3D Human Pose Construction Algorithm Based on 2D Node to 3D Model

This paper proposes an efficient method for converting 2D video images into 3D human poses based on deep learning and 3D model fitting. This algorithm extracts the key points by using a 2D pose estimation model and enhances the reliability of the estimation by using a confidence evaluation mechanism. For multi-human scenarios, this algorithm adopts the key point clustering technique to achieve the association of the individuals. The 2D key points are then mapped into 3D space using the SMPL model. The reconstruction accuracy is optimised using a deviation function. Finally, the 3D skeleton is normalised to improve the consistency of the results. The experimental results show that the method significantly improves the conversion accuracy and robustness with some practicality, and is able to accurately capture the character's movements while quickly constructing the 3D skeleton.