A Floor Division Method for Non-Manhattan Buildings Integrating Multi-Feature Clustering and Bidirectional Cloth Simulation

Objectives: Floor division is a core step in indoor 3D reconstruction, cross-floor inspection of unmanned platforms, and smart building management. Most existing methods rely on height distribution statistics and are designed for Manhattan-world structures. They often fail in non-Manhattan buildings with curved floors, inclined slabs, staggered stories, or large atriums, leading to missing floors, misclassification, and discontinuous boundaries. This study aims to propose a robust and accurate floor division method dedicated to complex non-Manhattan building point clouds. Methods: A floor division method for non-Manhattan buildings fusing multi-feature clustering and bidirectional cloth simulation is proposed. Firstly, fine preprocessing of point clouds is completed by voxel centroid downsampling, dual-dimensional pose correction and redundant data elimination. Then, for the preprocessed point clouds, weighted principal component analysis is adopted to conduct multi-feature clustering and segmentation, and the attribute distinction between ground and ceiling is realized combined with the coupling judgment of three factors including support, hierarchy and topology. Finally, the accurate extraction of irregular floor surfaces is achieved through the progressive process of floor interval division and bidirectional cloth simulation. Results: Experiments are conducted on eight typical complex scenes from the Matterport3D dataset. The results show that compared with the box-plot method and the bidirectional hierarchical separation method, the comprehensive F1 of the proposed method is improved by 38.39% and 33.02%, respectively. The average recall reaches 99.10%, the average precision reaches 99.85%, and the mean comprehensive F1 is 99.48%. The method remains stable and reliable even in extremely irregular structures such as staggered floors, curved surfaces, domes, and inclined slabs, with no obvious missing division, misclassification, or boundary fractures, demonstrating significantly stronger robustness than traditional methods. Conclusions: The proposed method integrating multi-feature clustering and bidirectional cloth simulation effectively overcomes the dependence of traditional algorithms on regular structures and uniform floor heights, and significantly improves the completeness, accuracy, and stability of floor division in non-Manhattan building scenarios. It can accurately handle complex curved surfaces, inclined floors, and strongly staggered structures, providing reliable technical support for practical applications including indoor 3D reconstruction, BIM modeling, cross-floor navigation of unmanned systems, and smart building operation and maintenance. It also has important theoretical value and application prospects for promoting automatic high-precision reconstruction in complex building environments.