Nowadays, there are many publicly available georeferenced data, like 3D CityGML models, that can be used as prior knowledge to perform accurate global localization. Iterative Closest Point (ICP) is a promising method for achieving this task, but it requires two point clouds that need to be partially overlapping in the initial state for better registration performance. Therefore, we investigated different detection and matching methods to automatically pre-register two non-overlapping point clouds based on a 2D overhead view and evaluated the registration results produced by an ICP algorithm. We used public data from the city of Aachen, Germany. A georeferenced point cloud was derived from the LOD2 CityGML model and a local point cloud was reconstructed from an image sequence using Structure from Motion (SFM). The evaluation results show that georeferenced LOD2 CityGML models can successfully be used for city-scale sub-meter global localization.