Action sports scoring method and device based on big data analysis

Abstract

The invention discloses an action sports scoring method and device based on big data analysis, and relates to the technical field of computer vision, and the method comprises the following steps: S1, constructing a human body posture tensor manifold; s2, generating aligned action track characteristics; s3, generating symmetrical positive definite manifold features; s4, generating deep manifold distribution parameters; s5, quantifying global distribution deviation characteristics; s6, calculating a residual vector based on the deep manifold distribution parameters and the standard action distribution parameters, inputting the residual vector into the improved AGCN model for processing, constructing an adaptive topology based on the residual vector, and performing aggregation analysis to obtain a joint-level physical angle error after multi-scale space-time convolution and manifold enhancement; and S7, outputting a comprehensive score vector. According to the method, the limitation that a traditional method only depends on local geometric features and ignores global topological structure constraints and statistical distribution priori is overcome, and an efficient solution is provided for intelligent scoring of sports actions.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and apparatus for scoring motion sports based on big data analysis. Background Technology

[0002] With the increasing granularity of sports training data collection, existing scoring systems face severe computational challenges in feature extraction and accurate quantification evaluation of massive amounts of high-dimensional motion data. Existing traditional motion scoring algorithms, such as improved 3D convolutional neural networks (3DCNN) or methods based on skeletal joint distances, while utilizing deep learning models to improve basic recognition efficiency, primarily rely on geometric coordinate errors in human posture within Euclidean space or similarity matching of single frames for scoring. This purely geometric distance-based approach ignores the implicit dynamic topological structure of motion sequences in high-dimensional manifold space (such as manifold curvature changes and phase distribution patterns of motion trajectories) and deep-seated motion law relationships, leading to distribution distortion in the high-dimensional manifold space when constructing scoring criteria. This limits the ability to accurately and globally evaluate complex and high-difficulty sports movements. Furthermore, classic scoring methods often struggle to fully utilize prior knowledge of statistical distributions in large-scale standard motion sample databases to constrain the evaluation space, resulting in quantification biases when dealing with subtle differences between individual athlete styles and standard specifications, increasing the risk of misjudgment.

[0003] Therefore, how to provide a motion sports scoring method and device based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a method and apparatus for sports motion scoring based on big data analysis. Through a big data manifold distribution deviation quantification mechanism based on Wasserstein distance, optimal transmission planning is performed on the athlete's current motion manifold features extracted from the improved AGCN network model and the constructed standard deep manifold distribution parameters. Wasserstein distance features containing the minimum transmission cost are extracted, and global distribution deviation features are generated based on this mapping. The Wasserstein distance is embedded as a core metric in a nonlinear normalization process. By solving for the optimal transmission plan, the minimum path that transforms the athlete's motion probability distribution to the standard probability distribution is found in the manifold space until a precise distance value is calculated. This mechanism effectively eliminates structural distortion caused by geometric distance-based evaluation by establishing a precise mapping path from the output of the improved AGCN network model to the global distribution deviation features. This ensures that the generated scoring results dynamically reflect the global distribution deviation between the athlete's motion and the standard specifications in the high-dimensional manifold space, achieving the technical effect of improving the accuracy of motion style quantification while maintaining high scoring performance. This invention overcomes the limitations of traditional methods that rely only on local geometric features, ignore global topological constraints, and statistical distribution priors, providing an efficient solution for intelligent sports motion scoring.

[0005] According to an embodiment of the present invention, a method and apparatus for motion sports scoring based on big data analysis includes the following steps:

[0006] S1. Collect motion video and extract key points of bones, model joint orientation as points on a special orthogonal group in three dimensions, and construct human posture tensor manifold using rotation matrix;

[0007] S2. Calculate the Karcher mean of the motion trajectory on the human posture tensor manifold, use geodesics to perform nonlinear resampling of the motion sequence, and output the aligned motion trajectory features.

[0008] S3. Project the motion trajectory features onto the tangent space using the Riemann exponential mapping, calculate the weighted covariance matrix using the manifold geodesic distance kernel function, extract the matrix eigenvalues ​​as micro curvature features, and output symmetric positive definite manifold features after positive definite processing.

