Anatomically block-based optimized human three-dimensional modeling method and electronic device

Abstract

This invention discloses a method and electronic device for human body 3D modeling based on anatomical block optimization; relating to the field of human body 3D modeling; it divides the parameter vector to be optimized into multiple anatomical blocks, and for each anatomical block, independently updates the corresponding search distribution based on the parameter sub-vectors belonging to that anatomical block and their respective sample weights of all samples in the original population, and then generates the next generation of the original population based on the updated independent search distribution of each block, so that the parameter optimization process of each anatomical block is independent and undependent, avoiding invalid searches caused by the coupling of irrelevant parameters when updating all parameters as a whole. Thus, it improves search efficiency and parameter estimation stability in high-dimensional human body parameter space.

Description

Technical Field

[0001] This invention relates to the field of human body 3D modeling, and in particular to a human body 3D modeling method and electronic device based on anatomical block optimization. Background Technology

[0002] 3D human pose and shape estimation is a key technology in fields such as human-computer interaction, virtual reality, and rehabilitation medicine. Current mainstream methods are usually based on parametric human models (such as SMPL), transforming the problem into optimizing low-dimensional shape and pose parameters, and solving it by minimizing the error between the model's reprojected contour and the observed image. However, this optimization process faces serious challenges: on the one hand, gradient descent methods (such as Adam), while converging quickly locally, are highly sensitive to initial values ​​and easily get trapped in local minima caused by limb occlusion or symmetry in the 85-dimensional high-dimensional parameter space; on the other hand, global optimization methods such as standard simulated annealing or covariance matrix adaptive evolution strategy (CMA-ES), while capable of escaping local optima, often treat all parameters as a single vector for uniform processing, failing to utilize the inherent kinematic hierarchy of the human body, resulting in low search efficiency and unstable covariance estimation under limited sample budgets.

[0003] Furthermore, existing evolutionary strategies face the problem of excessive parameter estimation burden when directly modeling the complete covariance matrix in high-dimensional space (7225 dimensions). While the natural evolutionary strategy (xNES) ensures positive covariance definiteness through matrix exponential mapping, its sample weight allocation relies on a symmetric sorting mechanism, which cannot effectively integrate asymmetric sample weights calculated based on the reprojection error, leading to instability in covariance learning in the presence of occlusion or noise. In addition, pose and shape estimation typically employ gradient-free global search methods, while texture reconstruction relies on gradient optimization through differentiable rendering. There is a lack of effective information transfer between these two stages—in particular, the texture optimization process fails to utilize the accurate human foreground mask (i.e., contour image) obtained in the geometric reconstruction stage to constrain the optimization region, instead calculating the rendering loss on the entire image, causing background pixels to severely interfere with the learning of specific subject textures. Summary of the Invention

[0004] To overcome the problems of unstable covariance estimation, low search efficiency, lack of utilization of anatomical hierarchy, and disconnect between geometry and texture reconstruction in existing technologies for high-dimensional human parameter optimization, this invention proposes a human 3D modeling method based on anatomical block optimization, including: Acquire a multi-view observation image set of the human body to be modeled. The image set contains observation images from multiple shooting angles, and the observation images include contour images. Obtain a human body model controlled by shape parameters and posture parameters, and concatenate the shape parameters and posture parameters into a parameter vector to be optimized; The parameter vector to be optimized is divided into multiple anatomical blocks, and an independent search distribution corresponding to each anatomical block is established; Generate an original population containing multiple samples, where each sample is an instance of the parameter vector to be optimized; The original population is iteratively optimized, and the following operations are performed in each iteration epoch: Based on the human body model, each sample in the original population is rendered as a predicted contour image from each shooting angle. For each sample, the reprojection error between each predicted contour image and the contour image at the corresponding viewpoint is calculated, and the total reprojection error of the sample is obtained by summing them up. The sample weight of each sample in the original population is calculated based on the total reprojection error of each sample. For each anatomical block, the independent search distribution of the anatomical block is adaptively updated based on the parameter subvectors of all samples in the original population belonging to that anatomical block and their respective sample weights, and the original population for the next iteration epoch is generated based on the updated independent search distribution. When the preset iteration termination condition is met, the iteration stops, and the parameters of the sample with the smallest total weight projection error are determined as the optimal attitude parameters and optimal shape parameters.

[0005] Furthermore, the step of dividing the parameter vector to be optimized into multiple anatomical blocks and establishing an independent search distribution corresponding to each anatomical block includes: Based on the hierarchical structure of human kinematics, the parameter vector to be optimized is divided into multiple anatomical blocks; the anatomical blocks include: the trunk block composed of the pelvis and spine, the arm blocks corresponding to the left and right arms respectively, the leg blocks corresponding to the left and right legs respectively, and the body shape block composed of shape parameters. For each anatomical block, an independent search distribution following a Gaussian distribution is established; the independent search distribution is configured with a corresponding covariance matrix, the dimension of which matches the dimension of the parameter sub-vector of the anatomical block.

[0006] Further, the step of calculating the reprojection error between each predicted contour image and the contour image at the corresponding viewpoint, and summing them to obtain the total reprojection error of the sample, includes: For the current sample, at each shooting viewpoint, the reprojection error between the predicted contour image and the observed contour image at that viewpoint is calculated based on the contour difference metric. The total reprojection error of the sample is obtained by summing the reprojection errors of the sample under all shooting angles.

[0007] Further, the calculation of the sample weight of each sample in the original population based on the total reprojection error of each sample includes: For the current iteration epoch, obtain the preset survival rate parameters and the total number of samples in the original population; Based on the survival rate parameter and the total number of samples, an annealing coefficient is determined to adjust the sharpness of the weight distribution; Based on the annealing coefficient and the total reprojection error of each sample, the sample weight of each sample is calculated using the Boltzmann distribution.

