Efficient Extrinsic Self-Calibration of Multiple IMUs using Measurement Subset Selection

Jongwon Lee, David Hanley, Timothy Bretl

Introduction

For autonomous systems like drones and mobile robots, precise extrinsic calibration—the determination of the relative position and orientation between sensors—is critical for accurate state estimation. When multiple Inertial Measurement Units (IMUs) are used, self-calibration using the robot's own motion is a preferred alternative to specialized calibration rigs. However, identifying informative motion segments within hours of operational data is a significant computational challenge. This paper presents an efficient method for measurement subset selection that reduces calibration time by two orders of magnitude without sacrificing accuracy.

The Challenge of Informative Motion

Not all motion is created equal for the purposes of calibration. Driving in a straight line or remaining stationary provides little to no information about the spatial relationship between sensors. To maintain efficiency and accuracy, researchers often use "greedy algorithms" to select a subset of the most informative data segments based on the Fisher Information Matrix (FIM).

The primary bottleneck in standard greedy selection is the "Evaluate" phase. Traditionally, the system must re-run a complex nonlinear least-squares optimization every time it considers adding a new data segment to the subset. For a dataset with $L$ segments, this results in $O(L^2)$ complexity, making the process prohibitively slow for long-duration datasets.

Proposed Method: Evaluation at Initial Parameters

The authors propose a streamlined approach: Greedy with Utility Evaluation at Initial Calibration Parameters. This method is based on the insight that the "utility" or information gain of a specific measurement is relatively robust to small errors in the initial estimates of the extrinsic parameters.

The Algorithm

Instead of the iterative re-calibration approach, the proposed algorithm follows these steps:

  1. Initial Guess: Perform a rapid initial orientation estimation.
  2. Single-Pass Evaluation: Evaluate the utility of all candidate data segments using the Fisher Information Matrix calculated at this initial guess.
  3. Selection: Select the top-performing segments to form a sparse subset.
  4. Final Optimization: Perform a single, heavy-duty nonlinear least-squares optimization using only the selected subset.

By decoupling the selection process from the final optimization, the computational complexity is reduced from $O(L^2)$ to $O(L)$.

Experimental Validation

The researchers validated the "Init-Param" approach using both extensive simulations and real-world hardware.

  • Simulation: 48 trajectories across five diverse datasets (including euroc_mav and kaist) were used to test the algorithm's robustness.
  • Hardware: A custom rig with two IMUs and a reference camera was used to collect real-world motion data.

The method was benchmarked against a "Baseline" (using all available data) and the "Original Greedy" algorithm.

Key Findings

1. Sensitivity and Robustness

A key contribution of the paper is the sensitivity analysis of the information gain. The researchers found that while utility evaluation is somewhat sensitive to orientation errors, it is remarkably insensitive to errors in position. This justifies the use of a quick orientation guess to drive the selection process.

2. Massive Efficiency Gains

The performance improvement was most striking in the kaist dataset (car-based motion). The total calibration time was slashed from 379.42 seconds (Baseline) to just 20.08 seconds (Proposed). Specifically, the "Evaluate" portion of the algorithm—the previous bottleneck—was reduced from hundreds of seconds to a mere 4.6 seconds.

3. "Less is More"

Interestingly, in scenarios like the kaist dataset where much of the motion is uninformative (straight-line driving), the subset selection method actually outperformed the full-data baseline. The baseline was "polluted" by low-utility data that made the optimization problem weakly observable. The proposed method achieved sub-centimeter and sub-degree accuracy using less than 3% of the total data.

Conclusion

This research demonstrates that efficient extrinsic self-calibration is possible by intelligently selecting data subsets. By proving that utility evaluation can be performed at initial parameter estimates, the authors have removed a major computational barrier. The resulting algorithm is fast, accurate, and has a low memory footprint, making it ideal for resource-constrained embedded systems in modern robotics.

Summary updated

BibTeX
@inproceedings{lee2024efficient,
  author = {Jongwon Lee and David Hanley and Timothy Bretl},
  title = {{Efficient Extrinsic Self-Calibration of Multiple IMUs using Measurement Subset Selection}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10802604},
}
Figure