A three-dimensional human body reconstruction method and system based on physiological parameter driving
By using a physiological parameter-driven 3D human body reconstruction method and optimizing the STAR model with physiological parameters, the problem of inaccurate 3D human body modeling caused by racial data bias is solved, and more accurate and biologically realistic 3D human body reconstruction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU VIPASSANA TECHNOLOGY CO LTD
- Filing Date
- 2026-03-05
- Publication Date
- 2026-06-09
AI Technical Summary
Existing 3D human reconstruction technologies often exhibit the "paper-thin effect" and anatomical fallacies when dealing with Asians or people with special body types. In particular, standard models such as SMPL and STAR models generate inaccurate 3D human models due to racial data bias.
A physiological parameter-driven 3D human body reconstruction method is adopted. By acquiring physiological parameters such as height, body mass index and bioelectrical impedance spectrum, the STAR human body model is optimized using optimizer and gradient masking technology. Local geometric refinement and surface micro-synthesis processing are introduced to generate a more accurate 3D human body model.
This solves the problem of inaccurate 3D human body modeling caused by racial data bias, avoids the "paper doll effect," and improves the accuracy and biological realism of 3D human body modeling.
Smart Images
Figure CN122176237A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional human body reconstruction technology, and in particular to a three-dimensional human body reconstruction method and system driven by physiological parameters. Background Technology
[0002] In the current digital wave, 3D human body reconstruction has rapidly expanded from simple computer graphics entertainment applications to core fields such as medical diagnosis, ergonomic simulation, virtual fitting, and the construction of metaverse digital avatars. However, existing technical approaches suffer from significant "modal fragmentation." While mainstream parametric models (such as the SMPL or STAR models) can generate realistic meshes, they are essentially based on statistical "surface fitting." Since the training data mainly comes from Western populations, these models often exhibit severe "paper-thin" effects or "anatomical fallacies" when dealing with Asian populations or specific body types (such as athletes with low body fat and high muscle mass, or those with central obesity).
[0003] Therefore, the applicant has developed a three-dimensional human body reconstruction method and system based on physiological parameters to solve the above problems. Summary of the Invention
[0004] This invention proposes a three-dimensional human body reconstruction method and system based on physiological parameters to solve the problem of severe "paper-thin human effect" when processing Asian people or people with special body types.
[0005] The present invention achieves the above objectives through the following technical solutions:
[0006] This invention provides a physiological parameter-driven three-dimensional human body reconstruction method, comprising:
[0007] Obtain human physiological parameters and a standard STAR human body model;
[0008] The preset optimizer is called to optimize the STAR human body model. The loss function of the optimizer includes the original loss term and the shoulder width loss term. The shoulder width loss term is the loss between the model shoulder width and the target shoulder width. The target shoulder width is obtained by weighted summation of height, body mass index and muscle correction factor. The muscle correction factor is non-linearly negatively correlated with the impedance value of the upper limb segment in the physiological parameters. The model shoulder width is the standard shoulder width of the STAR human body model.
[0009] During backpropagation, the parameters are updated by calling a preset parameter iteration update rule. The preset parameter iteration update rule is as follows: calculate the gradient of the model parameters according to the loss function, and use the optimization algorithm to update the shape parameters of the STAR human body model along the gradient descent direction.
[0010] Furthermore, the preset parameter iteration update rules include: constructing a gradient mask vector, setting the gradient mask vector to zero in the dimensions corresponding to the head and neck parameters, calculating the Hadamard product of the original gradient and the gradient mask vector, and using this product as the final update gradient to freeze the geometric deformation of the head and neck.
[0011] Furthermore, the physiological parameters are obtained by measuring the human body composition analysis instrument, which is used for continuous spectrum bioelectrical impedance spectrum scanning data, and also for obtaining human physiological parameters based on the bioelectrical impedance spectrum scanning data, based on the Cole-Cole model and Hanai mixture theory.
[0012] Furthermore, shoulder width loss item The expression is:
[0013] ;
[0014] ;
[0015] in, For the model's shoulder width, H represents the target shoulder width, H represents the subject's height, BMI represents body mass index, and α represents the target shoulder width. i γ i C i These are the regression coefficients and constant terms selected based on BMI segmentation, respectively.
[0016] It is a muscle repair factor.
[0017] Furthermore, the expression for the preset parameter iteration update rule is:
[0018] ;
[0019] ;
[0020] in, Update the gradient for the original parameters. To effectively update gradients, The preset gradient masking vector is N, which represents the parameter dimension of the STAR human model.
[0021] Furthermore, it also includes local geometric refinement processing, which includes:
[0022] Calculate the true volume of the human body based on intracellular fluid parameters or extracellular fluid parameters from physiological parameters.
