Toward Low-Overhead Fingerprint-Based Indoor Localization via Transfer Learning: Design, Implementation, and Evaluation

This work aims at proposing a transfer learning (TL)-based framework to enhance system scalability of fingerprint-based indoor localization by reducing offline training overhead without jeopardizing the localization accuracy. The basic principle is to reshape data distributions in the target domain based on the transferred knowledge from the source domains, so that those data belonging to the same cluster will be logically closer to each other, whereas others will be further apart from each other. Specifically, the TL-based framework consists of two parts, metric learning and metric transfer, which are used to learn the distance metrics from source domains and identify the most suitable metric for the target domain, respectively. Furthermore, this work implements a prototype of the fingerprint-based indoor localization system with the proposed TL-based framework embedded. Finally, extensive real-world experiments are conducted to demonstrate the effectiveness and the generality of the TL-based framework.