A Machine Learning-Based Anime Character Motion Capture Method and System

Abstract

The present invention relates to the technical field of motion capture, and discloses an animation character motion capture method and system based on machine learning, including: collecting the spatial three-dimensional point cloud of the animation character to be captured and purifying it to obtain a purified point cloud; performing a human body structure simulation on the animation character to be captured to obtain a simulated human body structure, collecting inertial data, calculating the shortest cumulative distance between the point cloud frame and the data point, and constructing an initial feature matrix; setting physical constraints for the simulated human body structure to correct the node positions of the simulated human body structure to obtain an optimized human body structure; constructing an observer for the optimized human body structure, analyzing the initial state of the optimized human body structure, constructing an action sequence for the optimized human body structure, and segmenting it to obtain action sub-units; identifying action features for matching and calculating the probability value of the matching action to implement the motion capture of the animation character to be captured. The present invention can improve the accuracy of motion capture of animation characters.

Description

Technical Field

[0001] The present invention relates to a method and system for animating character motion capture based on machine learning, belonging to the technical field of motion capture. Background Art

[0002] Currently, animating character motion capture based on machine learning is an important technical means in the field of animation production. It uses machine learning algorithms to process human motion data to drive animating characters to present realistic motion performances.

[0003] During the motion capture process, traditional motion capture methods often rely on sensors and markers. This approach has many limitations. For example, sensors need to be precisely fixed on specific parts of the human body, and markers may affect the accuracy of data collection due to occlusion or environmental interference. Also, in complex motion scenarios, traditional methods are difficult to accurately capture subtle motions and rapidly changing motions, resulting in a significant reduction in the smoothness and naturalness of animating character motions. At the same time, some existing machine learning models face the problem of a huge workload for data annotation when processing motion capture data. Due to the diversity and complexity of motion data, accurately annotating each motion sample requires a large amount of human and time costs. Moreover, some models perform poorly in terms of generalization ability. When encountering new motion types that are quite different from the training data, they cannot effectively capture and transform them, making the motion performances of animating characters not rich and diverse enough. In addition, different animating characters have unique body proportions and style characteristics. When traditional motion capture technologies adapt to these diverse character requirements, a large amount of manual adjustment and optimization is required, which is inefficient and has a low accuracy rate.

[0004] Therefore, there is an urgent need for a solution to improve the accuracy of animating character motion capture. Summary of the Invention

[0005] The present invention provides a method and system for animating character motion capture based on machine learning, whose main purpose is to improve the accuracy of security authentication on the premise of reducing the workload of path detection.

[0006] To achieve the above object, a method for animating character motion capture based on machine learning provided by the present invention includes:

[0007] Collect the spatial three-dimensional point cloud of the animating character to be captured, construct the point cloud radius of the spatial three-dimensional point cloud, calculate the statistical features of each point cloud in the spatial three-dimensional point cloud according to the point cloud radius, and perform point cloud purification on the spatial three-dimensional point cloud based on the statistical features to obtain a purified point cloud;

[0008] Perform a human body structure simulation on the anime character to be captured to obtain a simulated human body structure, collect the inertial data of the anime character to be captured, construct a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data, calculate the shortest cumulative distance between the point cloud frames and the data points using the distance matrix, align the data points based on the shortest cumulative distance to obtain aligned data points, and construct an initial feature matrix of the simulated human body structure using the aligned data points;

[0009] Set physical constraints for the simulated human body structure using the initial feature matrix, where the physical constraints include physical motion constraints and joint length constraints. Based on the physical constraints, construct a node position optimization function for the simulated human body structure, and use the node position optimization function to correct the node positions of the simulated human body structure to obtain an optimized human body structure;

[0010] Query the relative position relationships and joint angle change amounts of the joint points in the optimized human body structure to construct an observer for the optimized human body structure, analyze the initial state of the optimized human body structure using the observer, construct a state transition probability matrix for the optimized human body structure based on the initial state, construct an action sequence for the optimized human body structure using the state transition probability matrix, and segment the action sequence to obtain action sub-units;

[0011] Identify the action features of the action sub-units, construct action labels for the action features, match the action labels with a pre-configured action library to obtain matching actions, calculate the probability values of the matching actions, determine the target action based on the probability values, and perform action capture on the anime character to be captured based on the target action.

[0012] Optionally, the step of purifying the spatial three-dimensional point cloud based on the statistical features to obtain a purified point cloud includes:

[0013] Calculate the point cloud dispersion degree of the spatial three-dimensional point cloud based on the statistical features using the following formula:

[0014] ;

[0015] where, represents the point cloud dispersion degree, represents the set of neighborhood points of point p in the spatial three-dimensional point cloud, represents the neighborhood point of point p, represents the distance from point p to and represents the average distance of the spatial three-dimensional point cloud in the statistical features;

[0016] Based on the point cloud dispersion, perform point cloud purification on the three-dimensional spatial point cloud to obtain a purified point cloud.

[0017] Optionally, constructing the distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data includes:

[0018] Perform time dimension conversion on the point cloud frame and the data point to obtain a converted point cloud frame and a converted data point;

[0019] Determine the spatial dimensions of the converted point cloud frame and the converted data point;

[0020] Based on the spatial dimensions, construct the spatial coordinates corresponding to the converted point cloud frame and the converted data point to obtain point cloud frame coordinates and data point coordinates;

[0021] Based on the point cloud frame coordinates and the data point coordinates, calculate the Euclidean distance between the corresponding points of the converted point cloud frame and the converted data point to obtain a distance value;

[0022] Store the distance value into a pre-constructed blank matrix to obtain a distance matrix.

[0023] Optionally, calculating the shortest cumulative distance between the point cloud frame and the data point using the distance matrix includes:

[0024] Query a point cloud sequence point in the point cloud frame and a data sequence point in the data point;

[0025] Use the distance matrix to construct the shortest cumulative distance matrix between the point cloud sequence point and the data sequence point;

[0026] Based on the shortest cumulative distance matrix, calculate the shortest cumulative distance between the point cloud sequence point and the data sequence point using the following formula:

[0027] ;

[0028] Wherein, represents the shortest cumulative distance, represents the shortest cumulative distance matrix, represents the th distance matrix between the point cloud frame and the bth data point, represents the distance between the previous sequence point of the th point cloud frame and the previous sequence point of the th data point, represents the th distance between the previous sequence point of the point cloud frame and the th data point, the th point cloud frame to the The distance between the current data point and the previous sequence point.

