Research On The Normative Identification System Of Train Maintenance Personnel Safety Protection Based On Computer Vision

Railroad transportation carries the important task of achieving economic take-off,and the maintenance and repair of trains is the key to ensure the safe operation of railroad transportation,so train maintenance work has always been the focus of the industry.Train maintenance workshop has strict requirements on the behavior of maintenance personnel,but the traditional supervision and management solutions are not good enough to restrain the violations of maintenance personnel,so it is of great significance to create an efficient and intelligent supervision solution for train maintenance workshop.Based on image recognition technology,this thesis adopts machine learning and deep learning methods to realize a set of intelligent supervision system,which can efficiently supervise and identify the standardization of the safety protection apparatus worn by the maintenance personnel in real time,and promptly remind the personnel who wear irregularities to eliminate the safety hazards and guarantee the safety of the staff.(1)Detection and tracking of personnel is the basis for monitoring all violations.In this thesis,we propose to use DC-ECA-CenterNet,a lightweight target detection model,to achieve end-to-end personnel detection,and then implement a new personnel tracking algorithm(ASCenter Track)based on DC-ECA-CenterNet to achieve real-time and efficient personnel tracking by fusing the information of the previous picture frame.The backbone part of CenterNet,a Resnet-50 network model,is selected to achieve a balance between detection accuracy and speed;the standard convolution is replaced by deformation convolution to overcome the problem of missed detection due to various postures of maintenance personnel;the ECA-Net attention mechanism is fused with the detection head part of CenterNet to solve the problem of missed detection due to mutual occlusion of maintenance personnel.In order to solve the problem that the human target tracking algorithm(Center Track)often fails to track when people are blocked by each other or crowded,the fusion of target center distance and target frame area similarity(AS)as a new target association index effectively improves the tracking accuracy.Experiments show that the algorithm proposed in this thesis has a good effect of personnel detection and tracking in the maintenance workshop scene.(2)On the basis of completing the localization and tracking of personnel,the CBAMYOLOv5 s algorithm is used to achieve the detection of helmets and masks worn by maintenance personnel.The Convolutional Block Attention Module(CBAM)is incorporated in three different positions of the Backbone part of YOLOv5 s to enhance the extraction ability of Backbone for important features and improve the recognition accuracy of the model for helmets and masks;CIo U Loss is used to replace GIo U Loss to effectively improve the convergence speed of the network during training and also improve the regression localization accuracy of the network.The original NMS is replaced by DIo U-NMS to further improve the detection accuracy when targets are mutually occluded.The experimental results show that CBAM-YOLOv5 s has excellent performance and good detection effect.(3)Propose a combined maintenance personnel seat belt wearing detection algorithm.Firstly,an image area determination technique is proposed to determine the area in which the torso of the maintenance personnel is located and determine the detection target;then a preprocessing operation is performed on the image of the personnel in the operation area to highlight the information of the safety belt;finally,a HOG+SVM+connected domain area determination algorithm is proposed to determine whether the personnel wears the safety belt.The whole set of algorithm solution can accurately determine the personnel’s seat belt wearing situation and meet the actual project demand.(4)The program is written to orderly integrate all the trained algorithm models in this thesis and deploy them into the embedded device(Jetson TX2)to form a whole set of intelligent supervisory system.Then,the whole system is analyzed and the reliability of the whole solution of this thesis is evaluated.