Research on component-level segmentation and 3D reconstruction method of urban buildings driven by point cloud semantic cognition
Abstract: Component-level 3D reconstruction of buildings holds significant value in applications such as indoor navigation and facility maintenance. Point clouds, with their high-fidelity geometric information and rich spatial structural features, have become an important data source for 3D reconstruction. However, the unordered and unstructured nature of point clouds makes extracting precise component information and ensuring semantic-geometric consistency the core challenges in component-level segmentation and reconstruction. To address this, this paper proposes a semantic-geometry dual-driven reconstruction framework. First, based on semantic segmentation, a hierarchical component segmentation mechanism is constructed by integrating semantic information. The segmentation of complex components is effectively addressed through density clustering and triple geometric verification. Secondly, components are classified, and component-level reconstruction results are obtained via multi-strategy component reconstruction. Finally, a knowledge graph is introduced to parse spatial relationships and constrain component assembly parameters, thus achieving geometric and semantic consistency. Experimental results show that compared with baseline methods such as PointNet++ and PolyGNN, the proposed method significantly improves the accuracy of component-level reconstruction (mIoU increased by 16.5 % and assembly correctness rate reaching 96.2 %), providing more detailed 3D model support for applications such as smart cities and BIM modeling.
Keywords: Point cloud; Point cloud 3D reconstruction; Point cloud semantic segmentation; Component-level segmentation; Real-world 3D modeling
