Research On Intelligent Visual Recognition Algorithm Of Foreign Objects In High-speed Railway Ballastless Track

During the operation of the high-speed railway,foreign objects（such as broken elastic strips,stones,animal carcasses,scattered parts）falling on the ballastless track bed are easily rolled up by the "train wind" and hit the EMU,which is a serious threat to the safety of high-speed railways.At present,the inspection of foreign objects on ballastless tracks mainly relies on the method of manual patrolling,and it is difficult to guarantee the detection efficiency and accuracy.Therefore,it is urgent to study accurate and efficient intelligent detection methods to detect foreign objects on ballastless track in time to ensure the safe operation of high-speed railway.In recent years,integrated inspection trains have been widely used in national railway system,which has significantly improved the efficiency of railway inspection operations.Based on the ballastless track inspection images dynamically collected by integrated inspection trains,this research focuses on the problem of accurate and efficient foreign object detection on high-speed railway ballastless tracks,and studies vision-based intelligent foreign object recognition algorithm.The main research work is as follows:（1）In view of the relevant background of foreign object detection in ballastless track in the current high-speed railway operation and maintenance,this paper expounds the necessity of carrying out the research on foreign object vision intelligent algorithm of high-speed railway ballastless track.In addition,this paper makes an in-depth analysis of the technical difficulties in the research,investigate the technical status of the field of track inspection,deep learning,computer vision and foreign object intrusion recognition.What’s more,it clarifies the research on ballastless track foreign object vision intelligent algorithm based on supervised deep learning.（2）This paper proposes a track scene recognition algorithm based on image classification.Aiming at the problem that the background difference of inspection images in different railway track scenes affects the recognition effect of foreign objects on ballastless tracks,the image classification network is used to distinguish track scenes,and label smoothing is used to improve the generalization performance of the classification model.The experimental results show that the image classification algorithm used in this study can effectively discriminate the scenes of ballastless main line,ballastless turnout and ballasted line.（3）This paper proposes a ballastless track foreign object recognition algorithm based on regional positioning.Ballastless track foreign objects appear in random patterns and have complex backgrounds.In order to achieve accurate positioning and recognition of ballastless track foreign objects in inspection images,this study propose a ballastless track foreign object detection model based on YOLO v4.Due to the huge difference in the size and pattern of foreign objects,the feature fusion module of the object detection model is improved in this paper to improve the model’s perception ability for multi-scale feature representation.And the loss function is modified to improve the localization accuracy of foreign objects.During training,the strategy of transfer learning is used to strengthen the learning of multi-category foreign object features.The experimental results prove the effectiveness of the object detection method in the recognition of ballastless track foreign object.（4）In order to further improve recognition performance of foreign objects in ballastless track,this paper proposes a ballastless track foreign object recognition algorithm based on semantic segmentation.A semantic segmentation model named RFODLab is designed,and the pixel-level information of foreign objects is accurately obtained by using the model to segment the track image.In order to further effectively correlate the contextual semantic information of the inspection images,a channel attention mechanism is introduced into the backbone network of the model,which realizes the weighted constraints of the model on the regions to be identified.At the same time,due to the problem of unbalanced distribution of sample categories in the problem of foreign object recognition in ballastless tracks,the loss function of the segmentation model is improved to balance the distribution of categories.The experimental results show that the method based on semantic segmentation can realize the recognition of various types of ballastless track foreign objects such as broken elastic strips,animal carcasses,and falling fragments at the pixel level,and the recognition effect is more finegrained..（5）In view of the false alarm caused by different environmental facilities of different lines,this paper proposes a foreign object false alarm suppression recognition algorithm based on metric learning.The algorithm takes the manual selection of specific facilities of the line as the reference image and the test image to be identified as suspected foreign objects to measure the similarity between them.Firstly,feature extraction network is used to extract the feature of the benchmark image and the image to be tested.Then,the similarity between the images is calculated based on the extracted feature map through the metric learning network for the discrimination of foreign object false positives.The experimental results show that the method based on metric learning can effectively distinguish the false positive foreign object,which can be eliminated in subsequent operations.（6）Based on the track inspection image intelligent analysis software,this study integrates the track scene recognition algorithm,the ballastless track foreign object recognition algorithm based on regional positioning and semantic segmentation,and the foreign object false alarm suppression algorithm into the software system,and the system operation framework is designed to form a ballastless track so as to form the prototype system of intelligent recognition of ballastless track foreign object.In addition,the test and verification of the prototype system under actual conditions is carried out on the inspection image data collected from three sections of Lanzhou-Xinjiang,ShanghaiKunming and Shenyang-Dalian high-speed railway with different environments.What’s more,based on the verification results,this paper analyzes the causes of false positives in field applications,and evaluates the advantages and disadvantages of regional positioning and semantic segmentation foreign object recognition algorithm.