Location reference refining for semi-supervised location-labeling of crowdsourced wi-fi fingerprints

Abstract

Radio map construction, which is constructing signal patterns at each specific point in an area, is a prerequisite job for Wi-Fi Fingerprint-based Indoor Positioning System(IPS). However, radio map construction is known as a major drawback of Wi-Fi fingerprint-based IPS, because it requires manual calibration efforts to collect location labeled fingerprints. Several methods have been developed to reduce manual calibration efforts, but they require some amount of labeled fingerprints for initializing the learning process. Therefore, it still requires manual calibration efforts. Also, it can not achieve high accuracy, when the target space becomes large and complicated. In this research, we propose a radio map construction system using crowdsourced data that does not use manual collection data to reduce radio map construction cost. In order to obviate the need for manual calibration efforts, the system uses location-unlabeled fingerprints collected by numerous users' smartphones through implicit crowdsourcing. Only a small amount of fingerprints, which are Wi-Fi fingerprints labeled with the room where they have been collected, are key information for constructing an initial radio map. The system finds optimal placements of many location-unlabeled fingerprints by pattern recognition algorithm and improves quality of radio map. Experiments carried out in a large size indoor shopping mall revealed that the proposed method could successfully build a precise Wi-Fi radio map without manual calibration efforts to collect labeled samples.