Dance motion auxiliary generation method and system based on three-dimensional modeling

Abstract

The application discloses a dance action auxiliary generation method and system based on three-dimensional modeling, which comprises the following steps: constructing a three-dimensional human skeleton model; generating a joint action set of each joint node in a dance action; constructing a dance action auxiliary model based on the fusion of a time series convolution network and a graph convolution neural network; obtaining three-dimensional joint motion trajectory data in continuous frames; obtaining the correlation between dance actions by analyzing time series features and cooperative motion rules in dance actions; obtaining joint actions in the joint action set that have correlation with the current dance action, and generating a candidate action set; calculating the connection fluency score and the posture beauty score between the current dance action and the next dance action; and obtaining the candidate action with the highest total score by weighting as the next dance action. The application has the advantages that through accurate skeleton modeling and kinematic constraints, combined with deep learning technology, fluent and beautiful dance actions can be efficiently generated, and the application is suitable for dance creation and virtual performance.

Description

TECHNICAL FIELD

[0001] The present application relates to motion generation, particularly to a dance motion auxiliary generation method and system based on three-dimensional modeling. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, especially the application of deep learning algorithms in image and video processing, dance motion generation technology has made significant progress. These technologies can automatically generate new dance motions or assist in creation through learning and analysis of a large amount of dance motion data. For dancers and enthusiasts, the auxiliary generation method can provide personalized training programs and real-time motion feedback to help them improve their skills.

[0003] Current dance motion auxiliary generation methods on the market often have the shortcomings of low precision and unnatural motion generation. Many methods rely on 2D image processing, which cannot effectively capture complex three-dimensional human skeletons and spatial motion, resulting in generated motions lacking accuracy in details and easily having inappropriate joint positions or distorted poses. In addition, existing methods often ignore kinematic constraints and physical limitations, resulting in generated motions that are not physiologically reasonable and easily have poses or motions that exceed the human body's ability range. Although some methods incorporate deep learning technology, most focus only on the processing of temporal information, ignoring the spatial relationship between joints, resulting in generated dance motions lacking coherence and fluidity. More importantly, existing methods usually do not comprehensively consider the aesthetic features of motion, such as the flexibility, symmetry, and balance of poses, resulting in generated motions lacking in aesthetics. SUMMARY

[0004] To improve existing methods and systems, a dance motion auxiliary generation method and system based on three-dimensional modeling is provided. This method, through precise skeleton modeling and kinematic constraints, combined with deep learning technology, can efficiently generate smooth and aesthetically pleasing dance motions suitable for dance creation and virtual performance.

[0005] To achieve the above objectives, the technical solution adopted by the present application is as follows:

[0006] The dance motion auxiliary generation method based on three-dimensional modeling includes:

[0007] Obtaining initial pose data of a target dance and constructing a three-dimensional human skeleton model;

[0008] Based on the three-dimensional human skeleton model, for each joint node, based on its kinematic constraints, generating a set of allowed joint motions for the joint in the dance motion;

[0009] The dance motion auxiliary model based on the fusion of the time sequence convolution network and the graph convolution neural network is constructed, a large amount of labeled dance motion video data is input to train the dance motion auxiliary model, three-dimensional joint motion trajectory data in continuous frames are acquired, and based on the dance motion auxiliary model, the correlation between each dance motion is acquired by analyzing the time sequence characteristics of joint motion and the cooperative motion law in the dance motion.

[0010] Based on the trained dance motion auxiliary model, the three-dimensional human skeleton data of the current dance motion is input into the dance motion auxiliary model, the joint motion set is combined with the motion parameters allowed by each joint, the joint motion in the joint motion set that has correlation with the current dance motion is acquired, and a candidate motion set is generated.

[0011] Based on each joint motion in the candidate motion set, the fluency score and the posture aesthetic score between the current dance motion are calculated and acquired, and the candidate motion with the highest total score is output as the next dance motion through weighting.

[0012] Preferably, the initial posture data of the target dance is acquired, and a three-dimensional human skeleton model is constructed, specifically including:

[0013] The human key point data in each frame is extracted by using the existing posture estimation tool, including shoulders, elbows, knees, ankles, wrists, and three-dimensional coordinate data of them is acquired.

[0014] The three-dimensional coordinate data is preprocessed to generate dance posture data, and is mapped to the three-dimensional human skeleton model.

[0015] The posture of the skeleton in the three-dimensional human skeleton model is corrected to ensure that the angle and position of each joint meet the actual dance motion.

