Railway Track Foreign Object Intrusion Detection Based on Improved YOLOv5

Objects invading railway clearance pose a great threat to normal railway operations. Indetifying intruders within the railway clearance limit was of great significance for ensuring the safety of train operation. In response to the problems of large localization deviation and high false detection rates in the YOLOv5 benchmark model, YOLOv5 is improved by optimizing the anchor algorithm. Two classic anchor algorithms which are Auto anchor and K-Means clustering anchor are introduced to take place of the initial anchor of the YOLOv5. The experiment results show both Auto anchor and K-Means clustering anchor can effectively predict the actual size of orbital foreign objects, with K-Means clustering anchors being superior. It also indicates that optimizing the anchor that can improve the regression accuracy of prediction box in some degree, thereby improving mAP of the detection model.