[0009] S4. Map the features of the symmetric positive definite manifold to the tangent space to construct the query and key vectors. Use the manifold inner product for weighted aggregation, and map back to the manifold via the Riemann index to generate reconstructed instances. Output the deep manifold distribution parameters.

[0010] S5. Obtain standard motion distribution parameters based on a large-scale motion sample library, calculate the Wasserstein distance between the athlete's motion and the probability distribution corresponding to the standard motion distribution parameters, and quantify the global distribution deviation characteristics.

[0011] S6. Calculate the residual vector based on the deep manifold distribution parameters and standard motion distribution parameters, and input it into the improved AGCN model for processing. Construct an adaptive topology based on the residual vector, and after multi-scale spatiotemporal convolution and manifold enhancement, aggregate and analyze it into joint-level physical angle error.

[0012] S7. Tensor concatenation is performed based on micro curvature features, global distribution deviation features, and joint-level physical angle errors to generate an initial high-dimensional feature tensor. The energy field feature vector is generated by weighted modulation using the manifold topological constraint force matrix and then processed by a self-attention mechanism to output a comprehensive score vector.

[0013] Optionally, S1 specifically includes:

[0014] S11. Acquire motion video frames and extract the three-dimensional coordinates of skeletal key points, and output a sequence of skeletal key points containing spatial location information;

[0015] S12. Receive the sequence of key points of the skeleton, construct a local orthogonal coordinate system based on the vectors of the parent node and child node of the joint, obtain the joint rotation matrix by solving the coordinate transformation matrix between the basis vectors, represent the joint orientation state as manifold data points on a three-dimensional special orthogonal group, and output the joint rotation matrix data.

[0016] S13. Receive joint rotation matrix data, extract the corresponding unit quaternion parameters using matrix trace operation, and perform joint tensor quantization of the rotation matrix and unit quaternion in the channel dimension to construct a multi-channel human pose tensor manifold.

[0017] Optionally, S2 specifically includes:

[0018] S21. Receive the multi-channel human posture tensor manifold, and use the Riemann logarithm and exponential mapping operators to perform iterative operations in the manifold space to calculate the Karcher mean data of the manifold trajectory.

[0019] S22. Using the Karcher mean data as the geometric benchmark, the geodesic distance from the data points to the benchmark at each time point is calculated using the Riemann logarithmic mapping operator and a distribution curve is constructed. The interpolation interval is determined based on the target sampling rate.

[0020] S23. Within the interpolation interval, the Riemann logarithm and exponential mapping operators are used to project and reconstruct between the tangent space and the manifold space to generate aligned motion trajectory features with equal geodesic arc lengths for adjacent data points.

[0021] Optionally, S3 specifically includes:

[0022] S31. Receive motion trajectory features, calculate the arithmetic mean vector of the motion trajectory features and set it as the reference base point of the manifold tangent space. For each manifold data point in the motion trajectory features, use the Riemann logarithmic mapping operator to calculate the tangent space projection vector of the manifold data point relative to the reference base point. Arrange all tangent space projection vectors in time order to obtain a tangent space vector sequence that characterizes the attitude change trend.

[0023] S32. Receive the tangent space vector sequence, construct a Gaussian radial basis kernel function based on geodesic distance to calculate the weight coefficients, and combine it with the centered deviation vector to perform a weighted outer product summation, outputting a weighted covariance matrix that considers the characteristics of the manifold geometry.

[0024] S33. Perform eigenvalue decomposition on the weighted covariance matrix, correct negative eigenvalues ​​through positive thresholding to extract micro curvature features, and reconstruct manifold features that eliminate numerical singularity and satisfy positive definiteness.

[0025] Optionally, S4 specifically includes:

[0026] S41. Receive the symmetric positive definite manifold features, calculate the Riemann mean of the symmetric positive definite manifold features as the reference point for tangent space mapping, use the reference point to project the symmetric positive definite manifold features to the tangent space through the Riemann logarithmic mapping operator to obtain the tangent space feature matrix, process the tangent space feature matrix through linear transformation to generate the corresponding query vector and key vector, and use the query vector and key vector to calculate the similarity coefficient in the inner product space induced by the Riemann metric.

