QMEDNet: A quaternion-based multi-order differential encoder-decoder model for 3D human motion prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Traditional recurrent neural networks in Euclidean space struggle with high distortion and weak semantic expression for 3D human motion representation.
• Existing deep learning models for human motion prediction often face limitations in accurately capturing complex motion dynamics.

Purpose of the Study:

• To propose a novel model for 3D human motion prediction using hypercomplex vector spaces, specifically quaternion space.
• To improve the semantic expression and reduce distortion in human motion representation for deep learning models.

Main Methods:

• Developed a novel model, QMEDNet, based on quaternion representation for 3D human motion prediction.
• Fused multi-order differential information in both the encoder (using graph convolution) and decoder (using gated recurrent units).
• Mapped fused features into quaternion space for enhanced motion state processing.

Main Results:

• Demonstrated that multi-order information in quaternion space leads to a more reasonable description of 3D human motion.
• QMEDNet achieved superior performance compared to advanced short and long-term motion prediction methods.
• Validated the model's effectiveness on public datasets: Human 3.6M and CMU Mocap.

Conclusions:

• Quaternion-based representation significantly enhances 3D human motion prediction accuracy and semantic understanding.
• The proposed QMEDNet model offers a promising advancement in the field of human motion analysis and prediction.
• Future work can explore further applications of hypercomplex methods in motion-related AI tasks.