Modeling Static and Dynamic Joint Relationship for 3D Pose Estimation
In order to solve the problem that the existing three-dimensional(3D)pose estimation methods are difficult to accurately predict the position of key points of the human body in complex scenes,a 3D pose estimation method combining static and dynamic joint relationships was proposed.The purpose is to overcome the challenges of joint occlusion and posture singularity,and improve the accuracy of the model in complex scenes.The human joint map is obtained through mutual information calculation,which is used to guide the joint grouping and clustered based on the three-level degrees of freedom of the human structure.The static joint relationship model is constructed by using the cascade estimation and joint grouping feature sharing network,and the dynamic joint relationship model is designed by using the multi-group attention mechanism.In addition,a category balancing and pose reorganizing strategy is introduced to expand the variety of the data and to generalize the model's capability.The experiments demonstrates that the proposed model performs well on datasets such as Human3.6M,MPI-INF-3DHP and MPII.The results show that average error of key point positions for our model is reduced by at least 0.2 mm and the average accuracy is increased by at least 0.2％ when compared with currently available models,which can effectively improve the overall performance of 3D pose estimation,especially in complex scenes.
NETL
NSTL
万方数据
