A Hierarchical LSTM-Based Indoor Geomagnetic Localization Algorithm

The traditional wireless signals used for positioning, such as Wi-Fi and Bluetooth, are not stable enough for accurate indoor positioning due to the wireless signal multipath and time-varying effect. Compared with the wireless signals, the geomagnetic field signals inside buildings are influenced by ferromagnetic materials, which are significantly more stable for accurate indoor positioning. However, due to the location ambiguity problem, different positions may have similar geomagnetic readings, leading to significant positioning errors. Existing indoor geomagnetic positioning methods generally rely on single or short-sequence geomagnetic observations, which makes it difficult to discriminate between positions with similar geomagnetic values. Besides, geomagnetic anomalies can provide accurate position estimation within small-scale areas, but they cannot be utilized for large-scale localization due to the extensive existence of contour points and their rarity of remarkable geomagnetic anomalies. Therefore, we presents a geomagnetic indoor positioning algorithm based on two-level hierarchical LSTM (HLSTM) neural networks, which significantly increases the positioning accuracy by incorporating more historical geomagnetic observations. Based on the finding that geomagnetic signals are much stable, we adopt a sequence augmentation approach to generate a large number of geomagnetic trajectories for model training and testing instead of labor-extensive human collection. The HLSTM model is trained by the Pytorch framework with Cuda and cuDNN for parallel. Our proposed algorithm can obtain around 0.8m error of 67% probability on the self-collected ICT dataset and public MagPIE dataset. The results of the comparison experiment demonstrate that our proposed algorithm performs better accuracy and robustness when compared with other algorithms.