A New Statistical Method for Indoor Localization Using Unlabeled Crowdsourced Data

Most data-driven localization approaches rely on fingerprinting-based techniques, which require labor-intensive site surveys and periodic calibration to maintain accuracy. These efforts limit the scalability of fingerprinting-based methods. This paper presents a calibration-free approach for indoor localization using crowdsourced data without requiring location labels. The statistical information of the crowdsourced data is utilized to learn the signal propagation characteristics. We apply a cumulative distribution function (CDF) conversion to map signal strength measurements to distances from access points. This CDF conversion overcomes the limitations of the conventional logdistance path loss (LDPL) model and efficiently captures the effect of shadow fading and multipath. Based on the estimated distances, the target locations are determined using an improved trilateration algorithm. Our approach is fully unsupervised, requiring no location labels, and significantly reduces the need for site surveys. Experimental results demonstrate substantial improvements in localization performance compared to LDPL-based methods. Moreover, our localization accuracy using unlabeled data approaches that of the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k$</tex>-Nearest-Neighbor algorithm, which relies on labeled fingerprints.