Research on foreign object detection method in track area based on Mask-RCNN

Abstract

As the train speeds up, the damage caused by the collisions between trains and foreign objects are becoming more and more severe. Therefore, it is of great significance to monitor the intrusion of foreign objects in the track environment. In this paper, transfer learning is introduced into Mask-RCNN deep learning model. And the data set of rail image is used to train the model, which improves the effect of rail segmentation. The trained model is used to segment the rail in the picture, and the rail vulnerable area is divided based on the segmentation results. The sliding window ORB feature matching algorithm is used to calculate the similarity of vulnerable area. The detection of foreign objects in the area where is easy to invade is realized, and the detection reliability is improved. Experiments represent that this method has high accuracy, strong practicability, good robustness and universality.