An intelligent animation modeling method based on three-dimensional stereoscopic technology

Abstract

The application discloses a kind of intelligent animation modeling methods based on three-dimensional technology, it is related to computer graphics and three-dimensional animation production technical field, the application reconstructs hand-object interaction logic by three-dimensional technology, realize natural fine motion generation;In virtual anchor end, such as coffee cup scene, the spatial relationship of hand skeleton and prop point cloud is quantified as the topological connection diagram that can be calculated, rigid connection edge ensures the stable grip of functional area, fingertip and cup body have no suspended gap;Elastic adsorption edge allows flexible contact of non-functional area, when palm pressure changes, deformation gradient field drives prop surface microdeformation, synchronous joint torque dynamic adjustment, avoid hand rigid distortion when coffee pours in traditional scheme. The collaborative design of graph convolution network and physical engine further eliminates the risk of wearing mold: equidistance constraint of prop rigid area retains original topological structure, and soft contact area absorbs collision energy through deformation gradient field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer graphics and three-dimensional animation technology, and particularly relates to an intelligent animation modeling method based on three-dimensional technology. BACKGROUND

[0002] In a three-dimensional interactive scene such as a virtual anchor or VR social interaction, the demand for fine motion of characters and props is significantly increasing. For example, a virtual anchor needs to naturally hold a coffee cup, and a game character needs to pick up a weapon. The current mainstream solution uses an AI-generated animation tool, such as a text / image-driven modeling tool of Meshy or an AI bone binding plug-in of Blender, to automatically generate a basic holding motion by inputting a text instruction or a single image. Such a technology significantly reduces the threshold for animation production and supports rapid batch production.

[0003] In practical applications, it is found that the generated motion often has a three-dimensional space mismatch problem. The hand skeleton and the prop model are not accurately aligned in the three-dimensional space, causing the prop to float above the palm. The finger joints are not adapted to the geometry of the prop, such as a cylindrical cup handle, causing the fingertips to be embedded in the cup or suspended in the grip. The motion only relies on static key frames and cannot respond to dynamic interactions such as hand pressure deformation when pouring coffee. Such problems are particularly prominent in virtual live streaming and AR try-on scenes that require real-time interaction.

[0004] Some existing solutions use a physical engine post-processing, such as FinalIK of Unity, to introduce rigid body collision detection after AI-generated motion, and to forcibly adjust the hand position to avoid model penetration, but this easily causes joint angle mutation. Some solutions introduce a contact point optimization algorithm, such as a MetaHuman plug-in, manually label the prop grabbing point, and drive the hand joint to the target coordinate to attract, which still requires manual predefinition of the contact area. Although such methods can alleviate some space misalignment, they rely on additional calculation layers and do not fundamentally solve the three-dimensional topology adaptation problem. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides an intelligent animation modeling method based on three-dimensional technology to solve the problem that existing AI animation generation of handheld props causes suspension, model penetration, and lack of physical feedback due to three-dimensional space mismatch, and existing optimization solutions rely on manual intervention.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The embodiment of the present application provides an intelligent animation modeling method based on three-dimensional technology, which comprises,

[0009] Step S1, real-time acquisition of three-dimensional joint coordinates of the hand skeleton and three-dimensional point cloud data of the prop;

[0010] Step S2, based on the hand joint coordinates and the prop point cloud data, a hand-object contact relationship graph is constructed; wherein the nodes of the contact relationship graph include hand joint nodes and prop surface key points, and the edge weight is dynamically assigned by the Euclidean distance between the joint nodes and the surface points;

[0011] Step S3, inputting the contact relationship graph into a pre-trained graph convolution network, outputting the prop surface deformation gradient field and the hand joint rotation amount, and estimating the joint torque / force through inverse dynamics;

[0012] Step S4, according to the deformation gradient field, performing elastic deformation on the prop point cloud, and driving the hand skeleton to move according to the joint rotation amount, to generate a gripping motion sequence that satisfies the non-penetration constraint.

