A parameterized human body model reconstruction method based on a Chinese database

Abstract

The application discloses a parameterized human body model reconstruction method based on a Chinese database, three-dimensional human body grid data of multiple female naked bodies of different ages, body types and heights are collected to establish a database; point clouds are learned based on BPS, SMPL model grid points are respectively roughly registered to the vicinity of each three-dimensional human body grid point to generate corresponding rough registration models; then, a non-rigid grid registration algorithm based on an ICP algorithm is used to finely register each rough registration model to generate corresponding SMPLD models; each SMPLD model is corrected in a standard standing posture, and then PCA principal component analysis is performed on each SMPLD model after the standing posture correction, a plurality of principal component proportions obtained are used as body parameters of a parameterized human body model adapted to Chinese body shapes, original posture parameters of the SMPL model are used as posture parameters of the parameterized human body model adapted to Chinese body shapes, and a final parameterized human body model is constructed.

Description

Technical Field

[0001] This invention belongs to the technical field of image processing, specifically relating to a parametric human body model reconstruction method based on a Chinese database. Background Technology

[0002] The internet has brought convenience to people's lives, and online clothing shopping has become one of the main ways people consume in their daily lives. However, the inherent characteristics of clothing make it impossible to accurately describe its features using sizing, text, or photos, often resulting in ill-fitting clothes and a mismatch between the appearance and expectations. With the development of internet-based virtual reality technology, 3D virtual try-on technology can use computer vision, graphics, and related software to allow consumers to preview the realistic effect of trying on clothing, reducing their shopping time and improving fit and satisfaction. 3D human body modeling has always been a challenging and important part of the 3D virtual try-on field, essentially involving storing and representing the real human body in a 3D digital format on a computer.

[0003] Currently, computer vision-based methods can quickly and easily acquire high-quality 3D human body models. Based on a review of relevant literature, these methods can be categorized according to two criteria: The first category is unconstrained 3D human body modeling methods. These methods first acquire or estimate depth information using various means, and then reconstruct the complete geometric shape and texture information of the human body using point cloud registration. For example, some researchers use multi-view camera arrays for human body modeling. This single-moment multi-view 3D human body modeling method yields high-quality results and supports arbitrary clothing topology, but its real-time performance is poor, and the camera calibration process is complex. The second category is 3D human body reconstruction methods based on parametric (statistical) models. These methods mostly first estimate various parameters of human body models such as SCAPE and SMPL, and then supplement the model's texture details by fusing depth information from various sensors. Compared to the first category, this type of method has lower environmental requirements, and the reconstructed model has an animated skeleton and contains semantic information. For example, Cheng et al. investigated several mainstream parametric human body models, and their research showed that the SMPL model has better performance than SCAPE in terms of both speed and accuracy. Therefore, the SMPL model has gradually replaced the SCAPE model as the mainstream model. Because the model can represent human bodies of different shapes and poses with fewer dimensional parameters, and the linear functions used in the model are easy to optimize, many teams have proposed using the model for parametric human body modeling. These methods utilize the rich prior information on human body structure contained in the SMPL model as constraints to learn a mapping function from two-dimensional images to three-dimensional poses, thereby achieving model-data matching.

[0004] However, most of these parametric human body models are developed based on the fixed body shape of Westerners. In fact, there are certain differences between Eastern and Western human bodies: for example, in 2015, Yin Yan et al. compared data from a 2009 pilot survey of human body dimensions in China with data from the CAESAR survey of the United States completed in 2002, finding that Westerners are taller and stronger than Easterners, with longer arms, legs, and hands and feet; in terms of height, girth, and width, Westerners are larger than Easterners in anthropometry; however, the difference in head and neck length and upper body length is not significant. In 2020, Hu Xinrong et al., based on the proportional relationship between key structural parts of the human body and height provided by the 2019 European and American CAESAR human body dataset and Asian human body data obtained from the 2009 pilot survey of adult human body dimensions in China completed by the China National Institute of Standardization, found that the parameter values ​​of key structural parts of the Eastern human body, particularly the chest and hips, differed significantly. Due to these differences between Eastern and Western human bodies, parametric human body modeling algorithms cannot accurately estimate model parameters, making it difficult to guarantee the accuracy and continuity of 3D human body modeling results. Summary of the Invention

[0005] This invention provides a parametric human body model reconstruction method based on a Chinese database, which solves the technical problems in the prior art, such as the inability of parametric human body modeling algorithms to accurately estimate model parameters due to differences between Chinese and Western human bodies, making it difficult to guarantee the accuracy and continuity of 3D human body modeling results.