[0029] Optionally, constructing an initial feature matrix of the simulated human body structure by using the aligned data points, including:

[0030] Identifying the joint positions corresponding to human joints by using the aligned data points;

[0031] Querying the three-dimensional joint coordinates of the joints, and using the three-dimensional joint coordinates as the joint position features of the simulated human body structure;

[0032] Calculating the distances between adjacent joints of the joint positions, and determining the limb length features of the simulated human body structure based on the distances between adjacent joints;

[0033] Analyzing the distribution state of the data points in the field corresponding to the joint positions;

[0034] Determining the joint pose angle features of the simulated human body structure based on the distribution state;

[0035] Organizing the joint position features, the limb length features and the joint pose angle features to obtain sequence features;

[0036] Quantifying the sequence features to obtain an initial feature matrix.

[0037] Optionally, constructing a node position optimization function of the simulated human body structure based on the physical constraints, including:

[0038] Constructing a physical motion constraint function of the simulated human body structure based on the physical constraints;

[0039] Wherein, the physical motion constraint function can be expressed by the following formula:

[0040] ;

[0041] ;

[0042] Wherein, represents the physical motion constraint function, represents the angular range constraint value, m represents the number of corresponding joint nodes in the simulated human body structure, represents the joint angle variable of joint j, represents the minimum angle corresponding to, represents the maximum angle corresponding to;

[0043] Based on the physical constraints, a joint length constraint function for the simulated human body architecture is constructed, where the joint length constraint function can be expressed by the following formula:

[0044] ;

[0045] ;

[0046] where, represents the joint length constraint function, represents the joint length constraint value, represents the starting coordinate point of the joint in the simulated human body architecture, and L represents the joint length of the joint connected to the coordinate point and the coordinate point ;

[0047] Based on the physical motion constraint function and the joint length constraint function, a node position optimization function for the simulated human body architecture is determined, where the node position optimization function can be expressed by the following formula:

[0048] ;

[0049] where, represents the node position optimization function, represents the physical motion constraint function, represents the joint length constraint function.

[0050] Optionally, the analyzing the initial state of the optimized human body architecture using the observer includes:

[0051] Using the observer to construct a feature space for the optimized human body architecture;

[0052] Calculating the data point density of the feature space;

[0053] Based on the data point density, constructing a motion mode feature set for the optimized human body architecture

[0054] Using the motion mode feature set to calculate the state distribution of the optimized human body architecture;

[0055] Based on the state distribution, analyzing the initial state of the optimized human body architecture.

[0056] Optionally, the constructing the state transition probability matrix of the optimized human body architecture based on the initial state includes:

[0057] Setting an action sampling step for the initial state;

[0058] Based on the action sampling step, collecting a state transition data set for the initial state;

[0059] Construct an action analysis window of the state transition data set by using the action sampling step size;

[0060] Identify the hidden state corresponding to each step in the action sampling step size in the action analysis window to obtain a sequence of states;

[0061] Calculate the state transition frequency matrix of each state in the sequence of states;

[0062] Count the total number of occurrences of each state in the state transition frequency matrix;

[0063] Calculate the transition probability of each state based on the total number of occurrences by using the following formula:

[0064] ;

[0065] where, represents the transition probability, represents the total number of occurrences of state e in the state transition frequency matrix, represents the number of times from state e to state r;

[0066] Construct a state transition probability matrix of the optimized human body structure based on the transition probability.

[0067] Optionally, calculating the probability value of the matching action includes:

[0068] Calculate the cosine similarity between the matching action and the action feature corresponding to the action subunit;

[0069] Query the occurrence frequency of the matching action;

[0070] Calculate the action coordination rate of the matching action by using the following formula:

[0071] ;

[0072] where, represents the action coordination rate, represents a parameter for adjusting the steepness of the curve, represents the action coordination score of action c;

[0073] Calculate the probability value of the matching action according to the cosine similarity, the occurrence frequency, and the action coordination rate by using the following formula:

[0074] ;

[0075] ;

[0076] where, represents the cosine similarity, Indicates the frequency of occurrence of actions, Indicates the action coordination rate, Indicates the weight of the cosine similarity, Indicates the weight of the frequency of occurrence, Indicates the weight of the action coordination rate.

[0077] To solve the above problems, the present invention also provides an anime character action capture system based on machine learning, and the system includes:

[0078] A point cloud purification module, which is used to collect the three-dimensional spatial point cloud of the anime character to be captured, construct the point cloud radius of the three-dimensional spatial point cloud, calculate the statistical features of each point cloud in the three-dimensional spatial point cloud according to the point cloud radius, and perform point cloud purification on the three-dimensional spatial point cloud based on the statistical features to obtain a purified point cloud;

[0079] An architecture simulation module, which is used to simulate the human body architecture of the anime character to be captured to obtain a simulated human body architecture, collect the inertial data of the anime character to be captured, construct a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data, calculate the shortest cumulative distance between the point cloud frames and the data points using the distance matrix, perform data point alignment on the point cloud frames and the data points based on the shortest cumulative distance to obtain aligned data points, and construct an initial feature matrix of the simulated human body architecture using the aligned data points;

[0080] An architecture optimization module, which is used to set the physical constraints of the simulated human body architecture using the initial feature matrix, where the physical constraints include physical motion constraints and joint length constraints, construct a node position optimization function of the simulated human body architecture based on the physical constraints, and use the node position optimization function to correct the node positions of the simulated human body architecture to obtain an optimized human body architecture;

[0081] An action segmentation module, which is used to query the relative position relationship and joint angle change amount of the joint points in the optimized human body architecture to construct an observer of the optimized human body architecture, analyze the initial state of the optimized human body architecture using the observer, construct a state transition probability matrix of the optimized human body architecture based on the initial state, construct an action sequence of the optimized human body architecture using the state transition probability matrix, and perform action segmentation on the action sequence to obtain action sub-units;

[0082] An action capture module, which is used to identify the action features of the action sub-units, construct action labels of the action features, perform action matching between the action labels and a pre-configured action library to obtain matching actions, calculate the probability values of the matching actions, determine the target action based on the probability values, and perform action capture on the anime character to be captured based on the target action.