[0013] As a preferred scheme of the intelligent animation modeling method based on three-dimensional technology, in step S2, the edge weight function relationship of the hand-object contact relationship graph is constructed, and the process includes:

[0014] For hand joint j and prop surface point k, the Euclidean distance in three-dimensional space is calculated:

[0015] d jk =‖p j -q k ‖2,

[0016] Wherein, d jk represents the Euclidean distance between joint j and surface point k, unit: meter, p j represents the three-dimensional coordinate vector of joint j, unit: meter, q k represents the three-dimensional coordinate vector of prop surface point k, unit: meter, and ‖·‖2 represents the Euclidean norm operator;

[0017] When the surface point k is located in the prop grabbing functional area and the joint j is the end effector:

[0018] If d jk ≤τ rigid , then Otherwise

[0019] Wherein, represents the rigid connection edge weight, dimensionless, τ rigid = 0.02D represents the rigid connection threshold, unit: meter, and D represents the diagonal length of the prop bounding box, unit: meter;

[0020] When the surface point k is located in the non-functional area of the prop and satisfies d jk / D≤0.1:

[0021]

[0022] wherein, represents the elastic adsorption edge weight, dimensionless, λ represents the elastic attenuation coefficient, dimensionless;

[0023] Standardize all edge weights of the same node j:

[0024]

[0025] wherein, represents the normalized edge weight, dimensionless, w jk represents the original edge weight represents the k-nearest neighbor set of node j, n represents the index variable of the nearest neighbor point.

[0026] As a preferred scheme of the intelligent animation modeling method based on three-dimensional technology, in the step S2:

[0027] Match the topological connection of each hand joint to the prop surface point by k-nearest neighbor algorithm;

[0028] When the surface point is located in the prop grabbing functional area and the joint is the end effector, generate a rigid connection edge;

[0029] When the surface point is located in the non-functional area of the prop and the distance is less than 10% of the size of the prop, generate an elastic adsorption edge.

[0030] As a preferred scheme of the intelligent animation modeling method based on three-dimensional technology, in the step S3, the graph convolution network performs:

[0031] Calculate the attraction weight of the prop surface point to the hand joint through the graph attention mechanism;

[0032] Multiply the attraction weight with the parent-child constraint matrix of the joint kinematic chain to generate a rotation amount increment;

[0033] Fuse the basic motion layer and the increment through the residual connection.

[0034] As a preferred scheme of the intelligent animation modeling method based on three-dimensional technology, in the step S3, the graph attention mechanism includes:

[0035] Extract the local curvature feature of the prop surface point as the attention query vector;

[0036] Take the historical motion trajectory of the hand joint as the attention key vector;

[0037] The output attraction weight is a cosine similarity between the query vector and the key vector.

[0038] As a preferred scheme of the intelligent animation modeling method based on three-dimensional technology, the displacement vector and the joint quaternion generated by the output layer satisfy a cooperative constraint.

[0039] When the displacement vector drives the motion of the prop surface point, the topological structure of the rigid region of the prop is kept unchanged.

[0040] As a preferred scheme of the intelligent animation modeling method based on three-dimensional technology, in step S3, the output layer of the graph convolution network performs the following coupled calculation:

[0041] The attraction weight α output based on the graph attention mechanism jk , the joint rotation increment is calculated as:

[0042]

[0043] where Δq j represents the quaternion increment of the joint j, dimensionless, W q represents a rotation matrix, 4×3 matrix, dimensionless, α jk represents the attraction weight of the prop point k to the joint j, represents a surface point set connected to the joint j, C j represents a parent-child constraint matrix of the joint kinematic chain, 3×3 matrix, dimensionless, represents a curvature feature vector of the point k, three-dimensional vector, dimensionless;

[0044] The displacement vector of the prop surface point and the joint quaternion satisfy a rigid-flexible partition constraint:

[0045]

[0046] where u k represents the displacement vector of the surface point k, in meters, β jk represents a rigid region weight factor, dimensionless, taking a value of 1 for a functional connection point, or 0 for others, R(q j ) represents a rotation matrix derived from the quaternion q j , 3×3 matrix, dimensionless, represents the initial coordinates of the point k, in meters, p j represents the coordinates of the joint j, in meters, F k represents a deformation gradient field, 3×3 tensor, dimensionless, γ represents a flexible region coupling coefficient, dimensionless, represents a Jacobian matrix of the rotation matrix to the quaternion, 9×4 matrix, dimensionless;

[0047] The displacement of the rigid region is defined as satisfying the equidistance constraint:

[0048]

[0049] wherein, represents a set of point pairs of the prop rigid region, and m and n represent indexes of any two points in the rigid region.

