Study On Parametric Model Based 3D Human Pose Estimation

3D human pose estimation is widely used in the film industry,video game and human-computer interaction.However,presently commercialized solutions of 3D hu-man pose estimation rely on motion capture devices,which are expensive and restrict the exploration of further application for ordinary consumers.With the development of deep learning,many researchers begin to utilize convolutional neural networks to perform 3D human pose estimation.However,methods of deep learning require a large amount of data labeled with 3D ground truth joints which are tedious and difficult to obtain.To address the problems mentioned above,this thesis proposes an iterative optimization-based 3D human pose estimation method,which does not entail 3D joint labels.A parametric body shape model is incorporated as our representation of 3D pose,which enables synthesizing a detailed human body mesh by estimating low dimensional pa-rameters.The main contributions of this thesis can be summarized as follows:1)To address the issue that self-intersection often occurs in parametric body shape model based 3D human pose estimation,a self-intersection penalty term with linear time complexity is proposed.This paper proposes to detect self-intersection by travers-ing triangles only once,and then remove the self-intersection by deforming body mod-els appropriately.Compared with previous methods which approximate human body by simple geometries,this approach yields impressive high accuracy of self-intersection detection and prevention.2)A novel differentiable renderer with analytical derivatives is proposed to achieve higher accuracy and faster running speed in 3D human pose estimation.Previous dif-ferentiable renderers use numerical techniques to obtain the gradients.Nevertheless,numerically obtained gradients often cause low computational efficiency and accuracy.To resolve these issues,we introduce anti-aliasing to make rendering theoretically dif-ferentiable and facilitate the derivation of analytical gradients.The experimental results reveal that our proposed approach outperforms previous methods both in efficiency and accuracy.3)The proposed methods of self-intersection constraint and differentiable render-er are applied in 3D human pose estimation from mere silhouettes.The experimental results reveal that this approach yields appealing accuracy and high computational effi-ciency.