Learning behavior aware features across spaces for improved 3D human motion prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• 3D skeleton-based human motion prediction is crucial for human-machine interaction.
• Existing methods often struggle with information loss and distortions in Euclidean space representations.
• Novel approaches are needed to capture complex human motion dynamics effectively.

Purpose of the Study:

• To develop a novel network for improved 3D human motion prediction.
• To leverage both Euclidean and Geometric Algebra spaces for richer feature extraction.
• To enhance the accuracy and robustness of forecasting future human poses.

Main Methods:

• Proposed Cross-space Behavior-aware Feature Learning Networks (CBFL).
• Developed a Geometric Algebra Dependency-Aware Extractor using Geometric Algebra-based Fully Connected layers.
• Designed an Euclidean Kinematic-Aware Extractor with temporal-wise and spatial-wise Kinematic-Aware Attention and Feature Extraction.

Main Results:

• CBFL effectively captures spatial-temporal kinematic correlations and geometric algebraic dependencies.
• The integrated approach enhances motion dynamics representation.
• Experiments showed consistent performance improvement, reducing Mean Per Joint Position Error (MPJPE) by an average of 4.3% on the Human3.6M dataset.

Conclusions:

• The proposed CBFL model offers a significant advancement in 3D human motion prediction.
• Integrating Geometric Algebra space alongside Euclidean space provides a more comprehensive motion representation.
• This approach leads to more accurate and reliable forecasting of future human poses.