[0050] As a preferred scheme of the intelligent animation modeling method based on three-dimensional stereoscopic technology, the step S4 comprises:

[0051] After the deformation gradient field is applied to the prop point cloud, the embedding depth of the fingertip vertex and the prop surface is detected.

[0052] If the embedding depth exceeds a threshold value, an iteration of contact relationship graph reconstruction is triggered, comprising:

[0053] The attraction weight distribution coefficient is reset.

[0054] The k-nearest neighbor search range of the prop surface points is expanded.

[0055] A smooth constraint of the joint rotation amount is added.

[0056] As a preferred scheme of the intelligent animation modeling method based on three-dimensional stereoscopic technology, before the step S1, the method further comprises:

[0057] The physical parameters of the three-dimensional scene are acquired in real time, including the gravity direction and the position of the collision body.

[0058] In the step S4, the hand joint torque is dynamically scaled according to the ground elastic coefficient in the physical parameters.

[0059] As a preferred scheme of the intelligent animation modeling method based on three-dimensional stereoscopic technology, the method is integrated into three-dimensional animation software through a plug-in.

[0060] The plug-in intercepts the skeleton data and the prop model in the software, and covers the motion sequence generated in the step S4 to the original animation layer.

[0061] The skeleton trajectory of the non-contact part remains unchanged.

[0062] The application has the beneficial effects that: the application reconstructs hand-object interaction logic through three-dimensional technology, and realizes natural and fine motion generation; in a virtual anchor end scene of holding a coffee cup, the spatial relationship between the hand skeleton and the prop point cloud is quantized into a calculable topological connection graph, a rigid connection edge ensures stable gripping of a functional area (such as a cup handle), and there is no suspended gap between a fingertip and a cup body; an elastic adsorption edge allows flexible contact of a non-functional area (such as a cup wall), when the palm pressure changes, a deformation gradient field drives the surface micro-deformation of the prop, and the joint torque is dynamically adjusted, thereby avoiding hand rigidity distortion when the coffee is poured in the traditional scheme.

[0063] The collaborative design of the graph convolution network and the physical engine further eliminates the risk of penetration: the equidistant constraint of the rigid area of the prop retains the original topological structure (such as the invariance of the cup mouth circle), the soft contact area absorbs the collision energy through the deformation gradient field, the contact graph iteration is triggered when the fingertip is embedded in the depth threshold, the neighbor search range is adaptively expanded, and the joint trajectory is smoothed, thereby solving the penetration phenomenon in a complex grip.

[0064] At the system integration level, the plug-in architecture directly intercepts the skeleton and model data of three-dimensional software (such as Blender), only covers the action layer of the contact part, and keeps the original animation smoothness of the non-contact limb movement (such as anchor head rotation). The finally generated motion sequence has physical realism and artistic controllability-when the gravity direction changes, the wrist turning angle is automatically compensated, the position of the collision body is updated in real time, and the elbow obstacle avoidance path is updated, without the need for an artist to correct the penetration or suspended flaws frame by frame. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0066] Figure 1 The flowchart of the intelligent animation modeling method based on three-dimensional technology in embodiment 1. DETAILED DESCRIPTION

[0067] In order to make the above-mentioned purposes, features and advantages of the application more apparent and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification.

[0068] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the application, therefore the application is not limited by the specific embodiments disclosed below.

[0069] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0070] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides an intelligent animation modeling method based on three-dimensional stereoscopic technology, including the following steps:

[0071] Step S1: Real-time acquisition of the three-dimensional joint coordinates of the hand bones and the three-dimensional point cloud data of the props;

[0072] Step S2: Based on the hand joint coordinates and prop point cloud data, construct a hand-object contact relationship graph; wherein, the nodes of the contact relationship graph include hand joints and prop surface key points, and the edge weights are dynamically allocated by the Euclidean distance between the joints and surface points.

[0073] Step S3: Input the contact relationship graph into the pre-trained graph convolutional network, output the surface deformation gradient field of the prop and the rotation of the hand joint, and estimate the joint torque / resultant force through inverse dynamics;

[0074] Step S4: Elastically deform the prop point cloud according to the deformation gradient field, and drive the hand bones to move according to the joint rotation amount to generate a holding action sequence that satisfies the non-penetration constraint.

