Recognising railway infrastructure elements in videos and drawings using neural networks

Accurate data in the form of technical drawings of built assets are an essential requirement for the successful operation and reconstruction of the built environment. When the consistency between these data and the real-world situation cannot be ensured, the data are not reliable and need to be verified by comparing drawings and reality. Depending on the size and the number of assets, this may involve an enormous amount of manual effort. In this paper, an approach to supporting and automating this process by utilising machine learning concepts has been developed in the context of railway engineering. The research focuses on two aspects: The analysis of technical drawings to locate plan symbols and the recognition of infrastructure elements in video data of railway lines. Both tasks are time-intensive and errorprone processes when done manually. In this paper, it is described how the capabilities of convolutional neural networks are employed in analysing images from video data and of technical drawings, in order to detect automatically the location of railway infrastructure elements. The outcome of these two approaches can then be compared with catalogue elements and to check the consistency of corresponding technical drawings.