A Novel Indoor Fingerprint Localization System Based on Distance Metric Learning and AP Selection

A desirable fingerprint-based indoor localization (FIL) system aims to achieve an accurate positioning result within an acceptable time consumption, which is still challenging for application. Building a practical FIL system is a composite task of feature extraction and location estimation, resulting in related methods that is often hard to consider both the positioning accuracy and time consumption. This article proposes a novel FIL system that uses a combination of distance metric learning (DML) and access point (AP) selection method to tradeoff the positioning accuracy and time consumption. Specially, we first abstract the localization process to develop a mathematical model from the perspective of probability theory and reveal the significant impact of the received signal strength (RSS) similarity comparison on FIL. Then, we propose a perturbation theory-based AP selection method to select the best-position-discrimination AP subset from all to reduce the positioning time consumption. Meanwhile, we propose a DML-based method to extract the RSS distribution which involves the indoor environmental information, and further use it in RSS fingerprint similarity comparison to improve the positioning accuracy. We introduce the signal path-loss model into the proposed method for training to get the best similarity metric function. Finally, experimental results show that both the positioning accuracy and the time consumption are comparatively improved in the online phase by the proposed FIL system.