Meta-Transfer-Learning-Based Multimodal Human Pose Estimation for Lower Limbs

Area of Science:

• Robotics and Human-Computer Interaction
• Biomedical Engineering
• Machine Learning

Background:

• Accurate human pose estimation (HPE) is crucial for personalized interactive systems like cooperative robots and healthcare exoskeletons.
• Current methods require extensive datasets and frequent model updates, proving resource-intensive and time-consuming for individual adaptation.

Purpose of the Study:

• To develop a resource-efficient meta-transfer learning framework for accurate and stable human pose estimation (HPE).
• To enable rapid adaptation of HPE models to new individuals using minimal data through few-shot learning.

Main Methods:

• Integration of multimodal inputs: high-frequency surface electromyography (sEMG), visual-inertial odometry (VIO), and high-precision image data.
• A knowledge fusion strategy to enhance accuracy and stability by resolving data alignment issues.
• A few-shot learning approach for efficient real-time adaptation of encoders and decoders.

Main Results:

• The proposed framework achieves accurate and high-frequency human pose estimations, especially for intra-subject adaptation.
• Demonstrated efficient adaptation to new individuals with only a few samples.
• Successfully addressed data alignment issues through knowledge fusion.

Conclusions:

• The meta-transfer learning framework offers an effective solution for personalized motion analysis.
• Enables efficient, data-minimal adaptation for real-time applications in human-computer interaction and healthcare.
• Advances the field of human pose estimation by improving adaptability and reducing computational burden.