[0008] Furthermore, the adaptive update of the independent search distribution of the anatomical block includes: Calculate the normalized weights of each sample based on the sample weights of each sample in the original population, and determine the centering utility of the corresponding sample based on the normalized weights. Calculate the global weighted mean of the current iteration epoch based on the normalized weights of each sample; Based on the normalized weights of each sample, the original population is randomly resampled with replacement to obtain the resampled population. For the parameter subvectors of each sample in the current original population in each anatomical block, a whitening transformation is performed based on the global weighted mean of the current iteration epoch and the parameterization factor of the covariance matrix of the block in the previous iteration epoch to obtain the whitening bias of each anatomical block; wherein, the covariance matrix is ​​characterized by a parameterization factor that matches its dimension. For each anatomical block, the natural gradient matrix of the block is calculated based on the whitening bias of each sample in the original population in the current anatomical block and the centering utility of each sample in the original population. Based on the natural gradient matrix, the parameterization factor of the anatomical block in the previous iteration epoch is updated by exponential mapping to obtain the parameterization factor of the corresponding anatomical block in the current iteration epoch. Update the covariance matrix of the corresponding anatomical block based on the parameterization factor of the anatomical block in the current iteration epoch.

[0009] Furthermore, the step of generating the original population for the next iteration epoch based on the updated independent search distribution includes: Based on the updated covariance matrix of each anatomical block in the current iteration epoch, and combined with the preset activation weights of each anatomical block, a global perturbation covariance matrix is ​​constructed. Based on the global perturbation covariance matrix, the perturbation vector of each sample in the resampled population is set. Based on the perturbation vector, the corresponding resampled samples are perturbed to obtain the corresponding samples of the original population in the next iteration epoch.

[0010] Further, the whitening transformation of the parametric subvectors of each sample in each anatomical block in the current original population is performed based on the global weighted mean of the current iteration epoch and the parameterization factor of the covariance matrix of that block in the previous iteration epoch, to obtain the whitening bias of each anatomical block, including: For each anatomical block, calculate the local weighted mean of that block in the current iteration epoch. The local weighted mean is the parameter subvector corresponding to the global weighted mean in that anatomical block. Using the parameterization factor of the covariance matrix of the anatomical block in the previous iteration epoch, a whitening transformation is performed on the deviation between the parameter subvector of each sample in the current original population in the anatomical block and the local weighted mean of the block in the current iteration epoch, to obtain the whitening deviation of each sample in the current anatomical block.

[0011] Furthermore, during the execution of each iteration epoch, the following is also included: Determine whether there is a better sample in the current iteration epoch with a total reprojection error less than the global optimal sample. If so, update the global optimal sample and replace the worst sample with the largest total reprojection error in the original population of the current iteration epoch with the global optimal sample. The global optimal sample is the sample with the smallest total reprojection error in all historical iteration epochs.

[0012] Furthermore, the observed image further includes a color image; the method further includes: Based on the optimal pose parameters and optimal shape parameters, a reconstructed texture image is obtained by performing texture optimization on the color image in the observed image using differentiable rendering technology; including: Construct a target human body model based on the optimal posture parameters and optimal shape parameters; A global texture image bound to the target human body model is constructed. This global texture image is obtained by mapping the target human body model to the model surface through preset and fixed texture coordinates. Using a differentiable rendering function, the target human body model is rendered into a predicted color image at each shooting angle, and based on the contour image at the corresponding angle, the L2 error between each predicted color image and the color image in the corresponding observed image is calculated in the foreground region of the human body. Using the overall loss of L2 error corresponding to each shooting angle as the loss function, an optimizer is used to iteratively optimize the global texture image to obtain the reconstructed texture image.

[0013] To address the aforementioned technical problems, embodiments of the present invention also provide an electronic device, including: a processor and a memory storing a program, the program including instructions, which, when executed by the processor, cause the processor to perform the aforementioned anatomical block-based human body 3D modeling method.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention divides the parameter vector to be optimized into multiple anatomical blocks, and for each anatomical block, independently updates the corresponding search distribution based on the parameter sub-vectors belonging to the anatomical block of all samples in the original population and their respective sample weights. Then, it generates the next generation of the original population based on the updated independent search distribution of each block. This makes the parameter optimization process of each anatomical block independent and undependent, avoiding the invalid search caused by the coupling of irrelevant parameters when updating all parameters as a whole. This improves the search efficiency and the stability of parameter estimation in the high-dimensional human parameter space.

[0015] (2) This invention introduces parameterized factors to characterize the covariance matrix of each anatomical block and uses the technique of updating the parameterized factors of the previous iteration epoch based on the natural gradient matrix to achieve adaptive evolution of the covariance structure. At the same time, by combining whitening bias and centering utility to calculate the natural gradient matrix, the covariance update can effectively respond to the sample weight distribution, thereby improving the robustness and adaptability of the covariance learning process of high-dimensional anatomical blocks.

[0016] (3) This invention calculates the sample weights of each sample based on the total reprojection error and annealing coefficient according to the Boltzmann distribution, and further normalizes them into normalized weights for iterative updates. The normalized weights are used to: determine the centering utility corresponding to each sample, calculate the global weighted mean, perform population resampling, and construct a natural gradient matrix in combination with whitening bias. Then, based on the natural gradient matrix, the parameterization factors of the covariance matrix of each anatomical block in the previous iteration epoch are updated by exponential mapping, thereby realizing the adaptive evolution of the search distribution. This mechanism directly transforms the reprojection error into a quantitative evaluation of the sample contribution, effectively suppressing the negative impact of high-error samples (i.e., low-quality samples) on covariance learning, thereby enhancing the robustness of the algorithm when there is occlusion or noise interference in multi-view contour matching.

