Personalized modeling and posture deformation methods for whole-body anatomical structures in human biomechanical simulation

Abstract

This invention discloses a personalized modeling and posture deformation method for the whole-body anatomical structure in human biomechanical simulation, belonging to the field of personalized modeling technology. First, this invention obtains the relative rotation information between joints during human movement through a neural network model. Based on the hierarchical relationship of human joints and the topology of the skeleton, it realizes the posture deformation of the internal anatomical structure of the human body from bottom to top. By removing cross-interferences after posture adjustment using a de-crossing algorithm and introducing a smooth deformation mechanism using an optimized skinning algorithm, the realism of the personalized human model during movement is ensured. The model reconstruction results are optimized by combining body circumference information and mapping of muscle force lines, ultimately constructing a personalized human biomechanical simulation model containing the whole-body anatomical structure. This achieves realistic and reasonable deformation of the internal anatomical structure of the human body during posture deformation, making it applicable to biomechanical simulations of different individuals and postures.

Description

Technical Field

[0001] This invention belongs to the field of personalized modeling technology, specifically involving a personalized modeling method for the whole-body anatomical structure and a posture deformation method for human biomechanical simulation models. It is mainly applicable to the construction of personalized human biomechanical simulation models, achieving realistic and reasonable deformation of the internal anatomical structure when the posture changes. Background Technology

[0002] In recent years, the demand for personalized human body modeling has been increasing. Three-dimensional human body models, incorporating internal anatomical structures, are used in many fields, such as anatomy teaching, dynamics simulation, and electromagnetic radiation simulation. Personalized human body modeling is not only widely used in industrial simulation and the medical field, but is also gradually being applied in daily life. With the continuous development of virtual reality technology, the application of virtual characters is increasing. People hope to create personalized virtual characters that can simulate their own movements in real time, connecting the virtual and the real. For example, based on the changes in posture during human movement, personalized full-body anatomical structures can be simulated in real time using biomechanical modeling. When exercising at home or undergoing sports training without professional guidance, the different postures of the personalized human body model can be evaluated, and compared with the full-body anatomical structure under standard movements. This can identify incorrect movements and potential damage, providing real-time intelligent exercise guidance and standardizing people's daily exercise and healthy lifestyle.

[0003] Personalized modeling methods for human biomechanical simulation models generally involve motion capture systems, where users wear specialized sensing devices and collect data via multi-view cameras. While this traditional method can yield high-quality human model reconstructions and capture some human biomechanical features, it requires extensive manual calibration and time-consuming processing of data sampled from specific individuals, making it unsuitable for everyday life. Deep neural networks, which have emerged in recent years, excel in human pose estimation and 3D human reconstruction, often used in sports and fitness fields for 3D reconstruction of humans from videos or images. However, this method focuses more on the external skin model and the posture and movements of the human figure, neglecting the complex and diverse internal anatomical structures of the 3D human body. It lacks research on the deformation of internal anatomical structures and cannot effectively model human biomechanical movements. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a personalized modeling method for the whole-body anatomical structure and a posture deformation method for human biomechanical simulation models. The relative rotation information between joints during human movement is obtained through a neural network model. Based on the hierarchical relationship of human joints and the topology of the skeleton, posture transformation of the internal anatomical structures of the human body is realized from bottom to top. Simultaneously, a decrossing algorithm removes crossovers after posture adjustment, and an optimized skinning algorithm introduces a smooth deformation mechanism to ensure the realism of the personalized human model's movement. Muscle force lines are mapped onto the personalized human model to construct a personalized human biomechanical simulation model.

[0005] The technical solution of the present invention:

[0006] Research on a method for personalized modeling and posture deformation of whole-body anatomical structures for human biomechanical simulation models, the method includes the following steps:

[0007] Step A: Personalized Pose Deformation

[0008] Step A1, Anatomical Joints

[0009] First, based on the definition of human anatomy, 21 skeletal joints were defined for the motion skeleton of a personalized deformable digital human in its standard state: pelvis, lumbar vertebrae, thoracic vertebrae, cervical vertebrae, right shoulder, right elbow, right wrist, right fingertip, left shoulder, left elbow, left wrist, left fingertip, right hip, right knee, right toe, left hip, left knee, left toe, left clavicle, right clavicle, and top of the head. Based on these skeletal joint positions and combined with prior anatomical knowledge of each joint, the normal rotation angle of the skeleton was simulated, and the rationality of the results after skeletal rotation transformation was judged. The skeletal joints were then adjusted using modeling software. The above process was then repeated on the adjusted skeletal joints to finally obtain the positions of the anatomical joints.

