Joint Magnetometer-IMU Calibration via Maximum A Posteriori Estimation

1. Problem Statement

The paper addresses the challenge of jointly calibrating magnetometers and Inertial Measurement Units (IMUs). While individual sensors (accelerometers, gyroscopes, magnetometers) suffer from intrinsic errors (biases, scale factors, non-orthogonality), mounting multiple sensors on a single platform introduces extrinsic errors (misalignment between sensor coordinate frames).

Existing in-situ calibration methods (which do not require specialized equipment like turntables) generally fall into two categories:

1. Dynamic System-Based: Model temporal evolution using gyroscope data.
2. Static Orientation-Based: Treat orientations at different time steps independently.

Limitations of State-of-the-Art (SOTA) Methods:

• Kok et al. [10]: Formulates the problem as a Maximum Likelihood (ML) estimation using an Extended Kalman Filter (EKF) to approximate the likelihood. While accurate, it relies on numerical derivatives because analytical derivatives are difficult to derive. This necessitates running the EKF multiple times per optimization step, resulting in high computational complexity.
• Wu et al. [12]: Formulates the problem as a Maximum A Posteriori (MAP) estimation but approximates the posterior using an EKF with an augmented state. This reduces computational load but sacrifices accuracy because the solution depends heavily on the quality of the EKF linearization points and Gaussian approximations.

The core problem is to find a method that balances high calibration accuracy with computational efficiency, avoiding the numerical derivative bottleneck of [10] and the linearization sensitivity of [12].

2. Methodology

The authors propose a new Joint MAP Estimation framework that treats both the calibration parameters ($\theta$) and the full orientation trajectory ($x_{0:T}$) as unknowns to be estimated simultaneously.

A. Mathematical Formulation

The method maximizes the joint posterior probability $p(\theta, x_{0:T} | y_{1:T}, \tilde{u}_{0:T-1})$. By taking the negative logarithm, this becomes a Nonlinear Least Squares (NLS) problem:
$$\min_{\theta, x_{0:T}} L(\theta, x_{0:T})$$
The cost function $L$ consists of:

1. Prior terms: For calibration parameters and initial orientation.
2. Measurement residuals:
   • Accelerometer: $\| \tilde{s}_k + R_k^T g_n - o_a \|^2$
   • Magnetometer: $\| \tilde{m}_k - D_m R_k^T m(\alpha) - o_m \|^2$
   • Gyroscope (State Transition): $\| \tilde{u}_k - \frac{1}{\Delta T}\text{Log}(R_k^T R_{k+1}) - o_\omega \|^2$

Where:

• $\theta = \{o_a, o_\omega, o_m, D_m, \alpha\}$ (biases, distortion matrix, dip angle).
• $R_k \in SO(3)$ represents the orientation.
• $D_m$ is the magnetometer distortion matrix (factorized into intrinsic distortion and extrinsic rotation).

B. Key Technical Innovations

1. Analytic Derivatives: Unlike [10], the proposed method derives closed-form analytic derivatives for the cost function. This allows the use of efficient solvers (Gauss-Newton or Levenberg-Marquardt) without running an EKF repeatedly for numerical differentiation.
2. Manifold Optimization: Since orientation matrices lie on the $SO(3)$ manifold, the optimization is performed in the tangent space using exponential and logarithmic maps to ensure geometric consistency.
3. IMU Preintegration: To handle cases where magnetometer sampling rates are lower than IMU rates, the method utilizes IMU preintegration (from [17]) to aggregate gyroscope measurements between magnetometer updates, maintaining computational efficiency.
4. Initialization Strategy: The method employs a robust initialization pipeline:
   • Gyroscope bias from stationary data.
   • Intrinsic magnetometer calibration (ellipsoid fitting) to initialize distortion and bias.
   • Extrinsic calibration to align frames.

C. Computational Complexity

The proposed method scales as $O(N_{iter}(3T + \text{dim}(\theta)))$, where $T$ is the number of time steps.

• Kok et al. [10]: Scales as $O(N_{iter} \times \text{dim}(\theta) \times T \times 3^3)$ due to numerical derivatives requiring multiple EKF runs.
• Wu et al. [12]: Scales as $O(T \times (\text{dim}(\theta) + 3)^3)$.
  The proposed method's complexity is linear in $T$ and additive in dimension, making it significantly faster than [10] for large datasets.

3. Key Contributions

1. New Joint Calibration Framework: A MAP-based formulation that jointly estimates the full orientation trajectory and calibration parameters, avoiding the approximations inherent in marginal posterior methods.
2. Efficient Optimization: Derivation of analytic derivatives enabling fast convergence using standard NLS solvers, eliminating the need for computationally expensive numerical differentiation.
3. Theoretical Analysis: A rigorous comparison of the proposed method against SOTA methods ([10], [12]) regarding underlying assumptions (Gaussian approximations vs. exact posterior structure) and computational complexity bounds.
4. Comprehensive Evaluation: Extensive validation using both simulated data and real-world experiments involving 30 magnetometer-IMU pairs.

4. Experimental Results

A. Simulation Study

• Accuracy: The proposed method achieved the lowest Root Mean Square Error (RMSE) for all calibration parameters (biases, distortion matrix, etc.). It improved calibration accuracy by 20–30% compared to the other methods.
• Efficiency:
  • It was significantly faster than Kok et al.'s method (which required 20 CPU cores for parallel numerical differentiation).
  • It was slightly slower than Wu et al.'s method but offered superior accuracy.
• Robustness: The method remained robust to variations in sampling frequencies and sampling rate ratios between IMU and magnetometer.

B. Real-World Experiments

• Setup: Calibrated 30 magnetometer-IMU pairs on a sensor array using data collected outdoors in Linköping, Sweden.
• Speed: The proposed method calibrated all 30 pairs in under two minutes on a standard laptop. This is one order of magnitude faster than the most accurate SOTA method (Kok et al.) implemented by the authors.
• Navigation Performance: When used in a Magnetic Field-Aided Inertial Navigation System (MAINS):
  • Uncalibrated data resulted in significant position drift.
  • The proposed method reduced drift to a level comparable to Kok et al.'s method.
  • Wu et al.'s method resulted in noticeably higher position errors, confirming the accuracy limitations of EKF-based approximations.

5. Significance and Conclusion

This paper presents a breakthrough in sensor calibration by successfully bridging the gap between high accuracy and high computational efficiency.

• Practical Impact: The ability to calibrate 30 sensor pairs in under two minutes on consumer hardware makes this method highly suitable for large-scale deployment, such as in sensor arrays for robotics, augmented reality, or dense indoor mapping.
• Theoretical Value: It demonstrates that formulating the problem as a joint MAP estimation with analytic derivatives yields a more robust solution than marginalizing the orientation (as in [12]) or relying on numerical derivatives (as in [10]).
• Reproducibility: The authors have made all datasets and algorithms publicly available, fostering further research in sensor fusion and calibration.

In summary, the proposed method offers a reliable, fast, and highly accurate alternative for joint magnetometer-IMU calibration, outperforming existing state-of-the-art techniques in both speed and precision.