A method designed to reconstruct outdoor 3D building models automatically from a point cloud is presented in this paper. The proposed approach starts with building detection using spectral and spatial data from the UAV point cloud to remove non-building features. RANSAC, modified convex hull, and line growing algorithms are used to extract main roof planes and their boundaries. Roof planes are adjusted to each other using geometrical constraints, the height of each plane is estimated and a 3D model for the whole structure is constructed with LoD2. The key contribution of this approach is using a hybrid approach of model-driven with statistical analysis for modeling complex structures from a noisy point cloud. The reconstructed model shows that the workflow is sufficient to describe the whole building structure in the required LoD.