[0010] Step A2, Skeleton Topology

[0011] Based on the anatomical joints determined in step A1, it is also necessary to determine the inheritance relationship between the skeletal joints according to the skeletal topology, and represent the joint hierarchy as parent joints and child joints. During human movement, child joints only rotate relative to parent joints, without translation. Therefore, the Euclidean distance between parent and child joints remains constant, which is consistent with the objective fact that the length of the human skeleton is constant. By considering skeletal rotation from bottom to top, starting from the root node of the skeleton, the parent and child bones of each bone, as well as the root and tail nodes of each bone, are identified. The lowest-ranking child joint moves first, then moves relative to its parent joint, and then moves its child joints relative to its parent joint. The changes to the child joints are equivalent to accumulation. After all bones have rotated around the joints, the description of the corresponding human posture and movement is achieved.

[0012] Step A3, Remove Cross Correction

[0013] Based on the skeleton topology in step A2, after all bones from child bones to parent bones complete the global transformation, the human posture transformation result is obtained. Since the joints and rotation axes of bone rotation cannot strictly simulate the complex rotation and sliding process of human joints, it may also cause intersections between adjacent bones; or intersections between blood vessels, soft tissues and bones, and skin. Therefore, it is necessary to perform de-intersection operation to eliminate possible surface intersections and generate a model usable for simulation.

[0014] The method for de-crossing all anatomical surfaces in the human pose transformation result is as follows: First, let the two surfaces that need to be cross-checked be... and Both must ensure a closed mesh surface; traversal Let any vertex be an arbitrary vertex to be tested. Draw n rays starting from vertex v, and check the intersection of each ray with... The number of intersections, and checking whether each vertex is in Internal; if all n rays are inter... The number of intersection points of vertex v is odd, indicating that vertex v lies in... Inside, or outside;

[0015] turn up The points that need adjustment are as follows: If the object being tested is between soft tissue and bone, or between bones, then the points that need adjustment are... Central Internal points; if the object being tested is between soft tissue and skin, then the points that need to be adjusted are... Central External points; the points that need adjustment are located on the surface. External or internal; through each point that needs adjustment, do... For all the perpendiculars to the triangular facets, if the foot of the perpendicular falls inside the triangular facet, record the distance from the foot of the perpendicular to the point that needs adjustment. Select the point with the shortest distance among all distances as the updated coordinates of the point that needs adjustment. The updated coordinates of each point that needs adjustment are:

[0016]

[0017] in, Let k be the point that needs adjustment, and k be the foot of the perpendicular when the target point is at its shortest distance.

[0018] After failing to detect the point requiring adjustment, the surface... and Perform triangular face intersection detection to eliminate potential face intersections; traverse... Let all triangular faces be given. for For any triangular facet, detect the intersection of its three sides with the surrounding surface. Any triangular facet The intersection relationship; as long as Any one of the three edges is connected to... If they intersect, then they are considered as two triangular faces. and Intersecting; for All of the above The vertices of intersecting triangular faces need to be repositioned until no intersecting triangular faces are detected.

[0019] Step A4: Skin Optimization

[0020] First, the skinning result is achieved using a linear hybrid skinning method. The skin surface vertices obtained from statistical learning in the human silent pose are bound to each bone joint, with each skin surface vertex potentially attached to one or more bones (generally a maximum of four). Constrained least squares is used to estimate all skin surface vertices, generating linear skinning weights. Based on the bone pose transformation and the obtained linear skinning weights, the position transformation of the skin surface vertices is represented as follows:

[0021]

[0022] in, This represents the transformed vertex position, where N is the number of attached bones. It is the mixed weight matrix of the vertex to the nth attached bone. It is the skeleton transformation matrix. The initial position of the vertex;

[0023] Currently, the widely used skinning methods in experiments are linear blend skinning and dual quaternion blending, both of which have high computational efficiency in practical applications. Based on linear blend skinning, a method for skinning optimization through rotation centers is introduced to optimize the skinning deformation results. The rotation centers of each skin surface vertex are calculated, and points on the skin surface with similar weight distributions to the linear skinning are identified. Using the dual quaternion skinning transformation method, the rotational transformation in skinning deformation is calculated by superimposing rotation matrices from quaternions. Based on the linear transformation results of the calculated rotation centers of the skin surface vertices and the transformation results obtained by superimposing rotation matrices from quaternions, the translational transformation in skinning deformation is calculated. This process is expressed as:

