.png){kind=icon}
.png){kind=icon}
Return
Advancing human pose estimation through interdisciplinary physics-inspired deep learning models
.png){kind=icon}
PRE
AI
DOI：10.3389/fphy.2025.1558325.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Introduction Human pose estimation is a critical challenge in computer vision, with significant implications for robotics, augmented reality, and biomedical research. Current advancements in pose estimation face persistent obstacles, including occlusion, ambiguous spatial arrangements, and limited adaptability to diverse environments. Despite progress in deep learning, existing methods often struggle with integrating geometric priors and maintaining consistent performance across challenging datasets.Methods Addressing these gaps, we propose a novel framework that synergizes physics-inspired reasoning with deep learning. Our Spatially-Aware Pose Estimation Network (SAPENet) integrates principles of energy minimization to enforce geometric plausibility and spatiotemporal dynamics to maintain consistency across sequential frames. The framework leverages spatial attention mechanisms, multi-scale supervision, and structural priors to enhance feature representation and enforce physical constraints during training and inference. This is further augmented by the Pose Consistency_Aware Optimization Strategy (PCAOS), which incorporates adaptive confidence reweighting and multi-view consistency to mitigate domain-specific challenges like occlusion and articulated motion.Results and discussion Our experiments demonstrate that this interdisciplinary approach significantly improves pose estimation accuracy and robustness across standard benchmarks, achieving state-of-the-art results. The seamless integration of spatial reasoning and domain-informed physical priors establishes our methodology as a transformative advancement in the field of pose estimation.
Keywords:
pose estimation
spatial attention
structural priors
multi-scale supervision
adaptive optimization
Journal
F
IF:
2.1
Papers:
243
Citations:
0