Dual-Stream STGCN with Motion-Aware Grouping for Rehabilitation Action Quality Assessment.

Area of Science:

• Biomechanics
• Rehabilitation Engineering
• Computer Vision

Background:

• Action quality assessment (AQA) is crucial for objective feedback in sports training and rehabilitation.
• Current AQA methods struggle to model synergistic joint relationships in functional groups.
• Existing fixed or fully adaptive joint grouping lacks clinical insight and flexibility.

Purpose of the Study:

• To propose a dynamic, motion-aware joint grouping strategy for improved action quality assessment.
• To enhance the modeling of collaborative joint movements in rehabilitation exercises.
• To develop a flexible and clinically informed approach to AQA.

Main Methods:

• A two-stream architecture processes joint position and orientation data independently.
• A joint motion energy-driven learnable mask generator adaptively clusters joints into 6 functional groups.
• Two-stage attention interaction models intra-group temporal dynamics and inter-group spatial relationships.

Main Results:

• The proposed method achieved competitive results, outperforming current methods on the KIMORE dataset by reducing mean absolute deviation by 26.5%.
• The model demonstrated comparable performance on the UI-PRMD dataset.
• Ablation studies confirmed the effectiveness of the dynamic grouping and two-stream architecture.

Conclusions:

• The dynamic, motion-aware grouping strategy significantly enhances action quality assessment accuracy in rehabilitation.
• The method's flexibility and reliance on clinical principles offer a robust solution for evaluating human movement.
• The core principles are extendable to other fine-grained action-understanding tasks like surgical assessment.