Research On Appearance State Detection System Of Tarpaulin Train Based On Machine Vision

The safe operation of railway freight train is an important part of the development of rail transit.In the operation of receiving and sending railway trains in and out of the station,relevant operators need to check the train information and detect the running state of the train,but there are problems such as incomplete,unclear,uncertain and hidden safety hazards in the way of detecting the train state manually.Trains numerous models,this paper takes the appearance of the railway car tarpaulins state as the research object,combined with the target detection,image segmentation and image processing,such as advanced technology,cutting-edge machine vision algorithm is applied to implement tarpaulin car in and out of the station of irregularities in receiver operating on personnel,waist line and the analysis of damage detection,will eventually algorithm model for deployment in the embedded terminal,A tarpaulin train appearance state detection system is developed and its performance is verified.The main contents of this paper are as follows:(1)Firstly,study the illegal passengers and waist rope detection methods.The difficulty of detecting illegal passengers is mainly the lack of data.In this paper,the images of people and trains in VOC 2012 data set are fused to expand the data set of personnel detection.Aiming at the problem of low detection rate caused by existing algorithms on slender waist rope,large aspect ratio and redundant background,this paper uses the rotating rectangular frame based on long edge definition method to mark and improves YOLOv5,and proposes YOLOv5-RG algorithm to realize Angle prediction.Due to the inaccurate loss value caused by angle periodicity in angle prediction,Gaussian window function is used to process the predicted angle information so that the loss value can converge more stably in the training process.Through experimental verification,the average detection time of the improved YOLOv5-RG algorithm in this paper is 19 ms,and the average accuracy is increased by1.19%,which is more advantageous in accuracy and detection speed.(2)Secondly,the detection method of tarpaulin damage is studied.Firstly,the problem of insufficient tarpaulin damaged samples was solved by image processing.Then,the improved lightweight Deeplab V3+m algorithm is used to accurately segment the tarps.By adding a feature fusion operation in the decoding stage,and replacing the original backbone network with Mobile Net V2,the network is lighter and the detection speed is faster.Finally,the damage inside and edge of tarpaulin is detected by image processing.Experimental verification shows that the average detection time of the improved Deeplab V3+m algorithm in this paper is 38 ms,the average intersection ratio is improved by 1.69%,and the segmentation effect is better.The accuracy rate P,recall rate R and F1 values of the proposed tarpaulin damage algorithm all meet the requirements of tarpaulin damage detection.(3)Finally,the improved YOLOv5-RG model and Deeplab V3+m model in this paper are converted into 16-bit floating point Tensor RT model and deployed in embedded terminal device Jetson TX2.In order to meet the needs of actual deployment,a tarpaulin train appearance state detection system is developed to test system performance.The experimental results show that the system can meet the requirements of practical application for detecting illegal passengers,waist rope and tarpaulin damage.