Abstract:  The disturbance signals generated by ferromagnetic objects in a magnetic field provide a non-contact and passive means for target localization and tracking. However, traditional magnetic vector localization methods based on physical modeling suffer from degraded accuracy and limited robustness in complex environments. In this paper, we propose a magnetic anomaly tracking convolutional neural network (MagTrack-CNN) for high-precision spatial dynamic localization of ferromagnetic objects. The proposed method employs a dual-layer magnetic vector observation architecture to enhance localization sensitivity along the depth direction by incorporating vertical gradient information. A dual-branch independent prediction network is designed to mitigate gradient competition among spatial components, while a multi-scale feature extraction framework is constructed to effectively capture magnetic anomaly patterns at different spatial scales. Experimental results in stable magnetic field regions for localization scenarios involving a single uniformly magnetized target demonstrate that MagTrack-CNN achieves an overall positioning accuracy of 1.98 cm on the test set, and exhibits superior accuracy in z-axis localization. In dynamic localization tests across different trajectory types, the method consistently produces correct dynamic positioning. Meanwhile, noise sensitivity experiments show that under measurement noise levels of 0.5%-3.0%, the model maintains low errors (the mean absolute error  increases from 0.42 cm to 0.46 cm). Furthermore, the per-sample inference time remains within 2.3 ms, highlighting its promising potential for practical applications.

Key words:  , magnetic target localization, convolutional neural network, multi-scale feature, trajectory tracking