[0023] Calculate the difference between the initial volume of the STAR human body model and the actual volume of the human body;
[0024] The total energy function is constructed based on the data fitting term, the Laplacian smoothing term, and the ARAP energy function. The data fitting term is the difference. The data fitting term is used to drive the mesh surface of the STAR human model to expand towards the real volume boundary. The Laplacian smoothing term is used to suppress high-frequency noise in the displacement. The ARAP energy function is used to force each local triangular facet of the mesh to only rotate and translate during the deformation process.
[0025] The displacement of the mesh vertices is solved by minimizing the total energy function, the body shape parameters in the STAR human model are locked, and the mesh is nonlinearly deformed based on the displacement.
[0026] Furthermore, the total energy function The expression is:
[0027] ;
[0028] ;
[0029] For data fitting terms, For Laplace smoothing terms, For ARAP energy function, , These are the weighting coefficients of the Laplace smoothing term and the ARAP energy function, respectively. , : Vertex coordinates before deformation , These are all coordinates of the deformed mesh vertices. Let be the optimal rotation matrix for the local region to which the i-th mesh vertex belongs, where i represents the index of any vertex in the STAR human body model mesh, and j represents the index of the adjacent vertex that has a direct edge connection to vertex i. Describes the ring neighborhood set of vertex i. This represents the weight of the edge connecting vertex i and vertex j.
[0030] The specific algorithm and iterative process for solving the displacement of mesh vertices based on minimizing the total energy function includes using a local-global alternating optimization algorithm. This involves iteratively updating the mesh vertex coordinates by alternately fixing the vertex positions to solve for the optimal rotation matrix and fixing the rotation matrix to solve for the sparse linear equations. The specific process is as follows:
[0031] Regarding the total energy function constructed above Because the ARAP energy term includes rotation variables dependent on vertex positions, the objective function exhibits nonlinear characteristics. This invention employs a "local-global alternating optimization strategy" for iterative solution, with the specific steps as follows:
[0032] Local projection step:
[0033] In the k-th iteration, the positions p^(k-1) of the mesh vertices are initially fixed. At this point, for each local element in the mesh (such as a triangular patch or a ring neighborhood), the optimization problem degenerates into finding an optimal rotation matrix R_i that best matches the current deformation state. This can be solved analytically using singular value decomposition or extreme value decomposition.
[0034] Global solution step:
[0035] The rotation matrix R_i obtained from the fixed local step remains unchanged. At this point, the total energy function... It degenerates into a quadratic form function with respect to the vertex position p. Minimizing this quadratic form function is equivalent to solving a sparse linear system of equations:
[0036] ;
[0037] Where L is the Laplacian matrix of the grid or the system weighting matrix, and b is the right-hand vector containing the target data term and the rotation term. This system of linear equations can be solved efficiently using the Cholliski decomposition or the conjugate gradient method.
[0038] Convergence criterion:
[0039] The above local and global steps are executed alternately until the decrease in the total energy function is less than a preset threshold (e.g., 10^-5) or the preset maximum number of iterations (e.g., 10 times) is reached. At this point, the output vertex position is the final nonlinear deformation result.
[0040] Furthermore, it also includes surface micro-synthesis treatment, which includes:
[0041] The physiological sharpness index is calculated based on the phase angle, skeletal muscle index, and body fat percentage among the physiological parameters.
[0042] Based on the value range of the physiological sharpness index, linear interpolation is performed between the preset standardized normal map basis vector spaces to generate the final normal map. The preset standardized normal map basis vector spaces include high-frequency noise map, mid-frequency smooth map and low-frequency soft map.
[0043] Among them, the high-frequency noise map is used to simulate the texture of muscle fibers under extremely low sebum, the mid-frequency smooth map is used to simulate the outline of healthy muscles, and the low-frequency soft map is used to simulate the texture of skin covered by fat.
[0044] Furthermore, physiological sharpness index The calculation formula is:
[0045] ;
[0046] in, This represents the Sigmoid activation function or truncation function, used to restrict the output to the interval [0,1]. Indicates the normalized phase angle. Represents the normalized skeletal muscle index. This represents the normalized body fat percentage, where α, β, and γ are the phase angle weight, muscle mass weight, and fat inhibition weight coefficients, respectively. This is the bias constant.
[0047] Furthermore, based on the value range of the physiological sharpness index, linear interpolation is performed between preset standardized normal map basis vector spaces to generate the final normal map. The expression is:
[0048] ;
[0049] in, This represents a high-frequency noise graph. This represents the mid-frequency smoothing plot. This represents a low-frequency smoothing pattern. The index represents physiological sharpness, and Lerp represents linear interpolation.
