Cross-Domain Self-Supervised Complete Geometric Representation Learning for Real-Scanned Point Cloud Based...

Area of Science:

• Biomedical Engineering
• Computer Vision
• Rehabilitation Science

Background:

• Accurate lower-limb pose estimation is crucial for analyzing pathological gait.
• Single depth sensors offer potential for long-term monitoring in free-living environments.
• Existing methods struggle with partial geometric data from single viewpoints and require extensive ground truth data.

Purpose of the Study:

• To develop a novel cross-domain self-supervised framework for complete lower-limb geometric representation learning.
• To improve the accuracy of lower-limb pose estimation from single depth sensor data.
• To reduce the dependency on extensive ground truth data for training pose estimation models.

Main Methods:

• A cross-domain self-supervised learning framework was proposed.
• Knowledge transfer was utilized from unlabeled synthetic point clouds of full lower-limb surfaces.
• The method focused on learning complete geometric representations for accurate pose estimation.

Main Results:

• Significantly reduced the requirement for ground truth skeletons to only 1% during training.
• Achieved accurate and precise lower-limb pose estimation compared to existing methods.
• Successfully captured discriminative features for differentiating pathological gait patterns.

Conclusions:

• The proposed framework enables accurate lower-limb pose estimation from single depth sensors with minimal ground truth data.
• This approach facilitates more accessible and efficient pathological gait analysis for long-term monitoring.
• The method demonstrates superior performance in capturing subtle gait variations relevant to pathological conditions.