MagShield: Towards Better Robustness in Sparse Inertial Motion Capture Under Magnetic Disturbances¶

Conference: ICCV 2025 arXiv: 2506.22907 Code: Coming soon Area: Human Understanding Keywords: motion capture, IMU, magnetic disturbance, sensor fusion, human pose estimation

TL;DR¶

MagShield is proposed as the first method addressing magnetic disturbance in sparse inertial motion capture systems. It adopts a two-stage detect-then-correct strategy: detecting magnetic disturbances via joint analysis of multiple IMUs, and correcting orientation errors using a human motion prior network. The approach can be plug-and-played into existing sparse IMU motion capture systems to enhance robustness.

Background & Motivation¶

IMU-based motion capture systems are widely adopted for their lightweight and low-cost nature, and sparse IMU configurations (6 sensors) further reduce the barrier to use. However, a long-overlooked critical issue is that IMUs rely on magnetometers to measure the Earth's magnetic field for global orientation estimation. In environments with magnetic disturbances (e.g., indoor spaces, near electronic devices), the magnetometer may confuse ambient interference with the geomagnetic field, leading to erroneous orientation estimates and severely limiting practical deployment.

Existing magnetic disturbance detection methods rely solely on readings from individual IMUs, ignoring the unique characteristic that multiple IMUs are worn on the human body in motion capture scenarios. The core idea of MagShield is to exploit this prior knowledge—that IMUs are body-worn—across both the detection and correction stages.

Method¶

Overall Architecture¶

MagShield functions as an IMU orientation estimation module. It takes raw local sensor measurements from 6 IMUs (accelerometer, gyroscope, magnetometer) as input, and outputs processed global IMU readings (global acceleration, angular velocity, and orientation), which are then forwarded to a downstream inertial pose estimator (e.g., PNP, DynaIP). The pipeline consists of two stages:

Pose-Aware Multi-IMU Fusion Algorithm: Derives IMU readings from raw measurements, with magnetic disturbance detection at its core.
Neural Yaw Error Corrector: Further refines leaf-node IMU readings using human motion priors.

Key Designs¶

Pose-Aware Magnetic Disturbance Detector: The key observation is that even when magnetic field strength occasionally approaches 1 at coincidental points within a noisy field, the field rarely remains spatially consistent across nearby sensors. A spatial consistency check is therefore proposed: for each IMU, its \(k=3\) nearest-neighbor IMUs are identified based on the previously estimated body pose, and the method verifies whether their normalized magnetic field strengths all approach 1. The detection criterion is \(flag_i = \text{True}\) if and only if \(|\|\mathbf{m}_S^j\| - 1| < \epsilon_m, \forall j \in N(i)\), where \(\epsilon_m=0.15\). Compared to single-IMU threshold methods, leveraging multi-IMU spatial information substantially improves detection reliability.
Error State Kalman Filter Fusion: Each IMU independently uses an ESKF to fuse raw measurements. The state space includes orientation, accelerometer bias, and gyroscope bias. The prediction step updates orientation via angular velocity integration: \(\mathbf{R}_{G,t+1} = \mathbf{R}_{G,t} \text{Exp}((\boldsymbol{\omega}_{S,t} - \boldsymbol{\omega}_{S,t}^{bias})\delta t)\). The correction step applies two virtual observation (VO) updates: a gravity VO (activated when accelerometer magnitude approaches 9.8) and a geomagnetic VO (activated when \(flag\) is True). The magnetic field is projected onto the plane perpendicular to gravity before use, ensuring errors manifest solely as yaw errors.
Neural Yaw Error Corrector (YawCorrector): Assuming the root-node IMU is correct, the module estimates only the yaw error angles of the 5 leaf-node IMUs relative to the root node. An LSTM network regresses the error angles \(\boldsymbol{\Delta} \in \mathbb{R}^5\), taking as input the leaf-node orientations and accelerations expressed in the root-node reference frame, along with the gravity vector. Training data is synthesized by randomly placing magnets in virtual rooms to simulate magnetic disturbances; the synthetic field is superimposed on the geomagnetic field, then fed through the ESKF to obtain erroneous IMU orientations, from which ground-truth error angles \(\Delta_i^{gt} = \theta_i - \theta_{root}\) are computed. At inference, a weighted correction strategy is employed, where the weight \(w\) is dynamically adjusted based on disturbance detection results (increased upon detection, decreased otherwise), applying correction via \(\tilde{\mathbf{R}}_G^i = \mathbf{R}_g(-w\Delta_i)\mathbf{R}_G^i\).

