Semantic Correlation Attention-Based Multiorder Multiscale Feature Fusion Network for Human Motion Prediction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current human motion prediction models lack accuracy due to ignoring semantic correlations between body parts and temporal motion dynamics.
• Existing methods fail to capture the intricate relationships within human poses over time, limiting predictive capabilities.

Purpose of the Study:

• To propose a novel Semantic Correlation Attention-based Multiorder Multiscale Feature Fusion network (SCAFF) for enhanced human motion prediction.
• To address the limitations of current models by incorporating semantic correlations between body parts and motion time.

Main Methods:

• The proposed SCAFF network features an encoder-decoder architecture.
• The encoder utilizes a Multiorder Difference Calculation module (MODC) and stacked Semantic Correlation Attention-based Graph Calculation Operators (SCA-GCOs) for feature extraction.
• A Semantic Correlation Attention Module (SCAM) refines features by learning semantic correlations, while Multiorder and Multiscale Feature Fusion modules integrate extracted information.

Main Results:

• The SCAFF network successfully extracts multiscale features and captures temporal dynamics by considering semantic correlations.
• Experimental results on public datasets show that SCAFF significantly outperforms existing human motion prediction models.
• The model demonstrates superior accuracy in predicting future human states.

Conclusions:

• The study presents SCAFF as a pioneering approach in human motion prediction by integrating semantic correlations between body parts and motion time.
• The proposed method offers a significant advancement in the field, achieving state-of-the-art performance.
• SCAFF provides a more robust and accurate framework for understanding and predicting human movement.