Magnetic positioning Dataset

Traditional magnetic localization methods based on mathematical model and optimization algorithm often fail to achieve the global optimum due to their dependency on initial values of pose parameters. Despite deep learning can potentially address these limitations, the existing methods are restricted to capture both global position distribution and local spatial features, leading to diminished performance in tasks required precise pose information. To address these, we introduce PRPosNet, a novel CNN-Transformer hybrid network incorporated with a light-weight coordinate attention module. The proposed network utilizes CNN to extract positional features from magnetic field data, while leveraging the power of Transformers to handle the intricate long-range dependencies among these features. Additionally, the coordinate attention module is integrated toaugment the positional and orientational representation of the tri-axis magnetic induction intensity. To further enhance the stability, we have implemented high-pass and notch filtering to effectively suppressed noise within magnetic field data. Experimental results demonstrate that PRPosNet outperforms other state-of-the-art deep learning techniques in terms of positioning accuracy, with an average error of 0.91±0.69 mm in position and 0.38±0.38° in orientation. Moreover, we conduct an external validation and verify the robustness of the PRPosNet.