[0050] The present invention also provides a system for the aforementioned physiological parameter-driven three-dimensional human body reconstruction method, comprising:
[0051] An acquisition module is used to acquire physiological parameters of the human body and a standard STAR human body model;
[0052] The shoulder width correction module is used to call a preset optimizer to optimize the STAR human body model. The loss function of the optimizer includes the original loss term and the shoulder width loss term. The shoulder width loss term is the loss between the model shoulder width and the target shoulder width. The target shoulder width is obtained by weighted summation of height, body mass index and muscle correction factor. The muscle correction factor is non-linearly negatively correlated with the impedance value of the upper limb segment in the physiological parameters.
[0053] The gradient masking module is used to update parameters during backpropagation by calling a preset parameter iteration update rule. The preset parameter iteration update rule is: calculate the gradient of the model parameters according to the loss function, and use an optimization algorithm to update the shape parameters of the STAR human body model along the gradient descent direction.
[0054] The beneficial effects of this invention are as follows:
[0055] The present invention proposes a three-dimensional human body reconstruction method and system based on physiological parameters, which transforms the physiological parameters of the human body into geometric constraints and performs deformation correction on the STAR model from the inside out. This solves the problem of inaccurate three-dimensional human body modeling caused by racial data deviation in the standard STAR model, avoids the "paper-thin human effect", and improves the accuracy of three-dimensional human body modeling. Attached Figure Description
[0056] Figure 1 This is a flowchart of the physiological parameter-driven three-dimensional human body reconstruction method of this application. Detailed Implementation
[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0058] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0059] During the research, the applicant discovered that the SMPL and STAR models, based on principal component analysis (PCA), have the following significant drawbacks:
[0060] Racial bias: The model is primarily constructed using data from Europe and America. The relationship between height and frame width differs significantly between Asian populations and Western populations.
[0061] The "paper-thin" effect: Asian young women often have low BMI but normal height. In order to match their height, models are mistakenly made to widen their shoulders according to Western proportions; in order to maintain low weight (volume conservation), models are forced to drastically compress their chest depth (anteroposterior diameter), resulting in a side profile that is as thin as paper.
[0062] The phenomenon of "giant growth": While the height of modern Asian populations has increased rapidly, head circumference has not increased proportionally. The allometric growth pattern implied in the model incorrectly infers that "tall people = large head circumference," resulting in an imbalanced head circumference ratio in the generated Asian male models.
[0063] Back collapse: When a narrow waist constraint is forcibly set, traditional optimization algorithms will simply flatten the back mesh through mathematical compression, causing the anatomical structure to collapse.
[0064] Therefore, this application studies a three-dimensional human body reconstruction method and system based on physiological parameters to solve the above problems.
[0065] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0066] like Figure 1 As shown, a method and system for three-dimensional human body reconstruction based on physiological parameters includes:
[0067] This embodiment provides a physiological parameter-driven three-dimensional human body reconstruction method, including:
[0068] Obtain human physiological parameters and a standard STAR human body model;
[0069] The STAR human model is optimized by calling the preset optimizer. The loss function of the optimizer includes the original loss term and the shoulder width loss term. The shoulder width loss term is the loss between the model shoulder width and the target shoulder width. The target shoulder width is obtained by weighted summation of height, body mass index and muscle correction factor. The muscle correction factor is non-linearly negatively correlated with the impedance value of the upper limb segment in the physiological parameters.
[0070] Specifically, the calculation of the "muscle correction factor" is based on the principle of conductive volume in bioimpedance physics. Since muscle tissue is the main pathway for electric current, the cross-sectional area of the muscle in the upper limb segment is inversely proportional to the measured upper limb impedance modulus.
[0071] Given that shoulder width is positively correlated with the degree of upper limb muscle development, this invention constructs a correction function of the following form:
[0072] ;
[0073] Where: H is height, used to normalize limb length; |Z|_upper is the measured upper limb segmental impedance value; This is the "upper limb conductivity volume index," and the larger the value, the more developed the upper limb muscles are. is the average conductive volume index of the reference population; k is the regression coefficient, which is usually a positive value.
[0074] Logical conclusion: When the measured impedance |Z|_upper decreases, the conductivity volume index... The increase leads to an increase in the positive muscle correction factor δ_muscle, which drives the model to generate a wider shoulder geometry, establishing a nonlinear negative correlation between the correction factor and the impedance value.
[0075] During backpropagation, the parameters are updated by calling a preset parameter iteration update rule. The preset parameter iteration update rule is as follows: calculate the gradient of the model parameters according to the loss function, and use the optimization algorithm to update the shape parameters of the STAR human body model along the gradient descent direction.
[0076] In one embodiment, the preset parameter iteration update rule includes: constructing a gradient mask vector, setting the gradient mask vector to zero in the dimensions corresponding to the head and neck parameters, calculating the Hadamard product of the original gradient and the gradient mask vector, and using the product as the final update gradient to freeze the geometric deformation of the head and neck.