[0017] (4) In the iterative optimization process, the present invention explicitly maintains the historical global optimal sample, and when a sample with a smaller total reprojection error is detected in the current iteration epoch, the global optimal sample is updated and the worst sample in the current population is replaced with the global optimal sample, thereby ensuring that the historical optimal solution is always retained in the search process; after completing the geometric parameter optimization, based on the finally determined optimal pose parameters and optimal shape parameters, a global texture image bound to the target human body model is constructed, and a predicted color image is generated under each shooting angle using differentiable rendering technology; further, based on the human body foreground region defined by the contour image of the corresponding viewpoint, the L2 error between the predicted color image and the observed color image is calculated, and the overall loss of the L2 error of each viewpoint is used as the optimization target to iteratively optimize the global texture image, thereby achieving accurate alignment of texture and geometry under spatial mask constraints, effectively avoiding interference from the background region on texture optimization. Attached Figure Description

[0018] Figure 1 This is a flowchart of a human body 3D modeling method based on anatomical block optimization according to an embodiment of the present invention; Figure 2 These are observation images from multiple shooting angles in an embodiment of the present invention; Figure 3 This is a comparison image of contour and texture reconstruction in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0019] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0020] To overcome the problems of unstable covariance estimation, low search efficiency, failure to utilize anatomical hierarchical structure, and disconnect between geometry and texture reconstruction in existing technologies for high-dimensional human parameter optimization, such as... Figure 1 The present invention proposes a human body 3D modeling method based on anatomical block optimization, comprising: Acquire a multi-view observation image set of the human body to be modeled. The image set contains observation images from multiple shooting angles, and the observation images include contour images and color images. Figure 2 This invention illustrates a multi-view observation image set in one embodiment: it contains observation images from four different viewpoints, with each viewpoint corresponding to two rows of images: the upper row is a color image and the lower row is a binarized contour image; each column independently represents a complete observation pair under a viewpoint.

[0021] Obtain a human body model controlled by shape parameters and posture parameters, and concatenate the shape parameters and posture parameters into a parameter vector to be optimized; In this embodiment, the human body model adopts the parametric three-dimensional human body model SMPL (Skinned Multi-Person Linear model). This model is jointly controlled by shape parameters and pose parameters (represented by rotation vectors for relative rotation of each joint), and its default pose is T-pose (i.e., a neutral pose where all joint rotation vectors are zero).

[0022] To adapt to multi-view observation images, the system introduces global rigid body transformation parameters on the basis of the SMPL model, including global rotation matrix and translation vector, to transform the model from the local coordinate system to the world coordinate system.

[0023] During initialization, the system sets the attitude parameters to zero (i.e., T-pose) and applies small-amplitude Gaussian random perturbations to the shape parameters, global rotation matrix, and translation vector to generate an initial parameter vector to be optimized. This process requires no external initial attitude estimation or manual annotation.

[0024] The parameter vector to be optimized is divided into multiple anatomical blocks, and an independent search distribution corresponding to each anatomical block is established; The step of dividing the parameter vector to be optimized into multiple anatomical blocks and establishing an independent search distribution corresponding to each anatomical block includes: Based on the hierarchical structure of human kinematics, the parameter vector to be optimized is divided into multiple anatomical blocks; the anatomical blocks include: the trunk block composed of the pelvis and spine, the arm blocks corresponding to the left and right arms respectively, the leg blocks corresponding to the left and right legs respectively, and the body shape block composed of shape parameters. For each anatomical block, an independent search distribution following a Gaussian distribution is established; the independent search distribution is configured with a corresponding covariance matrix, the dimension of which matches the dimension of the parameter sub-vector of the anatomical block.

[0025] In one specific embodiment, the anatomical block is divided into six parts: (1) Trunk block, including root node and spine-related joints (corresponding to joint indices 0, 3, 6, 9, 12, 13, 14, 15 in SMPL model), where joint index 0 contains 3 global translation parameters and 3 rotation vector parameters, and each of the other joints contains 3 rotation vector parameters. The total dimension of its pose parameter sub-vectors is 27. (2) The left arm block includes the left shoulder, left elbow, left wrist and related joints of the left hand (index 16, 18, 20, 22), and its pose parameter subvector dimension is 12; (3) Right arm block, including right shoulder, right elbow, right wrist and related joints of the right hand (index 17, 19, 21, 23), with a pose parameter sub-vector dimension of 12; (4) The left leg block, which includes the left hip, left knee, left ankle and related joints of the left foot (index 1, 4, 7, 10), has a pose parameter subvector dimension of 12; (5) Right leg block, including the right hip, right knee, right ankle and related joints of the right foot (index 2, 5, 8, 11), with a dimension of 12; (6) Body block, which is composed of 10-dimensional shape parameter β.

[0026] Each block maintains its own covariance matrix independently, and the total number of covariance parameters is: Compared to directly modeling the complete covariance of an 85-dimensional fully parameterized vector (with a parameter count of...), The number of parameters is reduced by approximately 81.6%, significantly reducing the complexity and instability of covariance estimation.

[0027] Generate an original population containing multiple samples, where each sample is an instance of the parameter vector to be optimized; The original population is iteratively optimized, and the following operations are performed in each iteration epoch: Based on the human body model, each sample in the original population is rendered as a predicted contour image from each shooting angle. In this embodiment, the predicted contour image is a binary foreground mask image.

[0028] For each sample, the reprojection error between each predicted contour image and the contour image at the corresponding viewpoint is calculated, and the total reprojection error of the sample is obtained by summing them up. The process of calculating the reprojection error between each predicted contour image and the contour image at the corresponding viewpoint, and summing them to obtain the total reprojection error of the sample, includes: For the current sample, at each shooting viewpoint, the reprojection error between the predicted contour image and the observed contour image (i.e. the contour image in the observed image of the corresponding viewpoint) is calculated based on the contour difference metric. In one embodiment, the contour difference metric is calculated using symmetric difference area: for a predicted contour image and an observed contour image (both binary images) at a certain viewpoint, the reprojection error at that viewpoint is the total number of foreground pixels in the XOR result of the two. In other embodiments, IoU loss, Dice coefficient, or pixel-level L1 distance may also be used as the contour difference metric.

[0029] The total reprojection error of the sample is obtained by summing the reprojection errors of the sample under all shooting angles.