[0027] S42. Receive the tangent space feature matrix and similarity coefficients, perform normalization on the similarity coefficients to obtain the manifold inner product weights, use the manifold inner product weights to perform weighted summation and aggregation operations on the tangent space feature matrix, use the weighted aggregation mean vector as the reconstructed tangent space vector, and calculate the magnitude of the reconstructed tangent space vector based on the Riemann metric tensor.

[0028] S43. Receive the reconstructed tangent space vector and the reference point of the tangent space mapping. Using the reference point of the tangent space mapping as the origin, use the Riemann exponential mapping operator to map the reconstructed tangent space vector from the tangent space back to the manifold space, generate reconstructed instances, calculate the mean and covariance matrix of the reconstructed instances, and combine the mean and covariance matrix as deep manifold distribution parameters describing the characteristics of the feature distribution.

[0029] Optionally, S5 specifically includes:

[0030] S51. Construct a big data motion sample library, collect multi-dimensional standard athlete motion data and map it to the manifold space to form a motion sample library containing massive standard posture manifold features.

[0031] S52. Obtain standard motion distribution parameters, calculate the probability distribution model of standard motions using statistical analysis methods based on the motion sample library, and extract the statistical feature vector of the probability distribution model as standard motion distribution parameters.

[0032] S53. Obtain the manifold features of the athlete's current movement and calculate the Wasserstein distance between the probability distribution of the athlete's movement and the probability distribution corresponding to the standard movement distribution parameters.

[0033] S54. Based on the Wasserstein distance values, generate global distribution bias features by mapping.

[0034] Optionally, the improved AGCN network model includes a data-dependent topology construction layer, a multi-scale spatiotemporal convolutional layer, a manifold feature enhancement layer, an adaptive adjacency aggregation layer, and a physical error parsing layer.

[0035] The algorithm receives the deep manifold distribution parameters and the standard motion distribution parameters, calculates the initial deviation matrix of the deep manifold distribution parameters relative to the standard motion distribution parameters using the translation operator under the Riemannian geometric framework, performs a transpose operation on the initial deviation matrix, and adds the transpose matrix to the initial deviation matrix to generate a symmetric deviation matrix. The symmetric deviation matrix is ​​then mapped to the corresponding tangent space vector using the matrix logarithm operator to obtain the residual vector of the deep manifold distribution parameters relative to the standard motion distribution parameters.

[0036] The data-dependent topology construction layer is used to receive the residual vector and initialize it into a physical adjacency matrix according to the human skeleton connection relationship. It calculates the data adjacency matrix based on the cosine similarity of node features, generates an implicit adjacency matrix, and weightedly fuses it with the physical adjacency matrix to obtain an adaptive topology adjacency matrix.

[0037] The multi-scale spatiotemporal convolutional layer is used to input the residual vector and the adaptive topological adjacency matrix into the temporal and spatial graph convolutional units, respectively extracting temporal dynamic features and aggregating spatial semantic features, and fusing them to generate a multi-scale spatiotemporal feature map.

[0038] The manifold feature enhancement layer receives multi-scale spatiotemporal feature maps and maps them to the Lie algebra tangent space using a manifold projection operator to obtain the tangent space feature manifold. Eigenvalues ​​and eigenvectors of the tangent space feature manifold are extracted using an eigenvalue decomposition algorithm, and a diagonal spectral matrix composed of eigenvalues ​​is constructed. This diagonal spectral matrix is ​​used as the geometric constraint feature of the manifold. The geometric constraint feature of the manifold is then concatenated with the multi-scale spatiotemporal feature map to obtain the enhanced manifold feature tensor.

[0039] The adaptive adjacency aggregation layer is used to expand the enhanced manifold feature tensor along the node dimension, calculate the global attention coefficient, perform weighted pooling on the enhanced manifold feature tensor, and obtain the aggregated feature tensor through residual connection and nonlinear mapping of multilayer perceptron.

