Door-guided indoor space segmentation via progressive geometric analysis

Point cloud segmentation is crucial for Forensic Information Modeling (FIM) and Building Information Modeling (BIM) applications; however, existing methods either require extensive training data (deep learning) or struggle in complex architectural layouts (geometry-based). This paper presents a door-guided geometric framework that achieves robust indoor space segmentation without learned features. The approach introduces four preprocessing modules: 1) door gap detection through cross-sectional analysis (requiring only 4 manual clicks to identify all doors in a building), 2) corridor isolation via principal component analysis, 3) tile-based structural filtering, and 4) verticality-based wall extraction. These modules establish spatial boundaries before applying hierarchical watershed segmentation with multi-scale spillage prevention. Validated on the S3DIS Area 6 dataset (27 rooms, 1.17 million points), the framework achieved an average IoU of 96.5% and an F1-score of 98.0%, matching deep learning performance while eliminating training requirements. The purely geometric approach enables deployment in forensic engineering contexts where training data is unavailable and computational resources are limited, directly supporting damage assessment and structural investigation workflows.