3D Human Pose Estimation and Tracking Based on Feedback Mechanism
HUANG Qiancheng
CHEN Xianhua
WANG Lihui
TANG Jianhua
LIANG Zhiguang
Abstract：To effectively address the challenges of 3D human pose reconstruction in natural environments,this paper employs the MvPose matching mechanism to identify thesame targets across different views,and subsequently reconstruct multi-person 3D human poses.To achieve target tracking,an Extended Kalman Filter(EKF)is used,based on a variable-speed motion model,to predict the 3D human pose at the next moment.The predicted 3D human pose is then projected onto camera views,and the projected poses are associated with the targets identified by the matching mechanism.The Hungarian algorithm is used to obtain the optimal solution for target tracking.Furthermore,this paper proposes a method for reconstructing multi-person 3D human poses using feedback.Specifically,the pose tracker effectively fuses 2D human pose data to reconstruct 3D human poses.To improve the speed and accuracy of the 3D pose reconstruction,the epipolar RANSAC algorithm is used for robust estimation of matching point pairs to eliminate false matches.Subsequently,the final results are optimized using bundle adjustment.Experimental results show that while achieving 3D human pose tracking,the proposed feedback mechanism effectively reduces 3D pose reconstruction failures caused by multi-person interaction scenarios.
Keywords：3Dhuman pose estimationmulti-view geometrypose tracking
Publication Date：2025-08-30
Online Publishing Date：2025-09-17(First online date of this platform, not the publication date of the document)
Pages：8( 131-138 )