[0024]

[0025] Where t is the translation amount, It is a linear transformation of the center of rotation. It is a rotation matrix transformation of a superposition of four tuples;

[0026] Step B: Personalized modeling of the human biomechanical simulation model, combined with the actual body circumference dimensions measured in the actual biomechanical simulation, to make the personalized modeling results of the human biomechanical simulation model more accurate.

[0027] Step B1: Reconstruction of the 3D Human Body Model

[0028] First, feature extraction is performed on each frame of the image in the photo or video using a neural network model. The spatial coordinates of the anatomical joints defined in step A1 under different postures are then regressed. Then, the parent node defined for each bone joint in the skeleton topology in step A2 is combined with the rotation angle of each bone joint relative to its parent node to obtain the relative transformation matrix of each bone under different human postures.

[0029] Then, using the personalized pose deformation algorithm in step A, the relative transformation matrix of each bone is converted into an absolute transformation matrix. Based on the absolute transformation matrix of the bones and the human mean template obtained through statistical learning, the human body is transformed from a static pose to the current pose. Finally, the pose transformation results are decrossed and the skinning results are optimized to reconstruct the three-dimensional human body model of the athlete.

[0030] Step B2, Body circumference measurement

[0031] Based on the reconstructed 3D human body model of the athlete in step B1, body circumference is measured. First, on the human skin mean model obtained by statistical learning, the position information of the feature points when measuring each circumference of the human body is defined according to different measurement standards. Based on the deformation field constructed by the vertex displacement between the reconstructed 3D human body model and the human mean model, the deformation result of the body circumference feature points is interpolated by 3D spatial interpolation.

[0032] Since body circumference feature points are sparsely distributed in the human body model, to ensure the interpolation effect, the nearest points among all feature points in each circumference are selected and combined into a point set. Based on the deformation field of the vertices of the nearest points in each circumference within the point set, Gaussian interpolation is used to interpolate the circumference feature points. The absolute distance between two points is calculated as follows:

[0033]

[0034] Where d is a point a in three-dimensional space ( ) and b( Calculate the Euclidean distance between all vertices in the model and the current gravitational feature point. After arranging them in ascending order of distance, select the vertex coordinates of the closest points, take the union of the set to form a point set, and perform Gaussian interpolation.

[0035] Feature points such as the top of the head, pelvis, and soles of the feet are redefined. The model height is obtained based on the distance between these feature points after interpolation and registration. The actual height data of the athlete is collected to obtain the true proportion of the reconstructed model. The body posture information is obtained based on the actual weight data of the athlete, and the actual body circumference of the athlete is estimated. Based on the estimated body circumference, combined with the body circumference measured during biomechanical simulation, the results of personalized modeling are further optimized to make the reconstructed human body model more accurate.

[0036] Step B3, Muscle Force Line Mapping

[0037] In order for the constructed human biomechanical simulation model to perform biomechanical simulation, muscle force lines need to be mapped on the reconstructed three-dimensional human body model of the athlete; based on the musculoskeletal model in the existing biomechanical simulation experiment, the relative positional relationship of the father and son bones in the human anatomical model, and the relative positions of the muscle origin points and insertion points on each bone are sorted out.

[0038] During muscle force line mapping, the skeletal model obtained through statistical learning is first uniformly resampled. Based on the calibration points, the muscle and skeletal model point cloud is registered to the standard skeletal model of the digital human body obtained through statistical learning through nonlinear spatial transformation, thus completing the registration of the skeletal model and the mapping of muscle force lines.

[0039] The beneficial effects of this invention are as follows: This invention studies the construction of personalized human biomechanical simulation models and the optimization methods for model posture deformation. Personalized biomechanical models can be constructed for different individuals for use in biomechanical simulations. Under the optimized posture deformation algorithm, regardless of the angle of posture joint transformation, the personalized model can obtain realistic motion deformation results from internal anatomical structures to the external epidermis. Furthermore, by measuring the body circumference of athletes and constructing personalized human biomechanical simulation models through photos and videos, the invention assesses body condition and different posture movements, identifies incorrect movements during exercise and potential damage, and provides reasonable exercise training suggestions and intelligent exercise guidance. Attached Figure Description

[0040] Figure 1 This is a flowchart of the positioning method described in this invention.

