A Magnetic Field-based Ubiquitous 3D Tracking System With Physics-informed Deep Supervision Neural Network

Abstract Precise and consistent 3D tracking is fundamental to natural user interaction and underpins a wide range of applications. While significant progress has been made across multiple sensing technologies, sustaining robust performance over time remains challenging due to accumulated drift, line-of-sight, and privacy concerns. This study introduces a near-range, infrastructure-light 3D magnetic tracking system that enables robust tracking using a compact tri-axis receiver coil, specifically designed to maximize quality factor and ensure geometric symmetry, along with spatially distributed transmitter coils. Central to our inference approach is a physics-informed deep supervision network, embedded with two key innovations: a physics-constrained output formulation and a deep multimodal supervision strategy, delivering state-of-the-art inference results with an average ATE of 2.35 cm and RTE of 3.38 cm across a wide range of hand- and head-mounted tracking tasks. The system also supports high-accuracy gesture recognition, with F1 scores of 1.000 and 0.979 for wrist and digit gestures, respectively.