[0083] Compared with the problems described in the background art, the present invention first acquires the spatial three-dimensional point cloud of the anime character to be captured, constructs the point cloud radius and calculates the statistical features to purify the point cloud, so as to obtain the accurate original data of the character in the three-dimensional space and reduce the noise interference. Then, the human body structure of the character is simulated and inertial data is collected, the distance matrix is constructed to calculate the shortest cumulative distance to align the data points, and then the initial feature matrix of the simulated human body structure is constructed to realize the fusion and synchronization of different data, providing a high-quality data basis for subsequent motion capture; further, the present invention uses the initial feature matrix to set the physical constraints of the simulated human body structure, constructs the node position optimization function and corrects the node position to obtain the optimized human body structure, making the simulation conform to the physical laws of the real human body, enhancing the authenticity and reliability of the simulation; after that, the joint information of the optimized human body structure is queried to construct the observer, the initial state is analyzed and the state transition probability matrix is constructed to construct the action sequence, and then the action sequence is segmented into action subunits for in-depth analysis of the human motion structure. Finally, the action features of the action subunits are identified and action labels are constructed, which are matched with the pre-configured action library to obtain the matching actions, the probability values are calculated to determine the target actions, and the motion capture of the anime character is implemented based on the target actions. Considering various factors comprehensively, it is ensured that the selected actions meet the actual requirements, improving the vividness and naturalness of the anime character's actions, enhancing the quality of anime works, effectively solving many problems existing in traditional motion capture methods in terms of accuracy, data processing and action adaptation, and providing strong support and innovation for the development of anime character motion capture technology. Therefore, the machine learning-based anime character motion capture method and system provided by the embodiments of the present invention can improve the motion capture accuracy of anime characters. Brief Description of the Drawings

[0084] Figure 1 It is a schematic flow chart of a machine learning-based anime character motion capture method provided by an embodiment of the present invention;

[0085] Figure 2 It is a schematic module diagram for implementing the machine learning-based anime character motion capture method provided by an embodiment of the present invention.

[0086] The implementation, functional features and advantages of the present invention will be further described in conjunction with the embodiments with reference to the accompanying drawings. Detailed Embodiments

[0087] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0088] An embodiment of the present application provides a method for capturing the actions of anime characters based on machine learning. The execution entity of the method for capturing the actions of anime characters based on machine learning includes, but is not limited to, at least one of electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided in the embodiment of the present application. In other words, the method for capturing the actions of anime characters based on machine learning can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to: a single server, a server cluster, a cloud server, or a cloud server cluster, etc. Embodiment

[0089] Refer to Figure 1 As shown, it is a schematic flowchart of a method for capturing the actions of anime characters based on machine learning provided by an embodiment of the present invention. In this embodiment, the method for capturing the actions of anime characters based on machine learning includes:

[0090] S1. Collect the spatial three-dimensional point cloud of the anime character to be captured, construct the point cloud radius of the spatial three-dimensional point cloud, calculate the statistical features of each point cloud in the spatial three-dimensional point cloud according to the point cloud radius, and perform point cloud purification on the spatial three-dimensional point cloud based on the statistical features to obtain a purified point cloud.

[0091] In the embodiment of the present invention, by collecting the spatial three-dimensional point cloud of the anime character to be captured, the original data information of the anime character in the three-dimensional space, such as the shape and position of the character, etc., can be obtained, providing a basic basis for subsequent action capture.

[0092] Among them, the spatial three-dimensional point cloud refers to a set of a large number of discrete points obtained by a measuring device in the three-dimensional space, and can be collected by structured light measurement technology.

[0093] In the embodiment of the present invention, by constructing the point cloud radius of the spatial three-dimensional point cloud, it can be determined which points belong to the neighborhood of a certain point, and then the spatial three-dimensional point cloud can be divided into regions to facilitate the identification of the features of the point cloud.

[0094] Optionally, the point cloud radius can first perform a preliminary analysis on the collected three-dimensional point cloud data, such as calculating the distance distribution histogram between point clouds, observing its peaks and valleys, and then using a clustering algorithm (such as K-Means clustering) to cluster the point clouds, and calculating the average distance from the point clouds within each cluster to the center with each clustering center as a reference point.

[0095] In the embodiment of the present invention, by calculating the statistical features of each point cloud in the spatial three-dimensional point cloud according to the point cloud radius, the internal characteristics of the point cloud data can be better understood, providing a basis for determining whether the point cloud is a noise point or an outlier point, so as to realize the preliminary analysis and screening of the original point cloud data.

[0096] Among them, the statistical feature refers to the average distance and the standard deviation of the distance of the three-dimensional spatial point cloud.

[0097] Optionally, calculating the statistical features of each point cloud in the three-dimensional spatial point cloud according to the point cloud radius can be achieved by determining the neighborhood point set of each point based on the point cloud radius, calculating the sum of the distances between the point and its neighborhood points, and then dividing by the number of neighborhood points to obtain the average distance. For the standard deviation of the distance, first calculate the sum of the squares of the differences between the distances of the point and its neighborhood points and the average distance, and then take the square root after dividing by the number of neighborhood points.

[0098] In the embodiment of the present invention, by performing point cloud purification on the three-dimensional spatial point cloud based on the statistical features, the obtained purified point cloud can reduce the interference of noise and outliers on subsequent processing, and improve the accuracy and reliability of the entire motion capture process.

[0099] As an embodiment of the present invention, performing point cloud purification on the three-dimensional spatial point cloud based on the statistical features to obtain a purified point cloud includes: calculating the point cloud dispersion degree of the three-dimensional spatial point cloud based on the statistical features using the following formula:

[0100] ;

[0101] Among them, represents the point cloud dispersion degree, represents the set of neighborhood points of point p in the three-dimensional spatial point cloud, represents the neighborhood point of point p, represents the distance from point p to the distance of, represents the average distance of the three-dimensional spatial point cloud in the statistical features;

[0102] Based on the point cloud dispersion degree, perform point cloud purification on the three-dimensional spatial point cloud to obtain a purified point cloud.

[0103] Optionally, the process of performing point cloud purification on the three-dimensional spatial point cloud based on the point cloud dispersion degree to obtain a purified point cloud is as follows: first determine the dispersion degree threshold according to the shape characteristics of the anime character, data quality requirements, and subsequent application scenarios. For example, if motion capture is performed on an anime character with a relatively regular shape, for the point clouds of the main parts of the body (such as the torso and limbs), a relatively low dispersion degree threshold can be set because the point clouds of these parts are relatively evenly distributed; for complex structures such as hair, the threshold can be appropriately increased. Traverse all the points in the three-dimensional spatial point cloud. For each point, compare the calculated dispersion degree with the set threshold. If the dispersion degree of the point exceeds the threshold, determine that the point is a noise point or an outlier and remove it from the point cloud; if the dispersion degree is within the threshold range, retain the point. After such a screening process, the remaining points form the purified point cloud.

[0104] S2. Perform human body structure simulation on the to-be-captured anime character to obtain a simulated human body structure, collect the inertial data of the to-be-captured anime character, construct a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data, calculate the shortest cumulative distance between the point cloud frames and the data points using the distance matrix, perform data point alignment on the point cloud frames and the data points based on the shortest cumulative distance to obtain aligned data points, and construct an initial feature matrix of the simulated human body structure using the aligned data points.

[0105] In the embodiment of the present invention, through the above-mentioned human body structure simulation on the to-be-captured anime character to obtain a simulated human body structure, the character point cloud data can be presented in a more logical and structured human body framework form, which is convenient for corresponding to the real human body movement logic, so that the motion capture result can better conform to the actual human body movement law.

[0106] Among them, the simulated human body structure refers to a virtual model used to represent the human body structure and movement relationship. It is based on the connection mode of human bones and joints and constructs a structure similar to the real human bone framework in a computer environment.

