Advanced Deep Learning–Based Hybrid Rail Extraction Algorithm Leveraging LiDAR Technology

In the United States, around 1/3 of the rail network is operated by short lines. These railroads play an important role in the nation’s transportation system by serving as the feeder and distributor for the rail network, but often lack a digitized rail track inventory for timely and efficient rail asset management due to limited resources. Much research has been conducted to develop automatic rail extraction methods, since it is a critical step toward a comprehensive digitized rail track inventory. However, existing methods strongly rely on high-density point cloud data sets, sensor property and configuration, and assumptions on global features; therefore, their applications in short lines are limited, since rail tracks will travel through different terrains with various global features, and data sets owned by short lines are mostly low-density data sets with unknown sensor property and configuration. To address these limitations, this study proposes an automatic rail extraction method that can be applied to low-density data sets and is independent of sensor properties/configurations, and global features. The proposed method is tested on the grade-crossing data sets collected by the Federal Railroad Administration (FRA) with a low point density around the track bed area. The performance shows an average completeness of 97.1%, correctness of 99.7%, and quality of 96.8%. This approach helps short lines to establish their own digitized rail track inventory, allowing for effective operation planning and investment strategy, and builds the foundation for future geometry measurements and infrastructure management, thereby improving operational safety and efficiency without significant investment in high-end sensors and high-density data sets.