[0041] Figure 2 This is a flowchart of the digital human body personalized pose deformation algorithm construction process.

[0042] Figure 3 It is a flowchart for personalized modeling of human biomechanical simulation models.

[0043] Figure 4 The flowchart illustrates the posture deformation algorithm and the method for constructing a human biomechanical simulation model described in this invention, using a video of an athlete as an example. Detailed Implementation

[0044] The present invention proposes a method for personalized modeling of whole-body anatomical structures and a method for posture deformation in human biomechanical simulation models. Figure 1 As shown, the invention mainly consists of two parts: a method for personalized posture deformation of the digital human body; and a method for personalized modeling of the human biomechanical simulation model. The invention will be further explained in conjunction with specific implementation methods.

[0045] Step A: Research on personalized pose deformation algorithms, such as Figure 2 As shown.

[0046] Step A1: Definition of Anatomical Joints

[0047] Based on the definition of human anatomy, 21 skeletal joints were defined for the motion skeleton of a personalized deformable digital human in a standard state. These joints are: pelvis, lumbar vertebrae, thoracic vertebrae, cervical vertebrae, right shoulder, right elbow, right wrist, right fingertip, left shoulder, left elbow, left wrist, left fingertip, right hip, right knee, right toe, left hip, left knee, left toe, left clavicle, right clavicle, and top of the head.

[0048] Based on traditionally defined joint locations and incorporating prior knowledge of skeletal anatomy, the relative rotation between bones is simulated. By assessing the relative rotation, modeling software is used to adjust the joint locations. After testing the rationality of the new joint point bone rotation transformations, the anatomical joint locations are determined.

[0049] The number of joints selected in this invention is not fixed at 21; it can be added or removed according to requirements and algorithm performance. The anatomical feature points determined in this step serve as the relative rotation centers of each joint during subsequent posture changes.

[0050] Step A2, Skeleton Topology

[0051] Based on the definition of joints, it is also necessary to determine the inheritance relationship between related points according to the skeleton topology, and represent the joint hierarchy with parent joints and child joints. Starting from the root node of the skeleton, identify the parent and child bones of each bone, as well as the root and tail nodes of each bone, from top to bottom.

[0052] During posture deformation, the lowest-ranking sub-joint defined in the skeletal topology and joint hierarchy moves first, then moves relative to its parent joint, and then moves its sub-joints relative to its parent joint, accumulating the changes of the sub-joints; after all bones have rotated around the joints, the description of the corresponding human posture and movement is achieved.

[0053] Step A3, Remove Cross Correction

[0054] The methods for detecting whether patches intersect between surfaces and removing intersections between patches are as follows:

[0055] First, let the two surfaces that need to be cross-checked be... and Both refer to the curved surfaces of organs within the body's anatomical structure, and these surfaces must be closed mesh surfaces in three-dimensional space. (Traversal) Let any vertex be an arbitrary vertex to be tested. Draw n rays starting from v, and check the relationship between each ray and v. The number of intersections, and checking whether each vertex is in Internal. If all n rays are intersected... The number of intersection points of v is odd, indicating that v is in It is either inside or outside. In this project, n=10 can satisfy all test cases.

[0056] turn up The points that need adjustment. Taking blood vessels, bones, and skin as examples, if the object being tested is between blood vessels and bones or between bones, then the points that need adjustment are: Central Internal points; if the object being tested is between a blood vessel and the skin, then the points that need to be adjusted are... Central External points; the points that need adjustment are located on the surface. Whether it's the outside or the inside; determine if the point that needs adjustment is located on the surface. Is it external or internal? For each point that needs adjustment. ,go through Do Record the perpendiculars of all triangular faces if the foot of the perpendicular lies inside the triangle. The distance to the foot of the perpendicular. Take the shortest of all distances as... arrive Find the shortest distance, and then take the foot of the perpendicular k corresponding to the shortest distance as... The target point, then The coordinates should be updated to:

[0057]

[0058] If no point requiring adjustment is detected, perform a patch cross-check. Iterate through the area. Let all triangular faces be given. for For any given face, check the three edges of each face. Any of the above face pieces The intersection relationship. As long as any one of the three edges is with... If they intersect, then they are considered to be two and Intersecting. For All of the above The vertices of intersecting faces are then repositioned until no intersecting faces are detected, thus completing the de-intersection correction.

