Piecewise Linear Labeling Method for Speed-Adaptability Enhancement in Human Gait Phase Estimation.

Area of Science:

• Robotics and Biomechanics
• Human-Computer Interaction

Background:

• Accurate human gait phase estimation is crucial for synchronized control of wearable robotic devices like prostheses and exoskeletons.
• Data-driven, learning-based methods are increasingly used for gait phase estimation, but require precise ground truth labeling.

Purpose of the Study:

• To develop a novel labeling method for gait phase estimation that accounts for variable toe-off onset across different walking speeds.
• To improve the accuracy and speed adaptability of gait phase estimation models.

Main Methods:

• A piecewise linear labeling method was proposed to define ground truth for gait phase estimation.
• Whole-body marker data was used to compute angular positions and velocities of thigh and torso segments as input features.
• Three Long Short-Term Memory (LSTM) models (general, slow, normal-fast) were trained and compared using the new labeling method.

Main Results:

• The proposed piecewise linear labeling method significantly improved estimation accuracy when training the general LSTM model.
• Enhanced accuracy was observed particularly during the mid-stance phase of the gait cycle.
• The method demonstrated robust performance in accurately detecting critical gait events such as heel-strike and toe-off across various speeds.

Conclusions:

• The novel piecewise linear labeling method enhances speed adaptability in human gait phase estimation.
• This approach leads to improved accuracy for overall gait phase estimation and the precise detection of heel-strike and toe-off events.
• The findings support the application of this method for more effective control of wearable robotic systems.