IndoorGML-Constrained Navigation Across Multiple 3D Gaussian Splatting Scenes for Indoor Environments

Abstract

3D Gaussian Splatting (3DGS) has emerged as a compelling method for scene reconstruction by producing compact, photorealistic representations of physical environments in real-time. Its application to indoor spaces, however, introduces a fundamental challenge: reconstructing a multi-room building as a single coherent scene is undermined by the frequent occlusions that walls introduce, degrading the sparse reconstruction on which 3DGS training depends. Per-room reconstruction sidesteps this problem but produces a set of independent scenes in arbitrary coordinate frames that have no shared spatial reference.

This paper presents a pipeline that addresses these challenges by integrating IndoorGML's topological model as the organisational backbone for multiple independently reconstructed 3DGS scenes. The pipeline proceeds in four stages. First, per-room 3DGS scenes are reconstructed using Structure from Motion (SfM) and Gaussian training from video-captured image sets. Second, an Umeyama similarity transformation is estimated for each room between its (SfM) coordinate frame and a shared architectural floor plan reference, simultaneously resolving scale ambiguity and establishing a common metric coordinate system across all scenes. Third, the IndoorGML model is supplemented with additional attributes, and ingested into a PostGIS database that preserves the Cell–Node-Edge topological structure of the original schema. Fourth, a web-based viewer queries this database at runtime to enforce geometrically derived navigation constraints within each room, and to trigger topologically-guided transitions between rooms as the user crosses door boundaries.

The pipeline is demonstrated on three adjacent rooms at the Faculty of Architecture and the Built Environment, of the Delft University of Technology. The site comprises two lecture halls and a corridor space. The prototype achieves sustained real-time rendering performance above 60 frames per second on consumer-grade hardware, with navigation constraints successfully preventing the camera from reaching scene regions outside the training range. The results confirm that IndoorGML's topological model, when extended to reference 3DGS scene data and spatial transforms, provides a strong framework for multi-scene indoor navigation.