[0059] Step A4: Skin Optimization

[0060] First, the skinning result is achieved using a linear blending skinning method. The skin surface vertices obtained from statistical learning of the human body's silent pose are bound to the bone pose; each skin surface vertex can be bound to one or more bones. Constrained least squares is used to estimate all vertices, generating linear skinning weights, where the weights of all bones for a vertex sum to 1. Based on the bone pose transformation and the obtained skinning weights, the position transformation of the model vertices can be expressed as:

[0061]

[0062] in, This represents the transformed vertex position, where N is the number of attached bones. It is the mixed weight matrix of the vertex to the nth attached bone. It is the skeleton transformation matrix. This represents the initial position of the vertex.

[0063] Based on linear hybrid skinning, a method for skinning optimization through rotation centers is introduced to optimize the skinning deformation results. The rotation centers of each skin surface vertex are calculated, and points on the skin surface with similar linear skinning weight distributions are identified. Using the dual quaternion skinning transformation method, the rotational transformation in skinning deformation is calculated by superimposing rotation matrices from quaternions. Based on the linear transformation results of the calculated rotation centers of the skin surface vertices and the transformation results obtained by superimposing rotation matrices from quaternions, the translational transformation in skinning deformation is calculated. This process is expressed as:

[0064]

[0065] Where t is the translation amount. It is a linear transformation of the center of rotation. It is a rotation matrix transformation of a superposition of quaternions.

[0066] Step B: Personalized modeling of the human biomechanical simulation model, such as... Figure 3 As shown.

[0067] Personalized biomechanical simulation modeling for athletes involves optimizing the model reconstruction results through body circumference measurement and mapping of muscle force lines. First, a neural network model is used to detect anatomical joint information and joint rotation angles in real-time motion posture of the athlete. Based on the personalized human posture deformation algorithm in step A, a personalized 3D human model is reconstructed. The reconstructed human model is optimized by combining the athlete's actual circumference measurements with the estimated circumference measurements from the reconstructed model. Finally, muscle force line mapping is performed to achieve personalized biomechanical simulation modeling.

[0068] Step B1: Reconstruction of the 3D Human Body Model

[0069] The 3D human body model reconstruction uses Frankmocap and Vibe networks as base models, respectively, to extract feature points from photos and videos, obtaining relative transformation information between joints in different human postures in the photos or videos. Then, the researched posture deformation algorithm is used for cross-cutting correction and skinning result optimization to obtain the reconstructed 3D human body model of the athlete.

[0070] Step B2, Body circumference measurement

[0071] First, based on the human skin mean model obtained through statistical learning, the position information of the feature points when measuring various human circumferences is defined according to different measurement standards, which serves as the definition of each circumference size;

[0072] Based on the deformation field constructed by the vertex displacement between the reconstructed three-dimensional human body model surface and the average human body model surface in step B1, the deformation results of the body circumference feature points are interpolated by three-dimensional spatial interpolation, and the body circumference is measured.

[0073] Because body circumference feature points are sparsely distributed in human models, in order to ensure the interpolation effect...

[0074] Within each dimensional region, select the set of points that have the closest Euclidean distance to the current dimensional feature point, and use the deformed vectors of the vertices in the set to interpolate the dimensionality.

[0075] Feature points such as the top of the head, pelvis, and soles of the feet are redefined. The model height is obtained based on the distance between these feature points after interpolation and registration. The actual height data of the athlete is collected to obtain the true proportion of the reconstructed model. The body posture information is obtained based on the actual weight data of the athlete, and the actual body circumference of the athlete is estimated. Based on the estimated body circumference, combined with the body circumference measured during biomechanical simulation, the results of personalized modeling are further optimized to make the reconstructed human body model more accurate.

[0076] Step B3, Muscle Force Line Mapping

[0077] First, based on the musculoskeletal model in existing biomechanical simulation experiments, the parent-child relationships and positional information between bones, as well as the relative positions of muscle origins and insertions on the bones, are extracted. Since the number of vertices and faces in the skeletal model obtained through statistical learning is greater than that in the musculoskeletal model used in biomechanical simulation, the statistically learned skeletal model is first uniformly resampled. Then, the skeleton of the human anatomy model is registered to the standard skeletal model of the digital human body model through a nonlinear spatial transformation. Multiple sets of calibration points for the same locations are identified in the anatomical model and the standard template. Based on the spatial transformation between these calibration points, the point cloud to be registered is mapped to the corresponding spatial position in the target point cloud, completing the registration of the skeletal model and the mapping of muscle force lines.