[0107] Optionally, the simulated human body structure can be realized by using a pre-constructed general human model library (these models have standard human bone structures and joint connection modes). By matching and adjusting the approximate external dimensions of the anime character (such as height, limb length ratio, etc.) with the models, the general models are adapted to the anime character to achieve human body structure simulation. The pre-constructed general human model library can be configured using generative adversarial networks (GANs) or variational autoencoders (VAEs) in deep learning.

[0108] In the embodiment of the present invention, by collecting the inertial data of the to-be-captured anime character, the dynamic changes of the character in space can be obtained, such as the speed and direction change of limb swing, etc., so as to more comprehensively depict the movement process of the character, and further improve the accuracy and integrity of motion capture.

[0109] Among them, the inertial data refers to information about the motion state of an object, mainly including acceleration data and angular velocity data, which can be collected by an inertial measurement unit (IMU).

[0110] Furthermore, in the embodiment of the present invention, by constructing a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data, a quantitative correspondence of the spatial position relationship between two different types of data can be established to find the correlation between the two and realize the collaborative processing of data.

[0111] As an embodiment of the present invention, constructing the distance matrix between the point cloud frames and the data points in the purified point cloud includes: performing a time dimension conversion on the point cloud frames and the data points to obtain converted point cloud frames and converted data points, determining the spatial dimensions of the converted point cloud frames and the converted data points, constructing the corresponding spatial coordinates of the converted point cloud frames and the converted data points based on the spatial dimensions to obtain point cloud frame coordinates and data point coordinates, calculating the Euclidean distance between the corresponding points of the converted point cloud frames and the converted data points based on the point cloud frame coordinates and the data point coordinates to obtain distance values, and storing the distance values into a pre-constructed blank matrix to obtain the distance matrix.

[0112] Optionally, performing a time dimension conversion on the point cloud frames and the data points to obtain converted point cloud frames and converted data points can be achieved by unifying the representations of the point cloud frames and the data points on the time scale through interpolation, resampling, etc. according to information such as the timestamps of the data, so that they are under the same time dimension standard. The spatial dimension can be determined by analyzing the spatial range involved in the data to clarify whether it is a two-dimensional or three-dimensional space situation. Usually, for this type of data describing the shape of an object, it is mostly determined as a three-dimensional spatial dimension. The spatial coordinates can be obtained by extracting or converting the specific coordinate values of each point cloud frame and data point in this space from the corresponding data according to the determined spatial dimension (for example, three dimensions correspond to the X, Y, and Z axes).

[0113] In the embodiment of the present invention, by calculating the shortest cumulative distance between the point cloud frames and the data points using the distance matrix, the most reasonable matching method can be determined, so that the point cloud frames and the inertial data points can be aligned in the most fitting order and position subsequently, reducing the motion capture error caused by data asynchrony and mismatch, and ensuring the continuity and accuracy of motion capture.

[0114] It should be explained that the shortest cumulative distance refers to an optimal cumulative distance metric calculated through dynamic programming when aligning two time series, and this distance represents a "cost" for matching the two sequences in the time dimension. The cumulative distance corresponding to the path with the smallest cost is the shortest cumulative distance.

[0115] As an embodiment of the present invention, calculating the shortest cumulative distance between the point cloud frames and the data points using the distance matrix includes: querying a point cloud sequence point in the point cloud frames and a data sequence point in the data points, constructing the shortest cumulative distance matrix of the point cloud sequence point and the data sequence point using the distance matrix, and calculating the shortest cumulative distance between the point cloud sequence point and the data sequence point based on the shortest cumulative distance matrix using the following formula:

[0116] ;

[0117] Among them, represents the shortest cumulative distance, represents the shortest cumulative distance matrix, represents the distance matrix between the th point cloud frame and the th data point, distance between the previous sequential point of the th point cloud frame and the previous sequential point of the th data point.

[0118] Optionally, the point cloud sequential points can determine a specific point cloud sequential point from the point cloud frame dataset. This point cloud sequential point can be a representative point or a starting point selected in a certain order (such as chronological order or spatial position order). The data sequential points can select a corresponding starting data sequential point from the data point dataset. The shortest cumulative distance matrix can be obtained by gradually calculating the shortest cumulative distance at each position starting from the starting point based on the selected point cloud sequential points and data sequential points, using the distance matrix according to the idea of dynamic programming.

[0119] In the embodiment of the present invention, by aligning the point cloud frame and the data point based on the shortest cumulative distance to obtain the aligned data points, the fusion and synchronization of data from different sources can be realized, providing a high-quality data basis for more accurately constructing the initial feature matrix of the human body structure.

[0120] Optionally, the process of aligning the point cloud frame and the data point based on the shortest cumulative distance to obtain the aligned data points is as follows: starting from the lower right corner of the calculated shortest cumulative distance matrix, backtracking is performed according to the path information recorded when constructing the matrix (that is, whether each element comes from the left, above, or the upper left), and the shortest path from the starting point to the end point is obtained. The point pairs on this path represent the optimal alignment relationship between the point cloud frame and the data point in the time or space sequence. According to the point pair relationship on the shortest path, the points in the point cloud frame and the points in the data point are corresponded one by one.

[0121] In the embodiment of the present invention, by using the aligned data points to construct the initial feature matrix of the simulated human body structure, it can help users analyze the human body structure state, identify joint points, and perform action classification and other operations to provide standardized data input. It is an important link connecting the previous and the next in the entire motion capture process, and helps to promote the smooth development of subsequent motion capture and analysis work.

[0122] As an embodiment of the present invention, constructing an initial feature matrix of the simulated human body structure by using the aligned data points includes: identifying the joint positions corresponding to human joints by using the aligned data points, querying the three-dimensional coordinates of the joints at the joint positions, taking the three-dimensional coordinates of the joints as the joint position features of the simulated human body structure, calculating the distances between adjacent joints at the joint positions, determining the limb length features of the simulated human body structure based on the distances between adjacent joints, analyzing the distribution states of the data points in the corresponding fields of the joint positions, determining the joint posture angle features of the simulated human body structure based on the distribution states, organizing the data of the joint position features, the limb length features, and the joint posture angle features to obtain sequence features, and quantifying the sequence features to obtain an initial feature matrix.

