Automatic generation of 2D floor plans is a challenging task with widespread applications in construction management, such as monitoring and energy simulation. A common method for floor plan generation is building a 3D point cloud model of the environment and performing geometric operations to slice a horizontal layer for obtaining wall geometry. Therefore, point cloud consistency and smoothness have a significant impact on floor plan accuracy. Recent advances on mobile device depth sensing technology allow users to build point cloud models compactly, but factors such as depth noise and motion blur affect model quality negatively and thus the floor plan. This paper presents a dynamic layer extraction approach for floor plan generation. Horizontal and vertical planes in different scales are detected and ceiling of the environment is obtained utilizing sensor position and plane sizes. Then horizontal layers of the vertical wall points are sliced dynamically. A state-of-the-art geometric floor plan generation system is extended with the proposed approach and assessed on a dataset of indoor point cloud models retrieved by a mobile device. The quantitative results demonstrate that the proposed method improves the accuracy in terms of wall detection precision and recall by around 5 % and 33 %, respectively. It produces promising results on unstructured point clouds with missing parts and high levels of noise.