[0075] It should be noted that the 3D technology here refers to the 3D data acquisition, representation, and processing schemes used to achieve hand-prop interaction, including the acquisition and preprocessing of point cloud, mesh, and volume data, but not limited to specific devices and software implementations. The 3D joint coordinates of the hand skeleton refer to the position data of each joint in the scene coordinate system or its corresponding local coordinate system, output by existing skeleton rigging or gesture tracking modules, and unified based on the scene rendering coordinate system. The 3D point cloud data of the prop refers to a discrete set of samples from the prop surface, which can be obtained from offline model surface sampling or online reconstruction; when the original data is a mesh, its vertices or homogenized surface samples are used as the point cloud representation. Key points on the prop surface refer to representative surface sampling points used to construct the graph structure, obtained through uniform sampling, bias sampling based on geometric features, or anchor point sampling based on semantic tags; when a grasping function area exists, the sampling density can be increased in that area.

[0076] The hand-object contact graph refers to a weighted graph structure with hand joints and prop surface key points as nodes, and spatial relative relationships and semantic attributes between nodes as edge features. Edge weights are normalized to represent potential contact intensity. The graph can be undirected or directed, but undirected representation is used by default. It should be noted that the pre-trained graph convolutional network refers to a deep model trained on publicly available or self-built hand-prop interaction samples, capable of message passing and fusion on a graph structure. The network can include structures such as attention, residuals, and normalization. Training objectives include at least non-penetrating constraints, contact consistency, joint constraints, and temporal continuity. Parameters are not updated during the inference phase. It should be pointed out that the deformation gradient field refers to the displacement or local deformation description defined on the prop surface shell, used to express soft contact compliance during gripping. This application limits deformation to only act on the preset surface shell or buffer layer, and the solid parts labeled as rigid regions maintain their topological relationships. It should be noted that joint rotation refers to the minimum rotational representation used to update the bone posture, preferably using axis angle increments or quaternion increments, and is normalized and subjected to joint constraint after output. It should also be noted that the base motion layer refers to the original or redirected animation trajectory before input to this method. This method only overwrites the trajectory on joints determined to be related to hand-object contact; the trajectory of non-contact areas remains unchanged. Furthermore, the inverse dynamics estimation of joint torque / resultant force refers to the joint mechanical quantities obtained through a standard inverse dynamics process based on the bone's mass and inertial properties and external contact estimation. This estimation is used for adaptive adjustment of compliance or force closure assessment, not as the sole control basis. Finally, the non-penetrating gripping motion sequence here refers to the motion sequence obtained after detection and constraint according to a preset penetration tolerance. This means that penetration is evaluated and controlled within an allowable range using collision detection or signed distance functions to achieve a verifiable non-penetrating effect.

[0077] Furthermore, in step S2, the edge weight function relationship of the hand-object contact relationship graph is constructed. The process includes:

[0078] For hand joint point j and prop surface point k, calculate the Euclidean distance in three-dimensional space:

[0079] d jk =‖p j -q k ‖2,

[0080] Where, d jk p represents the Euclidean distance between joint point j and surface point k, in meters. j q represents the three-dimensional coordinate vector of joint point j, in meters. k Let represent the three-dimensional coordinate vector of point k on the surface of the prop, in meters, and ‖·‖2 represent the Euclidean norm operator;

[0081] When surface point k is located in the prop-grabbing functional area and joint point j is an end effector:

[0082] If d jk ≤τ rigid ,but otherwise

[0083] in, The weight of a rigidly connected edge is dimensionless, τ. rigid =0.02D represents the rigid connection threshold, in meters, and D represents the diagonal length of the prop enclosure, in meters;

[0084] When surface point k is located in the non-functional area of ​​the prop and satisfies d jk When / D≤0.1:

[0085]

[0086] in, λ represents the elastic adsorption edge weight, which is dimensionless, and λ represents the elastic decay coefficient, which is dimensionless and ranges from 8 to 12.

