AstroHSP: A hybrid supervision framework for robust monocular astronaut pose estimation.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Biomedical Engineering

Background:

• Accurate 3D human pose estimation is crucial for astronaut safety and ergonomics.
• Existing methods struggle with severe deformation, occlusion, and domain shift, especially without 3D ground truth.

Purpose of the Study:

• To develop a robust framework for accurate 3D human pose estimation in challenging, data-scarce environments.
• To address limitations in current pose estimation techniques for astronaut applications.

Main Methods:

• Proposed AstroHSP, a two-stage framework combining domain-adaptive 2D pose estimation and uncertainty-aware 3D pose lifting.
• Utilized mixed-domain training with Domain-Adaptive Batch Normalization for stable 2D predictions.
• Employed a dual-stream Transformer and conditional diffusion model for 3D pose refinement and uncertainty mitigation.
• Introduced a two-stage hybrid training strategy for fine-tuning without 3D ground truth.

Main Results:

• AstroHSP demonstrated superior performance across multiple datasets, including specialized scenarios.
• The framework achieved robust 3D pose generalization in highly constrained domains.
• Validated efficacy on the new AstPose dataset collected from real space missions.

Conclusions:

• AstroHSP provides a robust solution for 3D human pose estimation in challenging, real-world scenarios.
• The proposed methods effectively handle domain shift, occlusion, and lack of 3D annotations.
• This work advances astronaut safety and ergonomic analysis through improved pose estimation.