A topological reconstruction method for building point clouds integrating deep semantic segmentation and adjacency constraints

To address challenges in urban building digitization, such as data redundancy, structural ambiguity, and boundary inaccuracy, this paper proposes a topological reconstruction method integrating deep semantic segmentation and adjacency constraints. SoftGroup is used for instance-level segmentation, achieving high-precision component identification via joint network optimization. Based on segmentation results, an adjacency graph is constructed to model spatial topological relationships. To reduce boundary misconnections and geometric noise, an adjacency-guided boundary regularization strategy is introduced that combines face affiliation and spatial proximity within a global topological graph. Experiments on STPLS3D and SensatUrban show high shape fitting accuracy (SFA = 0.0046) and strong structural consistency (SCS > 0.9) for typical structures. The method also achieves robust recovery for complex elements such as conical roofs. Ablation studies confirm the effectiveness of boundary regularization in enhancing consistency. This efficient and accurate approach supports large-scale urban reconstruction and AR applications in real 3D environments.