[0087] Standardize the weights of all edges at the same key point j:

[0088]

[0089] in, w represents the normalized edge weights, which are dimensionless. jk Represents the original edge weights Let n represent the set of k nearest neighbors of key point j, where k = 5 and n represents the index variable of the nearest neighbor points;

[0090] Specifically, a dynamic mapping mechanism between Euclidean distance and edge weights is established here, and differentiated connections are achieved through functional area determination; rigid connections use step function constraints to grasp functional areas to ensure tool gripping stability, while elastic adsorption uses a Gaussian decay function, with the weight of non-functional areas decaying with the square of the distance through an exponential term, preserving flexible contact characteristics; normalization processing eliminates the influence of differences in the number of nearest neighbors, making the weight distribution conform to probabilistic characteristics; the adjustable range of the elastic decay coefficient λ adapts to the needs of props of different materials, such as a high λ value suitable for low-elasticity objects such as metals; the threshold design (2% and 10% of the prop size) takes into account both precise matching and flexible buffering, solving the clipping problem caused by traditional weight allocation methods;

[0091] It should be noted that the Euclidean distance here refers to the spatial distance between two nodes in a unified coordinate system. To eliminate the influence of scale, it can be standardized according to the prop size before being used for weight mapping. It should also be noted that the prop bounding box size refers to the scale of the smallest axis-aligned bounding box containing the prop model, used for distance standardization and threshold reference; it is recalculated according to the current scale when the model is scaled. Furthermore, functional region determination refers to the grasping or contact priority region determined based on the prop's semantic labels, manual annotation, or automatic segmentation results. If there is no semantic information, the entire surface is considered a non-functional region by default. Rigid connection edges refer to strong connections established between functional regions and end effectors, representing priority contact and pose subordination; their activation is determined by threshold rules or classifier output. Finally, elastic adsorption edges refer to flexible connections established within non-functional regions, representing a relationship where attraction increases with closer proximity. To avoid numerical instability, an upper limit is set on the adsorption strength at very close points. It should be noted that the nearest neighbor set refers to a group of surface key points selected for each incisor, obtained using spatial nearest neighbor search. To ensure stability, a fixed upper limit is set on the size of the set, and a point replenishment or backtracking strategy is adopted when the set is insufficient. It should also be noted that weight standardization refers to normalizing the weights of all outgoing edges of the same incisor to ensure a consistent sum, thereby ensuring comparability and numerical stability under different nearest neighbor sizes.

[0092] Furthermore, step S2 includes:

[0093] The k-nearest neighbor algorithm is used to match the topological connections from each hand joint point to the prop surface point.

[0094] When the surface point is located in the prop grabbing function area and the joint point is an end effector, a rigid connection edge is generated;

[0095] When a surface point is located in a non-functional area of ​​a prop and the distance is less than 10% of the prop's size, an elastic adsorption edge is generated.

[0096] Furthermore, the graph convolutional network in step S3 performs the following:

[0097] The attraction weights of prop surface points to hand joints are calculated using a graph attention mechanism.

[0098] Multiply the attraction weights by the parent-child constraint matrices of the joint kinematic chain to generate the rotation increment;

[0099] The basic action layer and the incremental layer are fused through residual connections;

[0100] It should be noted that the graph attention mechanism refers to the process of calculating and normalizing weights within the nearest neighbor range based on the query vector and key vector on the graph structure. The query vector originates from the local description of points on the prop surface, such as geometry and normals, while the key vector originates from the temporal context, such as the joint's historical pose, velocity, or acceleration. The parent-child constraint matrix refers to the kinematic mapping derived from the bone hierarchy relationship, used to propagate local influences along the joint hierarchy to the rotation update of the target joint. This can be implemented using Jacobian or equivalent differentiable mappings. Residual connection fusion refers to superimposing the network-predicted rotation increments and displacement corrections onto the bone pose of the basic motion layer to reduce drift and improve temporal stability. The historical motion trajectory refers to the joint state sequence collected within a fixed-length time window, used to construct the temporal context and suppress instantaneous noise.

[0101] Furthermore, graph attention mechanisms include:

[0102] Extract the local curvature features of points on the prop surface as the attention query vector;

[0103] The historical motion trajectory of the hand joints is used as the attention key vector;

[0104] The output attraction weight is the cosine similarity between the query vector and the key vector;

[0105] Furthermore, the output layer generates displacement vectors and joint quaternions that satisfy synergistic constraints;

[0106] When the displacement vector drives the movement of points on the prop surface, the topological structure of the rigid region of the prop remains unchanged;

[0107] Deformation is limited to the defined soft contact / grip buffer layer (such as surface shell, co-rotational FEM / shape matching);

[0108] It should be noted that the coordinated constraint of displacement vector and joint quaternion refers to simultaneously considering the consistency of prop surface displacement and hand joint rotation during solution or forward inference, so that the surface deformation of the soft contact area changes in tandem with the hand posture. It should also be noted that the rigid region refers to the prop entity portion that is not allowed to undergo relative deformation; within this region, only overall pose transformation is permitted, while the relative relationships between points remain unchanged. Furthermore, the soft contact / grip buffer layer refers to a deformable thin layer attached to the prop surface to simulate material compliance; its material properties can be obtained by mapping the prop material. Finally, it should be noted that shape matching and co-rotational finite element method refer to two representative algorithms for solving soft layer deformation; either one or an equivalent alternative can satisfy the requirements of this application for small deformation and stability.