[0077] This application aims to address the problems of "paper-thin" figures, "giant" figures, and lack of internal physiological details in the standard STAR model caused by racial data bias. This application establishes a "two-stage hybrid optimization framework" that transforms physiological parameters (extracellular fluid (ECW), intracellular fluid (ICW), phase angle, and impedance value) into geometric constraints, and performs deformation correction on the STAR model from the inside out.
[0078] The STAR model is essentially a statistically based parametric function M(β, theta), where β represents the body size parameter and theta represents the pose parameter. Therefore, this application introduces a gradient masking technique during the backpropagation process to optimize the β parameter. Specific principal component parameters controlling the head size are masked to 0. This mathematically "freezes" the statistical scaling of the head, maintaining its initial size based on actual head circumference measurements and preventing erroneous allometric growth with changes in height.
[0079] The STAR model suffers from allometric growth bias, meaning that it assumes a linear increase in head circumference as height increases (giantization), which does not align with Asian anatomy.
[0080] In one embodiment, the physiological parameters are obtained by a human body composition analyzer, which is used for continuous spectrum bioelectrical impedance spectroscopy scanning data, and also for obtaining human physiological parameters based on the bioelectrical impedance spectroscopy scanning data, based on the Cole-Cole model and Hanai mixture theory.
[0081] This application does not use the six discrete frequency points of traditional equipment, but instead performs bioelectrical impedance spectroscopy (BIS) scanning using multiple frequency points covering a wide bandwidth (e.g., 1kHz-1MHz) or a continuous spectrum sufficient to characterize the β-dispersion region. Based on the Cole-Cole biophysical model, the true impedance characteristics of extracellular fluid (ECW) and intracellular fluid (ICW) are accurately resolved. This application utilizes a high-precision programmable AC constant current source to perform continuous frequency sweeps or high-density discrete scans covering a wide bandwidth (preferably 1kHz to 1MHz). The sampling strategy is specially configured to accurately characterize the β-dispersion region. This frequency band is mainly dominated by the dielectric polarization effect of the cell membrane and contains rich cellular structural information. By acquiring a high signal-to-noise ratio response signal within this frequency band, the system ultimately outputs a complete complex impedance spectrum, thereby capturing subtle spectral fluctuations caused by differences in cell membrane integrity, providing high-confidence data support for subsequent Cole-Cole model fitting.
[0082] Traditional devices sample only at a few points, such as 1, 5, 50, 250, 500, and 1000 kHz. This sparse sampling forces algorithms to rely on empirical formulas to estimate the zero-frequency resistance (R0) and the infinite-frequency resistance (R∞). When subjects are in a state of edema, dehydration, or illness, this "estimation" will produce significant errors. This invention uses the Inbody770 device to achieve the above sampling and measurement process.
[0083] In one embodiment, the shoulder width loss term The expression is:
[0084] ;
[0085] ;
[0086] in, For the model's shoulder width, H represents the target shoulder width, H represents the subject's height, BMI represents body mass index, and α represents the target shoulder width. i γ i C i These are the regression coefficients and constant terms selected based on BMI segmentation, respectively.
[0087] It is a muscle repair factor.
[0088] To maintain volume conservation in individuals with low BMI, the standard STAR model tends to "flatten" the anteroposterior diameter of the thorax to match height, leading to a "paper-thin" effect. This application constructs a loss function that includes a shoulder width constraint term to optimize the body shape parameter β of the STAR model. The muscle correction factor is negatively correlated with the upper limb segment impedance value Z_arm; the lower the impedance, the more developed the muscles, and the larger the shoulder width correction value, used to compensate for individual differences in skeletal width.
[0089] The optimizer's loss function includes the original loss term and the shoulder width loss term. In the process of minimizing the loss function, the optimizer is forced to adjust the β parameter to increase the depth of the model's ribcage, thereby eliminating the "paper-thin" phenomenon.
[0090] In one embodiment, the expression for the preset parameter iteration update rule is:
[0091] ;
[0092] ;
[0093] in, Update the gradient for the original parameters. To effectively update gradients, The preset gradient masking vector is N, which represents the parameter dimension of the STAR human model.
[0094] In one embodiment, local geometry refinement processing is also included, which includes:
[0095] Calculate the true volume of the human body based on intracellular fluid parameters or extracellular fluid parameters from physiological parameters.
