While common digital road network graphs are able to represent real-world street network topology relations quite adequately, they are highly generalized with regard to the composition of a road. Irrespective of their actual number of lanes, roads are shown as just one single line. As many intelligent transportation systems (ITS) applications require or provide lane-specific data and services, this is no longer sufficient from a short- to medium-term perspective. In particular, automated driving requires high-accuracy graphs both in topology and in geometry to localize positions not only on the correct road, but also in the correct lane. In the following paper, a cost-effective methodology for deriving such lane-level road network graphs will be described. The methodology is applied to standard GNSS trajectories collected for three different road types (urban, interurban, motorway) by vehicles participating in real-world traffic situations (Floating Car Data). The methodology extracts the number and position of lane centrelines from pre-processed GNSS trajectories using a kernel density estimation (KDE) and distance relations. Results show that the proposed method can, depending on the quality of the input data, reliably model lane centrelines for different road settings.