Motion Inference Using Sparse Inertial Sensors, Self-Supervised Learning, and a New Dataset of Unscripted Human...

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Sensor Technology

Background:

• Wearable sensors are increasingly used for biomechanical monitoring, sports, rehabilitation, and human-computer interaction.
• Accurate human kinematics estimation typically requires numerous sensors, limiting practical applications.

Purpose of the Study:

• To achieve accurate full-body kinematics estimates using a minimal number of wearable sensors.
• To introduce a comprehensive dataset for training and evaluating motion inference models.

Main Methods:

• Developed the Virginia Tech Natural Motion Dataset (40+ hours of unscripted daily life motion).
• Employed self-supervised machine learning, specifically sequence-to-sequence and Transformer models, for kinematics inference.
• Utilized reduced sensor configurations (3-4 for upper body, 5-6 for full body).

Main Results:

• Achieved mean angular errors of 10-15 degrees for both upper body and full body kinematics.
• Demonstrated worst-case angular errors below 30 degrees.
• Validated model performance across different sensor placements and machine learning architectures.

Conclusions:

• Self-supervised learning enables accurate human kinematics prediction with significantly fewer sensors.
• The freely available dataset and code facilitate further research in wearable motion capture.
• This approach enhances the feasibility of wearable sensor applications in real-world scenarios.