[0030] The sample weight of each sample in the original population is calculated based on the total reprojection error of each sample. The calculation of the sample weight of each sample in the original population based on the total reprojection error of each sample includes: For the current iteration epoch, obtain the preset survival rate parameters and the total number of samples in the original population; Based on the survival rate parameter and the total number of samples, an annealing coefficient is determined to adjust the sharpness of the weight distribution; Based on the annealing coefficient and the total reprojection error of each sample, the sample weight of each sample is calculated using the Boltzmann distribution.

[0031] In this embodiment, the weights of each sample are calculated using the Boltzmann distribution, as shown in the following formula: ; in: Let be the sample weight of the i-th sample in the m-th iteration epoch; The annealing coefficient for the m-th iteration epoch is determined by the survival rate parameter, the total number of samples, and the total weight projection error of each sample, and is used uniformly for all samples within this epoch. In this embodiment, the formula for determining the annealing coefficient is: The equation is about The transcendental equations, in each iteration epoch, are obtained through a bisection method. Solve the above; This represents the survival rate parameter for the m-th iteration epoch; This represents the upper limit of the annealing coefficient (i.e., the upper bound of the search in the bisection method, which is a preset positive number). The total reprojection error of the i-th sample; y is the input set of multi-view observation images, which contains observation contour images from multiple viewpoints; For the i-th sample in the population, which is a specific instance of the parameter vector to be optimized, it consists of a set of specific shape parameters and orientation parameters.

[0032] Wherein, the total weight projection error This characterizes the cumulative contour difference (e.g., symmetry difference, IoU loss, or contour distance) of the i-th sample across all shooting viewpoints. The specific calculation logic is as follows: First, for each shooting viewpoint, using the corresponding predicted contour image, calculate the local difference measure between the predicted contour image and the observed contour image at that viewpoint; then, sum the local difference measures obtained for the sample across all shooting viewpoints to obtain the total reprojection error of the sample. .

[0033] For each anatomical block, an independent search distribution for that anatomical block is generated based on the parameter subvectors belonging to that anatomical block for all samples in the original population and their respective sample weights. (i.e., global weighted mean) In the The parameter subvectors corresponding to each anatomical block are: The covariance matrix is The population is adaptively updated using a multivariate normal distribution, and the original population for the next iteration epoch is generated based on the updated independent search distribution. The adaptive update of the independent search distribution of the anatomical block includes: The normalized weights of each sample in the original population are calculated, and the centering utility of the corresponding sample is determined based on the normalized weights. Specifically, the calculation is performed using the following formula: , ; In the formula, Indicates the first The normalized weights of each sample satisfy the following condition: ; This represents the total number of samples in the population. Indicates the first The centering utility of a sample, in physical terms, is the "excess contribution" of that sample relative to the uniform distribution baseline (i.e., 1 / N). because This means that the sum of the centering utilities of all samples is zero. This property guarantees that when the total weight projection error of all samples is equal, the normalized weight of each sample is also equal. This leads to each At this point, in the subsequent calculation of the natural gradient matrix... At that time, due to all Since the natural gradient matrix is ​​zero, the parameterization factor of the covariance is no longer updated, thus maintaining the current search distribution and avoiding unnecessary perturbations. Conversely, when the total reprojection error of some samples is significantly smaller (i.e., better), its corresponding... , while the poorer samples The resulting non-zero natural gradient will guide the covariance matrix to adaptively adjust towards a better sample cluster, thus improving search efficiency.

[0034] The global weighted mean of the current iteration epoch is calculated based on the normalized weights of each sample; the calculation formula is: ; In the formula, This represents the global weighted average value of the current iteration epoch, i.e., the m-th iteration epoch.

[0035] Based on the normalized weights of each sample, the original population is randomly resampled with replacement to obtain the resampled population. The process involves random resampling with replacement of the original population based on the normalized weights of each sample. The probability of each sample being selected is its normalized weight. This process is repeated N times to obtain the resampled population. N represents the total number of samples in the original population.

[0036] For the parameter subvectors of each sample in the current original population in each anatomical block, a whitening transformation is performed based on the global weighted mean of the current iteration epoch and the parameterization factor of the covariance matrix of the block in the previous iteration epoch to obtain the whitening bias of each anatomical block; wherein, the covariance matrix is ​​characterized by a parameterization factor that matches its dimension. The whitening transformation of the parametric subvectors of each sample in each anatomical block in the current original population is performed based on the parameterization factor of the global weighted mean of the current iteration epoch and the covariance matrix of the block in the previous iteration epoch, to obtain the whitening bias of each anatomical block, including: For each anatomical block, calculate the local weighted mean of that block in the current iteration epoch. The local weighted mean is the parameter subvector corresponding to the global weighted mean in that anatomical block. Using the parameterization factor of the covariance matrix of the anatomical block in the previous iteration epoch, a whitening transformation is performed on the deviation between the parameter subvector of each sample in the current original population in the anatomical block and the local weighted mean of the block in the current iteration epoch, to obtain the whitening deviation of each sample in the current anatomical block.

[0037] The formula for calculating whitening deviation in this embodiment includes: ; satisfy: ; In the formula: Indicates the index of the current iteration epoch ( ); Indicates the first Each anatomical region; Indicates that the i-th sample is in the... The parameter subvectors corresponding to each anatomical block; that is, the parameter vector instance to be optimized for the i-th sample. The parameter subvector corresponding to the k-th anatomical block; Indicates the first Global weighted mean in iterative epochs In the The parameter subvectors corresponding to each anatomical block; Indicates the previous iteration epoch. In the middle, the first Covariance matrix of each anatomical region The parameterization factor (e.g., Cholesky factor or square root factor). Indicates the epoch of the previous iteration. Anatomical block covariance parameterization factor The inverse of the matrix; This represents the standardized deviation vector obtained after whitening transformation (i.e., whitening deviation, used to eliminate the correlation between parameters and unify the dimensions). Indicates the first In the iterative epoch, the first The covariance matrix of each anatomical region is given by Define, where To update the obtained parameterization factor; Accordingly, in the previous iteration epoch, the 1st... The covariance matrix of each anatomical region is ; Additionally, for the first iteration epoch (i.e. The parameterization factor of the covariance matrix of the previous iteration epoch adopts a preset initial value.