Claims

1. A method for personalized modeling and posture deformation of the whole-body anatomical structure for human biomechanical simulation, characterized in that, The method includes the following steps: Step A: Personalized Pose Deformation Step A1, Anatomical Joints First, based on the definition of human anatomy, 21 skeletal joints were defined for the motion skeleton of a personalized deformable digital human in its standard state: pelvis, lumbar vertebrae, thoracic vertebrae, cervical vertebrae, right shoulder, right elbow, right wrist, right fingertip, left shoulder, left elbow, left wrist, left fingertip, right hip, right knee, right toe, left hip, left knee, left toe, left clavicle, right clavicle, and top of the head. Based on these skeletal joint positions and combined with prior anatomical knowledge of each joint, the normal rotation angle of the skeleton was simulated, and the rationality of the results after skeletal rotation transformation was judged. The skeletal joints were then adjusted using modeling software. The above process was then repeated on the adjusted skeletal joints to finally obtain the positions of the anatomical joints. Step A2, Skeleton Topology Based on the anatomical joints determined in step A1, it is also necessary to determine the inheritance relationship between the skeletal joints according to the skeletal topology, and represent the joint hierarchy as parent joints and child joints. During human movement, child joints only rotate relative to parent joints, without translation. Therefore, the Euclidean distance between parent and child joints remains constant, which is consistent with the objective fact that the length of the human skeleton is constant. By considering skeletal rotation from bottom to top, starting from the root node of the skeleton, the parent and child bones of each bone, as well as the root and tail nodes of each bone, are identified. The lowest-ranking child joint moves first, then moves relative to its parent joint, and then moves its child joints relative to its parent joint. The changes to the child joints are equivalent to accumulation. After all bones have rotated around the joints, the description of the corresponding human posture and movement is achieved. Step A3, Remove Cross Correction Based on the skeleton topology in step A2, after all bones from child bones to parent bones complete the global transformation, the human posture transformation result is obtained. Since the joints and rotation axes of bone rotation cannot strictly simulate the complex rotation and sliding process of human joints, it may also cause intersections between adjacent bones; or intersections between blood vessels, soft tissues and bones, and skin. Therefore, it is necessary to perform de-intersection operation to eliminate possible surface intersections and generate a model usable for simulation. The method for de-crossing all anatomical surfaces in the human pose transformation result is as follows: First, let the two surfaces that need to be cross-checked be... and Both must ensure a closed mesh surface; traversal Let any vertex be an arbitrary vertex to be tested. Draw n rays starting from vertex v, and check the intersection of each ray with the vertex v. The number of intersections, and checking whether each vertex is in Internal; if all n rays are inter... The number of intersection points of vertex v is odd, indicating that vertex v lies in... Inside, or outside; turn up The points that need adjustment are as follows: If the object being tested is between soft tissue and bone, or between bones, then the points that need adjustment are... Central Internal points; if the object being tested is between soft tissue and skin, then the points that need to be adjusted are... Central External points; the points that need adjustment are located on the surface. External or internal; through each point that needs adjustment, do... For all the perpendiculars to the triangular facets, if the foot of the perpendicular falls inside the triangular facet, record the distance from the foot of the perpendicular to the point that needs adjustment. Select the point with the shortest distance among all distances as the updated coordinates of the point that needs adjustment. The updated coordinates of each point that needs adjustment are: in, Let k be the point that needs adjustment, and k be the foot of the perpendicular when the target point is at its shortest distance. After failing to detect the point requiring adjustment, the surface... and Perform triangular face intersection detection to eliminate potential face intersections; traverse... Let all triangular faces be given. for For any given triangular facet, check the intersection of its three sides with the surrounding surface. Any triangular facet The intersection relationship; as long as Any one of the three edges is connected to... If they intersect, then they are considered as two triangular faces. and Intersecting; for All of the above The vertices of intersecting triangular faces need to be repositioned until no intersecting triangular faces are detected. Step A4: Skin Optimization First, the skinning result is achieved using a linear hybrid skinning method. Then, the skin surface vertices obtained from statistical learning in the human silent pose are bound to each bone joint, with each skin surface vertex potentially attached to one or more bones. Constrained least squares is used to estimate all skin surface vertices, generating linear skinning weights. Based on the bone pose transformation and the obtained linear skinning weights, the position transformation of the skin surface vertices is expressed as: