Research On The Method Of Object Detection In The Automatic System Of Hump Uncoupling Based On Deep Learning

China's railways extend in all directions and spread all over the country.They are not only the national long-distance travel but also the important means of transportation for goods.The growing economy has brought great tests to railway transportation.Apart from speeding up the train,it is also a direction to study how to improve the transportation efficiency.One step is to improve the efficiency of car disintegration in marshalling stations.The marshalling yard has to finish a large number of truck unmarshalling operations every day.At present,the automation system has been put into use to improve efficiency,but the hooking process is still done manually.It is necessary to realize complete automatic uncoupling.This paper studies the visual part of hook picking,that is,the detection of carriage and coupler handle.Because there are many types of cars and coupler handles,the traditional visual inspection method is difficult to design features,so this paper will adopt the method of depth learning.First,design the overall hook removal scheme,determine the hook removal process,analyze the requirements of the target capture system,and prepare for the next research.Aiming at the detection of cars type,an improved SSD detection model based on depth learning is proposed.The data set between carriage and carriage gap is newly built.The clustering algorithm is used to reset the number and aspect ratio of initial anchor frames and the data set is trained to obtain weight files.The comparison test with the original SSD shows that the improved SSD is superior to the original method as a whole,and the mAP is increased from 80.2%to 86.1%;The detection speed is increased to 41 frames per second.Finally,based on the detection results of the improved SSD model,the counting method of a car is designed by taking advantage of the fact that a car passes through with and without gaps.The test results show that the counting accuracy reaches 93.3%under the condition of clear video,which meets the counting requirements of the carriage.The coupler handle is in the shape of a thin strip and occupies a small area.The detection of the coupler handle involves the subsequent picking action of the manipulator.The contour of the handle needs to be detected to obtain an accurate handle position.In this paper,the image segmentation method based on depth learning is used to mark the contour of the coupler image.mask rcnn and deeplabv3+are used to segment the network training data set.Finally,the experimental test results are compared with the label map.The results show that the coupler can be detected by both methods,but the coupler handle profile obtained by deeplabv3+test is smoother and more complete,mIoU reaches 75.0%,and the detection speed is 3 seconds/picture.