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

Human pose estimation in orbit is critical for astronaut health monitoring, task assistance, and intelligent human-robot interaction aboard space stations. However, in microgravity, human poses exhibit arbitrary orientations and are often affected by severe occlusion and complex background interference, while the scarcity of annotated in-orbit data makes it difficult to directly transfer models trained on ground-based datasets. Existing semi-supervised methods also lack explicit constraints from human structural topology and pose-related physical priors, which often leads to unreasonable pseudo-labels and limits performance gains. To address these issues, we propose a physics-inspired semi-supervised pose estimation framework for microgravity scenarios. Specifically, a Canonical Orientation Constraint is introduced to alleviate orientation ambiguity; a Structure-aware Pseudo-Label Refinement module is designed to improve pseudo-label quality; and an Uncertainty-guided Rotational Consistency Framework is proposed to adaptively weight consistency learning under multi-view rotation augmentation. Within a Mean Teacher architecture, the proposed method jointly optimizes the supervised loss, orientation constraint, pseudo-label refinement, and rotational consistency objectives. Experiments on the Astro-Pose dataset show that the proposed method consistently outperforms both fully supervised and semi-supervised baselines under various extreme poses and occlusion conditions, improving AP from 47.6 to 55.6 and AR from 52.4 to 60.1, demonstrating its potential for space-station visual monitoring.