Georeferencing Railway Assets Detected in Camera Images And Updating High-Definition Railway Maps

Abstract

Geolocalisation of railway assets is a vital process in railway maintenance, necessitating high precision and often conducted manually via aerial maps and stereo vision cameras. This conventional approach, while reliable, is often time-consuming and expensive, driving an increasing demand for automation and improved accuracy through the use of mono cameras. This research seeks to transform this task's efficiency by automating the geolocalisation of railway assets such as level crossings, railway switches, and embedded rails using mono cameras. The study involves estimating the scale factor for mono cameras to recover depth information and then comparing the established OpenCV-based approach rotation matrix and the Algorithm Vanishing Point rotation matrix method with the novel method: Manual Vanishing Point rotation matrix methods. A custom algorithm was developed to georeference detected railway assets and assess the relative accuracy of the three methodologies. Upon evaluating the results against reference data (switch frogs), the Manual Vanishing Point rotation matrix method, a new proposition in this research, demonstrated superior performance, achieving a mean distance of 5.41m, a median distance of 5.33m, and a standard deviation of 1.72. Following this, the OpenCV-based approach rotation matrix registered a mean distance of 10.92m, a median distance of 9.00m, and a standard deviation of 7.68. Lastly, the Algorithm Vanishing Point rotation matrix method yielded a mean distance of 10.89m, a median distance of 11.00m, and a standard deviation of 2.01.