[0123] Optionally, the step of identifying the joint positions corresponding to human joints by using the aligned data points can establish a joint recognition model through learning and analyzing a large amount of human motion data according to the relative position relationships and motion patterns of human joints in three-dimensional space, and then input the aligned data points into the model. The model determines the point cloud regions corresponding to human joints according to features such as the spatial distribution, density change, and motion trend of the data points, so as to identify the joint positions. The step of calculating the distances between adjacent joints at the joint positions can, after determining the coordinates of each joint position, use the distance formula between two points in space to calculate the distances between adjacent joints. The step of analyzing the distribution states of the data points in the corresponding fields of the joint positions can take the joint positions as the centers, select the neighborhood data points within a certain range, and observe the distribution of these data points in space. The distribution states are analyzed by calculating indexes such as the vector angles between the data points and the connecting lines of the joint positions and the point cloud density changes. For example, when the knee joint bends, the distribution of the data points before and after the joint will show obvious angle changes and density differences, and the posture information such as the bending angle of the joint is determined according to these features. The step of organizing the data of the joint position features, the limb length features, and the joint posture angle features can be implemented by arranging and combining the position features, limb length features, and joint posture angle features of each joint in sequence according to the logical order of the human body structure, usually starting from the head joints and following the order of the spine, upper limbs, lower limbs, etc.

[0124] S3. Setting physical constraints of the simulated human body structure by using the initial feature matrix, where the physical constraints include physical motion constraints and joint length constraints. Based on the physical constraints, constructing a node position optimization function of the simulated human body structure, and using the node position optimization function to correct the node positions of the simulated human body structure to obtain an optimized human body structure.

[0125] In the embodiments of the present invention, setting the physical constraints of the simulated human body structure by using the initial feature matrix can introduce physical constraints, enabling the movement and morphological changes of the simulated human body structure to follow the physiological laws of the real human body, avoiding actions or body deformations that do not conform to common sense, improving the authenticity and reliability of the simulation, and laying a foundation for subsequent accurate motion capture and animation production.

[0126] Among them, the physical constraints include physical motion constraints and joint length constraints. Further explanation is that the physical constraints are physical limitations that define the range and manner of movement that human joints can perform in each direction, and the joint length constraint refers to setting the length limit of the bones connected to each joint according to the fixed length characteristics of the human bone structure.

[0127] Optionally, the physical constraints can extract the joint attitude angles from the initial feature matrix to determine the joint angle range constraints, and clarify the limb movement direction constraints according to the limb length and joint positions; then extract the limb length data to determine the standard value of the joint length, and establish an equation using the distance formula between two points to set the long physical constraints.

[0128] Furthermore, in the embodiments of the present invention, constructing the node position optimization function of the simulated human body structure based on the physical constraints can provide a quantitative means to control the change of the node positions of the simulated human body structure, enabling the automatic adjustment of the node positions according to these functional relationships in subsequent calculations and simulations to meet the physical constraint requirements. This helps to achieve automated and accurate human body structure simulation and optimization, improving the efficiency and accuracy of the entire motion capture and simulation system.

[0129] As an embodiment of the present invention, constructing the node position optimization function of the simulated human body structure based on the physical constraints includes: constructing the physical motion constraint function of the simulated human body structure based on the physical constraints, where the physical motion constraint function can be expressed by the following formula:

[0130] ;

[0131] ;

[0132] Wherein, represents the physical motion constraint function, represents the angle range constraint value, m represents the number of corresponding joint nodes in the simulated human body structure, represents the joint angle variable of joint j, represents the corresponding minimum angle, represents the corresponding maximum angle;

[0133] Based on the physical constraints, a joint length constraint function for the simulated human body architecture is constructed, where the joint length constraint function can be expressed by the following formula:

[0134] ;

[0135] ;

[0136] where, represents the joint length constraint function, represents the joint length constraint value, represents the starting coordinate point of the joint in the simulated human body architecture, and L represents the joint length of the joint connecting the coordinate point and the coordinate point ;

[0137] Based on the physical motion constraint function and the joint length constraint function, a node position optimization function for the simulated human body architecture is determined, where the node position optimization function can be expressed by the following formula:

[0138] ;

[0139] where, represents the node position optimization function, represents the physical motion constraint function, represents the joint length constraint function.

[0140] In the embodiment of the present invention, by using the node position optimization function to correct the node positions of the simulated human body architecture, an optimized human body architecture can be obtained, which can make the simulated human body architecture fully conform to the set physical constraints and eliminate the problems that do not conform to the physical characteristics of the real human body caused by reasons such as data errors, calculation deviations, or incomplete initial conditions.

[0141] Optionally, the process of using the node position optimization function to correct the node positions of the simulated human body architecture to obtain an optimized human body architecture is as follows: First, obtain the initial values of the node positions and joint angles of the simulated human body architecture, and determine the parameters such as the joint angle range and length in the node position optimization function. Then, select the gradient descent method to calculate the gradients of the optimization function with respect to the joint angle variable and the node coordinate variable. According to the set step size, iterate and update these variables in the opposite direction of the gradient. After each iteration, determine whether to converge by judging whether the function value changes or the gradient norm is less than the threshold, until convergence to obtain an optimized human body architecture that conforms to the physical constraints.

[0142] S4. Query the relative position relationship of joint points and the change amount of joint angles in the optimized human body structure to construct an observer for the optimized human body structure. Use the observer to analyze the initial state of the optimized human body structure, construct a state transition probability matrix for the optimized human body structure based on the initial state, construct an action sequence for the optimized human body structure using the state transition probability matrix, and perform action segmentation on the action sequence to obtain action sub-units.

[0143] In the embodiment of the present invention, by querying the relative position relationship of joint points and the change amount of joint angles in the optimized human body structure to construct an observer for the optimized human body structure, information such as the relative position relationship of joint points and the change amount of joint angles can be integrated and abstracted to form a set of symbols or variables that can represent the current state of the optimized human body structure, thereby facilitating the description of the state of the human body structure.

[0144] Among them, the observer refers to a symbol used to describe the state of the optimized human body structure. It integrates key information such as the relative position relationship of joint points and the change amount of joint angles queried from the optimized human body structure, and can be regarded as a "state label" that can concisely and effectively summarize the state characteristics of the human body structure at a certain moment.

[0145] Optionally, to query the relative position relationship of joint points and the change amount of joint angles in the optimized human body structure to construct an observer for the optimized human body structure, joint point coordinate data can be obtained through an action capture device or a sensor, and the relative position relationship can be obtained by calculating the coordinate difference between adjacent joint points; use an angle measurement tool or a vector calculation method based on coordinate data to measure the change amount of joint angles; finally, after integrating and quantifying these data, they are used as components of the observer to construct the observer.

[0146] Furthermore, in the embodiment of the present invention, by using the observer to analyze the initial state of the optimized human body structure, the starting point of the movement of the human body structure can be better understood, and then its subsequent changes and developments can be analyzed.

[0147] As an embodiment of the present invention, using the observer to analyze the initial state of the optimized human body structure includes: using the observer to construct a feature space for the optimized human body structure, calculating the data point density of the feature space, constructing a set of motion modal features for the optimized human body structure based on the data point density, using the set of motion modal features to calculate the state distribution of the optimized human body structure, and analyzing the initial state of the optimized human body structure based on the state distribution.

