Multi-View 3D Human Pose Estimation for Yoga and Exercise Posture Analysis
Year: 2026
Type: article
Abstract: Automated computer vision systems for yoga and physical exercise monitoring offer significant promise for remote healthcare, athletic training, and home fitness. While single-camera deep learning frameworks have enabled lightweight 2D landmark detection on commodity edge devices, their application to complex exercise evaluation remains fundamentally con-strained by visual self-occlusion, out-of-plane depth ambiguity, and a lack of biomechanical rigor. Multi-view 3D human pose estimation (HPE) provides the geometric foundation necessary to resolve these structural ambiguities. This survey provides a comprehensive analysis of the evolution of multi-view 3D HPE methodologies, tracing developments from classical algebraic and volumetric triangulation to cross-view attention transformers and state-space sequence modeling. We critically examine the persistent disconnect between laboratory 3D joint coordinate regression and actionable exercise feedback, investigate domain-specific challenges including non-upright contortions and dataset scarcity, and evaluate browser-side dual-stream processing via FitAlign, a heuristic two-stream posture-evaluation prototype. Finally, we establish a structured taxonomy of open research directions bridging geometric 3D vision, personalized biomechanics, and edge-native synchronization for accessible and reliable exercise monitoring.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7225436595