3D Human Pose Estimation Based On An Advanced Knowledge Semantic Modeling Method

3D Human Pose Estimation(3D-HPE)is a fundamental and challenging task in the field of computer vision.It aims to predict human pose information,such as the spatial position of body joints and body shape parameters,from monocular images or videos.3D-HPE has been widely used for many computer vision tasks,such as person re-identification,human body parsing,human action recognition,and it is a common but challenging task in computer vision to investigate end-to-end deep representational learning for 3D human pose estimation.However,current methods still suffer from the problem of large errors between the recognised 3D human pose nodes and the actual joint positions.Based on this,this topic proposes corresponding solutions by analysing the problems of existing research methods,and achieves more satisfactory results.Specifically,the main research content and contributions of this paper are as follows:(1)To address the problem that there is no effective way to fuse different scale time series in 3D body pose estimation algorithms,and the problem that it is too expensive to learn the overall temporal pose information,this paper proposes a 3D human pose estimation model based on temporal multi-feature fusion,which will help to improve the semantic relationship between each local temporal pose feature state and the overall temporal pose feature state distribution,so as to better understand and analyse the data,by setting the length of the temporal sequence and the fusion scale parameters.Experiments were conducted on the Human3.6M dataset and the results showed that the 3D human pose estimation model based on temporal multi-feature fusion was effective for modelling 3D human pose regression and yielded results that outperformed the benchmark model of Gast-Net.(2)Existing methods only consider pairwise interactions of spatial relationships of human articulation points,ignoring much a priori knowledge of the human body,and thus suffer from a major problem:the articulation points of the 2D skeleton of the input data are heavily overlapping,rapidly changing and of various scales,so that the end articulation points may be misguided from the a priori knowledge of the human kinematic space and lead to an accumulation of errors.In order to address this problem,a new topology of the skeletal kinematic space is developed by using the set of joints at different distances as features,which can describe the complex kinematic states of the skeleton.Based on these designs,a human topology-aware network,HANet,is proposed,which is able to progressively optimise the human body at multiple high-level semantic levels by using an incremental optimisation strategy joint node and skeletal space interactions that contain rich semantic information about the human body,thereby improving the accuracy of 3D human pose node estimation.