[0006] This invention can be achieved through the following technical solutions:

[0007] A parametric human body model reconstruction method based on a Chinese database includes the following steps:

[0008] I. Establishing a Database

[0009] Collect 3D human body mesh data of multiple female bodies of different ages, body types, and heights, and establish a database;

[0010] 2.3D Human Body Mesh Registration

[0011] First, the point cloud is efficiently learned based on BPS Basis Point Sets, and the SMPL model mesh points are quickly initially registered to the vicinity of each 3D human body mesh point in the database to generate the corresponding initial registration model. Then, a non-rigid mesh registration algorithm based on ICP is used to fine register each initial registration model, only optimizing the vertex displacement D component to generate the SMPLD model corresponding to the accurate registration.

[0012] III. Statistical Shape Analysis

[0013] First, the standing posture of each SMPLD model is corrected to unify it into a standard standing posture. Then, PCA principal component analysis is performed on each SMPLD model after standing posture correction. The proportions of multiple principal components obtained by PCA principal component analysis are used as the body parameters of the parametric human body model adapted to the Chinese body shape. The original posture parameters of the SMPL model are used as the posture parameters of the parametric human body model adapted to the Chinese body shape, and the final parametric human body model adapted to the Chinese body shape is constructed.

[0014] Furthermore, the fine registration method includes setting M(v) i ,x): This indicates that the standard human body surface 3D mesh vertices Mapping to model parameters The deformed 3D human body model is represented by the SMPLD model as x = {θ, β, D}, where θ is the pose parameter, β is the shape parameter, and D is the non-rigid deformation parameter, representing the offset of each vertex in the SMPL model reconstructed from {θ, β}.

[0015] Construct the point set in the SMPL model after reconstruction by x and 3D human body scan point set Loss function E:

[0016]

[0017] in, N represents the mesh surface of the SMPL model. v =6890, N f =13776, T v The number of grid points to be matched. This represents the initial mesh surface after initial registration by BPS, Lap represents the Laplacian operator, and dist(·) is the distance from a differentiable point to the surface. For continuous surface points, M′(·) represents the vertices of a discrete model. The defined model function M(·) is used for barycentric interpolation, d i The offsets of each vertex in the reconstructed SMPL model are given by α, β, and γ, which are dynamic weighting coefficients.

[0018] The ICP algorithm is used to search for the optimal loss function. In each iteration, the corresponding point pairs of the collected 3D human body model-SMPLD model are first calculated, and then the model parameters are updated using gradient or Gauss-Newton optimizers to minimize the distance between the scanned point and the corresponding model point.

[0019] Furthermore, pose correction includes the following steps:

[0020] 1) Using the skeletal skinning model in the SMPL model, three-dimensional joint reconstruction is performed on each registered SMPLD model to be corrected. 3d ;

[0021] 2) Select standard standing posture samples from the human body data database and calculate their J values. 3d And serve as a standard standing posture template {θ T};

[0022] 3) J for the SMPLD model to be corrected 3d Using a skeleton mapping algorithm, the relative position of the i-th joint to the standard standing posture template {θ} is calculated. T The difference in rotation angles of} and the corresponding rotation matrix Then, based on the LBS algorithm and the hybrid skin weight W, the SMPLD model after attitude transformation is calculated.

