On Deep Learning-Based Indoor Positioning and Uncertainty Estimation

With the advance of deep learning, several indoor fingerprint-based localization models have been proposed. While being able to learn the relationships between fingerprints and locations well, such models may suffer from environment change and data aging problems. To avoid a bad user experience, an uncertainty value can be provided to indicate the reliability of a localization estimate. We thus propose a multi-branch neural network that can conduct magnetic-based indoor localization and uncertainty estimation simultaneously. The main idea is to duplicate the main localization branch multiple times with different depths. A loss function is proposed to balance localization accuracy and uncertainty estimation. Through extensive experiments, we show that the proposed method outperforms Monte Carlo dropout approaches in AUCO by 72.9% and precision-recall AUC by more than 100%. Besides, the model uses much less parameters than the deep ensemble approach due to our shared-weight multi-branch design.