A Crowdsourcing-Based Wi-Fi Fingerprinting Mechanism Using Un-supervised Learning

摘要

In recent years, the Wi-Fi fingerprint-based indoor localization methods are widely applied to more and more ubiquitous applications. One of the key concerns is how to efficiently collect Wi-Fi fingerprint to reflect the harsh indoor environmental dynamics. However, continuous Wi-Fi fingerprinting confronts a contradiction: consumption in fingerprint collection and the real-time accuracy of fingerprint. We find that location fingerprint variations are related to crowd spatial distribution, and the distributions often varies periodically. Based on these observations, this paper proposes a crowdsourcing-based Wi-Fi fingerprinting mechanism using un-supervised learning, which exploit the historical data similar to the current fingerprint with particle filter method to enrich the data updating location fingerprint and generated updated location fingerprint with Gaussian process regression. Experimental results show that in our experimental environment, compared with the location fingerprints which are updated with only current data, the mean square error of the updated location fingerprints is reduced significantly.