Accurate Indoor Localization Using Magnetic Sequence Fingerprints with Deep Learning

Abstract

AbstractMagnetic field fingerprinting has been an interesting topic in indoor localization researches because of its advantages of being ubiquitous, energy-efficient and infrastructure-free. Most existing indoor magnetic field-based positioning methods use the raw three-dimensional magnetic field strength obtained by the magnetic sensor built in smartphones. However, they have to overcome the problem of ambiguity that originates from the nature of geomagnetic data, especially in the large-scale environment. In this paper, we first expand the dimension of magnetic data elements, and a sliding window mechanism is designed to construct magnetic sequence fingerprints to increase the distinguishability of magnetic field fingerprints. Moreover, an accurate indoor positioning model combining the advantages of one-dimensional convolutional neural network and long short-term memory network is designed to automatically learn the mapping between ground-truth positions and magnetic sequence fingerprints. To demonstrate the effectiveness of our proposed method, we perform a comprehensive experimental evaluation on three real-world datasets, and the results show that the proposed approach can remarkably improve positioning performance compared with other methods.KeywordsIndoor localizationMagnetic fieldMagnetic sequence fingerprintsSmartphoneDeep learning