Onboard Image Recognition Method for Detection of Foreign Objects on High Speed Railway Ballastless Tracks
An image recognition method was developed to detect foreign objects on ballastless tracks of high speed railways,thus addressing the recurrent issue of foreign object presence on these tracks during operations.This method is based on the improved DeepLab semantic segmentation model of foreign objects on ballastless tracks.The pixel-level information of foreign objects can be accurately captured by using the track image segmentation results from this model.To improve the foreign object detection rate and accuracy,a channel attention mechanism was introduced into the backbone network of the model to correlate image context information and enable weighted constraints of the model on areas to be detected.Furthermore,the loss function of the model was improved by balancing the category allocation proportion to address the unbalanced distribution of sample categories affecting the model in anomaly detection of ballastless tracks.The test results demonstrate that this method can detect various types of foreign objects on ballastless tracks at the pixel level,achieving a detection accuracy of 90％and a detection rate above 95％on the test set.
high speed railwayballastless trackforeign object on trackimage recognitionanomaly detectionsemantic segmentationattention mechanismloss function
NETL
NSTL
万方数据