[0040] The physical error analysis layer is used to map the aggregated feature tensor into a three-dimensional rotation vector through the axis-angle transformation layer, calculate the magnitude of the three-dimensional rotation vector as the joint rotation angle deviation, extract the unit vector direction of the three-dimensional rotation vector as the joint deviation orientation, and combine the joint rotation angle deviation and the joint deviation orientation to generate the joint-level physical angle error.

[0041] Optionally, S7 specifically includes:

[0042] S71. Obtain micro curvature features, global distribution deviation features and joint-level physical angle errors. Determine the splicing dimension based on each feature dimension and perform tensor concatenation to generate an initial high-dimensional feature tensor containing local geometric shape, overall probability distribution and physical motion constraint information.

[0043] S72. Receive the initial high-dimensional feature tensor, calculate the scalar potential energy value of each data point relative to the center of the manifold using the Gaussian radial basis kernel function, generate the manifold topological constraint force matrix based on the negative gradient direction of the potential energy, perform channel dimension weighted modulation on the initial high-dimensional feature tensor, and output the energy field feature vector carrying the geometric structure information of the manifold.

[0044] S73. The energy field feature vector is mapped to a query, key and value matrix through a self-attention weighting mechanism. The attention weight coefficients are calculated and the value matrix is ​​weighted and summed to output a score vector containing the comprehensive evaluation results.

[0045] According to an embodiment of the present invention, a motion sports scoring method and apparatus system based on big data analysis includes the following modules:

[0046] The Human Pose Tensor Manifold Construction Module is used to acquire motion videos and extract key points of the skeleton, model joint orientation as points on a special orthogonal group in 3D, and construct the human pose tensor manifold using rotation matrices.

[0047] The nonlinear resampling alignment module is used to calculate the Karcher mean of the motion trajectory on the human posture tensor manifold. It uses geodesics to perform nonlinear resampling on the motion sequence and outputs the aligned motion trajectory features.

[0048] The micro curvature feature extraction module is used to project the motion trajectory features onto the tangent space via Riemann exponential mapping, calculate the weighted covariance matrix using the manifold geodesic distance kernel function, extract the matrix eigenvalues ​​as micro curvature features, and output symmetric positive definite manifold features after positive definite processing.

[0049] The manifold feature reconstruction module is used to map the features of a symmetric positive definite manifold to the tangent space to construct query and key vectors, use the manifold inner product for weighted aggregation, and map back to the manifold via the Riemann index to generate reconstruction instances, outputting the deep manifold distribution parameters;

[0050] The global distribution deviation measurement module is used to calculate the Wasserstein distance between the probability distribution of an athlete's movement and that of a standard movement based on the deep manifold distribution parameters, and output the global distribution deviation feature.

[0051] The physical angle error analysis module is used to calculate the residual vector based on the deep manifold distribution parameters and the standard motion distribution parameters, and input it into the improved AGCN model for processing. Based on the residual vector, an adaptive topology is constructed, and after multi-scale spatiotemporal convolution and manifold enhancement, it is aggregated and analyzed into joint-level physical angle error.

[0052] The manifold energy field scoring module is used to cascade tensors based on micro curvature features, global distribution deviation features, and joint-level physical angle errors to generate an initial high-dimensional feature tensor. It then uses the manifold topological constraint force matrix to generate an energy field feature vector through weighted modulation, and finally processes it through a self-attention mechanism to output a comprehensive scoring vector.

[0053] The beneficial effects of this invention are:

[0054] (1) This invention employs an improved AGCN network model, which significantly enhances the feature extraction accuracy and topological structure representation ability of athlete movements in complex spatiotemporal scenarios. Utilizing the adaptive graph convolution mechanism of the improved AGCN network model, the optimal topological structure of the skeleton graph can be dynamically learned based on the input data, effectively capturing the implicit non-local spatiotemporal dependencies in the movement sequence. By introducing multi-view feature fusion and residual connection optimization, the model can accurately represent the dynamic change trend of movements in high-dimensional manifold space, thereby significantly improving the robustness and representation ability of movement features for complex postures and movement styles.