[0109] Furthermore, in step S3, the output layer of the graph convolutional network performs the following coupled computations:

[0110] The attraction weight α output based on the graph attention mechanism jk Calculate the joint rotation increment:

[0111]

[0112] Where, Δq j W represents the quaternion increment of joint j, dimensionless. q This represents the rotation transformation matrix, a 4×3 matrix, dimensionless, α jk This represents the attraction weight of prop point k to joint j. C represents the set of surface points connected to joint j. j The parent-child constraint matrix of the joint kinematic chain is a 3×3 matrix, dimensionless. The curvature eigenvector of point k is a three-dimensional, dimensionless vector.

[0113] Define the displacement vectors of points on the prop surface and the joint quaternions to satisfy the rigid-flexible partitioning constraint:

[0114]

[0115] Among them, u k β represents the displacement vector of surface point k, in meters. jk R(q) represents the rigid zone weight factor, which is dimensionless and takes a value of 1 to represent a functional connection point, or 0 to represent anything else. j ) represents the quaternion q j The derived rotation matrix is ​​a 3×3 matrix, dimensionless. p represents the initial coordinates of point k in meters. j F represents the coordinates of joint j, in meters. k The deformation gradient field is represented by a 3×3 tensor, which is dimensionless. γ represents the flexible coupling coefficient, which is dimensionless and ranges from 0.2 to 0.5. The Jacobian matrix representing the rotation matrix in relation to quaternions is a 9×4 matrix and is dimensionless.

[0116] The displacement of a rigid region must satisfy the equidistant constraint:

[0117]

[0118] in, Let m and n represent the set of point pairs representing the rigid region of the prop, where m and n represent the indices of any two points within the rigid region.

[0119] Specifically, the deformation gradient field and joint motion are constrained collaboratively through coupling equations; in the rigid region, the displacement vector is directly driven by the joint rotation matrix, using rigid body transformation formulas. Strictly maintain the topological structure; introduce a deformation gradient field F into the soft contact region. k The partial derivative with respect to joint rotation Coupling, with coefficient γ controlling the sensitivity of deformation to joint motion; equidistant constraints ensure that the point spacing in rigid regions remains constant, solving the problem of prop breakage or excessive twisting in traditional methods; the calculation of rotation increment integrates attraction weights and kinematic chain constraints, ensuring that the joint rotation Δq... j It accurately reflects the influence of the curvature characteristics of the prop surface, avoiding clipping caused by relative sliding between hand and object;

[0120] Here, the attraction weight refers to the weight calculated by the attention mechanism and normalized within the nearest neighbor range. It is used to measure the influence strength of each surface point on the target joint. The weight is continuously updated over time and a numerical stabilization strategy is set. It should be noted that rotational increment mapping refers to the process of mapping features that integrate geometric and semantic information to joint rotation updates. Preferably, a minimum rotation representation is generated first, followed by exponential mapping or normalization to obtain the final pose, ensuring numerical stability and composability. It should be pointed out that Jacobi refers to a linear approximate mapping describing the change in parameters due to rotation. It is used to establish the coupling relationship between deformation and joint updates in the soft contact region, and can be obtained through analytical methods or automatic differentiation. It should be noted that equidistant constraints refer to constraints that maintain the relative position between any two points in a rigid region, used to prevent props from breaking, compressing, or undergoing non-physical distortion during the calculation process.

[0121] Furthermore, step S4 includes:

[0122] After applying a deformation gradient field to the prop point cloud, the embedding depth between the fingertip vertex and the prop surface is detected.

