Machine Learning Approach for Detection of Track Assets for Railroad Health Monitoring with Drone Images

Abstract

With the advancements in technology made in the 21st century, object detection has attracted a lot of attention in recent years. It is probably the most well-known term within the domain of computer vision and it encounters some really interesting problems. In order to gain complete image understanding, we classify different images along with also trying precise estimation of the objects and the locations of these objects contained in each of the images. Object detection is a collection of these related tasks for identifying objects in digital photographs. In this work, we perform detection of different track assets in drone images. For this purpose we have utilized a pretrained model as they are convenient for serving the purpose. The classes in which the images can be categorized into are listed as 'Construction', 'Power Junction/Brick', 'Cement Slabs', 'Transformer wires', 'Garbage', 'Person'. We have used 'YOLO'v3 (You Only Look Once) for track asset detection. We have initially given an image as input to YOLOv3 and the framework then divides the input image into grids. Image classification and localization are then applied onto each grid. When an image is provided as an input, YOLOv3 framework predicts the bounding boxes and their corresponding class probabilities for objects (if any are found). © 2022 IEEE.