3D Human Pose Estimation Based On Knowledge Model

The research on 3D human pose estimation is more challenge in comparison with 2D human pose estimation,and is currently one of the most concerned research topics in the field of computer vision.Human pose estimation technology shows application prospects in the fields of virtual reality technology,sports motion analysis,medical rehabilitation training,autonomous driving,animation or photographic motion capture.Due to the lack of 3D information,direct regression of 3D human pose from 2D images is itself an ill-posed problem.Moreover,the current large-scale 3D human pose dataset used for model training is captured in experimental scenarios,which greatly limits the generalize ability of the model.In order to improve the generalization ability of the model,this study uses Bayesian probability models to characterize human posture,and introduces general knowledge to improve the performance of model posture estimation.This paper carried out the research on 3D human pose estimation in the following two aspects:(1)3D human posture estimation based on human body dynamics knowledge model.This study uses 3D human posture datasets to learn a human body dynamics knowledge model.Specifically,learn posture related human joint point relationships as the knowledge model,and employ it to assist the model learning,and finally improve the model accuracy and generalization performance.Based on human dynamics knowledge,a Bayesian Network is used to learn a human dynamics knowledge model Human3.6M(a 3D human pose annotation dataset in a laboratory environment).Then this knowledge model is used as prior knowledge in the learning of 3D human pose estimation model.The experimental results show that under the constraints of the human body dynamics knowledge model,the accuracy of 3D human pose estimation has been significantly improved,and the experimental improvement effect is significant.(2)3D human pose estimation based on constrained optimization theory.In this study,based on human body dynamics knowledge theory,constraints are optimized for the learned knowledge model of human body dynamics to make the knowledge representation model more consistent with the physiological structure of the human body,thereby effectively extracting features and estimating human posture more accurately.In the process of constructing a human posture knowledge model,constraints of human dynamics knowledge are added,that is,human joint constraint information is introduced as constraints for optimization,making the trained knowledge model more consistent with normal human posture in the real world,and finally significantly improving the model.The experimental results show that after adding joint constraints,the trained human knowledge representation model is more consistent with the human shape of the 3D world,and the accuracy of 3D human pose estimation is improved.