[0123] If the embedding depth exceeds the threshold, a contact graph reconstruction iteration is triggered, including:

[0124] Reset the attraction weighting coefficients;

[0125] Expand the k-nearest neighbor search range for prop surface points;

[0126] Add a smooth constraint on the joint rotation amount;

[0127] The embedding depth here refers to the penetration metric between the hand mesh and the prop surface, which can be obtained through distance field, nearest point, or engine collision feedback, and is used to trigger iterative corrections. It should be noted that contact graph reconstruction refers to recalculating the edge set, edge weights, and nearest neighbor range of the graph when the triggering conditions are met, and resetting or decaying historical states to escape local optima. It should also be pointed out that the smoothing constraint refers to regularizing the changes in joint angular velocity and angular acceleration to suppress numerical oscillations and improve temporal continuity; this constraint does not change the semantic actions of the original keyframes.

[0128] Furthermore, prior to step S1, the following is also included:

[0129] Real-time acquisition of physical parameters of the 3D scene, including gravity direction and collider position;

[0130] In step S4, the hand joint torque is dynamically scaled according to the ground elasticity coefficient in the physical parameters;

[0131] It should be noted that the physical parameters of the 3D scene refer to the environmental quantities used for dynamics solving in the rendering or engine environment, including the direction of gravity, the collision properties of the ground and props, and the elasticity and damping properties of materials. When this information is unavailable, this method uses the default environment settings for inference. It should also be pointed out that dynamic scaling of the ground elasticity coefficient refers to proportionally adjusting the joint mechanical quantities or compliance parameters based on the ground material properties to reflect motion stability under different contact flexibility conditions. The scaling strategy only affects joints or constraints related to contact and does not change the original trajectory of non-contact parts.

[0132] Furthermore, this method is integrated into 3D animation software via a plugin;

[0133] The plugin intercepts the skeletal data and prop models in the software and overlays the motion sequence generated in step S4 onto the original animation layer.

[0134] Among them, the skeletal trajectory of non-contact areas remains unchanged;

[0135] It should be noted that plugin integration into 3D animation software refers to embedding it into digital content creation software or real-time engines as an extension component or script, calling its public interfaces to complete data reading and writing and runtime execution. It should also be noted that intercepting skeletal data and prop models refers to obtaining the skeletal pose, layer information, and prop geometry data of the current frame through event callbacks or data interfaces, and completing coordinate alignment and scale unification before entering this method. Furthermore, overwriting the original animation layer means writing the generated holding motion into a new driving layer or pose channel without modifying the source file; the original animation can be fully restored when the user disables this function. Finally, ensuring that non-contact parts remain unchanged means that only the joint trajectories involving contact and holding are replaced during the output compositing stage; the remaining joint data is passed in its original order without modification to ensure compatibility with existing assets and toolchains.

[0136] It should be noted that the above definitions and implementation methods are used to meet the technical objectives of this application, namely, to improve the geometric fit, non-penetration, and temporal stability of the hand and prop within a real-time budget; without changing the technical essence, equivalent substitutions or functionally consistent improvements to component names, data structures, and solution strategies should be considered to fall within the protection scope of this application.

Claims

1. A smart animation modeling method based on three-dimensional stereoscopic technology, characterized in that, include, Step S1: Real-time acquisition of the three-dimensional joint coordinates of the hand bones and the three-dimensional point cloud data of the props; Step S2: Based on the coordinates of the hand joints and the point cloud data of the prop, construct a hand-object contact relationship diagram; wherein, the nodes of the contact relationship diagram include hand joints and key points on the prop surface, and the edge weights are dynamically allocated by the Euclidean distance between the joints and the surface points. Step S3: Input the contact relationship graph into a pre-trained graph convolutional network to output the surface deformation gradient field of the prop and the rotation of the hand joint. Estimate the joint torque / resultant force through inverse dynamics. The graph convolutional network in step S3 is executed as follows: The attraction weights of prop surface points to hand joints are calculated using a graph attention mechanism. The attraction weights are multiplied by the parent-child constraint matrix of the joint kinematic chain to generate the rotation increment; The basic action layer and the increment are fused through residual connections; The output layer generates displacement vectors and joint quaternions that satisfy synergistic constraints. When the displacement vector drives the movement of points on the prop surface, the topological structure of the rigid region of the prop remains unchanged. In step S3, the output layer of the graph convolutional network performs the following coupled computations: Attraction weights output based on graph attention mechanism Calculate the joint rotation increment: ， in, Indicates joint The quaternion increment is dimensionless. This represents the rotation transformation matrix, a 4×3 matrix, dimensionless. Indicates prop points For joints The attractiveness weight, Indicates joint The set of connected surface points The parent-child constraint matrix of the joint kinematic chain is a 3×3 matrix, dimensionless. Point The curvature eigenvector, a three-dimensional vector, dimensionless; Define the displacement vectors of points on the prop surface and the joint quaternions to satisfy the rigid-flexible partitioning constraint: ， in, Represents surface points The displacement vector, in meters. This represents the rigid zone weighting factor, which is dimensionless. A value of 1 indicates a functional connection point, or 0 indicates anything else. Represents quaternions The derived rotation matrix is ​​a 3×3 matrix, dimensionless. Point The initial coordinates, in meters. Indicates joint The coordinates, in meters. Represents the deformation gradient field, a 3×3 tensor, dimensionless. Represents the coupling coefficient of the flexible zone, dimensionless. The Jacobian matrix representing the rotation matrix in relation to quaternions is a 9×4 matrix and is dimensionless. The displacement of a rigid region must satisfy the equidistant constraint: ， ， in, The set of points representing the rigid regions of a prop. This represents the index of any two points within a rigid region. Step S4: Elastically deform the prop point cloud according to the deformation gradient field, and drive the hand bones to move according to the joint rotation amount to generate a holding action sequence that satisfies the non-penetration constraint.