[0148] Optionally, the feature space can use observation symbols such as the relative position relationship of joint points and the change amount of joint angles as dimensions, and combine them into a joint point feature vector describing the state of the human body structure, and is constructed using the special joint point feature vector. The data point density can use a density estimation algorithm (such as kernel density estimation) to determine the density around each data point in the feature space. The set of motion mode features can group the feature vectors in regions with similar densities according to the data point density to form a feature set representing different motion modes. Calculating the state distribution of the optimized human body structure can obtain the probability distribution of the optimized human body structure in different motion modes (states) by statistically analyzing the proportion of each set of motion mode features in the data. The initial state can select the state corresponding to the motion mode with the highest probability as the initial state of the optimized human body structure according to the state distribution.

[0149] In the embodiment of the present invention, the state transition probability matrix of the optimized human body structure constructed based on the initial state can quantitatively describe the possibility of the human body structure converting between different actions and postures, providing an important basis for predicting the subsequent actions and postures of the human body.

[0150] As an embodiment of the present invention, constructing the state transition probability matrix of the optimized human body structure based on the initial state includes: setting the action sampling step length of the initial state, collecting the state transition data set of the initial state based on the action sampling step length, constructing an action analysis window of the state transition data set using the action sampling step length, identifying the hidden state corresponding to each step length in the action sampling step length in the action analysis window to obtain a sequence of states, calculating the state transition frequency matrix of each state in the sequence of states, statistically analyzing the total number of times each state appears in the state transition frequency matrix, and calculating the transition probability of each state based on the total number using the following formula:

[0151] ;

[0152] Wherein, represents the transition probability, represents the total number of times state e appears in the state transition frequency matrix, represents the number of times from state e to state r;

[0153] Based on the transition probability, construct the state transition probability matrix of the optimized human body structure.

[0154] Optionally, the action sampling step size can determine an appropriate time interval as the sampling step size according to the characteristics of the action data and the analysis requirements. For example, sampling is performed every 0.2 seconds for subsequent data collection and analysis. The state transition data set can be composed of relevant action data in the initial state extracted from the data containing the actions of the optimized human body structure according to the set action sampling step size. Based on the action sampling step size, the action analysis window determines the window length (for example, one window contains 5 sampling steps), and the data set is divided accordingly to form individual action analysis windows. For each action corresponding to each sampling step within the action analysis window, the hidden state can be determined using the hidden Markov model, and these sequentially arranged hidden states are combined into a sequence state. The state transition frequency matrix can be obtained by statistically recording the number of times of transitioning from one state to another for the obtained sequence state. The state transition probability matrix can be completed by filling in the calculated state transition probabilities at the corresponding positions in a pre-set blank matrix.

[0155] In the embodiment of the present invention, by constructing the action sequence of the optimized human body structure using the state transition probability matrix, the movement process and behavior pattern of the human body structure within a period of time can be intuitively displayed, which is helpful for the overall understanding and grasp of complex human movements.

[0156] Optionally, when constructing the action sequence of the optimized human body structure using the state transition probability matrix, it can start from the initial state, randomly select the next state according to the state transition probabilities in the state transition probability matrix, and continuously repeat this process. Then, the sequentially selected states are connected in order, and each state corresponds to the movement information of a specific joint point. When combined, the action sequence of the optimized human body structure is constructed.

[0157] In the embodiment of the present invention, by segmenting the action sequence to obtain action sub-units, the continuous action sequence can be divided into several relatively independent action sub-units with specific semantics and functions according to certain criteria and methods, which is convenient for more detailed analysis of the structure and composition of human movements and is conducive to the classification, recognition, and research of different types of actions.

[0158] Optionally, when segmenting the action sequence to obtain action sub-units, it can be segmented according to the change of states in the action sequence. When a significant change in the state occurs, such as changing from "walking" to "jumping", the segmentation point is determined at the state transition, and the action sequence is truncated, thereby obtaining action sub-units with relatively independent semantics, such as walking sub-units, jumping sub-units, etc.

[0159] S5. Identify the action features of the action subunit, construct action labels for the action features, perform action matching between the action labels and a pre-configured action library to obtain matching actions, calculate the probability values of the matching actions, determine the target action based on the probability values, and perform action capture on the anime character to be captured based on the target action.

[0160] In the embodiment of the present invention, by identifying the action features of the action subunit, the action subunit can be visualized from a relatively abstract state sequence into a feature set that can be described and quantified.

[0161] Optionally, the action features can be obtained by using a deep learning model to extract key information such as the motion trajectory and speed change of the joints of the action subunit.

[0162] Furthermore, in the embodiment of the present invention, by constructing action labels for the action features, clear semantic labels can be assigned to each action subunit, facilitating subsequent data management, retrieval, and matching operations.

[0163] Optionally, the action labels can be automatically constructed using label scripts generated by java. For example, "walking" is 01 and "jumping" is 001, which can be specifically set according to the actual application.

[0164] In the embodiment of the present invention, by performing action matching between the action labels and a pre-configured action library to obtain matching actions, rapid classification and accurate positioning of the currently captured action subunit can be achieved.

[0165] Optionally, in the action library, the matching actions can be quickly screened out as matching actions through label indexing for actions with the same label.

[0166] In the embodiment of the present invention, by calculating the probability values of the matching actions, quantitative comparison and evaluation of multiple matching actions can be performed, scientifically measuring the suitability of each action to become the final target action, avoiding one-sidedness and uncertainty caused by simply relying on subjective judgment or simple rules to select actions, and providing an objective basis for accurately determining the action ultimately used for the anime character.

[0167] As an embodiment of the present invention, calculating the probability value of the matching action includes: calculating the cosine similarity between the matching action and the action features corresponding to the action subunit, querying the occurrence frequency of the matching action, and calculating the action coordination rate of the matching action using the following formula:

[0168] ;

[0169] Wherein, represents the action coordination rate, represents a parameter that adjusts the steepness of the curve, The action coordination score representing action c

[0170] According to the cosine similarity, the occurrence frequency, and the action coordination rate, use the following formula to calculate the probability value of the matching action:

[0171] ;

[0172] ;

[0173] wherein represents the cosine similarity represents the action occurrence frequency represents the action coordination rate represents the weight of the cosine similarity represents the weight of the occurrence frequency represents the weight of the action coordination rate

[0174] It should be explained that the action coordination score is obtained by using a machine learning model (such as a recurrent neural network) to learn the coordination pattern between action sequences, with the entire action sequence (including the current action subunit and the actions before and after it) and the matching action as inputs.

[0175] In the embodiment of the present invention, by determining the target action based on the probability value, it can be ensured that the finally selected action is the most in line with the actual capture situation and the requirements of anime production after comprehensively considering various factors, making the action performance of the anime character not only conform to the human motion logic but also be coordinated with the overall action style, scene, etc., improving the realism and naturalness of the anime character's actions and the quality of the anime work.