[0016] Preferably, based on the three-dimensional human skeleton model, for each joint node, based on the kinematic constraint condition, a set of motion parameters allowed by the joint in the dance motion is generated, specifically including:

[0017] Based on the physiological structure and physical limitation of each joint, the motion action constraint condition is constructed to acquire the degree of freedom, the constraint condition includes the maximum angle, the minimum angle and the rotation direction of the joint, and the degree of freedom includes the rotation degree of freedom and the translation degree of freedom.

[0018] Based on the constraint condition and the degree of freedom, a joint motion model is constructed to acquire all the actions that can be completed by each joint, generate a joint motion set, including discrete values of joint angles, speed and acceleration limitations and time sequence constraints.

[0019] Preferably, the construction is based on the fusion of a time convolution network and a graph convolution neural network dance action auxiliary model, a large amount of labeled dance action video data is input for dance action auxiliary model training, three-dimensional joint motion trajectory data in continuous frames are obtained, and the specific steps include:

[0020] The time convolution network captures the continuous frames of the dance action video data, obtains the three-dimensional position data of each joint in each frame, obtains the time sequence characteristics of the joint motion, and classifies and identifies the dance action category;

[0021] The graph convolution neural network obtains the adjacency relationship between the joints, regards the joints as nodes in the graph, and regards the skeleton relationship as the edge of the graph;

[0022] The time sequence characteristics obtained by the time convolution network and the spatial characteristics obtained by the graph convolution network are fused to construct a dance action auxiliary model.

[0023] Preferably, based on the trained dance action auxiliary model, the three-dimensional human skeleton data of the current dance action is input into the dance action auxiliary model, the motion parameters allowed by each joint in the joint action set are combined, the joint action in the joint action set that has relevance with the current dance action is obtained, and a candidate action set is generated. The specific steps include:

[0024] The three-dimensional human skeleton data of the current dance action is input into the trained dance action auxiliary model to obtain the joint action that has relevance with the current dance action;

[0025] Based on the motion parameters allowed by all joints in the joint action set, the joint action that has relevance with the current dance action is matched with the joint action in the joint action set to obtain a matching probability distribution;

[0026] Based on the joint action in the joint action set that is higher than the preset threshold in the matching probability distribution, a candidate action set is generated.

[0027] Preferably, based on the joint action in the candidate action set, the continuity and fluency score and the posture beauty score between the current dance action are calculated and obtained, and the candidate action with the highest total score is output as the next dance action through weighting. The specific steps include:

[0028] Based on the joint action in the candidate action set, the difference between the joint position of the current dance action and the position of the corresponding joint in the candidate action is obtained, and the position difference degree is calculated;

[0029] The speed continuity and fluency are calculated by calculating the speed of the joints in the current dance action and the candidate action;

[0030] Score the motion connection fluency based on the position difference degree and the speed connection fluency;

[0031] Score the posture beauty degree based on the symmetry, balance, stability of the motion and the stretch, tension of the posture;

[0032] Based on the motion connection fluency score and the posture beauty degree score, the weighted total score of each candidate motion is obtained through comprehensive weighting, and the candidate motion with the highest score is selected as the next dance motion.

[0033] Further, a dance motion auxiliary generation system based on three-dimensional modeling is proposed, comprising:

[0034] A three-dimensional human skeleton module: the three-dimensional human skeleton module is used to extract the initial posture data of the target dance and construct a three-dimensional human skeleton model, so as to ensure that the angle and position of each joint meet the actual dance motion;

[0035] A motion parameter module: the motion parameter module generates a set of allowed motion parameters for each joint based on kinematic constraints, including joint angle, speed and acceleration limit;

[0036] A dance motion auxiliary model module: the dance motion auxiliary model module is used to fuse a time series convolution network and a graph convolution neural network, analyze a large amount of labeled dance motion video data, extract time sequence features, and obtain the correlation between dance motions;

[0037] A candidate motion set module: the candidate motion set module is used to input the three-dimensional human skeleton data of the current dance motion, combine the joint motion set to generate a candidate motion set associated with the current motion;

[0038] A scoring module: the scoring module is used to calculate the connection fluency and posture beauty degree score of the candidate motion and the current dance motion, and output the candidate motion with the highest total score as the next dance motion through weighting;

[0039] A processor: the processor is used to process the calculation process of each formula and the construction calculation process of each model.

