Construction Quality Inspection by Local Point Cloud to BIM Registration and Component Association Based on Vision Foundation Model

Abstract

Construction quality inspection is essential for verifying conformity between physical installations and design specifications. Traditional inspection methods rely heavily on manual comparison between construction drawings and onsite conditions, making the process labor-intensive, subjective, and lacking in digital traceability. Although laser scanning technologies have been widely adopted for capturing as-built data, the automated registration of local LiDAR scans to BIM models and the association of onsite components with their designed counterparts remain underdeveloped. To address these challenges, this study proposes a novel framework for local point cloud to BIM registration and component association, leveraging vision foundation models to enable component-level inspection directly on construction sites. The proposed framework comprises four main stages. First, raw LiDAR scans are preprocessed and converted into binary visibility maps centered on the LiDAR sensor. Second, synthetic visibility maps are generated by virtually scanning the BIM model from corresponding viewpoints. Third, both real and synthetic visibility maps are encoded using a vision foundation model to extract high-dimensional features. These features are compared to generate similarity maps for coarse localization, which is further refined using the Iterative Closest Point (ICP) algorithm to achieve accurate registration. Finally, onsite components are identified using the vision foundation model and matched to their corresponding BIM components through bipartite graph matching, solved via the Hungarian algorithm. Experiments conducted on real-world construction sites demonstrate the robustness and accuracy of the proposed method, validating its effectiveness for practical applications in construction quality inspection and digital twin development