【作者】 宋园园；

【导师】 陈东岳；

【作者基本信息】 东北大学 ， 模式识别与智能系统， 2021， 硕士

【摘要】 三维人体结构化重构是指根据视觉传感信息为三维人体模型建立一种结构化参数模型,并根据参数估计结果与结构化人体模型进行人体三维重建的研究。三维人体结构化重构能够以少量的参数实现人体三维模型的快速重建,是模式识别与计算机视觉交叉领域的热点研究课题,在人机交互、特效制作、人体行为分析与理解、虚拟现实与增强现实等领域有着深刻的理论研究价值与广泛的应用前景。本文主要在上述背景下,研究基于单目静态图像的人体姿态与体型结构化参数的估计问题。具体研究工作继承了基于优化的三维人体重构技术的思路,并将其与基于回归的三维人体重构技术相结合,重点围绕三维人体网格模型头部姿态估计不准以及优化过程容易陷入局部极小值、训练模型对带有三维模型参数标签的人体数据集的强制依赖问题开展研究工作。1)针对三维人体网格模型头部姿态估计不准以及优化过程容易陷入局部极小值的问题,提出一种基于多模块融合的三维人体结构化重构方法。以往的算法中缺少对头部姿态的估计,容易造成人脸朝向出现偏差,重构出来的三维模型尤其是头部朝向与事实不符。针对这一问题,使用人脸面部关键点来调整头部姿态,在总的损失函数中加入头部投影的关键点误差函数,通过调整三维网格模型参数来改善头部朝向的问题。由于人体的复杂性,在算法的实际优化过程中存在许多极值点,容易导致最后的优化结果陷入局部极小值,造成模型重构的不准确,同时由于优化方法在时间效率上的劣势,本文通过基于回归的方法自适应地获取一个接近目标的合适初值以加速身体关节坐标的优化,在提高收敛速度的同时避免了优化过程中收敛到局部极小值。2)针对训练模型对带有三维模型参数标签的人体数据集的强制依赖问题,提出一种基于优化自监督的三维人体参数回归模型域适应方法。考虑到目前三维数据集标签的获取比较困难,本文将基于优化和基于回归的方法进行结合,将优化的结果作为回归的标签,以此来避免回归网络对三维标签数据集的强制依赖,同时回归网络的预测可以作为优化的初始值,二者是相辅相成,好的优化结果可以为回归网络提供更为精准的监督信号,使得网络预测出来的模型参数更为精准,从而为后面优化提供更接近目标的初始值。通过二者的融合不但可以解决模型对三维人体标签数据集的依赖问题,同时还可以提高模型的泛化能力,扩大应用背景。本研究将改进的优化与回归网络进行结合,为了评估算法在三维人体重构中的有效性,本文将与目前主流的人体重构算法进行对比。实验结果表明,本文算法的三维人体重构精度更高,在不同数据集甚至真实场景下的稳定性更好,说明本文的算法具有较高的研究意义和价值。更多还原

【Abstract】 The structural reconstruction of 3D human body refers to the research of establishing a structural parameter model for 3D human body model according to the visual sensing information,and carrying out 3D human body reconstruction according to the parameter estimation results and the structural human body model.3D structured body reconstruction to a small number of parameters to achieve the body rapid reconstruction of 3D model,is a hotspot in the field of pattern recognition and computer vision research topics,in human-computer interaction,special effects production,human behavior analysis and understanding,virtual reality and augmented reality has a profound theoretical research value and broad application prospects.Under the above background,this paper mainly studies the estimation of structural parameters of human body posture and body shape based on monocular static image.Inherited the specific research work based on the optimization of 3D reconstruction technology of body,with the regression based 3D reconstruction technology,the combination of body focus around the 3D grid model body head pose estimation are not allowed to be easy to fall into local minimum value,and optimize the process training model with three dimensional model parameters label of the human body data set mandatory dependence problem to carry out the research work.1)Aiming at the problem that the head pose estimation of the three-dimensional human body mesh model is inaccurate and the optimization process is easy to fall into the local minimum value,a three-dimensional human body structural reconstruction method based on multi-module fusion is proposed.Due to the lack of estimation of head posture in previous algorithms,the face orientation deviation is easy to occur,and the reconstructed 3D model,especially the head orientation,is inconsistent with the facts.To solve this problem,the key points of the face were used to adjust the head posture,and the error function of the key points of the head projection was added into the total loss function.Because of the complexity of the human body,in the process of practical optimization algorithm has many extreme value point,easy to cause the final optimization results in a local minimum value,causing inaccurate model reconstruction,at the same time because of the optimization method on the time efficiency of disadvantage,this paper based on the regression method is adaptive to obtain a suitable initial value close to the target to speed up the process of optimization,the body joint coordinates The convergence speed is increased and the convergence to local minimum is avoided in the optimization process.2)Aiming at the problem that the training model is forced to rely on the human body data set with the 3D model parameter label,a domain adaptation method of 3D human body parameter regression model based on optimization self-supervision is proposed.Given the current 3D data set of tags to get more difficult,this paper will be based on optimization and regression method based on combination of the optimized result as a return label,to avoid regression network rely on enforcement of 3D label data sets,and return to the network can be predicted as the initial value of optimization,the two complement each other,Good optimization results can provide more accurate monitoring signals for the regression network,and make the model parameters predicted by the network more accurate,so as to provide the initial value closer to the target for the subsequent optimization.The fusion of the two can not only solve the dependence problem of the model on the 3D human label data set,but also improve the generalization ability of the model and expand the application background.In this study,the improved optimization was combined with the regression network.In order to evaluate the effectiveness of the algorithm in 3D human reconstruction,this paper will compare it with the current mainstream human reconstruction algorithms.The experimental results show that the proposed algorithm has higher accuracy in 3D human body reconstruction and better stability in different data sets and even in real scenes,which indicates that the proposed algorithm has higher research significance and value.更多还原