Neural Scene Reconstruction for As-Built Compliance Verification in Construction Site Inspection

Automated construction compliance verification of as-built building components requires empirical evidence sourced directly from the physical job site. While prevailing methodologies rely on Building Information Modeling, they fail to capture actual as-built conditions, leaving components, equipment locations, and spatial clearances unverified. To address this gap, this paper presents an inspection methodology that sequentially leverages recent advancements in Neural Radiance Fields, automated rule checking, and streamlined video data acquisition to achieve verifiable quality assurance and inspection. The framework begins with consumer-grade site video data acquisition as the foundational input. These frames are processed to estimate camera trajectories and reconstruct the visible environment into a three-dimensional neural scene using a object registration and localization. Within this digital inspection space, semantic construction objects are isolated and systematically evaluated against predefined rule templates encoding regulatory logic related to as-built verification of site components. Finally, the framework links each compliance checking result back to an immutable record of visual evidence. Preliminary validation using a scaled 3D-printed components and an HVAC assembly demonstrates the workflow's effectiveness in object-level inspection and documentation of object features. While bounded by its limited scope for a small scale office and 3D print verification, this study establishes that integrating scene-reconstructing Neural Radiance Fields with automated rule checking provides a transparent baseline for vision-based virtual construction compliance and site inspection.