[0040] Compared with the prior art, the advantages of the present application are:

[0041] By constructing an accurate three-dimensional human skeleton model, this method can efficiently capture the details of dance movements, ensuring that the posture of each joint is consistent with the actual movement, providing high precision. Secondly, by combining kinematic constraints and physical limitations to generate the allowed action set of each joint, it can ensure that the generated dance movements conform to the physiological movement rules of the human body, avoiding unnatural or inappropriate movements. By fusing the temporal convolution network with the graph convolutional neural network, the method not only extracts temporal and spatial features from a large number of labeled videos, but also effectively identifies the relevance between dance movements, making the generated movements more smooth and coherent. Finally, by comprehensively scoring the smoothness of the candidate action and the aesthetic degree of the posture, it ensures that the output dance movements are not only natural and smooth, but also have artistic beauty. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 Schematic diagram of the method proposed by the present application;

[0043] Figure 2 Schematic diagram of the three-dimensional human skeleton model proposed by the present application;

[0044] Figure 3 Schematic diagram of the motion parameter set proposed by the present application;

[0045] Figure 4 Schematic diagram of the dance movement relevance proposed by the present application;

[0046] Figure 5 Schematic diagram of the candidate action set proposed by the present application;

[0047] Figure 6 Schematic diagram of the next dance movement output proposed by the present application. DETAILED DESCRIPTION

[0048] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be thought of by those skilled in the art.

[0049] The dance movement auxiliary generation system based on three-dimensional modeling includes:

[0050] Three-dimensional human skeleton module: the three-dimensional human skeleton module is used to extract the initial posture data of the target dance, and construct a three-dimensional human skeleton model, ensuring that the angle and position of each joint conform to the actual dance movement;

[0051] Motion parameter module: the motion parameter module generates the allowed motion parameter set of each joint based on kinematic constraints, including joint angle, velocity, and acceleration constraints;

[0052] The dance action auxiliary model module is configured to fuse a time sequence convolution network and a graph convolution neural network, analyze a large amount of labeled dance action video data, extract time sequence features, and obtain the correlation between dance actions.

[0053] The candidate action set module is configured to input three-dimensional human skeleton data of a current dance action, combine joint action sets, and generate a candidate action set associated with the current action.

[0054] The scoring module is configured to calculate the connection fluency score and the posture aesthetic score of the candidate action and the current dance action, and obtain the candidate action with the highest total score as the next dance action output through weighting.

[0055] The processor is configured to process the calculation process of each formula and the construction and calculation process of each model.

[0056] Referring to Figure 1 As shown in the figure, the dance action auxiliary generation method based on three-dimensional modeling includes the following steps.

[0057] Step 1: Obtain initial posture data of a target dance and construct a three-dimensional human skeleton model.

[0058] Step 2: Based on the three-dimensional human skeleton model, for each joint node, based on its kinematic constraint condition, generate a joint action set allowed for the joint in the dance action.

[0059] Step 3: Construct a dance action auxiliary model based on the fusion of a time sequence convolution network and a graph convolution neural network, input a large amount of labeled dance action video data for dance action auxiliary model training, obtain three-dimensional joint motion trajectory data in consecutive frames, and based on the dance action auxiliary model, obtain the correlation between dance actions by analyzing the time sequence features of joint motion and the cooperative motion law in the dance action.

[0060] Step 4: Based on the trained dance action auxiliary model, input three-dimensional human skeleton data of a current dance action into the dance action auxiliary model, combine the motion parameters allowed for each joint in the joint action set, obtain joint actions in the joint action set that have correlation with the current dance action, and generate a candidate action set.

[0061] Step 5: Based on each joint action in the candidate action set, calculate the connection fluency score and the posture aesthetic score between the current dance action, and obtain the candidate action with the highest total score as the next dance action output through weighting.

[0062] Referring to Figure 2 As shown in the figure, obtaining initial posture data of a target dance and constructing a three-dimensional human skeleton model specifically includes:

[0063] By using existing pose estimation tools, the key point data of the human body in each frame is extracted, including shoulders, elbows, knees, ankles, wrists, and their three-dimensional coordinate data is obtained;

[0064] The three-dimensional coordinate data is preprocessed to generate dance pose data and mapped to a three-dimensional human skeleton model;

[0065] The pose of the skeleton in the three-dimensional human skeleton model is corrected to ensure that the angle and position of each joint conform to the actual dance movement.