[0176] Optionally, the target action can compare the probability values and select the matching action corresponding to the maximum probability value as the target action.

[0177] In the embodiment of the present invention, by implementing action capture on the anime character to be captured based on the target action, the determined target action can be accurately mapped onto the anime character to be captured, enabling the anime character to perform corresponding joint movements, pose changes, etc. according to this action, realizing the transfer of action data from actual capture to the virtual character and completing the final implementation link of action capture.

[0178] Optionally, performing motion capture on the anime character to be captured based on the target action can deeply analyze the bone structure of the anime character to be captured, and accurately map the joint points in the target motion data to the character's bone joints. Then, coordinate transformation and scaling are performed to adapt the motion data to the character coordinate system and scale. Next, interpolation and smoothing processing are performed on the motion data to enhance the continuity and smoothness of the motion. Finally, the processed data is assigned to the character's bone joint attributes frame by frame through a programming interface to drive the character to smoothly present the animation effect according to the target action in the virtual scene. Embodiment

[0179] As Figure 2 shown, it is a functional module diagram of an anime character motion capture system based on machine learning according to the present invention.

[0180] The anime character motion capture system 200 based on machine learning according to the present invention can be installed in an electronic device. According to the implemented functions, the anime character motion capture system based on machine learning may include a point cloud purification module 201, an architecture simulation module 202, an architecture optimization module 203, an action segmentation module 204, and an action capture module 205. The modules in the present invention may also be referred to as units, which refer to a series of computer program segments that can be executed by a processor of an electronic device and can complete fixed functions, and are stored in the memory of the electronic device.

[0181] In the embodiments of the present invention, the functions of each module / unit are as follows:

[0182] The point cloud purification module is used to collect the three-dimensional spatial point cloud of the anime character to be captured, construct the point cloud radius of the three-dimensional spatial point cloud, calculate the statistical features of each point cloud in the three-dimensional spatial point cloud according to the point cloud radius, and perform point cloud purification on the three-dimensional spatial point cloud based on the statistical features to obtain a purified point cloud;

[0183] The architecture simulation module is used to perform human body architecture simulation on the anime character to be captured to obtain a simulated human body architecture, collect the inertial data of the anime character to be captured, construct a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data, calculate the shortest cumulative distance between the point cloud frames and the data points using the distance matrix, perform data point alignment on the point cloud frames and the data points based on the shortest cumulative distance to obtain aligned data points, and construct an initial feature matrix of the simulated human body architecture using the aligned data points;

[0184] An architecture optimization module, configured to set physical constraints of the simulated human body architecture by using the initial feature matrix, where the physical constraints include physical motion constraints and joint length constraints, and based on the physical constraints, construct a node position optimization function of the simulated human body architecture, and use the node position optimization function to correct the node positions of the simulated human body architecture to obtain an optimized human body architecture;

[0185] An action segmentation module, configured to query the relative position relationships and joint angle change amounts of the joint points in the optimized human body architecture to construct an observer of the optimized human body architecture, analyze the initial state of the optimized human body architecture by using the observer, construct a state transition probability matrix of the optimized human body architecture based on the initial state, construct an action sequence of the optimized human body architecture by using the state transition probability matrix, and segment the action sequence to obtain action sub-units;

[0186] An action capture module, configured to identify the action features of the action sub-units, construct action labels of the action features, match the action labels with a pre-configured action library to obtain matching actions, calculate probability values of the matching actions, determine a target action based on the probability values, and perform action capture on the anime character to be captured based on the target action.

[0187] Specifically, each module in the machine learning-based anime character action capture system 200 in the embodiments of the present invention adopts the same technical means as those in the above-mentioned Figure 1 machine learning-based anime character action capture method described above, and can produce the same technical effects, which will not be elaborated here.

[0188] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for capturing the actions of anime characters based on machine learning, characterized in that, The method includes: Collecting the spatial three-dimensional point cloud of the anime character to be captured, constructing the point cloud radius of the spatial three-dimensional point cloud, calculating the statistical features of each point cloud in the spatial three-dimensional point cloud according to the point cloud radius, and performing point cloud purification on the spatial three-dimensional point cloud based on the statistical features to obtain a purified point cloud; Performing human body structure simulation on the anime character to be captured to obtain a simulated human body structure, collecting the inertial data of the anime character to be captured, constructing a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data, calculating the shortest cumulative distance between the point cloud frames and the data points using the distance matrix, performing data point alignment on the point cloud frames and the data points based on the shortest cumulative distance to obtain aligned data points, and constructing an initial feature matrix of the simulated human body structure using the aligned data points; Setting physical constraints for the simulated human body structure using the initial feature matrix, where the physical constraints include physical motion constraints and joint length constraints, constructing a node position optimization function for the simulated human body structure based on the physical constraints, and correcting the node positions of the simulated human body structure using the node position optimization function to obtain an optimized human body structure; Querying the relative position relationship and joint angle change amount of the joint points in the optimized human body structure to construct an observer for the optimized human body structure, analyzing the initial state of the optimized human body structure using the observer, constructing a state transition probability matrix for the optimized human body structure based on the initial state, constructing an action sequence for the optimized human body structure using the state transition probability matrix, and segmenting the action sequence to obtain action sub-units; Identifying the action features of the action sub-units, constructing action labels for the action features, matching the action labels with a pre-configured action library to obtain matching actions, calculating the probability values of the matching actions, determining the target action based on the probability values, and performing action capture on the anime character to be captured based on the target action.

2. The method for capturing the actions of anime characters based on machine learning according to claim 1, wherein The performing point cloud purification on the spatial three-dimensional point cloud based on the statistical features to obtain a purified point cloud includes: Calculating the point cloud dispersion degree of the spatial three-dimensional point cloud based on the statistical features using the following formula: ； Among them, represents the point cloud dispersion degree, represents the set of neighborhood points of point p in the three-dimensional spatial point cloud, Q represents the neighborhood points of point p, represents the distance from point p to Q, represents the average distance of the three-dimensional spatial point cloud in the statistical features; Performing point cloud purification on the spatial three-dimensional point cloud based on the point cloud dispersion degree to obtain a purified point cloud.

3. The method for animating character motion capture based on machine learning according to claim 1, characterized in that, The constructing a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data includes: Performing time dimension conversion on the point cloud frames and the data points to obtain converted point cloud frames and converted data points; Determining the spatial dimensions of the converted point cloud frames and the converted data points; Constructing corresponding spatial coordinates for the converted point cloud frames and the converted data points based on the spatial dimensions to obtain point cloud frame coordinates and data point coordinates; Calculating the Euclidean distance between the corresponding points of the converted point cloud frames and the converted data points based on the point cloud frame coordinates and the data point coordinates to obtain distance values; Storing the distance values into a pre-constructed blank matrix to obtain a distance matrix.

