IMPRESS: Incomplete Human Motion Prediction via Motion Recovery and Structural-Semantic Fusion.

Area of Science:

• Computer Vision
• Robotics
• Human-Robot Interaction

Background:

• Human motion prediction is crucial for computer vision and human-robot interaction.
• Existing methods struggle with incomplete data and capturing diverse body part relationships.
• These limitations hinder accurate human motion prediction performance.

Purpose of the Study:

• To propose a novel method, IMPRESS (Incomplete human Motion Prediction through motion REcovery and Structure-Semantic fusion), for robust human motion prediction.
• To address challenges posed by missing observations and improve the modeling of inter-segmental dynamics.
• To enhance prediction accuracy by capturing both structural and semantic relationships within human motion.

Main Methods:

• Developed a wavelet-based self-attention module for motion recovery, distinguishing high-frequency motion details and low-frequency global trends.
• Designed a structure-semantic fusion graph convolutional network to model diverse relations among body parts.
• Implemented a dual-channel sliding window attention mechanism to capture motion periodicity and ensure smoother predictions.

Main Results:

• IMPRESS demonstrated superior performance on benchmark datasets (Human3.6M, CMU-MoCap) under both complete and incomplete observation scenarios.
• The wavelet-based attention effectively captured motion details and global trends.
• The structure-semantic fusion network successfully enhanced the modeling of body part interdependencies.

Conclusions:

• IMPRESS offers a significant advancement in human motion prediction, particularly in challenging conditions with missing data.
• The proposed method effectively recovers motion and models complex body part relationships.
• IMPRESS achieves state-of-the-art results, paving the way for more reliable human motion analysis in real-world applications.