[0038] In this embodiment, the initial values ​​of the parameterization factors for each block are set as follows: ,in, This is the initial scale parameter, and its value covers a reasonable range of variation of the pose parameters in the corresponding anatomical block.

[0039] Since each block has a small dimension (e.g., the torso block is 24-dimensional, and the limbs and body shape blocks are no more than 12-dimensional), the computational cost of covariance update is controllable and significantly lower than that of the full-parameter joint modeling scheme, which is conducive to the efficient operation of the algorithm on conventional computing platforms.

[0040] For each anatomical block, the natural gradient matrix of that block is calculated based on the whitening bias of each sample in the original population within the current anatomical block and the centering utility of each sample in the original population; the calculation formula is as follows: ; In the formula: The superscript T indicates the transpose operation of a matrix or vector; For dimension (i.e., the first) The identity matrix (dimension of the anatomical block parameter subvectors); Indicates the epoch of the current iteration. The natural gradient matrix of each anatomical block.

[0041] This invention introduces parameterized factors to characterize the covariance matrix of each anatomical block and employs a technique of exponentially mapping and updating the parameterized factors of the previous iteration based on the natural gradient matrix, thereby achieving adaptive evolution of the covariance structure. Simultaneously, by combining whitening bias and centering utility to calculate the natural gradient matrix, the covariance update can effectively respond to the sample weight distribution, improving the robustness and adaptability of the high-dimensional anatomical block covariance learning process.

[0042] Based on the natural gradient matrix, the parameterization factor of the anatomical block in the previous iteration epoch is updated by exponential mapping to obtain the parameterization factor of the corresponding anatomical block in the current iteration epoch; the calculation formula is: ; In the formula, The learning rate represents the update of the covariance matrix. This represents the matrix exponential function.

[0043] It is worth noting that the parameterization factor is updated using a matrix exponential form, i.e.: ; It can rigorously guarantee the covariance matrix of each anatomical block at any iteration epoch m. It always remains symmetric positive definite (SPD).

[0044] This is because: initial parameterization factor It is a non-singular matrix, while its natural gradient matrix is... For a symmetric matrix, its matrix exponent is... It must be a symmetric positive definite matrix; the product of two non-singular matrices is also a non-singular matrix, thus ensuring... Non-singular, thus making It is always symmetric positive definite.

[0045] This SPD guarantee does not depend on the centering utility. The specific calculation method is as follows: If the whitening deviation is constructed into a symmetric matrix, the above properties hold.

[0046] Therefore, the covariance update mechanism adopted in this embodiment theoretically avoids the risk of covariance matrix degradation or non-positive definiteness, and ensures the numerical stability of the search distribution and the convergence of the algorithm.

[0047] Update the covariance matrix of the corresponding anatomical block based on the parameterization factor of the anatomical block in the current iteration epoch.

[0048] This invention calculates the sample weights of each sample based on the total reprojection error and annealing coefficient according to the Boltzmann distribution, and further normalizes them into normalized weights for iterative updates. These normalized weights are used to: determine the centering utility of each sample, calculate the global weighted mean, perform population resampling, and construct a natural gradient matrix in conjunction with whitening bias. Then, based on the natural gradient matrix, the parameterization factors of the covariance matrix of each anatomical block in the previous iteration epoch are updated exponentially, achieving adaptive evolution of the search distribution. This mechanism directly transforms the reprojection error into a quantitative evaluation of the sample's contribution, effectively suppressing the negative impact of high-error samples (i.e., low-quality samples) on covariance learning, thereby enhancing the algorithm's robustness in multi-view contour matching when there is occlusion or noise interference.

[0049] The process of generating the original population for the next iteration epoch based on the updated independent search distribution includes: Based on the updated covariance matrix of each anatomical block in the current iteration epoch, and combined with the preset activation weights of each anatomical block, a global perturbation covariance matrix is ​​constructed; the calculation formulas used include: ; ; ; In the formula: Indicates the first In the iterative epoch, the first The activation weight of each anatomical block, the value of which varies with epoch. The increase exhibits a monotonically non-decreasing change: proximal blocks (such as trunk blocks) take over in early epochs. Remote blocks (such as arm blocks, leg blocks) from The process begins by gradually increasing the number of epochs to 1 and maintaining it after the number of epochs reaches the preset activation threshold. Thus, the optimization process prioritizes adjusting the near-end degrees of freedom that have a greater impact on the overall pose, and then gradually introduces the far-end details after they converge, avoiding errors in the search direction caused by premature coupling of high-dimensional parameters.

[0050] This represents the operation of assembling a matrix in a block-diagonal manner; It is a block diagonal matrix consisting of K scalar-scaled identity matrices, with a total dimension of . K represents the number of anatomical blocks; the block diagonal matrix is... The diagonal pieces are ; This represents the global perturbation covariance matrix.

[0051] The perturbation vector for each sample in the resampled population is defined based on the global perturbation covariance matrix. The corresponding resampled samples are then perturbed based on these perturbation vectors to obtain the corresponding samples of the original population in the next iteration epoch. The techniques employed include: ; In the formula: This represents the survival rate parameter at the t-th iteration epoch, used to control the decay of the perturbation amplitude; It represents the cumulative product of survival rates from the first iteration epoch to the current m-th iteration epoch; it is monotonically non-increasing with each epoch and is used to simulate the annealing cooling process, so that the perturbation amplitude gradually decreases. Let i represent the perturbation vector of the i-th sample; it is obtained by using a matrix with a mean of zero (...). The samples were obtained by independent random sampling from a multivariate Gaussian distribution; for different sample indices , and They are independent of each other; This represents the i-th sample in the population after resampling. This represents the i-th sample of the original population in the next iteration epoch after the perturbation update.

[0052] During the execution of each iteration epoch, the following is also included: Determine whether there is a better sample in the current iteration epoch with a total reprojection error less than the global optimal sample. If so, update the global optimal sample and replace the worst sample with the largest total reprojection error in the original population of the current iteration epoch with the global optimal sample. The global optimal sample is the sample with the smallest total reprojection error in all historical iteration epochs.

