A monocular 3D human pose estimation method fusing part-aware and skeletal topology prior
By employing a dual-path temporal attention mechanism and skeletal topological constraints, human motion features are decoupled and anatomical constraints are explicitly applied, thereby improving the accuracy and reliability of monocular 3D human pose estimation and solving the problems of motion heterogeneity and insufficient anatomical constraints in existing technologies.
CN122156873APending Publication Date: 2026-06-05NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-05
AI Technical Summary
Technical Problem
Existing monocular 3D human pose estimation methods are inadequate in handling human motion heterogeneity and lack of anatomical structural constraints, resulting in low pose estimation accuracy and physical reliability.
Method used
A dual-path temporal attention mechanism is employed to decouple global and local motion features, and anatomical constraints are explicitly strengthened by applying skeletal topological constraints through a graph convolutional network, thereby improving the accuracy and rationality of pose estimation.
Benefits of technology
It achieves high-precision monocular 3D human pose estimation, improves the ability to model complex movements and the physical reliability of poses, and solves the problem of anatomical inconsistencies in existing methods.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN122156873A_ABST
Abstract
The application discloses a monocular three-dimensional human pose estimation method fusing part perception and bone topology prior, and belongs to the field of computer vision. The application aims to solve the problem that existing methods ignore human motion heterogeneity and lack explicit anatomical constraints. The method comprises the following steps: obtaining 2D joint coordinates of a monocular video sequence and generating initial features; extracting features by using a parallel double-path time attention mechanism, wherein a global path captures overall motion, and a local path independently models features divided into five semantic component groups, and a cross-part graph convolution network is used to capture limb coordination, and two features are fused through an adaptive gating unit; and finally, features are transformed to a bone domain through a bone structure refining module, a graph convolution network is used to explicitly correct bone length and connection relationship based on a topology structure, and a refined three-dimensional pose is generated. The application enhances the physical rationality and accuracy of the predicted pose.
Need to check novelty before this filing date? Find Prior Art