This paper presents a novel image-based indoor localization approach that leverages semantic segmentation techniques by combining synthetic images generated from building information modelling (BIM) and deep learning (DL) methods. Unlike traditional image-based methods that compare overall visual features, our approach focuses on stable building components-such as floors, ceilings, walls, doors, and windows-that remain largely consistent over time. To enhance localization accuracy, we introduce a new similarity metric based on the intersection-over-union (IoU) ratio for each building component. This metric accounts for the spatial distribution of building elements, allowing for more robust comparisons between semantic masks derived from real-captured images and those generated from synthetic BIM models. The method reduces the dependence on detailed BIM models, enhancing reliability by focusing on the geometric and semantic consistency of key building elements. The feasibility of the approach is demonstrated through a case study with an institutional building, showcasing promising localization performance despite low-fidelity BIM models.
Urban infrastructures and buildings today face the intertwined pressures of extreme disasters, cascading effects, carbon constraints, and social uncertainty