This paper presents a multimodal 3D data fusion workflow tailored for registering mobile laser scanning (MLS) point clouds to CityGML semantic building models (MLS-to-CityGML) in urban street environments. The proposed method first extracts stable footprint features from MLS fragments using a multi-band consensus filtering strategy, which suppresses noise and non-building elements in MLS observations. In parallel, 2D target features are derived from CityGML Level of Detail 2 (LoD2) building geometry by extracting and cleaning ground-contact footprint segments. The resulting representations are aligned in 2D using raster-based distance-transform matching, and the estimated pose is subsequently transferred to 3D through vertical alignment and refined using the plane-voxel generalized iterative closest point (PV-GICP) algorithm. Finally, MLS drift analysis is conducted based on adaptive fragmentation. Evaluation across five urban scenarios shows that the proposed multi-band consensus filter provides reliable coarse initialization. The mean horizontal residual of all evaluated scenes achieves 0.041m after coarse registration, while subsequent PV-GICP refinement further reduces the overall mean residual from 0.065m to 0.015 m. The experimental results demonstrate the workflow as a robust coarse-to-fine registration strategy for structured urban environments, thereby providing a reliable georeferencing basis for updating city-scale semantic models.