Long Short-Term Indoor Positioning System via Evolving Knowledge Transfer

Abstract

Traditional fingerprint-based positioning approaches work well on static data; they cannot handle scenarios where the data distribution, the feature space and even the signal source evolve over time, which are ubiquitous in real-world applications. One straightforward approach for circumventing these difficulties is to repeat labeled data calibration for maintaining an up-to-date fingerprint database, which is usually infeasible or expensive in large-scale indoor environments. In this paper, we propose a Long Short-Term indoor Positioning (LSTP) framework that enables adaptation at different time scales with low human-effort, and thus extends the effectiveness of existing fingerprinting techniques for a more generalized environment. Specifically, LSTP mainly considers the distribution discrepancy caused by continuous environmental dynamics in short-term positioning and the feature space heterogeneity in long-term positioning. To address the first challenge, we design an incremental ensemble localization model which leverages multiple source classifiers to resolve distribution differences in an online manner. To address the second challenge, we seek to borrow knowledge learned from an earlier time period with plenty of labeled samples for the current time period, thus reducing the required number of new calibration samples. By fully capturing the transferable spatial information across different time periods with multi-level constraints (sample, feature, and model levels), we can study a discriminative domain-invariant space from which we can make better predictions. The experiments on three real-world datasets demonstrate the superiority of the proposed framework, which outperforms the state-of-the-art systems by 18% in mean accuracy.