Benchmarking Foot Trajectory Estimation Methods for Mobile Gait Analysis

Julius Hannink, Malte Ollenschläger, Felix Kluge, Nils Roth, Jochen Klucken, Björn M. Eskofier

TL;DR

This paper benchmarks different methods for estimating foot trajectory from wearable inertial sensors, a crucial step for objective gait analysis in medical applications. By evaluating various orientation estimation and double integration techniques on a common dataset, they identify the most suitable processing pipeline, highlighting the trade-offs between accuracy, precision, and computational cost.

What Problem Does This Paper Address?

Assessing gait is vital for diagnosing and monitoring neurological conditions like Parkinson's disease. While wearable inertial measurement units (IMUs) offer a promising mobile solution, accurately reconstructing foot trajectory from IMU data remains a challenge. Existing research often uses diverse datasets, making direct comparison of different estimation methods difficult. This paper aims to bridge this gap by providing a fair, unified evaluation of several key processing steps for foot trajectory estimation using a standardized dataset and a marker-based motion capture system as a ground truth.

Methods

The paper focuses on two critical stages in the foot trajectory estimation pipeline: Orientation Estimation and Double Integration.

For Orientation Estimation, they compare three methods:

  • Gyroscope Integration: A straightforward method that updates orientation based solely on angular velocity.
  • Madgwick CF: A complementary filter that fuses accelerometer and gyroscope data, with a specific update path for static conditions.
  • Euston CF: Another complementary filter, similar to Madgwick's but with a different error correction mechanism and tunable gain parameters.

For Double Integration, which converts acceleration to position, they evaluate three approaches:

  • Direct Integration: A basic method using the trapezoidal rule for integration, followed by linear drift removal.
  • Direct and Reverse Integration: This method fuses a forward and a time-reversed integration to better handle boundary conditions, particularly the zero-velocity assumption at the start and end of a stride.
  • Analytic Integration: This approach decomposes the acceleration signal into a Fourier or B-Spline basis and then analytically integrates these basis functions.

All methods are evaluated against a VICON motion capture system.

Datasets

The study utilizes a dataset previously acquired by Kanzler et al. [20]. This dataset consists of:

  • 735 strides from 16 healthy subjects.
  • IMU data captured from sensors placed on the lateral side of the shoe.
  • Synchronized ground truth data from a VICON motion capture system, providing precise foot position and orientation.
  • The dataset was collected under various walking conditions to ensure a wide range of stride lengths and velocities.

Key Insights and Ideas

  • No single "best" method: The optimal choice for orientation estimation and double integration depends on the specific application's requirements, particularly the trade-off between accuracy and computational cost.
  • Complementary filters offer marginal gains: While Madgwick and Euston CFs provide some improvement in orientation estimation accuracy, they come with a significant computational overhead compared to simple gyroscope integration. For real-time applications, gyroscope integration might be preferred.
  • Direct and Reverse Integration excels for clearance: This method significantly improves precision in estimating foot clearance compared to direct or analytic integration, likely due to its robust handling of boundary conditions.
  • Analytic integration doesn't outperform numeric: Contrary to theoretical expectations, the analytic integration scheme did not yield superior results to numeric integration methods, suggesting that its complexity might not be justified in this context.
  • Parameter-free gyroscope integration is viable: For single-stride gait analysis, simple gyroscope integration is as valid as more complex sensor fusion techniques, offering broader applicability due to its parameter-free nature.

Key Results

  • Orientation Estimation Ranking (Best to Worst): Madgwick CF → Gyroscope Integration → Euston CF.
  • Madgwick CF achieved the best accuracy but was ~4x slower than Gyroscope Integration.
  • Gyroscope Integration was a strong contender, offering a good balance of accuracy and speed, and is parameter-free.
  • Euston CF was the least accurate among the tested methods.
  • Double Integration Ranking (Best to Worst): Direct & Reverse Integration → Direct Integration → Analytic Integration.
  • Direct & Reverse Integration achieved the highest precision, particularly for clearance estimation (0.84 ± 3.98 cm).
  • Analytic Integration was the slowest and did not outperform numeric integration methods.
  • Computational Cost: Gyroscope integration was the fastest orientation estimation method (6.13 ms/stride). Direct Integration was the fastest double integration method (20.91 ms/stride). The combined pipeline using Madgwick CF and Direct Integration was the fastest overall.

Limitations and Future Work

  • Dataset Size and Generalization: The authors acknowledge that their dataset, while comprehensive for healthy subjects, might not be large enough to fully guarantee generalization to other populations or conditions.
  • Patient Data: The current study focused on healthy subjects. Future work should involve acquiring and comparing estimation techniques on patient data (e.g., Parkinson's disease patients) to assess performance in clinical populations.
  • Real-time Performance: While execution times were measured, further investigation into real-time implementation and its challenges is suggested.
  • Analytic Integration: The theoretical advantages of analytic integration were not realized in practice. Further research could explore alternative basis functions or optimization strategies for this method.

Summary updated

BibTeX
@inproceedings{hannink2017benchmarking,
  author = {Julius Hannink and Malte Ollenschläger and Felix Kluge and Nils Roth and Jochen Klucken and Björn M. Eskofier},
  title = {{Benchmarking Foot Trajectory Estimation Methods for Mobile Gait Analysis}},
  booktitle = {Sensors},
  year = {2017},
  url = {https://openreview.net/forum?id=lY4vzD2BD0J},
}
Figure