[0096] Calculate the difference between the initial volume of the STAR human body model and the actual volume of the human body;
[0097] The total energy function is constructed based on the data fitting term, the Laplacian smoothing term, and the ARAP energy function. The data fitting term is the difference value used to drive the mesh surface of the STAR human model to expand towards the true volume boundary. The Laplacian smoothing term is used to suppress high-frequency noise in the displacement and prevent unnatural spikes from appearing on the model surface. The ARAP energy function is used to force each local triangular facet of the mesh to undergo only rotation and translation during deformation.
[0098] The displacement of the mesh vertices is solved by minimizing the total energy function, the body shape parameters in the STAR human model are locked, and the mesh is nonlinearly deformed based on the displacement.
[0099] When nonlinear deformation of the mesh is required based on visceral fat or local muscle mass (estimated by ICW / ECW), simply moving vertices can lead to mesh collapse or spikes. In this stage, β is locked, and the mesh vertex coordinates are directly optimized, introducing an energy term that is as rigid as possible (ARAP) as a regularization constraint.
[0100] In one embodiment, the total energy function The expression is:
[0101] ;
[0102] ;
[0103] For data fitting terms, For Laplace smoothing terms, For ARAP energy function, , These are the weighting coefficients of the Laplace smoothing term and the ARAP energy function, respectively. , : Vertex coordinates before deformation , These are all coordinates of the deformed mesh vertices. Let be the optimal rotation matrix for the local region to which the i-th mesh vertex belongs, where i represents the index of any vertex in the STAR human body model mesh, and j represents the index of the adjacent vertex that has a direct edge connection to vertex i. Describes the ring neighborhood set of vertex i. This represents the weight of the edge connecting vertex i and vertex j, usually the cotangent weight.
[0104] The formula mandates that each local triangular facet of the mesh can only rotate and translate during deformation, and shearing or excessive stretching is strictly prohibited. When attempting to compress the abdomen to match a "slim waist" during 3D human reconstruction, ARAP energy generates a reverse supporting force, forcing the back mesh to maintain the rigidity of its anatomical structure and preventing the back from collapsing inward.
[0105] In one embodiment, the method further includes a surface microstructure synthesis process, which includes:
[0106] The physiological sharpness index is calculated based on the phase angle, skeletal muscle index, and body fat percentage among the physiological parameters.
[0107] Based on the value range of the physiological sharpness index, linear interpolation is performed between the preset standardized normal map basis vector spaces to generate the final normal map. The preset standardized normal map basis vector spaces include high-frequency noise map, mid-frequency smooth map and low-frequency soft map.
[0108] Among them, the high-frequency noise map is used to simulate the texture of muscle fibers under extremely low sebum, representing high-frequency geometric details such as muscle bundle separation and vascular dilation under extremely low sebum; the mid-frequency smooth map is used to simulate the outline of healthy muscles, representing the outline of healthy muscle blocks with moderate texture; the low-frequency soft map is used to simulate the texture of skin under fat coverage, representing the smooth and soft texture under subcutaneous fat coverage, mainly containing low-frequency geometric information.
[0109] Geometric meshes cannot represent the microscopic textures of "muscle definition" or "swelling." Existing techniques typically assume that muscle mass primarily affects the vertex displacement of a 3D mesh (i.e., body shape). However, our research has found a strong correlation between local density variations in muscle mass and high-frequency geometric details of the body surface. Therefore, incorporating muscle mass parameters into the normal map generation logic enables the simulation of high-precision body surface textures with extremely low computational cost and without increasing the mesh facet count—a significant improvement over traditional 3D reconstruction workflows. Normal mapping is a lighting rendering technique that simulates high-frequency details (such as wrinkles and blood vessels) by perturbing the surface normal vectors without altering the vertex positions of the geometric model.
[0110] In one embodiment, the physiological sharpness index The calculation formula is:
[0111] ;
[0112] in, This represents the Sigmoid activation function or truncation function, used to restrict the output to the interval [0,1]. Indicates the normalized phase angle. The normalized skeletal muscle index represents the strength of muscle support. Representing normalized body fat percentage, used as a negative masking factor, α, β, and γ are the phase angle weight, muscle mass weight, and fat inhibition weight coefficients, respectively. This is a bias constant used to ensure the weights. It is within the valid interpolation range.
[0113] Wherein, β is the positive gain coefficient: the phase angle reflects cell compactness, and skeletal muscle mass reflects muscle bundle volume. The higher the values of both, the clearer the surface texture (sharpening).
[0114] γ is the negative inhibition coefficient: body fat percentage reflects the thickness of subcutaneous fat. The higher the value, the stronger the smoothing / covering effect (softening) on muscle texture.
[0115] This application no longer relies on discrete threshold judgments in a single dimension, but instead establishes a continuous "bioelectric-visual mapping function". The core of this function lies in calculating a comprehensive "physiological sharpness index", which can dynamically quantify the "compactness" and "clarity" of human tissue in vision, thereby determining the mixed weight of the surface normal map of the 3D model.