[0023] Furthermore, the final expression for the parametric human body model adapted to the Chinese human form is as follows:

[0024]

[0025] Among them, V i F represents the average value of each vertex in the grid, i.e., the value of each vertex in the corrected SMPLD model. i Triangular mesh, JM 6890×24 W is the weighted coefficient matrix of the 24 skeleton points in the SMPL model. 6890×24 This is the weighting coefficient matrix for bone and skin deformation. There are 24 joint rotation coefficients. These are global displacement parameters, i.e., attitude parameters. For scaling parameters, PC i These are the shape parameters, i.e., the reconstruction coefficients obtained from PCA principal component analysis.

[0026] The beneficial technical effects of this invention are as follows:

[0027] 1) Since the SMPL model is a parametric model generated based on Western human body data, its shape has certain compatibility issues with Chinese human body shapes. Therefore, we consider using Chinese human bodies as the research object, using a 3D human body scanner to scan the subjects, thereby obtaining a certain number of 3D human body models, constructing a multi-shape dataset that conforms to the characteristics of Chinese human bodies, and establishing a parametric 3D human body model that can reflect the characteristics of Chinese human bodies through PCA.

[0028] 2) A two-step registration strategy of "coarse-fine" is proposed: First, the point cloud is efficiently learned based on BPS to quickly and initially configure the SMPL model grid points near the scanned grid points, achieving coarse registration; then, a non-rigid grid registration algorithm based on ICP is used for fine registration, optimizing only the vertex displacement D component to generate an accurately matched SMPLD model. Furthermore, the deep learning-based BPS coarse registration used in this invention ensures relatively high matching accuracy while averaging only 0.5 seconds per sample. Meanwhile, the fine registration process based on pyTorch 1.17.0+cu11.0 takes only about 50 seconds per sample on average. Therefore, the "coarse-fine" two-step registration method based on BPS coarse registration and non-rigid template fitting not only ensures the registration accuracy of the nonlinear model but also greatly improves the registration efficiency.

[0029] 3) By comparing the parameterized human body model constructed using the reconstruction method of the present invention with the SMPL average template, it can be seen that the parameterized model based on the Chinese dataset has smaller fitting error and higher accuracy. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0031] Figure 2 This is a schematic diagram showing the age, height, and body type distribution of the samples collected in this invention;

[0032] Figure 3 This is a schematic diagram of the screening Poisson surface reconstruction and its downsampling results according to the present invention;

[0033] Figure 4 A schematic diagram of the SMPL model using the LoopReg algorithm based on deep learning for registration;

[0034] Figure 5 This is a schematic diagram showing the registration result achieved using the BPS coarse registration method of the present invention;

[0035] Figure 6 This is a schematic diagram showing the registration result achieved using the fine registration method of the present invention;

[0036] Figure 7 This is the energy iteration curve diagram for the fine registration of the present invention;

[0037] Figure 8 This is a schematic diagram of the average matching error of grid points for coarse and fine registration of 152 samples in this invention.

[0038] Figure 9 This is a standard standing posture sample and its skeleton diagram for the present invention;

[0039] Figure 10This is a schematic diagram comparing the pose correction results before and after the present invention;

[0040] Figure 11 This is a visualization of the first 10 principal component analyses of the PCA principal component analysis results of this invention;

[0041] Figure 12 A schematic diagram comparing the existing SMPL average model and the model constructed using the reconstruction method of this invention;

[0042] Figure 13 This diagram illustrates a comparison of the reconstruction results of the parametric model based on the Chinese dataset and the SMPL model of the present invention. Detailed Implementation

[0043] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0044] In order to construct a parametric human body model that can adapt to the physical characteristics of Chinese people, such as Figure 1 As shown, this invention first collected 152 body samples from adult women in East China, then removed braids, performed Poisson reconstruction, downsampling, and foot plane cutting. Considering both registration accuracy and efficiency, this invention proposes a two-step "coarse-fine" registration strategy. First, it efficiently learns the point cloud based on BPS, quickly initializing the SMPL model mesh points near the scanned mesh points. Then, it uses a non-rigid mesh registration algorithm based on ICP (Iterative Closest Points) for fine registration, optimizing only the vertex displacement D component to generate a precisely matched SMPLD model. After registration, pose correction and principal component analysis were performed on the registered data to obtain PCs with Chinese anatomy characteristics. i The shape parameters are then determined, and finally, the model is reconstructed. Details are as follows:

[0045] The SMPL model is a linear model based on vertex skinning. It is a parametric human deformation model generated from a large amount of human scan data through statistical learning. This model can represent different human postures and body shapes by changing parameters, and can simulate the protrusions and depressions of human muscles and other tissues during limb movement, thereby realizing human modeling in any posture.

