PhysAstro-Pose: Physics-Inspired Semi-Supervised Human Pose Estimation in Microgravity Environments.

Area of Science:

• Computer Vision
• Robotics
• Aerospace Engineering

Background:

• Human pose estimation is vital for astronaut health and task assistance in space stations.
• Microgravity poses challenges like arbitrary orientations, occlusion, and limited in-orbit data.
• Existing methods struggle with unique microgravity conditions and lack physical constraints.

Purpose of the Study:

• To develop a physics-inspired semi-supervised framework for accurate human pose estimation in microgravity.
• To address orientation ambiguity, improve pseudo-label quality, and enhance consistency learning.
• To overcome limitations of ground-based models and existing semi-supervised approaches.

Main Methods:

• Proposed a physics-inspired semi-supervised pose estimation framework.
• Introduced a Canonical Orientation Constraint to resolve orientation ambiguity.
• Designed a Structure-aware Pseudo-Label Refinement module and an Uncertainty-guided Rotational Consistency Framework.
• Utilized a Mean Teacher architecture for joint optimization of multiple objectives.

Main Results:

• The proposed method significantly outperformed fully supervised and semi-supervised baselines on the Astro-Pose dataset.
• Achieved improvements in Average Precision (AP) from 47.6 to 55.6 and Average Recall (AR) from 52.4 to 60.1.
• Demonstrated robust performance under extreme poses and occlusion conditions.

Conclusions:

• The physics-inspired framework effectively addresses challenges in microgravity human pose estimation.
• The method shows significant potential for enhancing visual monitoring systems in space stations.
• Improved pose estimation accuracy supports astronaut health monitoring and human-robot interaction.