Loss & Training¶

YawCorrector training loss: \(\mathcal{L} = \|\boldsymbol{\Delta} - \boldsymbol{\Delta}^{gt}\|_2\)
Training data: synthetic data generated solely from the AMASS dataset
Network architecture: linear input layer → 2-layer LSTM (hidden dimension 256) → linear output layer, ReLU activation
Training configuration: 40% dropout, batch size 256, Adam optimizer
Sensor fusion achieves up to 1000 fps; YawCorrector achieves up to 400 fps; integrated with the most computationally expensive downstream system PNP, the pipeline still runs at 60 fps

Key Experimental Results¶

Main Results (MagIMU Dataset — Local Pose)¶

Method	SIP Err (°)	Ang Err (°)	Pos Err (cm)	Mesh Err (cm)
PNP + baseline	26.63	24.65	8.99	10.86
PNP + detector	25.67	23.20	8.56	10.26
PNP + ours	24.19	20.23	8.17	9.74
DynaIP + baseline	31.55	28.79	9.10	11.17
DynaIP + detector	30.79	27.26	8.64	10.67
DynaIP + ours	28.68	22.12	8.05	9.84

MagShield reduces error by approximately 10% across all metrics, with the most notable improvement in joint rotation error—up to 25% reduction (PNP: 24.65→20.23; DynaIP: 28.79→22.12).

Ablation Study (Comparison of IMU Synthesis Methods)¶

Method	SIP Err (°)	Ang Err (°)	Pos Err (cm)	Mesh Err (cm)
PNP + w/o syn mf	24.41	20.53	8.26	9.86
PNP + ours	24.19	20.23	8.17	9.74
DynaIP + w/o syn mf	28.78	22.51	8.12	9.91
DynaIP + ours	28.68	22.12	8.05	9.84

Physics-based magnetic field simulation for training data synthesis outperforms simple noise augmentation, as it more faithfully replicates the actual effect of magnetic disturbances on IMU measurements.

Key Findings¶

Yaw errors typically do not cause significant deviations in joint or vertex positions (e.g., arm twisting while standing results in negligible positional change), explaining why angular error improvements are most pronounced.
Although YawCorrector only corrects leaf nodes, it indirectly improves global translation estimation accuracy through improved local pose quality.
The method introduces no negative impact in clean magnetic environments (evaluated on the TotalCapture dataset), enabling its use as an always-on component.
The approach generalizes across two different inertial pose estimators (PNP and DynaIP), demonstrating strong compatibility.

Highlights & Insights¶

MagShield is the first work specifically targeting magnetic disturbance in sparse inertial motion capture, addressing a previously overlooked research gap.
The two-stage detect-then-correct design is modular and clear, enabling plug-and-play integration with various existing systems.
The multi-IMU spatial consistency approach for magnetic disturbance detection is concise and effective.
The magnetic disturbance data synthesis method is generalizable and addresses the scarcity of such training data.
The weighted correction strategy prevents over-intervention under normal magnetic conditions, balancing accuracy and robustness.

Limitations & Future Work¶

Magnetic disturbance detection relies solely on IMU positional information, without exploiting relative rotations or local magnetic field directional relationships.
Magnetometer magnetization (permanent bias induced by strong magnets) is not considered; the method may fail under such conditions.
The assumption that the root-node IMU is correct may not hold under extreme disturbances.
An end-to-end learned approach could be explored as an alternative to the two-stage pipeline.

The multi-sensor spatial consistency checking idea is generalizable to other multi-sensor scenarios.
The physics-simulation-based training data synthesis approach can inspire sensor noise modeling in other domains.
The weighted correction strategy (dynamically adjusting correction strength based on detector output) constitutes a general post-processing paradigm.

Rating¶

Novelty: ⭐⭐⭐⭐ First work to address magnetic disturbance in sparse IMU motion capture; multi-IMU joint detection is novel.
Experimental Thoroughness: ⭐⭐⭐⭐ Introduces the MagIMU dataset, validates across two downstream systems, provides complete ablation and clean-environment testing.
Writing Quality: ⭐⭐⭐⭐ Problem formulation is clear, method derivation is complete, and physical mechanisms are well explained.
Value: ⭐⭐⭐⭐ Addresses a real-world application pain point; plug-and-play design has high engineering value; real-time performance meets practical requirements.