[0046] The SMPL model takes shape parameter β and pose parameter θ as input and outputs a human body model with N=6890 vertices and K=23 joints, and a 3D human skeleton with parent-child structure. The pose parameter θ represents 75 vectors of the overall human body pose and the relative angles of 24 joints. The shape parameter β represents 10 vectors of human height, weight, head-to-body ratio, etc., each vector representing a shape basis, which is obtained by performing PCA on the 3D human body dataset.

[0047] I. Establishing a Database

[0048] Three-dimensional human body mesh data of multiple female bodies of different ages, body types, and heights were collected to establish a database.

[0049] This invention collected lightweight body samples from 152 adult women in East China. The age, height, and body type distribution of the samples are as follows: Figure 2 As shown.

[0050] During data collection, subjects were required to stand on designated footprint markers, looking straight ahead, with their arms naturally extended outwards from their sides. Filtering, as the first step in point cloud processing, often significantly impacts subsequent processing. Therefore, only by customizing noise points, outliers, etc., during filtering preprocessing according to subsequent processing requirements can registration, feature extraction, surface reconstruction, and visualization be performed more effectively. Considering that braids (hair) in some individuals' samples might affect the accuracy of subsequent model fitting, the Meshlab open-source software can be used for pre-processing manual trimming to remove braids. Furthermore, due to inherent limitations of line laser 3D scanning equipment, such as occlusion and quasi-parallelism to the emission light plane, various holes of different sizes exist in the human body mesh. These can be pre-reconstructed and repaired using a shielded Poisson reconstruction algorithm. Simultaneously, 3D point clouds often contain a large amount of redundant data. Balancing the accuracy and speed of subsequent model matching, a downsampling method using quadratic edge folding extraction was employed to reduce the number of faces in the collected mesh to 20K, resulting in an average original point cloud mesh count of over 600K.

[0051] Screening of Poisson surface reconstruction and its downsampling results as follows Figure 3 As shown, in view of the distortion of the foot plane after the curved surface is reconstructed, the present invention also performs cutting processing on the floor plane.

[0052] 2.3D Human Body Mesh Registration

[0053] Algorithms based on non-rigid template registration typically use shape, pose, and vertex displacement as variables to construct an objective function for optimizing statistical parameters between the scanned mesh and the deformable template. Then, nonlinear optimization is used to obtain the optimal solution. However, this method requires the initial position of the deformable template to be sufficiently close to the scanned mesh; otherwise, due to significant pose and shape errors, the nonlinear search may get trapped in a local minimum. Therefore, considering both registration accuracy and efficiency, this invention proposes a two-step "coarse-fine" registration strategy. First, based on BPS Basis Point Sets, the point cloud is efficiently learned, and the SMPL model mesh points are quickly and initially registered to the vicinity of each 3D human body mesh point in the database, generating the corresponding initial registration model, i.e., the coarse registration model. Then, a non-rigid mesh registration algorithm based on the ICP algorithm is used to finely register each initial registration model, optimizing only the vertex displacement D component to generate the SMPLD model corresponding to the precise registration. The fine registration process is as follows:

[0054] Let M(v) i ,x): This indicates that the standard human body surface 3D mesh vertices A human body model mapped to 3D points deformed according to model parameters x∈χ′. This represents the set of vertices in a standard human body mesh. During registration, it refers to the set of vertices in the scanned 3D human body mesh data. Let the SMPLD model be represented as x = {θ, β, D}, where θ is the pose parameter, β is the shape parameter, D is the non-rigid deformation parameter, and D represents the offset of each vertex in the SMPL model reconstructed from {θ, β}.

