Comparing sparse inertial sensor setups for sagittal-plane walking and running reconstructions

Eva Dorschky, Marlies Nitschke, Matthias Mayer, Ive Weygers, Heiko Gaßner, Thomas Seel, Bjoern M. Eskofier, Anne D. KoelewijnFrontiers in Bioengineering and Biotechnology 2025 · DOI

TL;DR

This paper demonstrates that accurate sagittal-plane gait analysis (kinematics and kinetics) is achievable with sparse inertial sensor setups, comparable to full lower-body configurations. By using an optimal control approach with musculoskeletal models, they show that strategically placed sensors on the feet, pelvis, or thighs can provide robust biomechanical insights, paving the way for more unobtrusive, real-world motion analysis.

What Problem Does This Paper Address?

Accurate assessment of human movement, including spatiotemporal, kinematic (joint angles), and kinetic (forces, moments) variables, is crucial for clinical diagnostics, rehabilitation, and sports performance analysis. Traditional methods often rely on laboratory-based optical motion capture (OMC), which is expensive, time-consuming, and limits data collection to controlled environments. Inertial Measurement Units (IMUs) offer a more portable and cost-effective solution, but achieving comprehensive biomechanical analysis with IMUs typically requires a large number of sensors placed on many body segments. This can be obtrusive, affecting natural movement and increasing the chance of placement errors. The core problem this paper addresses is how to achieve accurate and comprehensive gait analysis using a minimal number of IMUs, thereby increasing usability and enabling "in-the-wild" data collection.

Methods

The researchers employed an optimal control approach to reconstruct sagittal-plane walking and running motions. This method involves creating a dynamic simulation of a musculoskeletal model and then optimizing the simulation parameters to simultaneously:

  1. Minimize muscular effort: This encourages physiologically realistic movements.
  2. Minimize tracking error: This ensures the simulated IMU data (accelerometer and gyroscope readings) closely matches the measured IMU data.

This approach inherently addresses issues like sensor drift and soft-tissue artifacts by directly estimating kinetics and kinematics without relying on step-wise inverse dynamics calculations. They tested seven different sensor configurations, ranging from two IMUs on the feet to a full lower-body setup with seven IMUs. The performance of these sparse setups was compared against a gold-standard OMC system.

Datasets

The study utilized a previously collected dataset from ten healthy male participants performing walking and running at six different speeds. This dataset included synchronized data from custom-built IMUs and an optical motion capture system with force plates. The data was pre-processed and is publicly available. No new datasets were introduced; rather, existing data was used to evaluate various IMU configurations.

Key Insights and Ideas

  • Less can be more (sometimes): A full sensor setup doesn't always guarantee the best results. Strategic placement of fewer sensors can yield comparable or even better accuracy for specific variables.
  • Pelvis and Thighs Matter: Including sensors on the pelvis or thighs, in addition to the feet, significantly improves the accuracy of kinematic and kinetic estimations, particularly in reducing forward trunk lean and improving spatiotemporal variables.
  • Foot-only is Limited: While foot-worn IMUs are unobtrusive, relying solely on them leads to the largest errors, especially for knee angles during running.
  • Application-Specific Optimization: The "best" sensor setup depends on the specific biomechanical variables of interest. For instance, some setups might be better for kinematics, while others excel at kinetics.
  • Dynamic Consistency is Key: The optimal control approach ensures that the reconstructed movements are physically plausible and follow the laws of physics, a benefit not always guaranteed by purely data-driven machine learning models.

Key Results

  • Spatio-temporal variables: RMSDs were similar across most setups, but the two-IMU foot-only setup (F) showed the highest errors. Adding a pelvis sensor (FP) improved accuracy.
  • Trunk Lean: Setups without a pelvis sensor (F, FS, FT) resulted in significantly more forward trunk lean during both walking and running compared to setups with a pelvis sensor.
  • Joint Angles and Moments: Generally, setups including pelvis or thigh sensors (FP, FT, FSP, FTP, FSTP) showed smaller errors than the foot-only setup (F). For example, the largest error in knee angle during running was 18 degrees for the two-IMU setup (F) versus 11 degrees for the full setup (FSTP).
  • Relationships (R²): Strong or moderate relationships (R² > 0.5) were found on average for most kinematic and kinetic variables across setups, except for hip and knee moments in some configurations.
  • Quantitative Errors: Mean RMSDs ranged from 4.8-18 degrees for joint angles, 1.0-5.4 BW BH% for joint moments, and 0.03-0.49 BW for ground reaction forces.
  • Computational Cost: While optimal control simulations are computationally intensive (48 ± 3 min per simulation), the number of sensors did not drastically alter the computation time, except for the foot-only setup during walking.