[0053] When the preset iteration termination condition is met, the iteration stops, and the parameters of the sample with the smallest total weight projection error are determined as the optimal attitude parameters and optimal shape parameters.

[0054] The method further includes: Based on the optimal pose parameters and optimal shape parameters, a reconstructed texture image is obtained by performing texture optimization on the color image in the observed image using differentiable rendering technology; including: Construct a target human body model based on the optimal posture parameters and optimal shape parameters; A global texture image bound to the target human body model is constructed. This global texture image is obtained by mapping the target human body model to the model surface through preset and fixed texture coordinates. Using a differentiable rendering function, the target human body model is rendered into a predicted color image at each shooting angle, and based on the contour image at the corresponding angle, the L2 error between each predicted color image and the color image in the corresponding observed image is calculated in the foreground region of the human body. Using the overall loss of L2 error corresponding to each shooting angle as the loss function, the Adam optimizer is used to iteratively optimize the global texture image to obtain the reconstructed texture image.

[0055] This invention explicitly maintains historical global optimal samples during iterative optimization. When a sample with a smaller total reprojection error is detected in the current iteration epoch, the global optimal sample is updated, and the worst sample in the current population is replaced with the global optimal sample, thus ensuring that historical optimal solutions are always retained in the search process. After completing the geometric parameter optimization, a global texture image bound to the target human body model is constructed based on the finally determined optimal pose and shape parameters, and a predicted color image is generated at each shooting angle using differentiable rendering technology. Furthermore, based on the human body foreground region defined by the contour image of the corresponding viewpoint, the L2 error between the predicted color image and the observed color image is calculated, and the global texture image is iteratively optimized with the overall loss of L2 error at each viewpoint as the optimization objective. This achieves precise alignment of texture and geometry under spatial mask constraints, effectively avoiding interference from the background region on texture optimization.

[0056] In one embodiment of the present invention, after obtaining the optimal attitude parameters and optimal shape parameters Next, a texture reconstruction step is performed to generate a reconstructed texture image. This specifically includes: First, a global texture image is constructed that is bound to the target human body model (e.g., the SMPL model). The texture image is mapped onto the model surface using preset and fixed texture coordinates (UV coordinates) of the target human model in a resting pose. Since the texture coordinates of the SMPL model are bound to the resting pose, once the association between the vertices and texture pixels on the model surface is established, the vertex-texture pixel correspondence remains unchanged even if the model undergoes geometric deformation (driven by the pose parameter θ), thus ensuring the consistency of the texture under different poses.

[0057] Subsequently, based on the optimal pose parameters and optimal shape parameters, the textured target human body model is rendered using a differentiable rendering function. Rendering is performed from various shooting angles to generate a predicted color image. Differentiable rendering functions can be implemented using SoftRas, DIB-R, or other methods.

[0058] To optimize the global texture image, the following loss function is defined and minimized: ; In the formula: , representing the global texture image (i.e., texel image) to be optimized, which typically has a dimension greater than 65000; This indicates the search for the texture image that minimizes the objective function. ; , representing the total number of images observed from multiple perspectives; Differentiable rendering function, used to render a human model into a two-dimensional color image under given camera and lighting conditions, supports gradient backpropagation; A 3D human body mesh with textured coordinates defined by the target human body model, wherein: , representing shape parameters (optimal), which control the human body shape (such as height, weight, and build). , representing the posture parameters (optimal), controlling the rotation of human joints (expressed in axis angles); , representing preset and fixed texture coordinates (UV coordinates); , represents the lighting parameters under the i-th viewpoint, used to describe the lighting conditions (such as ambient light, directional light, etc.) during rendering. , represents the camera intrinsic parameter matrix corresponding to the i-th viewpoint, which includes imaging parameters such as focal length and principal point; , represents the camera rotation matrix (part of the extrinsic parameters) of the i-th viewpoint, describing the camera orientation; , represents the camera translation vector (part of the extrinsic parameters) of the i-th viewpoint, describing the camera position; , represents the color image in the observation image acquired from the i-th viewpoint, and serves as the supervised ground truth; , represents the silhouette mask of the contour image in the observation image corresponding to the i-th viewpoint. It is a binary image with the foreground (human body region) set to 1 and the background set to 0. , representing the L2 norm, is used to calculate the pixel-level color difference between the predicted image and the observed image in the foreground region; The entire expression is Differentiable: Due to differentiable rendering functions and fixed... The existence of mapping, the loss function with respect to the texture image The gradient can be calculated efficiently; Adam optimizer: used for high-dimensional texture variables Perform iterative optimization (e.g., 300 steps) to minimize the above loss function, and finally obtain the reconstructed texture image.

[0059] In this embodiment, Figure 3 This is a comparison image of contour and texture reconstruction using the method of the present invention; Figure 3 From left to right: The first sub-image is a comparison image between the predicted contour image corresponding to the initial parameter vector to be optimized and the contour image of the corresponding view in the multi-view observation image set. The second sub-image is: The comparison image between the generated predicted contour image and the contour image of the corresponding view in the multi-view observation image set after the m-th iteration epoch optimization. The third sub-image is: When the preset iteration termination condition is met, the predicted contour image corresponding to the sample with the smallest total reprojection error is compared with the contour image of the corresponding view in the multi-view observation image set. The fourth sub-image is: Based on the optimal pose parameters and optimal shape parameters, an intermediate image is generated by texture optimization of the color image corresponding to the viewpoint in the multi-view observation image set through differentiable rendering. The fifth sub-image is the reconstructed texture image.

[0060] The sixth sub-image is a color image of the corresponding viewpoint in the multi-view observation image set.

[0061] In conclusion, Figure 3This invention demonstrates that, through contour-driven iterative optimization and differentiable texture refinement, it can gradually converge to a reconstruction result with highly consistent geometry and appearance without initial pose estimation.