[0055] The registration method is to find a series of corresponding point sets in the SMPL model reconstructed from x. and 3D human body scan point set Find the loss function E and minimize it:

[0056]

[0057] in, N represents the mesh surface of the SMPL model. v =6890, N f =13776, T v The number of grid points to be matched. This represents the initial mesh surface after initial registration by BPS, Lap represents the Laplacian operator, and dist(·) is the distance from a differentiable point to the surface. For continuous surface points, M′(·) represents the vertices of a discrete model. The defined model function M(·) is used for barycentric interpolation, d iThe offsets of each vertex in the reconstructed SMPL model are given by α, β, and γ, which are dynamic weighting coefficients.

[0058] The ICP algorithm is used to perform an optimal search for the above loss function to minimize it. In each iteration, the corresponding point pairs of the acquired 3D human body model-SMPLD model are first calculated, and then the model parameters are updated using gradient or Gauss-Newton optimizers to minimize the distance between the scanned point and the corresponding model point.

[0059] To verify the feasibility of the "coarse-fine" two-step registration method of the present invention, the LoopReg algorithm based on deep learning was used to predict and fit the SMPL model to the above-mentioned three-dimensional human body scan data, as follows: Figure 4 As shown, the left column is the fused image of the original mesh (gray) and the coarse matching mesh (black); the second column is a side view of the first column; the third column is the coarse registration triangular mesh image; and the right column is the registration error image, where the mesh point colors correspond to the registration errors. The error color table is shown in the top image. The average coarse registration errors of the samples in the image are 2.08 cm and 1.88 cm, respectively. The mesh output of the BPS coarse registration method of this invention for registering 3D human body scan data and SMPL models is shown below. Figure 5 As shown, the average error of vertex matching of the sample grid in the figure is 0.91 and 0.79 cm, which is more than twice the average accuracy compared with the LoopReg algorithm.

[0060] This implementation uses PyTorch (1.7.1) + CU110 development kit on the Python platform to achieve non-rigid template fine registration. The non-linear optimization adopts the Adam method with learning rate lr = 0.005 and betas = (0.9, 0.999). Figure 7 for Figure 6 The energy iteration curve of fine registration is shown in the figure. Series 1 corresponds to the fitting results of the first row of samples, and Series 2 corresponds to the fitting results of the second row of samples. As can be seen from the figure, the objective function basically converges after 30 iterations. Figure 6 This is a schematic diagram of the fine registration results. The average matching error of the fine registration for the first and second rows of samples is 0.56 cm and 0.54 cm, respectively, which is nearly 50% lower than the average matching error of the coarse registration.

[0061] The average matching error of grid points for coarse and fine registration of 152 samples is as follows: Figure 8 As shown, the maximum, minimum, and average values ​​of the coarse registration average matching error based on BPS are 0.0115, 0.0075, and 0.0090m, respectively, while the corresponding values ​​for the fine registration based on non-rigid model matching are 0.0060, 0.0048, and 0.0054m.

[0062] The hardware platform configuration for the "coarse-fine" two-step registration code is: Intel i7-9750H, 16GB RAM, Win64 system, and NVIDIA GeForce RTX 2070 with Max-Q Design GPU with 8GB RAM. All critical code runs on the GPU. On the same platform, traditional non-rigid matching after 3D human body scanning takes approximately 15 minutes per sample (including pose pre-registration and nonlinear iteration). Deep learning-based BPS pre-registration, while maintaining relatively high matching accuracy, takes only 0.5 seconds per sample on average, and the fine registration process, based on pyTorch 1.17.0 + cu11.0, takes only about 50 seconds per sample on average. Therefore, the "coarse-fine" two-step registration method based on BPS pre-registration and non-rigid template fitting not only ensures the registration accuracy of the nonlinear model but also significantly improves registration efficiency.