[0066] Specifically, using existing pose estimation tools to extract human key points from the target dance video or sequence frames, including shoulders, elbows, knees, ankles, wrists, and mapping the preprocessed three-dimensional coordinates to the skeleton, i.e. constructing a three-dimensional human skeleton, the human skeleton is defined by the relative positions between key points, and the angle between each pair of joints is calculated by the relative positions between joints;

[0067] In order to ensure that the pose of the skeleton model conforms to the actual dance movement, the data of each frame needs to be corrected, and the joint angle in each frame needs to be corrected to ensure that the angle change between each joint and adjacent joint of the skeleton is smooth and natural.

[0068] Referring to Figure 3 Based on the three-dimensional human skeleton model, for each joint node, based on its kinematic constraint conditions, a set of motion parameters allowed for the joint in the dance movement is generated, including:

[0069] Based on the physiological structure and physical limitations of each joint, motion action constraints are constructed to obtain degrees of freedom, the constraints include the maximum angle, minimum angle, and rotation direction of the joint, and the degrees of freedom include rotational degrees of freedom and translational degrees of freedom;

[0070] Based on the constraint conditions and degrees of freedom, a joint motion model is constructed to obtain all the actions that can be completed by each joint, and a joint action set is generated, including discrete values of joint angles, speed and acceleration limits, and timing constraints.

[0071] Specifically, the motion of each joint is limited by physics and physiology, and different joints have different ranges of motion, for example, the maximum bending angle of the elbow joint may be 150 degrees and the minimum angle is 0 degrees, the rotation direction and rotation axis of the joint are also constrained, for example, the rotation of the shoulder joint can be along multiple axes, but the knee joint can only rotate along one axis. The specific rotation direction restriction can be represented by a rotation matrix or a quaternion;

[0072] The degree of freedom of each joint defines the number of directions in which the joint can move freely, the rotational degree of freedom is the rotational degree of freedom of the joint around an axis, for example, the elbow joint and the knee joint each have one rotational degree of freedom, the shoulder joint has three rotational degrees of freedom, and the translational degree of freedom is the displacement degree of freedom of the joint, which describes the change in the position of the joint, for example, the hip joint usually has three translational degrees of freedom in three-dimensional space.

[0073] Referring to Figure 4 As shown, a dance action auxiliary model based on the fusion of a time series convolutional network and a graph convolutional neural network is constructed, a large amount of labeled dance action video data is input to train the dance action auxiliary model, and three-dimensional joint motion trajectory data in consecutive frames are obtained, specifically including:

[0074] The time series convolutional network captures consecutive frames of dance action video data, obtains three-dimensional position data of each joint in each frame, obtains time series features of joint motion, and classifies and identifies dance action categories;

[0075] The graph convolutional neural network obtains the adjacency relationship between the joints, regards the joints as nodes in the graph, and regards the skeleton connection relationship as the edge of the graph;

[0076] The time series features obtained by the time series convolutional network are fused with the spatial features obtained by the graph convolutional network to construct a dance action auxiliary model.

[0077] Specifically, the time series convolutional network extracts time features through convolution operation, for the joint data in each frame of video, the TCN extracts time series features through the following formula, the formula is:

[0078]

[0079] Wherein, is the time series feature of joint j at time t, is the convolution kernel, is the bias term, K is the size of the convolution kernel, the time series features of the joint at different time steps are obtained by sliding window calculation, is the original input data of joint j;

[0080] Through the continuous time series convolutional layer, the motion features of each joint in each frame can be obtained from the video data, and the extracted time series features will be used for subsequent dance action classification and analysis;

[0081] Each joint in the human skeleton is regarded as a node of the graph, the connection relationship between the joints is regarded as the edge of the graph, a graph is constructed, the edge set represents the spatial connection relationship between the joints, the graph convolutional network performs convolution operation on the graph nodes through the adjacency matrix, extracts the spatial features between the joints, and considers the spatial topological structure of the skeleton;

[0082] The outputs of the time convolution network and the graph convolution network are fused to construct a more powerful dance action auxiliary model, and the similarity between different dance actions is measured to identify the cooperative motion relationship between different joints.

