AIRLoc: An IoU-Guided Adaptive Fusion Method for Floorplan-Based Visual Localization

Abstract

Floorplan-based indoor visual localization is essential for autonomous navigation and augmented reality using only RGB sensors. Existing methods often suffer from two limitations: they rely on local ray sampling that overlooks aggregated region-level consistency, and they lack adaptive mechanisms to fuse multimodal features under varying scene complexity. In this letter, we propose AIRLoc, an adaptive framework that addresses these issues through three key components. First, a piecewise planar inverse-depth smoothing method regularizes ray-wall geometry, suppressing noise while preserving structural boundaries. Second, a field-of-view region-overlap Intersection-over-Union (IoU) matching mechanism treats the ray sequence within the camera’s field of view as a closed region, capturing aggregated spatial overlap that point-wise matching misses. Third, a newly trained image-map-conditioned adaptive weighting network dynamically predicts fusion weights for depth, semantic, and IoU modalities. Experiments on Structured3D and ZInD benchmarks demonstrate that our method achieves superior recall, validating its robustness in challenging indoor environments.