Self-Calibrating Indoor Localization with Crowdsourcing Fingerprints and Transfer Learning
Precise indoor localization is one of the key requirements for fifth Generation (5G) and beyond, concerning various wireless communication systems, whose applications span different vertical sectors. Although many highly accurate methods based on signal fingerprints have been lately proposed for localization, their vast majority faces the problem of degrading performance when deployed in indoor systems, where the propagation environment changes rapidly. In order to address this issue, the crowdsourcing approach has been adopted, according to which the fingerprints are frequently updated in the respective database via user reporting. However, the late crowdsourcing techniques require precise indoor floor plans and fail to provide satisfactory accuracy. In this paper, we propose a low-complexity self-calibrating indoor crowdsourcing localization system that combines historical with frequently updated fingerprints for high precision user positioning. We present a multi-kernel transfer learning approach which exploits the inner relationship between the original and updated channel measurements. Our indoor laboratory experimental results with the proposed approach and using Nexus 5 smartphones at 2.4GHz with 20MHz bandwidth have shown the feasibility of about one meter level accuracy with a reasonable fingerprint update overhead.