Human Motion Enhancement and Restoration via Unconstrained Human Structure Learning

Area of Science:

• Computer Vision
• Biomechanical Engineering
• Machine Learning

Background:

• Human motion capture (HMC) is increasingly used in civilian applications like gaming and sports science.
• Affordable HMC sensors often yield inaccurate motion data, limiting their practical use.
• Existing methods may require extensive data or prior knowledge of human kinematics.

Purpose of the Study:

• To develop a novel, unsupervised method for human motion reconstruction and enhancement.
• To improve the accuracy and reduce the cost of motion capture data acquisition.
• To leverage spatio-temporal attention-based graph convolutional networks (ST-ATGCNs) for motion data refinement.

Main Methods:

• Utilizing ST-ATGCNs to learn human skeleton structure and motion dynamics.
• Implementing an unsupervised approach for motion data restoration.
• Validating the method on diverse motion datasets and real-world sensor data (e.g., SONY mocopi).

Main Results:

• Demonstrated significant enhancement of low-precision motion capture data quality.
• Showcased the ST-ATGCNs' ability to learn motion logic without prior kinematic constraints.
• Achieved effective motion data restoration and improved accuracy.

Conclusions:

• The ST-ATGCN approach offers a cost-effective solution for high-quality motion capture.
• This method enhances both the accessibility and precision of human motion data.
• The findings suggest a promising direction for advancing HMC technology for broader applications.