in, This represents the transformed vertex position, where N is the number of attached bones. It is the mixed weight matrix of the vertex to the nth attached bone. It is the skeleton transformation matrix. The initial position of the vertex; Based on linear hybrid skinning, a method for skinning optimization through rotation centers is introduced to optimize the skinning deformation results. The rotation centers of each skin surface vertex are calculated, and points on the skin surface with similar linear skinning weight distributions are identified. Using the dual quaternion skinning transformation method, the rotational transformation in skinning deformation is calculated by superimposing rotation matrices from quaternions. Based on the linear transformation results of the calculated rotation centers of the skin surface vertices and the transformation results obtained by superimposing rotation matrices from quaternions, the translational transformation in skinning deformation is calculated. This process is expressed as: Where t is the translation amount, It is a linear transformation of the center of rotation. It is a rotation matrix transformation of a superposition of four tuples; Step B: Personalized modeling of the human biomechanical simulation model, combined with the actual body circumference dimensions measured in the actual biomechanical simulation, to make the personalized modeling results of the human biomechanical simulation model more accurate. Step B1: Reconstruction of the 3D Human Body Model First, feature extraction is performed on each frame of the image in the photo or video using a neural network model. The spatial coordinates of the anatomical joints defined in step A1 under different postures are then regressed. Then, the parent node defined for each bone joint in the skeleton topology in step A2 is combined with the rotation angle of each bone joint relative to its parent node to obtain the relative transformation matrix of each bone under different human postures. Then, using the personalized pose deformation algorithm in step A, the relative transformation matrix of each bone is converted into an absolute transformation matrix. Based on the absolute transformation matrix of the bones and the human mean template obtained through statistical learning, the human body is transformed from a static pose to the current pose. Finally, the pose transformation results are decrossed and the skinning results are optimized to reconstruct the three-dimensional human body model of the athlete. Step B2, Body circumference measurement Based on the reconstructed 3D human body model of the athlete in step B1, body circumference is measured. First, on the human skin mean model obtained by statistical learning, the position information of the feature points when measuring each circumference of the human body is defined according to different measurement standards. Based on the deformation field constructed by the vertex displacement between the reconstructed 3D human body model and the human mean model, the deformation result of the body circumference feature points is interpolated by 3D spatial interpolation. Since body circumference feature points are sparsely distributed in the human body model, to ensure the interpolation effect, the nearest points among all feature points in each circumference are selected and combined into a point set. Based on the deformation field of the vertices of the nearest points in each circumference within the point set, Gaussian interpolation is used to interpolate the circumference feature points. The absolute distance between two points is calculated as follows: Where d is a point a in three-dimensional space ( ) and b( Calculate the Euclidean distance between all vertices in the model and the current gravitational feature point. After arranging them in ascending order of distance, select the vertex coordinates of the closest points, take the union of the set to form a point set, and perform Gaussian interpolation. Feature points such as the top of the head, pelvis, and soles of the feet are redefined. The model height is obtained based on the distance between these feature points after interpolation and registration. The actual height data of the athlete is collected to obtain the true proportion of the reconstructed model. The body posture information is obtained based on the actual weight data of the athlete, and the actual body circumference of the athlete is estimated. Based on the estimated body circumference, combined with the body circumference measured during biomechanical simulation, the results of personalized modeling are further optimized to make the reconstructed human body model more accurate. Step B3, Muscle Force Line Mapping In order for the constructed human biomechanical simulation model to perform biomechanical simulation, muscle force lines need to be mapped on the reconstructed three-dimensional human body model of the athlete; based on the musculoskeletal model in the existing biomechanical simulation experiment, the relative positional relationship of the father and son bones in the human anatomical model, and the relative positions of the muscle origin points and insertion points on each bone are sorted out. During muscle force line mapping, the skeletal model obtained through statistical learning is first uniformly resampled. Based on the calibration points, the muscle and skeletal model point cloud is registered to the standard skeletal model of the digital human body obtained through statistical learning through nonlinear spatial transformation, thus completing the registration of the skeletal model and the mapping of muscle force lines.

Images