Researchers at Liaoning Technical University have developed a semantic information fusion (SIF) network that improves the accuracy of 3D human pose estimation, particularly in scenarios involving self-occlusion and complex poses. The method addresses a key limitation in existing multi-hypothesis approaches: insufficient learning of dependencies between feasible pose solutions, which often leads to suboptimal fusion results.
Key takeaways
- The network comprises three modules: a hierarchical feature extraction (HFE) module that models joint structure and extracts multi-level semantic features; a feature refinement module (FRM) that enhances intra-hypothesis correlations and temporal dependencies; and a hierarchical feature fusion (HFF) module with an association calculation sub-module that learns inter-hypothesis dependencies for effective cross-hypothesis information transfer.
- A novel multi-head enhanced self-attention (MESA) mechanism adds summation and sigmoid operations to standard attention, improving joint position correlation within each hypothesis and aiding dependency learning.
- Validated on Human3.6M, MPI-INF-3DHP, and HumanEva-I datasets, the method demonstrates higher accuracy and robustness against self-occlusion and complex poses compared to existing techniques.
For industrial automation, this research offers a pathway to more reliable human motion capture in human-robot collaboration, gesture-based control, and safety monitoring, where accurate pose estimation under challenging conditions is critical.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-01-13 · “融合双层语义信息的3维人体姿态估计网络”