Experimental evaluation of regression model-based walking speed estimation using lower body-mounted IMU

Area of Science:

• Biomechanics
• Wearable technology
• Sensor data analysis

Background:

• Accurate walking speed estimation is crucial for clinical assessments and rehabilitation.
• Inertial sensors offer a portable solution for gait analysis.
• Optimizing sensor placement and regression models is key to improving accuracy.

Purpose of the Study:

• To concurrently compare the accuracy of regression model-based walking speed estimation.
• To evaluate the impact of different variables, features, sensor locations, and regression methods.
• To identify optimal configurations for reliable gait analysis using lower body inertial sensors.

Main Methods:

• 15 healthy subjects performed free walking trials.
• Lower body inertial sensors were mounted at the waist and ankle.
• Variables included external acceleration, and features were extracted from time and frequency domains.
• Gaussian process regression and least squares regression with Lasso were compared.

Main Results:

• Gaussian process regression demonstrated higher accuracy than least squares regression with Lasso.
• External acceleration as a variable improved estimation accuracy.
• Waist and ankle-mounted sensors achieved similar accuracies (4.5% and 4.9%) using both time- and frequency-domain features.
• Accuracy decreased more significantly for waist-mounted sensors using only frequency-domain features.

Conclusions:

• Gaussian process regression is a superior method for walking speed estimation with inertial sensors.
• External acceleration is a valuable variable for enhancing gait analysis accuracy.
• Waist and ankle sensor placements provide comparable results when using comprehensive feature sets.
• Sensor placement and feature selection critically influence estimation accuracy, particularly with frequency-domain features.