2. The intelligent animation modeling method based on three-dimensional technology as described in claim 1, characterized in that, In step S2, the edge weight function relationship of the hand-object contact relationship graph is constructed. The process includes: For hand joints and prop surface points Calculate the Euclidean distance in three-dimensional space: ， in, Indicates key points With surface points Euclidean distance, in meters. Indicates key points The three-dimensional coordinate vector, in meters. Indicates the surface point of the prop The three-dimensional coordinate vector, in meters. Operators representing Euclidean norms; When surface points Located in the item grabbing function area and at the key point When it is an end effector: like ,but ,otherwise , in, This represents the weight of a rigidly connected edge, which is dimensionless. This represents the rigid connection threshold, in meters. This indicates the diagonal length of the prop enclosure, in meters. When surface points Located in the non-functional area of ​​the prop and meeting the requirements hour: ， in, Represents the elastic adsorption edge weight, dimensionless. This represents the elastic damping coefficient, which is dimensionless. For the same joint point Standardize all edge weights: ， in, This represents the normalized edge weights, which are dimensionless. Represents the original edge weights Indicates key points The set of k nearest neighbors, The index variable represents the nearest neighbor.

3. The intelligent animation modeling method based on three-dimensional technology as described in claim 2, characterized in that, Step S2 includes: The k-nearest neighbor algorithm is used to match the topological connections from each hand joint point to the prop surface point. When the surface point is located in the prop grabbing function area and the joint point is an end effector, a rigid connection edge is generated; When a surface point is located in a non-functional area of ​​a prop and the distance is less than 10% of the prop's size, an elastic adsorption edge is generated.

4. The intelligent animation modeling method based on three-dimensional stereoscopic technology as described in claim 1, characterized in that, The graph attention mechanism includes: Extract the local curvature features of points on the prop surface as the attention query vector; The historical motion trajectory of the hand joints is used as the attention key vector; The output attraction weight is the cosine similarity between the query vector and the key vector.

5. The intelligent animation modeling method based on three-dimensional technology as described in claim 1, characterized in that, Step S4 includes: After applying a deformation gradient field to the prop point cloud, the embedding depth between the fingertip vertex and the prop surface is detected. If the embedding depth exceeds the threshold, a contact graph reconstruction iteration is triggered, including: Reset the attraction weighting coefficients; Expand the k-nearest neighbor search range for prop surface points; Add a smooth constraint on the joint rotation amount.

6. The intelligent animation modeling method based on three-dimensional technology as described in claim 1, characterized in that, The steps preceding step S1 also include: Real-time acquisition of physical parameters of the 3D scene, including gravity direction and collider position; In step S4, the hand joint torque is dynamically scaled according to the ground elasticity coefficient in the physical parameters.

7. The intelligent animation modeling method based on three-dimensional stereoscopic technology as described in claim 1, characterized in that, This method is integrated into 3D animation software via a plugin; The plugin intercepts the skeletal data and prop models in the software and overlays the motion sequence generated in step S4 onto the original animation layer. Among them, the skeletal trajectory of non-contact areas remains unchanged.

Images