Limitations and Future Work

  • Model Simplifications: The musculoskeletal model used a simplified foot-ground contact model (two points), which may affect the accuracy of horizontal GRF estimations. Future work could explore more sophisticated contact models.
  • Participant Variability: The study noted significant individual differences in simulation quality, particularly for kinetic variables. Further analysis is needed to understand these variations, potentially through more personalized musculoskeletal models.
  • Sensor Placement Uncertainty: The current method relies on manual sensor placement. Future research could investigate estimating sensor placement and orientation directly within the optimal control or machine learning framework.
  • Generalizability: While the study focused on sagittal-plane motion, extending this to 3D analysis with sparse sensors is a logical next step.
  • Real-world Validation: Further validation studies are needed to assess the usability and technical/clinical validity of these sparse sensor setups in patient populations with movement disorders.
  • Processing Differences: The authors acknowledge potential differences in data processing between their IMU-based optimal control approach and the OMC gold standard, which could influence direct comparisons. Future work could involve comparing optimal control simulations against other optimal control simulations.

Summary updated

Abstract

Estimating spatiotemporal, kinematic, and kinetic movement variables with little obtrusion to the user is critical for clinical and sports applications. One possible approach is using a sparse inertial sensor setup, where sensors are not placed on all relevant body segments. Here, we investigated if movement variables can be estimated similarly accurate from sparse sensor setups as from a full lower-body sensor setup. We estimated the variables by solving optimal control problems with sagittal plane lower-body musculoskeletal models, in which we minimized an objective that combined tracking of accelerometer and gyroscope data with minimizing muscular effort. We created simulations for 10 participants at three walking and three running speeds, using seven sensor setups with between two and seven sensors located at the feet, shank, thighs, and/or pelvis. We found that differences between variables estimated from inertial sensors and those from optical motion capture were small for all sensor setups. Including all sensors did not necessarily lead to the smallest root mean square deviations (RMSDs) and highest coefficients of determination ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:msup><mml:mrow><mml:mtext>R</mml:mtext></mml:mrow> <mml:mrow><mml:mn>2</mml:mn></mml:mrow> </mml:msup> </mml:mrow> </mml:math> ). Setups without a pelvis sensor led to too much forward trunk lean and inaccurate spatiotemporal variables. Mean RMSDs were highest for the setup with two foot-worn inertial sensors (largest error in knee angle during running: 18 deg vs. 11 deg for the full lower-body setup), and ranged between 4.8-18 deg for the joint angles, between 1.0-5.4 BW BH% for the joint moments, and between 0.03 BW-0.49 BW for the ground reaction forces. We found strong or moderate relationships ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:msup><mml:mrow><mml:mtext>R</mml:mtext></mml:mrow> <mml:mrow><mml:mn>2</mml:mn></mml:mrow> </mml:msup> <mml:mo>></mml:mo> <mml:mn>0.5</mml:mn></mml:mrow> </mml:math> ) on average for all kinematic and kinetic variables, except for the hip and knee moment for five out of the seven setups. The large range of the coefficient of determination for most kinetic variables indicated individual differences in simulation quality. Therefore, we conclude that we can perform a comprehensive sagittal-plane motion analysis with sparse sensor setups as accurately as with a full sensor setup with sensors on the feet and on either the pelvis or the thighs. Such a sparse sensor setup enables comprehensive movement analysis outside the laboratory, by increasing usability of inertial sensors.
BibTeX
@inproceedings{dorschky2025comparing,
  author = {Eva Dorschky and Marlies Nitschke and Matthias Mayer and Ive Weygers and Heiko Gaßner and Thomas Seel and Bjoern M. Eskofier and Anne D. Koelewijn},
  title = {{Comparing sparse inertial sensor setups for sagittal-plane walking and running reconstructions}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2025},
  doi = {10.3389/fbioe.2025.1507162},
}
Figure