[0063] III. Statistical Shape Analysis

[0064] The shape space of a 3D human body model is defined by an average template and principal component shape orientations. It is obtained by performing PCA calculations on the shapes of registered samples on a multi-shape dataset.

[0065] Considering that the human standing posture differs from the standard standing posture during actual 3D scanning, the standard standing posture is as follows: Figure 9 As shown, pose correction should be performed before conducting shape statistical analysis to minimize the proportion of pose differences in the principal components of the PCA statistical results. The specific method is as follows:

[0066] Standard standing posture samples were selected from a large human body data database, and their J values ​​were calculated. 3d And serve as a standard standing posture template {θ T};

[0067] 3) J for the SMPLD model to be corrected 3d Using a skeleton mapping algorithm, the relative position of the i-th joint to the standard standing posture template {θ} is calculated. T The difference in rotation angles of} and the corresponding rotation matrix Then, based on the LBS algorithm and the hybrid skin weight W, the SMPLD model after attitude transformation is calculated.

[0068] The results before and after pose correction are as follows: Figure 10 As shown, the first column is the fine registration diagram, the second column is the pose correction diagram, and the third and fourth columns are fusion diagrams. White represents before correction, and black represents after correction. The purpose of principal component analysis is to maximize the interpretable variance of vertex offset relative to the average shape, given a finite shape orientation. The cumulative variance contribution of the first 10 principal components is 89.18097% > 85%. Figure 11 The diagram shows the human body shape when each PCA parameter (±5) varies individually; the middle diagram represents the average body shape. Figure 11 We can see that: PC1 represents the size and weight of the entire human body. Initially (average model), it is 0. Positive changes result in a smaller, heavier body; negative changes result in a larger, thinner body; positive changes result in a forward lean. PC2 also shows positive changes in a smaller, heavier body; negative changes result in a larger, thinner body; negative changes result in a forward lean. PC3 indicates a larger belly and thinner, longer arms and legs, allowing adjustment of the upper and lower body proportions. PC4 represents horizontal compression and stretching of the human body; positive compression results in lowered shoulders. PC5 represents horizontal compression and stretching of the human body; positive compression results in wider shoulders. PC6 indicates a forward lean of the upper body and overall weight gain. PC7 indicates a leftward lean of the upper body and overall weight loss. PC8 represents vertical compression and stretching of the human body; negative compression results in a smaller chest and larger abdomen and buttocks; positive compression results in a larger chest and smaller abdomen and buttocks. PC10 represents the size of the chest, abdomen, and buttocks; initially, it is 0. Negative changes result in larger sizes, positive changes result in smaller sizes.

[0069] IV. Final Model Reconstruction

[0070] The final expression for the parametric human body model adapted to the Chinese human form is as follows:

[0071]

[0072] Among them, V i F represents the average value of each vertex in the grid, i.e., the value of each vertex in the corrected SMPLD model. i Triangular mesh, JM 6890×24 W is the weighted coefficient matrix of the 24 skeleton points in the SMPL model. 6890×24 This is the weighting coefficient matrix for bone and skin deformation. There are 24 joint rotation coefficients. These are global displacement parameters, i.e., attitude parameters. For scaling parameters, PC i These are the shape parameters, i.e., the reconstruction coefficients obtained from PCA principal component analysis.

[0073] In the existing technology, the SMPL average template, i.e., the first and last two columns are black and have undergone pose correction, is compared with the parametric model average template based on the Chinese dataset, i.e., the second and last two columns are white. Figure 12 As shown in the figure, there are significant differences in the body shape between adult Chinese women and adult women in Europe and America.

[0074] The reconstruction results are compared between the parametric model based on the Chinese dataset of this invention and the existing SMPL model. The reconstruction results are as follows: Figure 13As shown. From left to right, the first column: the parametric model based on the Chinese dataset; the second column: the SMPL model; the third and fourth columns: the fusion of the two reconstructed models (black: SMPL model, white: the parametric model of this invention).

