The demand for three-dimensional (3D) parametric models, such as Building Information Modeling (BIM), has significantly increased due to the numerous applications in smart cities and digital twin technologies. BIM is also a valuable tool for efficiently managing reconstructed structures during the operation and maintenance phases. However, the current process of converting scanned point cloud data (PCD) into BIM is mainly manual design or semi-automatic development design through Revit, encountering significant challenges in accurately representing complex interior layouts and extracting details from point clouds. This results in inefficient and less accurate BIM model generation. To address these challenges, A novel method from scanning PCD to BIM has been proposed, utilizing deep learning-based point cloud segmentation neural network and unsupervised clustering algorithms to obtain different components. This process can automatically generate structured 3D models based on the geometric structures of the original point cloud data. Initially, an enhanced deep learning-based point cloud segmentation neural network is employed. Subsequently, an efficient workflow for reconstructing 3D building models of structured interior scenes is developed, capable of handling large-scale data, including multiple room layouts in Manhattan, and enables the automatic reconstruction of 3D models. It ensures higher fidelity in representing complex interior layouts and capturing detailed architectural elements, benefiting a variety of applications in smart city and digital twin implementations.