Markerless 3D pose estimation system for highly reflective aquatic environments using multi-view polarization imaging and dual-physics constrained network (DP-CRN)
The three-dimensional pose measurement of water sports holds significant value for competitive training and sports injury prevention. However, the coupling interference induced by water surface specular reflection and human wet body surface highlights substantially degrades the accuracy of conventional unlabeled methods relying solely on RGB imagery. This paper presents an markerless 3D pose estimation system that integrates multi-view polarization imaging with a Dual-Physics Constrained Network (DP-CRN). The system constructs a multi-view synchronous acquisition platform comprising six DoFP polarization cameras and establishes a coupled water-body reflection imaging model. A polarization decoupling algorithm for dynamic water surfaces is designed based on the Fresnel-Mueller matrix, enabling adaptive highlight suppression through frame-by-frame tracking of the time-varying water surface normal direction. The DP-CRN incorporates dual constraints from optics and biomechanics is constructed, integrating the polarization Fresnel equation, bone length conservation, and joint kinematic range constraints into the loss function. Additionally, a cross-angle polarization Stokes consistency loss is introduced to drive end-to-end 3D pose optimization at the physical level. On an experimental dataset encompassing four categories of water sports and approximately 450,000 frames, the proposed method achieves an average per-joint position error of 64.7 mm and a PCK@100 accuracy of 75.8%, representing an improvement of approximately 14%-17% over existing state-of-the-art approaches, while inter-frame jitter is reduced by 32%. Real-time inference at 108 fps is achieved on a single GPU. Ablation experiments demonstrate that water surface polarization decoupling constitutes the primary contributor to system accuracy, and the dual constraints of optics and biomechanics yield complementary gains in spatial accuracy and temporal stability. Future work will extend the proposed framework to fully uncontrolled open-water environments, more diverse camera configurations, and broader athlete populations, while incorporating underwater refraction correction and time-series diffusion models to improve robustness under full submersion and severe splash occlusion. Ablation experiments demonstrate that water surface polarization decoupling constitutes the primary contributor to system accuracy, and the dual constraints of optics and biomechanics yield complementary gains in spatial accuracy and temporal stability.