Dimensionality Reduction of Human Gait for Prosthetic Control

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Controlling lower limb prostheses is complex.
• Dimensionality reduction techniques are needed to simplify human gait data.
• The optimal dimensionality reduction method for gait data is unclear.

Purpose of the Study:

• Compare Principal Component Analysis (PCA) and pose-based autoencoders (Pose-AE) for gait data.
• Evaluate PCA, Pose-AE, and a movement-based autoencoder (Move-AE) for capturing time-varying gait properties.
• Assess methods for movement classification and individual identification to inform prosthetic control.

Main Methods:

• Applied PCA and Pose-AE to human kinematics data during walking and stair climbing.
• Trained Move-AE on full human movement trajectories.
• Compared performance of PCA, Pose-AE, and Move-AE on movement classification and individual identification tasks.

Main Results:

• Pose-AE demonstrated higher Variance Accounted For (VAF) than PCA for dimensionality reduction across various movements.
• Move-AE significantly outperformed both PCA and Pose-AE in movement classification and individual identification.
• Autoencoders proved more suitable than PCA for gait data dimensionality reduction.

Conclusions:

• Autoencoders are more effective than PCA for dimensionality reduction of human gait data.
• Move-AE can encode comprehensive movement representations beneficial for prosthetic control.
• This research provides a pathway for improved lower limb prosthesis control systems.