Machine Learning for Human Motion Intention Detection.

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Exoskeleton control requires precise understanding of user intent to prevent injury.
• Current methods often struggle with the timing of intent recognition, leading to operational issues.
• Identifying gait transitions offers a novel approach for in-time detection.

Purpose of the Study:

• To investigate the feasibility of detecting gait transitions for improved exoskeleton control.
• To evaluate the effectiveness of machine learning models in identifying gait changes.
• To assess the speed and generalizability of transition detection.

Main Methods:

• Utilized Inertial Measurement Unit (IMU) sensor data for gait analysis.
• Trained and tested linear Feedforward Neural Networks and Long Short-Term Memory networks.
• Employed gait data from five subjects for model development and validation.

Main Results:

• Machine learning networks successfully distinguished gait transition periods from continuous motion.
• Rapid detection of walking-to-sitting transitions (0.17s) was achieved, suitable for control applications.
• Detection of standing-to-walking transitions showed longer latency (up to 1.2s).
• Models demonstrated cross-subject generalizability without performance degradation.

Conclusions:

• Gait transition detection is a viable strategy for real-time exoskeleton control.
• Machine learning models offer promising results for in-time gait change identification.
• The developed approach shows potential for robust and adaptable human-exoskeleton interaction.