[0062] The human body 3D modeling method provided by this invention is based on a gradient-free collaborative optimization framework based on anatomical blocks, which effectively overcomes the technical bottlenecks of existing technologies in joint estimation of human posture and shape, such as low search efficiency, unstable high-dimensional covariance learning, and easy loss of historical optimal solutions.

[0063] Specifically, the fundamental differences and innovations of this invention compared to the prior art are reflected in the following aspects: First, in terms of parameter space organization, this invention divides the parameter vector to be optimized, composed of shape and posture parameters, into multiple anatomical blocks and establishes an independent search distribution for each block. These anatomical blocks are divided according to the hierarchical structure of human kinematics: the trunk block composed of the pelvis and spine, the arm blocks corresponding to the left and right arms respectively, the leg blocks corresponding to the left and right legs respectively, and the body shape block composed of shape parameters. Each block is configured with an independent Gaussian search distribution (based on covariance matrix representation). This design abandons the existing technology's mode of treating all parameters as a single whole for global search, and significantly improves the physical rationality and convergence efficiency of the search by decoupling the parameter optimization process of different body parts.

[0064] Secondly, in population evaluation and weight allocation, this invention abandons the symmetric weighting method that relies solely on sample ranking in traditional evolutionary strategies. Instead, it introduces a survival rate parameter and an annealing coefficient based on the total reprojection error of each sample, and calculates asymmetric sample weights using a Boltzmann distribution. This mechanism allows high-quality samples with smaller errors to receive higher weights, thus playing a dominant role in subsequent distribution updates and achieving refined exploration.

[0065] Furthermore, regarding the adaptive update mechanism for the search distribution, this invention proposes a stable and efficient update process. This process includes: calculating normalized weights and centering utility, determining the global weighted mean, performing resampling with replacement, performing whitening transformation to obtain the whitening bias of each anatomical block, and finally calculating the natural gradient matrix based on the whitening bias and centering utility. Crucially, this invention utilizes a parameterized factor to characterize the covariance matrix and updates this parameterized factor through an exponential mapping. This design avoids the numerical instability caused by directly manipulating high-dimensional covariance matrices, ensuring the robustness and positive definiteness of covariance evolution under finite sample conditions.

[0066] Finally, regarding the optimal solution maintenance strategy, this invention introduces an explicit elite retention mechanism. In each iteration epoch, the system determines whether there is a new sample superior to the current global best sample; if so, the global best sample is updated, and the worst sample in the current population is replaced with this global best sample. This mechanism forcibly retains the best solution found during the historical search process, effectively preventing the loss of the optimal solution due to random perturbations, and greatly enhancing the convergence stability of the algorithm.

[0067] In summary, this invention does not turn to gradient-based optimization. Instead, within the gradient-free global search paradigm, it constructs an efficient, stable, and physically meaningful human geometric parameter optimization scheme through the synergistic effect of four core technologies: independent search distribution of anatomical blocks, asymmetric reweighting based on Boltzmann distribution, stable covariance update based on natural gradient and exponential mapping, and an explicit elite retention mechanism. This results in a comprehensive improvement in accuracy and robustness over existing gradient-free methods.

[0068] This invention divides the parameter vector to be optimized into multiple anatomical blocks. For each anatomical block, the corresponding search distribution is adaptively updated independently based on the parameter sub-vectors belonging to that anatomical block and their respective sample weights of all samples in the original population. Then, the next generation of the original population is generated based on the updated independent search distribution of each block. This makes the parameter optimization process of each anatomical block independent and undependent, avoiding the invalid search caused by the coupling of irrelevant parameters when updating all parameters as a whole. Thus, it improves the search efficiency and the stability of parameter estimation in the high-dimensional human body parameter space.

[0069] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of this invention.

[0070] The present invention also provides a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of the present invention.

[0071] This invention also provides a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of this invention.

[0072] refer to Figure 4The present invention will now describe a structural block diagram of an electronic device that can serve as a server or client in embodiments of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0073] like Figure 4 As shown, the electronic device includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the electronic device. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0074] Multiple components in the electronic device are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information into the electronic device. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0075] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the methods and processes described above. For example, in some embodiments, the method embodiments of the present invention may be implemented as a computer program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on an electronic device via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 may be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0076] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0077] In the context of embodiments of the present invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0078] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of the components in a specific posture (as shown in the attached figures). If the specific posture changes, the directional indication will also change accordingly. Furthermore, descriptions involving "first," "second," or "a" in the present invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. In the description of the present invention, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly defined. In the present invention, unless otherwise explicitly specified and defined, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can be a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal communication of two components or the interaction between two components, unless otherwise explicitly defined. For those skilled in the art, the specific meanings of the above terms in this invention can be understood according to the specific circumstances. Furthermore, the technical solutions of the various embodiments of this invention can be combined with each other, but only on the basis that those skilled in the art can implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

Claims

1. A method for three-dimensional human body modeling based on anatomical block optimization, characterized in that, include: Acquire a multi-view observation image set of the human body to be modeled. The image set contains observation images from multiple shooting angles, and the observation images include contour images. Obtain a human body model controlled by shape parameters and posture parameters, and concatenate the shape parameters and posture parameters into a parameter vector to be optimized; The parameter vector to be optimized is divided into multiple anatomical blocks, and an independent search distribution corresponding to each anatomical block is established; Generate an original population containing multiple samples, where each sample is an instance of the parameter vector to be optimized; The original population is iteratively optimized, and the following operations are performed in each iteration epoch: Based on the human body model, each sample in the original population is rendered as a predicted contour image from each shooting angle. For each sample, the reprojection error between each predicted contour image and the contour image at the corresponding viewpoint is calculated, and the total reprojection error of the sample is obtained by summing them up. The sample weight of each sample in the original population is calculated based on the total reprojection error of each sample. For each anatomical block, the independent search distribution of the anatomical block is adaptively updated based on the parameter subvectors of all samples in the original population belonging to that anatomical block and their respective sample weights, and the original population for the next iteration epoch is generated based on the updated independent search distribution. When the preset iteration termination condition is met, the iteration stops, and the parameters of the sample with the smallest total weight projection error are determined as the optimal attitude parameters and optimal shape parameters.

