Walking speed estimation using foot-mounted inertial sensors: comparing machine learning and strap-down integration...

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Accurate estimation of walking speed is crucial for gait analysis and rehabilitation.
• Foot-mounted inertial measurement units (IMUs) offer a practical solution for unobtrusive gait monitoring.
• Existing methods like strap-down integration (SDI) have limitations in accuracy and variability.

Purpose of the Study:

• To compare the efficacy of machine learning (ML) and strap-down integration (SDI) for stride-by-stride walking speed estimation.
• To evaluate the performance of different gait event detection methods within SDI.
• To assess the impact of personalization on ML-based walking speed estimation.

Main Methods:

• Walking speed was calculated by dividing stride length by stride time derived from foot-mounted IMU data.
• SDI methods employed hidden Markov model (HMM)-based or threshold-based gait event detection.
• An ML method utilized a linear regression model for stride length estimation, validated via leave-one-subject-out cross-validation and subject-specific calibration.

Main Results:

• Both ML and SDI methods achieved comparable root mean square estimation errors (2.0-4.2%).
• The ML method demonstrated lower intra-subject variability and higher inter-subject variability compared to SDI.
• Personalization of the ML model reduced its intra-subject variability.

Conclusions:

• Machine learning offers a robust alternative to SDI for estimating walking speed from IMU data.
• The ML method's ability to reduce variability and benefit from personalization makes it highly promising for gait analysis.
• Further research into ML algorithms can enhance the accuracy and reliability of wearable gait monitoring systems.