Ultra-Robust Real-Time Estimation of Gait Phase

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Accurate gait phase estimation is crucial for analyzing human locomotion and developing assistive devices.
• Existing methods often struggle with diverse walking conditions and lack robustness.

Purpose of the Study:

• To develop an ultra-robust and accurate gait phase estimator using a time-delay neural network (D67).
• To evaluate the estimator's performance across various walking speeds and conditions, including interactions with exoskeletons.

Main Methods:

• Training a time-delay neural network (D67) on hip and knee joint angle data from 14 participants.
• Testing the D67 on diverse gait data, including normal walking, varied speeds, and challenging conditions like stop-start and speed changes.
• Comparing D67 performance against state-of-the-art techniques and evaluating its robustness with an active exoskeleton.

Main Results:

• Achieved average Root Mean Square Error (RMSE) of 1.74% (treadmill) and 2.35% (overground) for spatial gait phase estimation.
• Demonstrated Mean Absolute Error (MAE) of 1.70% (treadmill) and 2.74% (overground) for temporal heel-strike detection.
• Exhibited uniform performance across participants and gait conditions, proving robustness to speed variations, limping, and sudden starts/stops.

Conclusions:

• The D67 provides a highly accurate and robust gait phase estimation solution.
• It demonstrates superior or comparable performance to existing methods, even without ground contact sensors.
• The D67 shows significant potential for real-world applications, including integration with active exoskeletons.