[0083] Referring to Figure 5 As shown, based on the completed dance action auxiliary model, the three-dimensional human skeleton data of the current dance action is input into the dance action auxiliary model, and the motion parameters allowed by each joint in the joint action set are combined to obtain the joint action in the joint action set that is associated with the current dance action, and a candidate action set is generated, which specifically includes:

[0084] The three-dimensional human skeleton data of the current dance action is input into the completed dance action auxiliary model to obtain the joint action associated with the current dance action;

[0085] Based on the motion parameters allowed by all joints in the joint action set, the joint action associated with the current dance action is matched with the joint action set to obtain a matching probability distribution;

[0086] Based on the joint actions in the joint action set that exceed the preset threshold in the matching probability distribution, a candidate action set is generated.

[0087] Specifically, based on the current joint action data and all allowed motion parameters in the joint action set, action matching is performed, the matching degree of the current joint action and each joint action in the action set is calculated to generate a matching probability distribution, and by setting a preset threshold, joint actions with a matching degree higher than the threshold are screened out to generate a candidate action set;

[0088] The candidate action set includes all joint actions with a matching degree higher than the threshold. These actions are the most likely joint motions of the current dance action after matching the three-dimensional skeleton data of the current dance action with the motion parameters in the joint action set.

[0089] Referring to Figure 6 As shown, based on the joint actions in the candidate action set, the continuity and fluency score and the posture aesthetics score between the current dance action and the candidate action are calculated and obtained, and the candidate action with the highest total score is obtained by weighting as the next dance action output, which specifically includes:

[0090] Based on the joint actions in the candidate action set, the difference between the joint position of the current dance action and the position of the corresponding joint in the candidate action is obtained, and the position difference degree is calculated;

[0091] The velocity continuity and fluency are calculated by calculating the velocity of the joints in the current dance action and the candidate action;

[0092] score the motion connection fluency based on the position difference degree and the speed connection fluency;

[0093] score the posture aesthetics based on the motion symmetry, balance, stability, and stretchability and tension of the posture;

[0094] based on the motion connection fluency score and the posture aesthetics score, obtain a weighted sum score of each candidate motion through comprehensive weighting, and select the candidate motion with the highest score as the next dance motion.

[0095] Specifically, the position difference degree is used to measure the difference in joint position between the current dance motion and the candidate motion, and the total position difference degree of the entire motion is obtained by calculating the Euclidean distance and weighting the average of the difference degrees of all joints. The speed connection fluency reflects the smoothness of the connection between the current dance motion and the candidate motion in terms of speed.

[0096] The symmetry of the motion is evaluated by calculating the position difference degree between the left and right joints, and the stretchability and tension are measured by the angle change between the joints. The posture aesthetics is scored based on the above.

[0097] The final motion score is obtained by weighting the motion connection fluency score and the posture aesthetics score. By sorting the weighted sum scores of all candidate motions, the candidate motion with the highest score is selected as the next dance motion.

[0098] In summary, the advantages of the present application are as follows: by constructing an accurate three-dimensional human skeleton model, the method can efficiently capture the details of dance motions, ensure that the posture of each joint is consistent with the actual motion, and provide high precision. Secondly, by combining kinematic constraints and physical limitations to generate a set of allowed motions for each joint, it can ensure that the generated dance motions conform to the physiological movement of the human body, avoiding unnatural or inappropriate motions. By fusing the time convolution network and the graph convolution neural network, the method can not only extract time and spatial features from a large number of labeled videos, but also effectively identify the relevance between dance motions, making the generated motions more fluent and coherent. Finally, by comprehensively scoring the connection fluency of the candidate motion and the aesthetics of the posture, the output dance motion is not only natural and fluent, but also has artistic beauty.

[0099] It should be noted that the above-mentioned order of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above describes a specific embodiment of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0100] The various embodiments in the specification are described in progressive manner, and the same or similar parts between the various embodiments can be mutually referred to, and each embodiment focuses on the difference from other embodiments.

