FPS+/48FPS+ can be achieved on the two platforms respectively,which meets the real-time requirements of the application.