Reduced-effort generation of indoor radio maps using crowdsourcing and manifold alignment

Abstract

Constructing radio maps for indoor environments using either measurements or signal propagation models suffers from extensive work loads and/or significant radio map inaccuracies. In this paper, we propose a Wi-Fi radio map generation and update solution in indoor environments using a propagation modeling simulator, a limited number of labeled calibration fingerprints and many crowdsourced unlabeled measurements, using manifold alignment. This semi-supervised, dimensionality reduction transfer learning method estimates the location of unlabeled crowdsourced radio signal observations by aligning them with the calibration fingerprints and the simulated radio map in a low-dimensional space. It thus reduces the calibration cost and can be easily deployed in any indoor environment given its floor plan and few calibration fingerprints. In addition to simple deployment, our solution can easily re-construct new radio maps in cases of changes in time, device or floor plans, which significantly reduces the large re-calibration load. Testing results show that the proposed solution can achieve as low as 6.4 to 6 dBm root mean square error in radio signal estimation with 15% to 30% of the full fingerprinting load.