Regional Perceptrons Ensemble for Permanent Magnetic Localization

Permanent magnetic localization (PML) offers a low-cost and convenient approach to wireless motion tracking. Recent neural network methods have successfully improved the practicality of PML in real-world applications. However, the neural networks are prone to overfitting because ambient noise often influences the magnetic measurements, which are highly sensitive to the magnetic tracer’s location. Methods: To address this problem, we propose a spatial–angular ensemble (SAE) of regional perceptrons for PML. First, the SAE constructs sub-region localization (SRL) modules to estimate the tracer’s state in various spatial and angular regions. Second, it equips the spatial-aware (SA) and angular-aware (AA) modules with the ability to select the estimates from the SRL modules. Lastly, we design an attentional feature fusion (AFF) module to dynamically integrate these localization results. The SRL, SA, and AA modules are implemented by stacking the gated multilayer perceptrons (gMLPs), which can adaptively suppress the noisy features. Furthermore, we present a linear interpolation strategy to impose local smooth regularization during neural network training, thereby enhancing the model’s generalization performance. Results: Based on the assessments of our experiment system, the proposed SAE with linear interpolation (SAE-LIN) significantly outperforms the current approaches in terms of magnetic location accuracy, and its computing cost satisfies the requirements for real-time processing. Conclusion: This article provides a feasible neural network method to reduce the impact of noises and regional differences in magnetic measurements, potentially improving the accuracy of PML.