[0101] The above description is merely preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating dance motion assistance based on three-dimensional modeling, characterized by, The method comprises the following steps: Obtain initial pose data of the target dance and construct a three-dimensional human skeleton model; Based on the three-dimensional human skeleton model, for each joint node, generate a set of joint actions allowed in the dance action based on its kinematic constraint conditions; Capture each frame of the dance action video data through a time convolution network, obtain the three-dimensional position data of each joint in each frame, obtain the time sequence characteristics of the joint motion, and classify and identify the dance action category; Obtain the adjacency relationship between the joints through a graph convolution neural network, treat the joints as nodes in the graph, and treat the skeleton correlation as the edges of the graph; Fuse the time sequence characteristics obtained by the time convolution network and the spatial characteristics obtained by the graph convolution network, construct a dance action auxiliary model, input a large amount of labeled dance action video data to train the dance action auxiliary model, obtain the three-dimensional joint motion trajectory data in the continuous frames, and based on the dance action auxiliary model, analyze the time sequence characteristics of the joint motion and the cooperative motion law in the dance action to obtain the correlation between the dance actions; Based on the trained dance action auxiliary model, input the three-dimensional human skeleton data of the current dance action into the dance action auxiliary model, combine the motion parameters allowed by each joint in the joint action set, obtain the joint action in the joint action set that has correlation with the current dance action, and generate a candidate action set; Based on each joint action in the candidate action set, obtain the difference between the joint position of the current dance action and the position of the corresponding joint in the candidate action, and calculate the position difference degree; Calculate the speed of the joints in the current dance action and the candidate action to obtain the speed connection fluency; Score the action connection fluency based on the position difference degree and the speed connection fluency; Score the pose aesthetic degree based on the symmetry, balance, stability of the action and the stretch, tension of the pose; Based on the action connection fluency score and the pose aesthetic degree score, obtain the weighted total score of each candidate action through comprehensive weighting, and select the candidate action with the highest score as the next dance action. 2.The dance motion auxiliary generation method based on three-dimensional modeling according to claim 1, wherein, The method for obtaining initial pose data of the target dance and constructing a three-dimensional human skeleton model comprises the following steps: Extract human key point data in each frame, including shoulders, elbows, knees, ankles, wrists, and obtain their three-dimensional coordinate data through existing pose estimation tools; Preprocess the three-dimensional coordinate data to generate dance pose data and map it to the three-dimensional human skeleton model; Correct the pose of the skeleton in the three-dimensional human skeleton model to ensure that the angle and position of each joint meet the actual dance action. 3.The dance motion auxiliary generation method based on three-dimensional modeling according to claim 1, wherein, The method for generating a set of motion parameters allowed by each joint in the dance action based on the three-dimensional human skeleton model and the kinematic constraint conditions of each joint node comprises the following steps: Construct motion action constraint conditions based on the physiological structure and physical limitations of each joint to obtain degrees of freedom, the constraint conditions include the maximum angle, minimum angle and rotation direction of the joint, and the degrees of freedom include rotation degrees of freedom and translation degrees of freedom; Based on the constraint condition and the degree of freedom, a joint motion model is constructed, all actions that can be completed by each joint are obtained, and a joint action set is generated, including discrete values of joint angles, speed and acceleration limits, and timing constraints. 4.The dance motion auxiliary generation method based on three-dimensional modeling according to claim 1, wherein, The three-dimensional human skeleton data of the current dance action is input into the trained dance action auxiliary model, and the motion parameters allowed by each joint in the joint action set are combined to obtain the joint actions related to the current dance action in the joint action set, and a candidate action set is generated, specifically including: The three-dimensional human skeleton data of the current dance action is input into the trained dance action auxiliary model, and the motion parameters allowed by each joint in the joint action set are combined to obtain the joint actions related to the current dance action in the joint action set, and a candidate action set is generated, specifically including: Based on the motion parameters allowed by each joint in the joint action set, the joint actions related to the current dance action are matched with the joint actions in the joint action set to obtain a matching probability distribution; Based on the joint actions in the joint action set that are higher than the preset threshold in the matching probability distribution, a candidate action set is generated.

5. A system for generating dance motions based on three-dimensional modeling, for implementing the method for generating dance motions based on three-dimensional modeling according to any one of claims 1 to 4, characterized in that, Including: The three-dimensional human skeleton module is used to extract the initial posture data of the target dance and construct a three-dimensional human skeleton model, ensuring that the angle and position of each joint meet the actual dance action; The motion parameter module generates the motion parameter set allowed by each joint based on the kinematic constraint condition, including joint angle, speed, and acceleration limit; The dance action auxiliary model module is used to fuse the timing convolution network and the graph convolution neural network, analyze a large number of labeled dance action video data, extract time sequence features, and obtain the relevance between dance actions; The candidate action set module is used to input the three-dimensional human skeleton data of the current dance action and generate a candidate action set related to the current action in combination with the joint action set; The scoring module is used to calculate the connection fluency and posture aesthetics scores of the candidate action and the current dance action, and obtain the candidate action with the highest total score as the next dance action output by weighting; The processor is used to process the calculation process of each formula and the construction calculation process of each model.

Images