Locomotive Speed Sensor Image Region Detection Based On Two-step Learning Strategy

Trains are an important tool for modern transportation.Whether it is a cargo train or a passenger train,train maintenance is an indispensable task for safety management.The maintenance of speed sensors for locomotives(i.e.,train heads)is one of the most important tasks.Different vehicles have different speed requirements,train speed is an important reference for train drivers and an important guarantee for the safety of the passengers.The train driver determines whether to accelerate or decelerate the train according to the current speed collected by the speed sensors.If the speed is not controlled,the safety cannot be guaranteed.With the increase of population mobility and the increasing number of goods transported,the demand for trains is also increasing.A large amount of manpower is required,the efficiency is low,and it is easy to cause life injury with traditional train maintenance method.In this paper,we identify and process image with speed sensor collected from spot based on the technology of image processing and deep learning.A method of image region detection and classification based on a two-step learning strategy is proposed,from which to we can detect and discriminate whether there is a speed sensor or not in the test images.The aim of this paper is to detect whether there is a speed sensor in the sample images collected from the spot.The feature of the sample is that,the size of the image is large,while the target(speed sensor)in the image is small,which is similar to the background color around the target at the same time.The samples image include daytime and night scene,and background is complex.There is often a lot of pollution noise such as mud.The main work of this paper is as follows:(1)A two-step learning strategy is proposed to detect the image.Firstly,the Region of Interest(ROI)is determined by the method of target region localization based on deep learning.Then the improved convolution neural network model is used to classify and recognize the ROI.The automatic judgment of speed sensor is realized,and the accuracy of judgment is improved.(2)Two kinds of advanced target detection You Only Look Once(YOLO)and Single Shot Multi box Detector(SSD)algorithms are used to locate ROI.At the same time,VGG-16 structure network is used to locate ROI directly.Three localization algorithms of ROI are realized.Among the three methods,target area localization algorithm based on VGG-16 has the best performance.(3)Speed sensor detection is carried out after the installation areas of locomotive speed sensors are correctly located.In this step,two detection algorithms are adopted: LeNet-5 convolution neural network model and improved LeNet-5 convolution neural network model.By comparing the accuracy of the detection results of the two algorithms,it is found that the improved LeNet-5 convolution neural network model is more effective according to the characteristics of the samples.After a large number of tests,the classification and recognition accuracy of the proposed algorithm can meet the spot experimental requirements,the results show that the algorithm has good practicability.