[0116] In one embodiment, linear interpolation is performed between preset standardized normal map basis vector spaces based on the value range of the physiological sharpness index to generate the final normal map. The expression is:
[0117] ;
[0118] in, This represents a high-frequency noise graph. This represents the mid-frequency smoothing plot. This represents a low-frequency smoothing pattern. The index represents physiological sharpness, and Lerp represents linear interpolation.
[0119] The above methods can achieve the transformation from "soft" ( From ")" to "robust ()" From ")" to "silk pulling" A smooth visual transition. For example, when a user's phase angle is detected to be high but their body fat percentage is also slightly high, Item will inhibit Excessive growth of blood vessel textures prevents the system from incorrectly generating overly dry blood vessel textures, thus ensuring that the final rendering result conforms to physiological facts.
[0120] In this embodiment, the normal map base , , It is not based on subjective artistic creation, but rather on standardized extraction products from real human body scan datasets. Its construction process includes the following technical steps:
[0121] 1. Construction of the reference sample set:
[0122] A reference population database containing different body composition characteristics (e.g., N=500) is pre-constructed. High-resolution 3D surface point clouds of each subject are acquired using a high-precision structured light scanner (such as ArtecEva or an equivalent precision device), with a mesh face number of no less than 2 million, to ensure that high-frequency geometric details such as skin texture, subcutaneous vascular protrusions, and muscle fiber orientation are captured.
[0123] 2. Clustering and extraction of basis types:
[0124] Based on the subjects' bioimpedance measurements (especially phase angle, skeletal muscle index, and body fat percentage), the scan data were divided into three feature clusters:
[0125] The sample cluster corresponds to a phase angle >8°, a body fat percentage <10%, and a skeletal muscle index at a high percentile (e.g., >75%). High-frequency normal information is extracted from the scan model using a high-pass filtering algorithm, preserving muscle bundle separation features.
[0126] Sample clusters with phase angles of 6°-8°, body fat percentages of 10%-20%, and moderate skeletal muscle indices were selected. Mid-frequency normal information was extracted, preserving muscle contours while smoothing microscopic textures.
[0127] This refers to sample clusters with a phase angle <6° or a body fat percentage >20% and a skeletal muscle index at a low percentile (e.g., <25%). Gaussian blurring of the normal information is used to simulate the occlusion effect of the subcutaneous fat layer on muscle texture.
[0128] 3. Normalized mapping:
[0129] The average normal distribution of the three feature clusters extracted above is baked onto a standardized SMPL topology using topological reprojection technology to generate a standardized normal map base with unified UV coordinates.
[0130] Through the above steps, this invention establishes a quantifiable mathematical correspondence between biophysical parameters (impedance) and microscopic geometric features (normals). The "mixing" process is essentially a linear interpolation operation between the three discrete digital feature clusters based on real-time measured phase angle values, rather than a simple image superposition.
[0131] The following section provides a detailed explanation of this technical method through specific application scenarios.
[0132] Scenario 1: High muscle mass accompanied by high body fat (e.g., heavy-weight fighter physique).
[0133] In such samples, although the skeletal muscle mass is high, resulting in a significant positive contribution to the "muscle enhancement term" in the formula, the extremely high body fat percentage causes the "fat coverage term" to have a much larger negative deduction.
[0134] Calculation results: Overall weight It will be pulled down to a lower range.
[0135] Rendering effect: The system will primarily sample And mix a small amount The resulting model visually presents a grand overall volume (determined by mesh deformation), but the surface texture is smooth, and no obvious muscle bundle separation is visible. This aligns with the anatomical fact that "the thickness of the subcutaneous fat layer conceals strong muscles."
[0136] Scenario 2: Moderate muscle mass accompanied by extremely low body fat (e.g., endurance long-distance runner physique).
[0137] In such samples, although the skeletal muscle mass is only at a moderate level, the contribution of the "muscle enhancement" term is only average; however, due to the extremely low body fat percentage, the value of the "fat coverage" term in the formula is extremely small (i.e., very little deduction).
[0138] Calculation results: Overall weight It remains in the higher range.
[0139] Rendering effect: The system will tend to sample. The final generated model, though small in size (determined by the mesh), will display clearly defined blood vessels and sharp muscle lines on its surface. This aligns with the physiological characteristic that "an extremely thin subcutaneous fat layer reveals deep muscle texture."
[0140] In summary, by introducing body fat percentage as a negative regularization factor, this invention effectively solves the common problem of "mismatch between body shape and texture" in traditional 3D reconstruction (such as incorrectly rendering obese individuals with high muscle mass as bodybuilders), and significantly improves the biological realism of digital twin models.