2. The human body 3D modeling method based on anatomical block optimization according to claim 1, characterized in that, The step of dividing the parameter vector to be optimized into multiple anatomical blocks and establishing an independent search distribution corresponding to each anatomical block includes: Based on the hierarchical structure of human kinematics, the parameter vector to be optimized is divided into multiple anatomical blocks; the anatomical blocks include: the trunk block composed of the pelvis and spine, the arm blocks corresponding to the left and right arms respectively, the leg blocks corresponding to the left and right legs respectively, and the body shape block composed of shape parameters. For each anatomical block, an independent search distribution following a Gaussian distribution is established; the independent search distribution is configured with a corresponding covariance matrix, the dimension of which matches the dimension of the parameter sub-vector of the anatomical block.

3. The human body 3D modeling method based on anatomical block optimization according to claim 1, characterized in that, The process of calculating the reprojection error between each predicted contour image and the contour image at the corresponding viewpoint, and summing them to obtain the total reprojection error of the sample, includes: For the current sample, at each shooting viewpoint, the reprojection error between the predicted contour image and the observed contour image at that viewpoint is calculated based on the contour difference metric. The total reprojection error of the sample is obtained by summing the reprojection errors of the sample under all shooting angles.

4. The human body 3D modeling method based on anatomical block optimization according to claim 1, characterized in that, The calculation of the sample weight of each sample in the original population based on the total reprojection error of each sample includes: For the current iteration epoch, obtain the preset survival rate parameters and the total number of samples in the original population; Based on the survival rate parameter and the total number of samples, an annealing coefficient is determined to adjust the sharpness of the weight distribution; Based on the annealing coefficient and the total reprojection error of each sample, the sample weight of each sample is calculated using the Boltzmann distribution.

5. The human body 3D modeling method based on anatomical block optimization according to claim 1, characterized in that, The adaptive update of the independent search distribution of the anatomical block includes: Calculate the normalized weights of each sample based on the sample weights of each sample in the original population, and determine the centering utility of the corresponding sample based on the normalized weights. Calculate the global weighted mean of the current iteration epoch based on the normalized weights of each sample; Based on the normalized weights of each sample, the original population is randomly resampled with replacement to obtain the resampled population. For the parameter subvectors of each sample in the current original population in each anatomical block, a whitening transformation is performed based on the global weighted mean of the current iteration epoch and the parameterization factor of the covariance matrix of the block in the previous iteration epoch to obtain the whitening bias of each anatomical block; wherein, the covariance matrix is ​​characterized by a parameterization factor that matches its dimension. For each anatomical block, the natural gradient matrix of the block is calculated based on the whitening bias of each sample in the original population in the current anatomical block and the centering utility of each sample in the original population. Based on the natural gradient matrix, the parameterization factor of the anatomical block in the previous iteration epoch is updated by exponential mapping to obtain the parameterization factor of the corresponding anatomical block in the current iteration epoch. Update the covariance matrix of the corresponding anatomical block based on the parameterization factor of the anatomical block in the current iteration epoch.

6. The human body 3D modeling method based on anatomical block optimization according to claim 5, characterized in that, The process of generating the original population for the next iteration epoch based on the updated independent search distribution includes: Based on the updated covariance matrix of each anatomical block in the current iteration epoch, and combined with the preset activation weights of each anatomical block, a global perturbation covariance matrix is ​​constructed. Based on the global perturbation covariance matrix, the perturbation vector of each sample in the resampled population is set. Based on the perturbation vector, the corresponding resampled samples are perturbed to obtain the corresponding samples of the original population in the next iteration epoch.

7. A method for three-dimensional human body modeling based on anatomical segmentation optimization according to claim 5, characterized in that, The whitening transformation of the parametric subvectors of each sample in each anatomical block in the current original population is performed based on the parameterization factor of the global weighted mean of the current iteration epoch and the covariance matrix of the block in the previous iteration epoch, to obtain the whitening bias of each anatomical block, including: For each anatomical block, calculate the local weighted mean of that block in the current iteration epoch. The local weighted mean is the parameter subvector corresponding to the global weighted mean in that anatomical block. Using the parameterization factor of the covariance matrix of the anatomical block in the previous iteration epoch, a whitening transformation is performed on the deviation between the parameter subvector of each sample in the current original population in the anatomical block and the local weighted mean of the block in the current iteration epoch, to obtain the whitening deviation of each sample in the current anatomical block.

8. The human body 3D modeling method based on anatomical block optimization according to claim 1, characterized in that, During the execution of each iteration epoch, the following is also included: Determine whether there is a better sample in the current iteration epoch with a total reprojection error less than the global optimal sample. If so, update the global optimal sample and replace the worst sample with the largest total reprojection error in the original population of the current iteration epoch with the global optimal sample. The global optimal sample is the sample with the smallest total reprojection error in all historical iteration epochs.

9. A method for three-dimensional human body modeling based on anatomical segmentation optimization according to claim 1, characterized in that, The observed image further includes a color image; the method further includes: Based on the optimal pose parameters and optimal shape parameters, a reconstructed texture image is obtained by texture optimization of the color image in the observed image using differentiable rendering technology; including: Construct a target human body model based on the optimal posture parameters and optimal shape parameters; A global texture image bound to the target human body model is constructed. This global texture image is obtained by mapping the target human body model to the model surface through preset and fixed texture coordinates. Using a differentiable rendering function, the target human body model is rendered into a predicted color image at each shooting angle, and based on the contour image at the corresponding angle, the L2 error between each predicted color image and the color image in the corresponding observed image is calculated in the foreground region of the human body. Using the overall loss of L2 error corresponding to each shooting angle as the loss function, an optimizer is used to iteratively optimize the global texture image to obtain the reconstructed texture image.

10. An electronic device, comprising: A processor, and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform a human body three-dimensional modeling method based on anatomical block optimization according to any one of claims 1 to 9.

Images