Robust Stride Detector from Ankle-Mounted Inertial Sensors for Pedestrian Navigation and Activity Recognition with Machine Learning Approaches

Bertrand Beaufils, Frédéric Chazal, Marc Grelet, Bertrand MichelSensors 2019 · DOI

TL;DR

This paper introduces a novel stride detection and activity recognition system using ankle-mounted inertial sensors. By combining a machine learning approach with a ZUPT-inspired technique, the system achieves high accuracy in detecting various strides, even atypical ones, and robustly reconstructs pedestrian trajectories for applications like gait analysis in clinical trials.

What Problem Does This Paper Address?

Accurate pedestrian navigation and activity recognition are crucial for various applications, including healthcare and personal assistance. While Global Navigation Satellite Systems (GNSS) are common for outdoor navigation, they fail in indoor or obstructed environments. Inertial Measurement Units (IMUs) offer a lightweight and deployable solution, but integrating their data leads to significant drift errors. Traditional methods like Zero Velocity Update (ZUPT) rely on detecting foot-ground contact, which is challenging for ankle-mounted sensors where zero velocity is rarely observed. Furthermore, existing stride detection methods often struggle with atypical gaits (e.g., small steps, stairs) and are not robust in real-world scenarios with varied movements. This paper addresses the need for a robust, ankle-mounted stride detection and activity recognition system that can handle diverse gaits and challenging environments, particularly for clinical applications.

Methods

The proposed approach is a multi-stage pipeline leveraging machine learning.

  • Terrestrial Reference Frame Computation: The system first removes gravity from the accelerometer data and computes a terrestrial reference frame. This is crucial for consistent data processing regardless of the sensor's orientation.
  • Pseudo-Speed Computation: Using the processed inertial data, a "pseudo-speed" is calculated. This metric, derived from integrating acceleration, is used to identify candidate intervals that might represent strides.
  • Stride Detection (Machine Learning): A Gradient Boosting Tree (GBT) classifier is employed to distinguish true strides from non-stride movements within the candidate intervals. This is a key innovation, moving beyond fixed thresholds to a data-driven approach.
  • Trajectory Reconstruction: Inspired by ZUPT, a lever arm model estimates ankle speed at stride boundaries. This speed information is then fused with an extended Kalman filter to reconstruct the pedestrian's trajectory, mitigating integration drift.
  • Activity Recognition (Machine Learning): The computed stride trajectories are then used as features for another GBT classifier. This multi-class classifier identifies the activity associated with each stride (e.g., walking, running, climbing stairs, atypical strides).

A significant departure from prior work is the adaptive windowing approach, where the system infers the window size rather than using a fixed-size sliding window, which can lead to errors at activity transitions.

Datasets

The paper describes the creation of a custom dataset for training and evaluation.

  • Training Data: A dataset was built by recording individuals of various ages and heights performing several activities while wearing the WATA device. This data was manually labeled for stride detection (stride vs. non-stride) and activity recognition (atypical stride, walking, running, climbing stairs, descending stairs). The database contains approximately 6000 positive and 6000 negative stride intervals.
  • Evaluation Data:
    • Controlled Environments: Motion capture (MOCAP) sessions were conducted with seven participants performing various walking paces, small steps, and side steps. This provided ground truth for stride detection accuracy.
  • Daily Activities: Tests were performed in scenarios mimicking daily life, including sitting, bicycling, car rides, carrying the device in hand, backpack, or pocket.
  • Clinical Trials: Data from DMD patients during hospital tests (four-stairs test, 6-minute walk, running) and a healthy child's home recording were used to evaluate HAR performance in real-world, uncontrolled environments.

Key Insights and Ideas

  • Ankle-Mounted Sensors are Viable: The research demonstrates that robust stride detection and activity recognition are achievable even with ankle-mounted IMUs, overcoming the limitation of not observing zero velocity.
  • Machine Learning for Robustness: Employing GBT for both stride detection and activity recognition proves highly effective in handling the variability of human gaits and complex movement patterns, outperforming traditional methods.
  • Adaptive Windowing is Key: Moving away from fixed-size sliding windows and inferring interval sizes adaptively significantly improves performance, especially at activity transitions.
  • Pseudo-Speed as a Powerful Feature: The "pseudo-speed" metric, derived from processed inertial data, is identified as a crucial feature for detecting stride boundaries, particularly in challenging scenarios like fast side-stepping.
  • Trajectory as a Rich Feature for HAR: The computed stride trajectory, not just simple kinematic features, provides rich information for accurately classifying activities, even for subtle differences in gait.

Key Results

  • Stride Detection:
    • 100% accuracy for walking strides (over 5600 detected).
  • ~98% accuracy for atypical strides (over 2000 detected), including small steps and side steps.
  • Near-zero false positive rates in critical daily situations (sitting, bicycling, car rides, etc.).
  • Trajectory Reconstruction:
    • Absolute mean error of ~3 cm for walking stride length.
  • Absolute mean error of ~5 cm for atypical strides.
  • Demonstrated good performance in complex indoor environments with narrow spaces and stairs.
  • Activity Recognition:
    • Overall accuracy of ~99.4% for adult recordings.
  • Over 97% success rate for DMD patient recordings, which is considered challenging due to gait variability.
  • Successfully classified five activity types: "atypical stride," "walking," "running," "climbing stairs," and "descending stairs."

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the difference between "atypical stride" and "walking" can be subtle, leading to some misclassifications, especially for small forward steps.
  • The performance on DMD patient data, while good, was slightly lower than on adult data, potentially due to the disproportionate representation of DMD strides in the training set and the inherent difficulty in classifying their unique gaits.
  • The accuracy of stride length estimation for atypical strides is slightly lower than for normal walking, attributed to the lever arm model's reduced consistency for ankle speed estimation at stride boundaries.
  • Future Work:
    • The authors believe their methodology is ready for application in long-term home recordings for clinical trials, enabling the computation of clinical outcomes related to hospital tests.
  • The ability to run is identified as a key physical capacity for DMD patients, suggesting future work could focus on more precise running stride analysis.
  • The system's robustness in uncontrolled environments and its potential for continuous monitoring in home settings are highlighted as key advantages for future clinical applications.

Summary updated

BibTeX
@article{beaufils2019robust,
  author = {Bertrand Beaufils and Frédéric Chazal and Marc Grelet and Bertrand Michel},
  title = {{Robust Stride Detector from Ankle-Mounted Inertial Sensors for Pedestrian Navigation and Activity Recognition with Machine Learning Approaches}},
  journal = {Sensors},
  year = {2019},
  doi = {10.3390/s19204491},
}
Figure