[0141] The final output normal vector is a weighted sum of the base map, and the effect is that as the phase angle value increases (representing the improvement of cell membrane integrity), the lighting and shadow rendering effect on the model surface will automatically and smoothly transition from "soft" to "tight".
[0142] The system of the present invention for the physiological parameter-driven three-dimensional human body reconstruction method comprises:
[0143] An acquisition module is used to acquire physiological parameters of the human body and a standard STAR human body model;
[0144] The shoulder width correction module is used to call a preset optimizer to optimize the STAR human body model. The loss function of the optimizer includes the original loss term and the shoulder width loss term. The shoulder width loss term is the loss between the model shoulder width and the target shoulder width. The target shoulder width is obtained by weighted summation of height, body mass index and muscle correction factor. The muscle correction factor is non-linearly negatively correlated with the impedance value of the upper limb segment in the physiological parameters.
[0145] The gradient masking module is used to update parameters during backpropagation by calling a preset parameter iteration update rule. The preset parameter iteration update rule is: calculate the gradient of the model parameters according to the loss function, and use an optimization algorithm to update the shape parameters of the STAR human body model along the gradient descent direction.
[0146] A system for a physiological parameter-driven 3D human body reconstruction method also includes a local geometry refinement module, which is used for:
[0147] Calculate the true volume of the human body based on intracellular fluid parameters or extracellular fluid parameters from physiological parameters.
[0148] Calculate the difference between the initial volume of the STAR human body model and the actual volume of the human body;
[0149] The total energy function is constructed based on the data fitting term, the Laplacian smoothing term, and the ARAP energy function. The data fitting term is the difference. The data fitting term is used to drive the mesh surface of the STAR human model to expand towards the real volume boundary. The Laplacian smoothing term is used to suppress high-frequency noise in the displacement. The ARAP energy function is used to force each local triangular facet of the mesh to only rotate and translate during the deformation process.
[0150] The displacement of the mesh vertices is solved by minimizing the total energy function, the body shape parameters in the STAR human model are locked, and the mesh is nonlinearly deformed based on the displacement.
[0151] A system for a physiological parameter-driven three-dimensional human body reconstruction method further includes a surface microstructure synthesis processing module, which is used for:
[0152] The physiological sharpness index is calculated based on the phase angle, skeletal muscle index, and body fat percentage among the physiological parameters.
[0153] Based on the value range of the physiological sharpness index, linear interpolation is performed between the preset standardized normal map basis vector spaces to generate the final normal map. The preset standardized normal map basis vector spaces include high-frequency noise map, mid-frequency smooth map and low-frequency soft map.
[0154] Among them, the high-frequency noise map is used to simulate the texture of muscle fibers under extremely low sebum, the mid-frequency smooth map is used to simulate the outline of healthy muscles, and the low-frequency soft map is used to simulate the texture of skin covered by fat.
[0155] The processing logic of this invention is as follows:
[0156] Initialization: Invoke the standard STAR model to generate the base mesh.
[0157] Phase 1 (Global Shape Constraints): Using shoulder width correction and gradient masking, the β parameter is optimized through gradient descent to correct the proportion between the skeleton and the overall outline.
[0158] The second stage (local geometric refinement): the β parameter is locked, and vertex-level optimization is entered. The local volume calculated using ARAP anti-collapse and bioimpedance is used to directly drive the displacement of mesh vertices.
[0159] The third stage (surface micro-synthesis): phase angle texture mapping, generating surface normal details based on physiological function indicators.
[0160] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for three-dimensional human body reconstruction based on physiological parameters, characterized in that, include: Obtain human physiological parameters and a standard STAR human body model; The STAR human model is optimized by calling the preset optimizer. The loss function of the optimizer includes the original loss term and the shoulder width loss term. The shoulder width loss term is the loss between the model shoulder width and the target shoulder width. The target shoulder width is obtained by weighted summation of height, body mass index and muscle correction factor. The muscle correction factor is non-linearly negatively correlated with the impedance value of the upper limb segment in the physiological parameters. During backpropagation, the parameters are updated by calling a preset parameter iteration update rule. The preset parameter iteration update rule is as follows: calculate the gradient of the model parameters according to the loss function, and use the optimization algorithm to update the shape parameters of the STAR human body model along the gradient descent direction.
2. The method for three-dimensional human reconstruction based on physiological parameters according to claim 1, characterized in that, The preset parameter iteration update rules include: constructing a gradient mask vector, setting the gradient mask vector to zero in the dimensions corresponding to the head and neck parameters, calculating the Hadamard product of the original gradient and the gradient mask vector, and using this product as the final update gradient to freeze the geometric deformation of the head and neck.
