Adaptive 3DGS-SLAM-driven incremental online geometric digital twinning complex indoor built environments

Digital twinning indoor built environments creates virtual replicas that enable continuous monitoring and data-driven management throughout asset lifecycles, supporting automated inspection and real-time operational decision-making. Existing geometric modeling methods struggle with modeling quality and efficiency in real-time 3D reconstruction and incremental updating. This study developed an adaptive 3D Gaussian Splatting-Simultaneous Localization and Mapping (3DGS-SLAM) approach for incremental online geometric digital twinning in complex indoor built environments that addresses these challenges through a multi-stage process: utilizing visual alignment to track camera poses against the existing 3DGS model, identifying significant differences between rendered and newly captured images for keyframe selection, evaluating the contribution of each Gaussian primitive to these differences through parallel programming, and adaptively modifying only those critical Gaussians representing scene changes locally. The adaptive 3DGS-SLAM achieves online 3D reconstruction and incremental updating using streaming video data with advanced processing speed and visual fidelity, suppressing catastrophic forgetting and visual tearing problems. Computer experiment validation on the ReplicaCAD dataset demonstrates the effectiveness of the proposed method under controlled conditions, achieving a 57% reduction in average processing time and 15% improvement in rendering quality compared to the baseline MonoGS framework. The case study on real scenes further confirms that the proposed approach achieves higher rendering quality while reducing processing time by 21.5%, substantially reducing visual voids and tearing during incremental updates with localized model reconstruction only.