Novel method refines kinematic human pose estimation using joint angles
A new approach to refining marker-free human pose estimation (HPE) has been proposed, addressing persistent errors in keypoint recognition and trajectory fluctuations. The method introduces a robust joint angle-based description of kinematic poses, approximates temporal variation of joint angles with high-order Fourier series to generate reliable ground truth, and employs a bidirectional recurrent network as a post-processing module. Trained on a high-quality dataset constructed via this technique, the network improves the accuracy of single-image HPE models. The research is detailed in arXiv:2507.11075.
Key facts
- arXiv:2507.11075
- Announce Type: replace-cross
- Marker-free human pose estimation (HPE) has increasing applications
- Current HPE suffers from keypoint recognition errors and trajectory fluctuations
- Existing deep learning models limited by inaccurate training datasets
- Proposed method uses joint angle-based description
- Temporal variation of joint angles approximated using high-order Fourier series
- Bidirectional recurrent network designed as post-processing module
Entities
Institutions
- arXiv