Human Motion Prediction via Dual-Attention and Multi-Granularity Temporal Convolutional Networks

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Intelligent devices are increasingly integrated into daily life, necessitating precise human motion analysis for effective human-device interaction.
• Current human motion prediction methods struggle to capture complex spatial and temporal dynamics in motion data, limiting prediction accuracy.

Purpose of the Study:

• To develop a novel human motion prediction method that effectively captures spatial correlations and temporal dependencies.
• To improve the accuracy and reliability of human motion prediction for enhanced human-device interaction.

Main Methods:

• Proposed a dual-attention (DA) model integrating joint and channel attention to extract spatial features from 3D human motion data.
• Developed multi-granularity temporal convolutional networks (MgTCNs) with varied receptive fields to capture intricate temporal dependencies in motion sequences.
• Utilized Human3.6M and CMU-Mocap datasets for comprehensive evaluation.

Main Results:

• The proposed DA-MgTCNs method significantly outperformed existing approaches in both short-term and long-term human motion prediction.
• Demonstrated superior performance in accurately predicting future human poses and movements based on sequential data.
• Validated the effectiveness of the dual-attention mechanism and multi-granularity temporal modeling.

Conclusions:

• The novel DA-MgTCNs method offers a significant advancement in human motion prediction accuracy.
• This approach facilitates more harmonious coexistence and efficient interaction between humans and intelligent devices.
• The findings highlight the potential of advanced deep learning architectures for complex motion analysis.