A Semantic-Grid Structural Completion Method for Indoor Space Segmentation from 3D Point Clouds

Indoor space segmentation is essential for indoor navigation, 3D reconstruction, and Building Information Modeling (BIM). However, reliable segmentation from unstructured 3D point clouds remains challenging due to structural voids caused by occlusion and noise, as well as the difficulty of distinguishing permanent structural elements from dense non-structural clutter. To address these issues, this paper proposes a semantic-grid structural completion method for indoor space segmentation from 3D point clouds. The method first integrates RandLA-Net-based semantic segmentation with geometric similarity correction to improve structural consistency. Subsequently, a semantic-grid structural completion algorithm detects and fills structural voids under height constraints; this process employs dual-grid structural marking with a 2D semantic occupancy grid and a 3D voxel grid to identify missing observations and generates synthetic points with inherited semantic labels to restore structural integrity within the scene. A density-aware height difference filtering method is then applied to remove non-structural clutter and clearly separate structural elements from the rest of the scene. Finally, indoor spaces are delineated through connectivity-based segmentation and inverse distance-weighted label propagation. Experiments on public datasets, including S3DIS, UZH and Structured3D, demonstrate that the proposed method consistently outperforms existing approaches, achieving a mean F1 Score of 0.99, an Intersection over Union (IoU) of 0.98, and a Segmentation Error Rate (SER) of 0 in most scenarios, particularly in occlusion-affected and structurally complex indoor environments.