Indoor Positioning Based on Hybrid Domain Transfer Learning

Abstract

In multiple sources jointly indoor positioning context, in the offline phase, it is easier to obtain enough sample features from multiple sources to form multiple source domains, but we have difficulties in obtaining enough sample features in real positioning phase due to sensors fault or changing environments, thus leading to the missing data in a target domain. How to efficiently transfer the knowledge from the source domains to the target domain is the key issue for high accurate sensing. To address the above challenge, we propose a hybrid domains transfer learning (HDTL) (i.e., transfer knowledge between source domains and the single target domain) for high accurate target positioning. Specifically, HDTL first divide the multiple source domains into homogeneous and heterogeneous feature according to the single target domain; then we can calculate the mapping from the homogeneous feature to the heterogeneous feature in the source domains by using shared homogeneous knowledge of the source domain and target domain as a bridge. Next, we can complement heterogeneous data in the target domain by using this mapping in order to guarantee that the features in the source domains and target domain exist in a common homogeneous space. Finally, the location of the target can be calculated in this new feature space. HDTL can localize targets accurately in presence of sensors faults by resorting to the transfer knowledge from the source domains, and is thus superior to some existing transfer learning algorithms. The experimental results show that HDTL can acquire better accuracy in multiple source jointly positioning environments as compared with some existing stand-alone positioning environments.