Attitude-Aided Linear Calibration of Triaxial Accelerometers

Technical Summary: Attitude-Aided Linear Calibration of Triaxial Accelerometers (ALAC)

Problem Statement
Triaxial MEMS accelerometers are critical for inertial sensing and navigation but suffer from deterministic errors (offset, scale factor, non-orthogonality, and misalignment) and stochastic noise. Existing calibration methods face a trade-off:

• Reference-based methods (e.g., using high-precision turntables) rely on strict sensor alignment with the reference frame and often require complex iterative optimization or specific fixture setups, limiting their applicability to low-cost or in-field scenarios.
• Self-calibration methods (relying solely on accelerometer data) avoid external equipment but typically require nonlinear iterative optimization (e.g., Levenberg-Marquardt, genetic algorithms). These approaches are prone to initialization sensitivity, convergence issues, and local minima, making them computationally expensive and less robust for online applications.

There is a need for a calibration method that is computationally efficient (non-iterative), robust to platform misalignment, and capable of operating on any system providing attitude information (e.g., robotic arms, IMUs, turntables) without requiring strict sensor-to-platform alignment.

Methodology: Attitude-Aided Linear Accelerometer Calibration (ALAC)
The authors propose ALAC, a closed-form, linear calibration algorithm that utilizes external attitude information to estimate sensor errors. The core methodology involves:

1. Unified Calibration Model: The authors introduce a Combined Error Matrix (CEM), denoted as $C$, which unifies the sensor's scale factors ($S$), non-orthogonality ($T$), and alignment rotation ($^e_sR$) into a single $3 \times 3$ matrix. The model relates raw accelerometer measurements ($acc$) to the external reference frame acceleration ($^eA$) via:
   $$^eA = C \cdot acc - ^eA_0$$
   where $^eA_0$ is the bias expressed in the external frame.

2. Attitude-Aided Constraint: Under static conditions, the true acceleration is zero, and the sensor measures only gravity. By incorporating the known orientation of the platform (e.g., from a robot's forward kinematics or an IMU's attitude estimation), the gravity vector in the external frame is constrained. This leads to a homogeneous equation:
   $$C \cdot acc - ^eA_0 - ^e_bR \cdot ^bA_g = 0$$
   where $^e_bR$ is the known rotation from the base to the sensor, and $^bA_g$ is the gravity vector in the base frame (with a known norm $g$).

3. Constrained Homogeneous Least Squares (CHLS): The unknown parameters (elements of $C$, bias $^eA_0$, and gravity vector $^bA_g$) are stacked into a single vector $x$. The calibration is formulated as a CHLS problem:
   $$\min_x \| \bar{A}x \|_2^2 \quad \text{s.t.} \quad \| Dx \|_2 = 1$$
   This problem is solved in closed-form using standard linear algebra techniques, specifically Singular Value Decomposition (SVD) or Generalized Eigenvalue Problems (GEVP). No iterative optimization is required.

4. Parameter Extraction: Once the CEM $C$ is estimated, the individual calibration parameters (scale factors, non-orthogonality, and rotation matrix) are analytically extracted using a Gram-Schmidt-like decomposition process. The sign ambiguity inherent in homogeneous solutions is resolved by ensuring the resulting rotation matrix has a determinant of $+1$.

5. Recursive Extension: The linear formulation allows for a recursive extension, enabling online/in-field calibration where parameters are updated as new data arrives, facilitating dynamic drift compensation.

Key Contributions

• Unified Linear Model: The formulation of the CEM allows for the simultaneous estimation of scale, non-orthogonality, and alignment rotation within a single linear framework, eliminating the need for iterative solvers.
• Attitude-Aided Flexibility: The method is applicable to any platform providing orientation data (robotic arms, turntables, IMUs) and does not require predefined sensor orientations or strict mechanical alignment, unlike traditional reference-based methods.
• Closed-Form Solution: The algorithm provides a direct, non-iterative solution, ensuring computational efficiency and avoiding convergence issues associated with nonlinear optimization.
• Robustness: The method jointly estimates the bias and gravity vector, implicitly accounting for platform misalignment errors.

Experimental Results
The authors validated ALAC using synthetic data, a robot-mounted IMU (MPU6050 on an EFFORT manipulator), and a public quasi-static IMU dataset.

• Static Robot Experiments:

  • ALAC achieved a Root Mean Square Error (RMSE) of 0.00913 m/s² and a compensation error of 0.338 m/s² using 24 static orientations.
  • It significantly outperformed reference-based baselines (e.g., TLS, quasi-Newton, UKF) which suffered from large errors due to alignment assumptions or convergence instability.
  • ALAC's accuracy was comparable to iterative self-calibration methods (Newton, LM, Ellipsoid) but achieved this without iterative optimization.
  • A "six-position" sampling strategy (24 orientations) proved more effective than random sampling, converging faster and providing better numerical conditioning.

• Quasi-Static IMU Experiments:

  • On raw, noisy trajectory data, ALAC (both offline and online modes) achieved the lowest RMSE and compensation error among all evaluated baselines, including online Kalman filter methods and iterative self-calibration.
  • On filtered data, ALAC's accuracy was on par with the best iterative self-calibration methods.
  • The recursive (online) implementation demonstrated superior real-time stability and convergence compared to offline baselines.

Significance and Claims
The paper claims that ALAC provides a practical, robust, and efficient calibration scheme for MEMS-based inertial platforms. Its primary significance lies in bridging the gap between high-precision laboratory calibration and low-cost, in-field applications. By leveraging attitude information, it removes the dependency on expensive turntables and strict alignment, while its linear, closed-form nature makes it suitable for real-time, online calibration where iterative methods are too slow or unstable. The authors assert that the method is particularly well-suited for factory calibration and in-field sensor fusion applications where system uncertainties (sensor noise, attitude errors, mechanical misalignment) are present.