Up-to-date digital network maps are essential to building, maintaining, and upgrading railway systems. It is an enormous process to digitize different railway networks because a majority of them were built decades ago without precise measurements. Recent research has investigated model-based methods to detect rails for mainline and high-speed trains. However, these methods cannot be applied on various rail types flexibly, such as the embedded rails for trams; also, they perform poorly on the sparse regions in a 3D LiDAR scan where the density of points decreases as the distance from the scanner increases. Therefore, a flexible yet robust method to detect and extract a railway network is required. This thesis proposes a learning-based method without modeling to detect rail tracks on 3D scans by segmenting them on top-view projected images and predicting their center lines with deep neural networks. We also seek to refine the estimated center lines by fitting splines with RANSAC. We successfully show that this method is able to detect tracks with high accuracy. Qualitative results present the rail track segmentation and the center prediction, as well as the best approximation of a center line by a spline. Quantitative results show the metric for segmentation and the distribution of distances between the predicted centers and the ground-truth centers in both 3D space and the projected 2D images.