Personalization of Wearable Sensor-Based Joint Kinematics Estimation Using Computer Vision for Hip Exoskeleton...

Area of Science:

• Biomechanics
• Biomedical Engineering
• Computer Vision

Background:

• Accurate lower-limb joint kinematics are crucial for patient monitoring, rehabilitation, and exoskeleton control.
• Current deep learning (DL) models often need extensive datasets for adaptation to new gait patterns.
• Computer vision human pose estimation models are deployable but not suitable for camera-restricted environments.

Purpose of the Study:

• To develop a computer vision-based DL adaptation framework for real-time joint kinematics estimation.
• To enable accurate gait analysis with minimal data requirements and without professional motion capture.
• To adapt existing models for clinical populations and novel users.

Main Methods:

• Proposed a computer vision-based DL adaptation framework utilizing transfer learning.
• Adapted a temporal convolutional network (TCN) using a small dataset (1-2 gait cycles).
• Validated the framework on stiff knee gait data.

Main Results:

• The adapted TCN model reduced root mean square error by 9.7% compared to a model trained only on able-bodied data.
• The model achieved a 19.9% reduction in root mean square error compared to a model trained solely on stiff knee data.
• Demonstrated feasibility of smartphone camera-trained DL models for real-time kinematics estimation.

Conclusions:

• The proposed framework enables real-time joint kinematics estimation with minimal data requirements.
• This approach facilitates adaptation to novel users and clinical populations, such as those with stiff knee gait.
• The technology has potential applications in wearable robotics and remote patient monitoring.