Multilayer Joint Gait-Pose Manifolds for Human Gait Motion Modeling

Area of Science:

• Computer Vision
• Robotics
• Biomechanics

Background:

• Human gait and pose modeling are crucial for understanding human motion.
• Existing Gaussian process (GP) models face challenges with complex gait variations and limited data.

Purpose of the Study:

• To introduce novel multilayer joint gait-pose manifolds (multilayer JGPMs) for enhanced human motion modeling.
• To explore topological priors (cylindrical and toroidal) for coupling gait and pose manifolds.
• To develop techniques for efficient model learning with limited data.

Main Methods:

• Utilized a topologically-constrained Gaussian process (GP) latent variable model.
• Introduced training data diversification to generate varied stride data.
• Implemented topology-aware local learning to accelerate model training.

Main Results:

• Demonstrated the effectiveness of multilayer JGPMs on Carnegie Mellon University motion capture data.
• Showcased superior performance compared to existing GP-based motion models.
• Validated the suitability of cylindrical and toroidal topological priors.

Conclusions:

• Multilayer JGPMs offer a robust framework for complex human gait and pose modeling.
• The proposed data diversification and local learning techniques improve model efficiency.
• This work advances the state-of-the-art in human motion analysis and synthesis.