20 February 2026 in SSRN Electronic Journal
Scaffolding safety inspection in construction sites increasingly relies on 3D point cloud perception, yet automated checking remains limited because building code clauses target structural configurations rather than isolated components. This work addresses the grounding challenge of mapping clause level abstract construction concepts to concrete, verifiable structural instances in real world point clouds. We propose a hierarchical grounding pipeline that progressively elevates semantics from component level instance and semantic parsing to orientation aware member recognition, and finally to pose and spatial context aware capsule graph reasoning for structural functional concept inference. On real world scaffolding scans, our approach attains 0.9148 OA and 0.8713 Macro-F1 on concept reasoning and achieves 78.16% end-to-end correct grounding on Level-3 instances. Compared with direct grounding, it improves Instance F1@0.50 by +7.5%. The grounded concept instances provide explicit structural evidence to localize inspection targets, enabling more reliable and scalable clause-oriented safety checking in complex site layouts.