Template-Based Step Detection with Inertial Measurement Units

Laurent Oudre, Rémi Barrois, Thomas Moreau, Charles Truong, Aliénor Vienne-Jumeau, Damien Ricard, Nicolas Vayatis, Pierre-Paul Vidal

TL;DR

This paper introduces a novel template-based algorithm for detecting steps using inertial measurement units (IMUs). The method leverages a library of diverse step templates to achieve high accuracy (98% recall and precision) across healthy and pathological subjects, outperforming traditional signal processing approaches.

What Problem Does This Paper Address?

Accurate and robust step detection is crucial for monitoring human locomotion, especially in clinical settings where gait abnormalities can indicate underlying health issues. Traditional methods often rely on signal processing techniques like filtering and thresholding, which are highly sensitive to parameter tuning and struggle with the variability introduced by different pathologies, ages, and walking conditions (e.g., initiation, termination, or turning steps). This makes it difficult to develop a single algorithm that performs well across a heterogeneous population, limiting its clinical applicability. The authors aim to bridge this gap by developing a method that is adaptable to diverse gait patterns and pathologies.

Methods

The core of the proposed method is a template-based matching approach. Instead of relying on fixed thresholds or signal processing rules, the algorithm uses a predefined library of 55 manually curated step templates. These templates represent various types of steps (normal, initiation, termination, U-turn) observed in healthy and pathological subjects.

The detection process involves:

  1. Pearson Correlation Coefficient Calculation: For each template and each of the three IMU signal components (vertical acceleration, Z-axis acceleration, and Y-axis angular velocity), the algorithm calculates the Pearson correlation coefficient against the input signal across all possible time shifts. This measure is chosen for its independence from scale and orientation, and its ability to capture shape similarities.
  2. Local Maxima Search: Potential step occurrences are identified by finding local maxima in the correlation coefficients.
  3. Greedy Selection: The algorithm iteratively selects the highest correlation coefficient. If the corresponding template can be placed at that time without overlapping a previously detected step, the step is confirmed. The time window of the detected step is then marked as "forbidden" for subsequent iterations. This greedy approach continues until no more potential steps can be found or the correlation values fall below a threshold ($\lambda$).
  4. Post-processing: A final step removes false detections by discarding steps whose standard deviation is significantly lower than that of the template used for detection, indicating periods of inactivity.

Datasets

The study utilized a substantial database comprising 1020 recordings from 230 subjects. This dataset included:

  • Healthy subjects
  • Orthopedic disease patients (lower limb osteoarthrosis, cruciate ligament injury)
  • Neurological disease patients (hemispheric stroke, Parkinson's disease, toxic peripheral neuropathy, radiation-induced leukoencephalopathy)

The database contained a total of 40,465 annotated steps, including normal steps, initiation/termination steps, and U-turn steps. The annotations were performed by experts trained on specialized gait validation tools (GAITRite and CodaMotion), with a reported manual annotation error of approximately 1-2 samples.

Key Insights and Ideas

  • Diversity is Key: The central idea is that a single "ideal" step model is insufficient. A library of diverse templates is necessary to capture the wide range of gait patterns encountered in real-world and clinical scenarios, especially across different pathologies.
  • Emulating Clinical Observation: The template-based approach is framed as an attempt to emulate the clinical observation process, where experts recognize different gait patterns.
  • Robustness over Precision in Parameter Tuning: The template method, by its nature, is less sensitive to the precise tuning of parameters compared to traditional signal processing methods. This makes it more adaptable to varied conditions.
  • U-turn Steps are Challenging: The study highlights U-turn steps as particularly difficult to detect accurately, likely due to their lower amplitude and altered dynamics, which can be mistaken for noise or inactivity.

Key Results

The template-based method demonstrated excellent performance across the diverse dataset:

  • Overall: 98.34% recall and 98.30% precision.
  • Comparison to Pan-Tompkins: The proposed method significantly outperformed the state-of-the-art Pan-Tompkins algorithm (97.82% recall, 95.72% precision) especially on neurological and orthopedic subjects, where Pan-Tompkins showed a tendency for over-detection.
  • Step Type Performance: Normal steps achieved the highest accuracy (99.58% recall, 99.04% precision). U-turn steps showed the lowest performance (83.87% recall, 90.76% precision), though still promising.
  • Time Estimation: For correctly detected steps, the median absolute errors for start and end times were low (around 4-8 samples or 0.04-0.08s), and duration errors were around 7-15 samples (0.07-0.15s), showing good accuracy in temporal localization.
  • Computational Cost: The algorithm is computationally efficient, processing over 8.5 hours of data in less than 30 seconds.

Limitations and Future Work

  • U-turn Step Detection: The paper acknowledges that U-turn steps, especially those with very low amplitude or where the foot drags, remain a challenge. The current template library has a duration constraint that might miss very short or very long steps.
  • Template Library Optimization: The current template library was manually curated. Future work could explore:
  • Automatic Template Generation: Using unsupervised machine learning techniques (like dictionary learning) to automatically extract typical steps from data.
  • Semi-supervised Learning: Developing methods to automatically adapt the library to specific pathologies or signal types.
  • Expanding the Library: Including more templates for outlier cases like very short steps or pathological U-turns.
  • Free-living Conditions: The authors suggest the method's robustness and low computational cost make it suitable for extension to free-living monitoring, including unconstrained walking.
  • Sensor Placement: While the current study focused on foot-worn sensors, adapting the method for other sensor placements (e.g., waist-worn) is also a potential avenue.

Summary updated

Abstract

This article presents a method for step detection from accelerometer and gyrometer signals recorded with Inertial Measurement Units (IMUs). The principle of our step detection algorithm is to recognize the start and end times of the steps in the signal thanks to a predefined library of templates. The algorithm is tested on a database of 1020 recordings, composed of healthy subjects and patients with various neurological or orthopedic troubles. Simulations on more than 40,000 steps show that the template-based method achieves remarkable results with a 98% recall and a 98% precision. The method adapts well to pathological subjects and can be used in a medical context for robust step estimation and gait characterization.
BibTeX
@inproceedings{oudre2018template,
  author = {Laurent Oudre and Rémi Barrois and Thomas Moreau and Charles Truong and Aliénor Vienne-Jumeau and Damien Ricard and Nicolas Vayatis and Pierre-Paul Vidal},
  title = {{Template-Based Step Detection with Inertial Measurement Units}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=td6G4Jq6FJh},
}

Loading related content...

Figure