基于人体模型——SMPL的人体三维测量方法及应用

Human Body 3D Measurement Method and Application Based on Human Body Model—SMPL

摘要: 人体外廓非接触式测量是一种通过使用传感技术来测量人体的外廓参数或特征，而无需直接接触人体表面获得人体外廓数据的方法。但要得到较为精准的数据需要采用数学建模方法，传统如曲线拟合、三维重建、三角剖分等来创建人体的模型，其中三维重建是三维人体测量的主流方法，包括单视图RGB图像、多视图RGB图像、单视图深度图、多视图深度图重建等。受人体形态的影响，不同个体之间，三维重建后的模型点云个数和位置都不固定，不能自动定位到准确的测量点进行测量。而对于大规模人体数据测量，需要通过对人体模型进行标准化，即对模型进行编码来固定人体特征点的序列号，从而实现系统自动测量的目的。SMPL (Skinned Multi-Person Linear model)是一种用于建模人体形状和姿态的编码模型。其蒙皮表面通过形状参数进行对应的顶点映射，可以实现不同体态的表面编码。本文我们使用单视图深度图传感器获取深度图，利用DoubleFusion进行人体三维重建，并得到对应形态的SMPL顶点坐标(6890个)，进而根据SMPL顶点的三维坐标，可以对人体关键尺寸进行三维计算，得到人体的测量结果。我们的方法在提高模型的精度和逼真度方面取得了进展，包括增加细节以及与服装模型适配等。该模型适用于人体三维测量、虚拟试衣、三维动画、动作捕捉行业，在电子商务、时尚行业、虚拟和增强现实等领域具有广泛的应用前景。

Abstract: Non-contact measurement of the human body is a technique employed to assess various parame-ters or characteristics without direct physical contact, utilizing sensing technology. However, for enhanced data accuracy, the application of mathematical modeling methods, such as curve fitting, 3D reconstruction, and triangulation, is imperative. 3D reconstruction stands out as the primary approach in 3D anthropometry, encompassing techniques such as single-view RGB imaging, mul-ti-view RGB imaging, single-view depth mapping, and multi-view depth map reconstruction. The inherent variability in human body shapes poses a challenge: the number and position of model vertices following three-dimensional reconstruction differ across individuals, rendering automatic identification of precise measurement points unattainable. In the context of large-scale human body data measurement, it is necessary to standardize the human body model, that is, code the model to fix the serial number of human body feature points, so as to achieve the purpose of automatic measurement of the system. SMPL (Skinned Multi-Person Linear model) is a coding model for mod-eling human shape and posture. The surface of the skin is mapped to the corresponding vertex by the shape parameters, and the surface coding of different states can be realized. In this study, we employ a single-view depth map sensor to obtain the depth map and use DoubleFusion to recon-struct the human body in three dimensions, and get the corresponding SMPL vertex coordinates (6890). Leveraging the three-dimensional coordinates of SMPL vertices, key dimensions of the hu-man body can be calculated across three dimensions, leading to the acquisition of precise body measurements. Our method represents a significant advancement in enhancing model accuracy and fidelity, encompassing the incorporation of finer details and adaptation to clothing models. The resulting model proves well-suited for three-dimensional human body measurements, virtual fit-ting, applications in three-dimensional animation, and motion capture industries. Moreover, it holds substantial promise for diverse applications in the realms of e-commerce, the fashion indus-try, and virtual and augmented reality.