【作者】 刘彬；

【导师】 刘秀平；

【作者基本信息】 大连理工大学 ， 计算数学， 2021， 博士

【摘要】 近十年来,随着虚拟试衣,个性化服装定制,身体健康监控系统等应用需求的快速增长,简单、高效的三维人体数字化表示技术受到科研人员越来越多的关注。传统的基于大型激光扫描仪的三维人体重建技术为三维人体数字化表示提供了有力的数据保障,方兴未艾的机器学习技术特别是深度学习技术则为三维人体数字化表示提供了新的理论和算法工具。大致来说,三维人体数字化表示涉及以下内容的研究与应用:数据预处理、人体形状表示空间以及高效的三维人体重建。数据预处理包含对扫描得到的三维物体进行噪声过滤以及姿势归一化等技术,旨在获得一个高质量的易于进行形状统计建模的数据集;人体形状表示空间是指对三维人体数据集进行统计建模得到其压缩形式的参数化表示模型,方便数据存储、传输和处理;高效的三维人体重建是指在人体形状表示空间协助下,预测和分析图片或者视频中的人体形状,可应用于虚拟试衣或者个性化的服装设计。为了给三维人体数字化表示提供新的理论依据和技术方法,本论文对上述几个方面分别做了调研,主要贡献如下:(1)针对三维扫描设备在物体重建过程中遭受到的噪声影响,本文提出了一种传播的网格法向滤波算法。该算法通过深入剖析双边滤波的基本原理,将滤波算子中权重的设计重新定义在网格表面自身的流形结构中,从而能够在物体表面沿着最短路径的方向利用累积的法向差异去计算高斯核函数中的距离度量。总的来说,该算法弥补了网格滤波原理中的漏洞,同时也增强了网格滤波算法的可解释性,实现了在过滤网格表面噪声的同时能够有效地维持住网格原有细节的目的。(2)针对规范人体姿势的问题,本文介绍了一种适用于非刚性物体的姿势归一化方法。该方法基于形状变形框架,能够促使网格上所有不相邻的点对在变形的过程中远离,同时又可以将局部变形约束在保刚性能量的正则化策略中,有效地保护了三维物体的局部细节,极大降低了自身的几何畸变,从而实现将姿势不同的非刚性物体展开到规范的姿势下。提出的模型在交替迭代的算法优化框架下,模型的求解被转化为解线性系统,简单且高效。此外,为了进一步增强算法的计算有效性和鲁棒性,本文分别在迭代的内部实施了并行策略以及在迭代的外部使用了由粗到细的级联策略。(3)针对三维人体形状表示空间问题,本文从稀疏表示的角度出发建立了一种新的人体形状表示统计模型。首先将人体网格分割成大量的可供稀疏表示学习的训练信号,然后使用在线L0算法学习出用于重建原始人体数据集中所有训练信号的字典矩阵和稀疏表示系数,最后利用重构出的信号组装出原始的三维人体。相较于传统的基于主成分分析(PCA)的统计模型,本文提出的算法具有更高的压缩率、更低的形状近似误差,能够降低三维人体数据存储和传输压力。重要的是,从局部训练信号中学习出的字典具备一定的通用性,可以重建出与原始人体模型拓扑不一样的其他形状。(4)针对高效的三维人体重建问题,本文试图从两个正交的人体轮廓(即正面和侧面视图)重建三维人体形状。通过使用监督学习的方式,本文设计的卷积神经网络(CNN)架构不仅可以自动提取正视图和侧视图的判别特征,还可以通过融合手段精确提取他们之间的混合特征,并使用全连接的神经网络建立混合特征和人体形状表示空间系数之间的映射函数。最后利用学习出的映射函数和基于PCA或稀疏表示的人体形状统计模型,重建出高精度的三维人体形状。除此之外,本文提出的技术可被应用于自动化服装设计领域中。更多还原

【Abstract】 In recent years,with the growth of demand in applications such as virtual try-on,customized design and body health monitoring,simple and effective 3D human body modeling techniques have caught more and more attention.Traditional 3D human body reconstruction techniques based on large laser scanner provides a great deal of body data for 3D human body digital representation,the machine learning technologies in the ascending stage,especially deep learning,provide a new theoretical and algorithmic tool for 3D human body digital representation.Generally speaking,3D human digital representation mainly involves the following aspects:data preprocessing,human shape representation space and efficient 3D human body reconstruction.Data preprocessing refers to techniques such as noise filtering and posture standardization of the scanned 3D objects in order to obtain a high-quality and easy to use human body data set;The human body shape representation space refers to the statistical modeling of the 3D human body data to obtain the parameterized representation model in its compressed form,which is convenient for data storage,transmission and processing;Efficient 3D human body reconstruction refers to the prediction and analysis of the human body shape in pictures or videos with the assistance of the human body shape representation space,which is applied to virtual try-on or personalized clothing customization.In order to provide new theoretical basis and technique methods for 3D human body digital representation,this paper has conducted investigations on the above aspects and the main contributions are as follows:(1)Aiming at the problem of mesh denoising in the process of object reconstruction based on 3D scanner,a propagated mesh normal filtering algorithm is proposed.By examining the basic principle of bilateral filtering,this algorithm redefines the weight design of the filtering operator in the manifold structure of the mesh surface,so that it can use the accumulated normal difference to calculate the distance measure in the Gaussian kernel function along the shortest path on the mesh surface.In general,the algorithm supplements the loopholes in the principle of mesh filtering,and enhances the interpretability of mesh surface filtering algorithm,which can effectively maintain the original details of the mesh while filtering the surface noise.(2)Aiming at the problem of standardizing human posture,this paper aims to provide a posture standardization technique suitable for all non-rigid objects.This method is based on the shape deformation framework in digital geometry,and can unfold any non-rigid 3D object into a standard posture.The proposed model can promote all non-adjacent vertex pairs in the mesh to move away during the deformation process,and the local deformation is limited to the regularization strategy of preserving rigidity energy,which effectively protects the local details of 3D object and greatly reduces its own geometric distortion.Under the framework of alternating iterative algorithm optimization,the solution of this model only needs to solve the linear equations,which is simple and efficient.Implementing the parallel strategy within the iterative process further increases the computational effectiveness of the algorithm,and adopting the cascade strategy ensures its robustness.(3)Aiming at the problem of 3D human shape representation space,this paper establishes a new statistical model of human shape representation from the perspective of sparse representation.Compared with the traditional statistical model based on principal component analysis(PCA),our method has a higher compression rate,lower shape approximation error,and reduces the pressure of storing and transmitting 3D human body data.First,the human body mesh is divided into a large number of training signals for sparse representation learning.Then use the online L0 algorithm to learn the dictionary matrix and sparse representation coefficients that can reconstruct all the training signals in the original human body data set.Finally,the reconstructed signal is used to assemble the original three-dimensional human body.Further experiments found that the learned dictionary matrix has approximate generalization and can reconstruct other surface that are different from the original human body mesh.(4)Aiming at the efficient 3D human body reconstruction,this paper attempts to reconstruct the 3D human body shape from two orthogonal human body silhouettes(ie,front and side views).By using supervised learning,the convolutional neural network(CNN)architecture proposed in this paper can not only automatically extract the discriminative features from the front and side view,but also accurately extract the mixed features between them through fusion module,and uses the fully connected layer to establish the mapping function between the mixed feature and the human body shape representation space coefficient.Then using the learned mapping function and the human body shape statistical model,such as the shape statistical model based on PCA or sparse representation,to reconstruct a high-precision 3D human body mesh.Finally,this technology is applied to the field of automated clothing customization.更多还原