An Automated Rail Extraction Framework for Low-Density LiDAR Data Without Sensor Configuration Information

Around one-third of the U.S. rail network is owned and operated by shortlines (Class II and III railroads). The infrastructure conditions of these railroads are marginal, and their revenue and number of employees are insufficient. To conduct proper and timely infrastructure management, it is critical for shortlines to establish a reliable and cost-effective inventory of their existing rail tracks. Significant efforts have been made to develop and employ automatic rail extraction methods as it is a critical step to establishing a rail infrastructure inventory. However, existing methods rely heavily on high-density point cloud datasets with known sensor properties and configurations or sophisticated imaging sensors. These requirements prevent their deployments in shortlines because point cloud data available to shortlines are often in low-density and with unknown specifications constrained by their financial situations. To address those limitations, we propose an automatic, configuration-independent coarse-to-fine extraction method for low-density LiDAR data. The proposed extraction framework is only developed based on high-level geometric features of the rail that can even be captured by low-density point cloud data with unknown sensor properties and configurations. The proposed method is evaluated on the existing grade-crossing dataset collected by the Federal Railroad Administration for grade-crossing safety evaluation with a point cloud density of only 293 points/
$\text{m}^{{2}}$
. The overall performance shows average completeness of 96.97%, correctness of 99.71%, and quality of 96.67% for all extraction scenarios. The promising results provide shortlines a capability to extract rails for any geometry measurements based on any low-density LiDAR data available to them.