4. The method for capturing the actions of anime characters based on machine learning according to claim 1, wherein, The calculating the shortest cumulative distance between the point cloud frames and the data points using the distance matrix includes: Query a point cloud sequence point in the point cloud frame and a data sequence point in the data points; Construct the shortest cumulative distance matrix of the point cloud sequence point and the data sequence point by using the distance matrix; Based on the shortest cumulative distance matrix, calculate the shortest cumulative distance between the point cloud sequence point and the data sequence point by using the following formula: ； Among them, represents the shortest cumulative distance, represents the shortest cumulative distance matrix, represents the th point cloud frame and the th data point distance matrix, represents the distance between the previous sequence point of the th point cloud frame and the previous sequence point of the th data point, represents the distance between the previous sequence point of the th point cloud frame and the th data point, represents the distance between the th point cloud frame and the previous sequence point of the th data point.

5. The method for animating character motion capture based on machine learning according to claim 1, wherein Construct the initial feature matrix of the simulated human body structure by using the aligned data points, including: Identify the joint positions corresponding to the human joints by using the aligned data points; Query the three-dimensional joint coordinates of the joint positions, and use the three-dimensional joint coordinates as the joint position features of the simulated human body structure; Calculate the adjacent joint distances of the joint positions, and determine the limb length features of the simulated human body structure based on the adjacent joint distances; Analyze the distribution state of the data points in the corresponding field of the joint positions; Determine the joint pose angle features of the simulated human body structure based on the distribution state; Organize the data of the joint position features, the limb length features and the joint pose angle features to obtain sequence features; Quantify the sequence features to obtain the initial feature matrix.

6. The method for animating character motion capture based on machine learning according to claim 1, characterized in that Construct the node position optimization function of the simulated human body structure based on the physical constraints, including: Construct the physical motion constraint function of the simulated human body structure based on the physical constraints; Among them, the physical motion constraint function can be expressed by the following formula: ； ； Among them, represents the physical motion constraint function, represents the angular range constraint value, m represents the number of corresponding joint nodes in the simulated human body structure, represents the joint angle variable of joint j, represents the corresponding minimum angle, represents the corresponding maximum angle; Construct the joint length constraint function of the simulated human body structure based on the physical constraints, where the joint length constraint function can be expressed by the following formula: ； ； Among them, represents the joint length constraint function, represents the joint length constraint value, represents the serial number of the coordinate point of the segment in the simulated human body structure, and L represents the coordinate point and the coordinate point the joint length of the joint connected to it; Determine the node position optimization function of the simulated human body structure based on the physical motion constraint function and the joint length constraint function, where the node position optimization function can be expressed by the following formula: ； Among them, represents the node position optimization function, represents the physical motion constraint function, represents the joint length constraint function.

7. The method for capturing the actions of anime characters based on machine learning according to claim 1, characterized in that Analyze the initial state of the optimized human body structure by using the observer, including: Construct the feature space of the optimized human body structure by using the observer; Calculate the data point density of the feature space; Construct the motion mode feature set of the optimized human body structure based on the data point density Calculate the state distribution of the optimized human body structure by using the motion mode feature set; Analyze the initial state of the optimized human body structure based on the state distribution.

8. The method for animating character motion capture based on machine learning according to claim 1, characterized in that, Construct the state transition probability matrix of the optimized human body structure based on the initial state, including: Set the action sampling step length of the initial state; Collect the state transition data set of the initial state based on the action sampling step length; Construct an action analysis window of the state transition data set by using the action sampling step length; Identify the hidden states of each step in the action sampling step length in the action analysis window to obtain sequence states; Calculate the state transition frequency matrix of each state in the sequence states; Count the total number of times each state appears in the state transition frequency matrix; Calculate the transition probability of each state by using the following formula based on the total number of times; ； Among them, represents the transition probability, represents the total number of times state e appears in the state transition frequency matrix, represents the number of times from state e to state r; Construct the state transition probability matrix of the optimized human body structure based on the transition probability.

9. The method for capturing the actions of anime characters based on machine learning according to claim 1, wherein, Calculate the probability value of the matching action, including: Calculate the cosine similarity between the matching action and the corresponding action feature of the action subunit; Query the occurrence frequency of the matching action; Calculate the action coordination rate of the matching action using the following formula: ； Among them, represents the action coordination rate, represents a parameter for adjusting the steepness of the curve, represents the action coordination score of action c; Calculate the probability value of the matching action using the following formula based on the cosine similarity, the occurrence frequency, and the action coordination rate; ； ； Among them, represents the cosine similarity, represents the occurrence frequency, represents the action coordination rate, represents the weight of the cosine similarity, represents the weight of the occurrence frequency, represents the weight of the action coordination rate.

10. An anime character motion capture system based on machine learning, characterized in that, The system includes: A point cloud purification module, configured to collect the spatial three-dimensional point cloud of the anime character to be captured, construct the point cloud radius of the spatial three-dimensional point cloud, calculate the statistical features of each point cloud in the spatial three-dimensional point cloud according to the point cloud radius, and perform point cloud purification on the spatial three-dimensional point cloud based on the statistical features to obtain a purified point cloud; An architecture simulation module, configured to perform human body architecture simulation on the anime character to be captured to obtain a simulated human body architecture, collect the inertial data of the anime character to be captured, construct a distance matrix between the point cloud frames in the purified point cloud and the data points in the inertial data, calculate the shortest cumulative distance between the point cloud frames and the data points using the distance matrix, perform data point alignment on the point cloud frames and the data points based on the shortest cumulative distance to obtain aligned data points, and construct an initial feature matrix of the simulated human body architecture using the aligned data points; An architecture optimization module, configured to set the physical constraints of the simulated human body architecture using the initial feature matrix, where the physical constraints include physical motion constraints and joint length constraints, construct a node position optimization function of the simulated human body architecture based on the physical constraints, and correct the node positions of the simulated human body architecture using the node position optimization function to obtain an optimized human body architecture; An action segmentation module, configured to query the relative position relationship and the joint angle change amount of the joint points in the optimized human body architecture to construct an observer of the optimized human body architecture, analyze the initial state of the optimized human body architecture using the observer, construct a state transition probability matrix of the optimized human body architecture based on the initial state, construct an action sequence of the optimized human body architecture using the state transition probability matrix, and perform action segmentation on the action sequence to obtain action subunits; An action capture module, configured to identify the action features of the action subunits, construct action labels of the action features, perform action matching between the action labels and a pre-configured action library to obtain matching actions, calculate the probability values of the matching actions, determine a target action based on the probability values, and perform action capture on the anime character to be captured based on the target action.

Images