Rail extraction is one of the basic tasks for video-based driver support in railways, being used, for example, to determine regions of interest for detecting obstacles and signs. In this paper, we introduce an approach that performs rail extraction by matching edge features to candidate rail patterns modeled as sequences of parabola segments. Patterns are precomputed for areas near the camera and generated on the fly for more distant regions. The proposed approach addresses the challenges posed by the open and complex environment, such as bad visibility, shadows, motion blur, clutter, occlusion, and confusing structures. Moreover, it does not require explicit knowledge about train speed or camera position, running fast enough for practical use without specialized hardware. The feasibility of the proposed approach was verified in experiments using videos captured under real operation conditions.