PRPosNet: A CNN–Transformer Model With Coordinate Attention for Precision and Robust Magnetic Pose Estimation

Based on mathematical model and optimization algorithms, traditional magnetic localization methods often fail to achieve the global optimum due to their reliance on the initial values of pose parameters. Despite the potential of deep learning in addressing this issue, the existing methods are unsuitable for capturing both global position distribution and local spatial features, which compromises their performance when precise pose information is required. Therefore, a novel convolutional neural network (CNN)–Transformer model named PRPosNet, is proposed in this article. It consists of a pose-related feature extraction (PRFE) module, a spatial feature augmentation module (SFAM), and a sensor interaction dependency modeling (SIDM) module. After being filtered through high-pass and notch filters, the magnetic field intensity data is processed through a CNN-based architecture to capture localized position and orientation details. Subsequently, the pose representation of the tri-axis magnetic induction intensity is improved through the integrated coordinate attention in the SFAM. A dedicated Transformer-based module is used to integrate global coupling and long-range dependency among the pose-related features. The experimental results demonstrate that PRPosNet outperforms the existing deep learning techniques in positioning accuracy, with an average error of $0.91~\pm ~0.69$ mm in position and $0.38^{\circ }~\pm ~0.38^{\circ }$ in orientation. Furthermore, external validation verifies the robustness of PRPosNet.