[0075] We selected detailed differences in the reconstructed models for comparison: a) the left elbow and armpit, b) the left leg, and c) the right leg. It is clear that the parametric model reconstruction based on the Chinese dataset better matches the human body. The average joint error and average vertex error data from the fitting results of the two models are compared in the table below.

[0076]

[0077] As shown in the table above, the parametric model fitting results based on the Chinese dataset show a reduction in both joint and vertex errors. For Experiment 1, the average joint and vertex errors decreased by 26.2% and 20.0%, respectively; for Experiment 2, they decreased by 19.4% and 16.1%, respectively; and for Experiment 3, they decreased by 21.7% and 12.5%, respectively. Overall, the parametric model reconstruction based on the Chinese dataset significantly improves the model fitting accuracy.

[0078] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A parametric human body model reconstruction method based on a Chinese database, characterized in that... Includes the following steps: I. Establishing a Database Collect three-dimensional human body mesh data of multiple Chinese women of different ages, body types, and heights, and establish a database; 2.3D Human Body Mesh Registration First, the point cloud is efficiently learned based on BPS Basis Point Sets, and the SMPL model mesh points are coarsely registered to the vicinity of each 3D human body mesh point in the database to generate the corresponding coarse registration model. Then, a non-rigid mesh registration algorithm based on ICP algorithm is used to finely register each coarse registration model, only optimizing the vertex displacement D component to generate the SMPLD model corresponding to the precise registration. III. Statistical Shape Analysis First, the standing posture of each SMPLD model is corrected to unify it into a standard standing posture. Then, PCA principal component analysis is performed on each SMPLD model after standing posture correction. The results are used as the body parameters of the parametric human body model adapted to the Chinese body shape. The original posture parameters of the SMPL model are used as the posture parameters of the parametric human body model adapted to the Chinese body shape to construct the final parametric human body model adapted to the Chinese body shape. The final expression for the parametric human body model adapted to the Chinese human form is as follows: in, For grid vertices, It is a triangular grid. This is the weighted coefficient matrix for the 24 skeleton points in the SMPL model. This is the weighting coefficient matrix for bone and skin deformation. There are 24 joint rotation coefficients. These are global displacement parameters. For scaling parameters, Reconstruction coefficients obtained from PCA principal component analysis.

2. The parametric human body model reconstruction method based on a Chinese database according to claim 1, characterized in that: The precise registration method includes setting This indicates that the sampling standard surface will be used. 3D mesh vertices Mapping to model parameters The deformed 3D point human body model, denoted by SMPLD model as follows: Where θ is the attitude parameter, β is the shape parameter, and D is the non-rigid deformation parameter, representing the result of... The offsets of each vertex in the reconstructed SMPL model. Build by Point set in the reconstructed SMPL model and 3D human body scan point set loss function : in, This represents the mesh surface of the SMPL model. =6890, , The number of grid points to be matched. This represents the initial mesh surface after coarse registration using BPS. Represents the Laplace operator. It is the distance from a differentiable point to the surface. They are continuous surface points. Represents a vertex Defined model function Perform centroid interpolation. This represents the offset of each vertex in the reconstructed SMPL model. , and These are dynamic weighting coefficients. The ICP algorithm is used to search for the optimal loss function. In each iteration, the corresponding point pairs of the collected 3D human body model-SMPLD model are first calculated, and then the model parameters are updated using gradient or Gauss-Newton optimizers to minimize the distance between the scanned point and the corresponding model point.

3. The parametric human body model reconstruction method based on a Chinese database according to claim 1, characterized in that... Pose correction includes the following steps: 1) Reconstruct the 3D joints of each registered SMPLD model to be corrected using the skeleton skinning model in the SMPL model. ; 2) Select standard standing posture samples from the human body data database and calculate their... And serve as a standard standing posture template ; 3) For the SMPLD model to be corrected The skeleton mapping algorithm is used to calculate the first... Each corresponding joint is a standard standing posture template. Rotation angle difference and the corresponding rotation matrix Then, based on the LBS algorithm and hybrid skin weights Calculate the SMPLD model after attitude transformation.

Images