3. The method for three-dimensional human reconstruction based on physiological parameters according to claim 1, characterized in that, Shoulder width loss item The expression is: ； ； in, For the model's shoulder width, H represents the target shoulder width, H represents the subject's height, BMI represents body mass index, and α represents the target shoulder width. i γ i C i These are the regression coefficients and constant terms selected based on BMI segmentation, respectively. It is a muscle repair factor.
4. The method for three-dimensional human reconstruction based on physiological parameters according to claim 1, characterized in that, The expression for the preset parameter iteration update rule is: ； ； in, Update the gradient for the original parameters. To effectively update the gradient, M is a preset gradient mask vector, and N represents the parameter dimension of the STAR human model.
5. The method for three-dimensional human reconstruction based on physiological parameters according to claim 1, characterized in that, It also includes local geometric refinement processing, which includes: Calculate the true volume of the human body based on intracellular fluid parameters or extracellular fluid parameters from physiological parameters. Calculate the difference between the initial volume of the STAR human body model and the actual volume of the human body; The total energy function is constructed based on the data fitting term, the Laplacian smoothing term, and the ARAP energy function. The data fitting term is the difference. The data fitting term is used to drive the mesh surface of the STAR human model to expand towards the real volume boundary. The Laplacian smoothing term is used to suppress high-frequency noise in the displacement. The ARAP energy function is used to force each local triangular facet of the mesh to only rotate and translate during the deformation process. The displacement of the mesh vertices is solved by minimizing the total energy function, the body shape parameters in the STAR human model are locked, and the mesh is nonlinearly deformed based on the displacement.
6. The method for three-dimensional human reconstruction based on physiological parameters according to claim 5, characterized in that, Total energy function The expression is: ； ； For data fitting terms, For Laplace smoothing terms, For ARAP energy function, , These are the weighting coefficients of the Laplace smoothing term and the ARAP energy function, respectively. , : Vertex coordinates before deformation , These are all coordinates of the deformed mesh vertices. Let be the optimal rotation matrix for the local region to which the i-th mesh vertex belongs, where i represents the index of any vertex in the STAR human body model mesh, and j represents the index of the adjacent vertex that has a direct edge connection to vertex i. Describes the ring neighborhood set of vertex i. This represents the weight of the edge connecting vertex i and vertex j.
7. The method for three-dimensional human body reconstruction based on physiological parameters according to claim 1, characterized in that, It also includes surface micro-synthesis treatment, which includes: The physiological sharpness index is calculated based on the phase angle, skeletal muscle index, and body fat percentage among the physiological parameters. Based on the value range of the physiological sharpness index, linear interpolation is performed between the preset standardized normal map basis vector spaces to generate the final normal map. The preset standardized normal map basis vector spaces include high-frequency noise map, mid-frequency smooth map and low-frequency soft map. Among them, the high-frequency noise map is used to simulate the texture of muscle fibers under extremely low sebum, the mid-frequency smooth map is used to simulate the outline of healthy muscles, and the low-frequency soft map is used to simulate the texture of skin covered by fat.
8. The method for three-dimensional human reconstruction based on physiological parameters according to claim 7, characterized in that, Physiological sharpness index The calculation formula is: ； in, This represents the Sigmoid activation function or truncation function, used to restrict the output to the interval [0,1]. Indicates the normalized phase angle. Represents the normalized skeletal muscle index. This represents the normalized body fat percentage, where α, β, and γ are the phase angle weight, muscle mass weight, and fat inhibition weight coefficients, respectively. This is the bias constant.
9. A method for three-dimensional human body reconstruction based on physiological parameters according to claim 7 or 8, characterized in that, Based on the range of the physiological sharpness index, linear interpolation is performed between preset standardized normal map basis vector spaces to generate the final normal map. The expression is: ； in, This represents a high-frequency noise graph. This represents the mid-frequency smoothing plot. This represents a low-frequency smoothing pattern. The index represents physiological sharpness, and Lerp represents linear interpolation.
10. A system for a physiological parameter-driven three-dimensional human body reconstruction method according to any one of claims 1-9, characterized in that, include: An acquisition module is used to acquire physiological parameters of the human body and a standard STAR human body model; The shoulder width correction module is used to call a preset optimizer to optimize the STAR human body model. The loss function of the optimizer includes the original loss term and the shoulder width loss term. The shoulder width loss term is the loss between the model shoulder width and the target shoulder width. The target shoulder width is obtained by weighted summation of height, body mass index and muscle correction factor. The muscle correction factor is non-linearly negatively correlated with the impedance value of the upper limb segment in the physiological parameters. The gradient masking module is used to update parameters during backpropagation by calling a preset parameter iteration update rule. The preset parameter iteration update rule is: calculate the gradient of the model parameters according to the loss function, and use an optimization algorithm to update the shape parameters of the STAR human body model along the gradient descent direction.