[0055] (2) This invention effectively solves the problem that traditional scoring methods only focus on local geometric errors while ignoring global statistical distribution characteristics by using a large-scale manifold distribution deviation quantification mechanism based on Wasserstein distance. By calculating the Wasserstein distance between the athlete's movement probability distribution and the standard deep manifold distribution parameters, the optimal transport planning is introduced into the scoring process, and the geodesic distance is used to measure the global distribution deviation in the manifold space. This method overcomes the limitations of traditional point-to-point matching or simple mean square error calculation when dealing with high-dimensional data distributions. It can accurately reflect the overall deviation of the athlete's movements from the standard specifications in the manifold structure, realize a deep-level quantitative evaluation of movement quality, and provide a scientific basis for intelligent guidance of sports training. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is an overall flowchart of a motion sports scoring method based on big data analysis proposed in this invention;

[0058] Figure 2 This is a flowchart illustrating the working principle of the improved AGCN model for a motion sports scoring method based on big data analysis proposed in this invention.

[0059] Figure 3 This is a schematic diagram of the structure of a motion sports scoring device based on big data analysis proposed in this invention. Detailed Implementation

[0060] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figure 1 and Figure 2 A motion sports scoring method based on big data analysis includes the following steps:

[0062] S1. Collect motion video and extract key points of bones, model joint orientation as points on a special orthogonal group in three dimensions, and construct human posture tensor manifold using rotation matrix;

[0063] S2. Calculate the Karcher mean of the motion trajectory on the human posture tensor manifold, use geodesics to perform nonlinear resampling of the motion sequence, and output the aligned motion trajectory features.

[0064] S3. Project the motion trajectory features onto the tangent space using the Riemann exponential mapping, calculate the weighted covariance matrix using the manifold geodesic distance kernel function, extract the matrix eigenvalues ​​as micro curvature features, and output symmetric positive definite manifold features after positive definite processing.

[0065] S4. Map the features of the symmetric positive definite manifold to the tangent space to construct the query and key vectors. Use the manifold inner product for weighted aggregation, and map back to the manifold via the Riemann index to generate reconstructed instances. Output the deep manifold distribution parameters.

[0066] S5. Obtain standard motion distribution parameters based on a large-scale motion sample library, calculate the Wasserstein distance between the athlete's motion and the probability distribution corresponding to the standard motion distribution parameters, and quantify the global distribution deviation characteristics.

[0067] S6. Calculate the residual vector based on the deep manifold distribution parameters and standard motion distribution parameters, and input it into the improved AGCN model for processing. Construct an adaptive topology based on the residual vector, and after multi-scale spatiotemporal convolution and manifold enhancement, aggregate and analyze it into joint-level physical angle error.

[0068] S7. Tensor concatenation is performed based on micro curvature features, global distribution deviation features, and joint-level physical angle errors to generate an initial high-dimensional feature tensor. The energy field feature vector is generated by weighted modulation using the manifold topological constraint force matrix and then processed by a self-attention mechanism to output a comprehensive score vector.

[0069] In this embodiment, S1 specifically includes:

[0070] S11. Read the input motion video file, decode the video stream frame by frame to obtain the original image data, extract the two-dimensional pixel coordinates of the human skeleton nodes in each frame, combine the camera calibration parameters to back-project the two-dimensional pixel coordinates to the three-dimensional world coordinate system, calculate the three-dimensional spatial coordinate values ​​of a total of 17 skeletal key points such as the root node, left knee joint, and right elbow joint, and combine the three-dimensional coordinate data arranged in the order of joint number into a sequence of skeletal key points containing spatial position information.

[0071] S12. Receive the sequence of skeletal key points. For each joint node, calculate the vector pointing to the target joint as the child node vector based on the root node coordinates as the origin, and the vector pointing to the root node as the parent node vector. Construct the basis vector of the local orthogonal coordinate system using the three component values ​​of the parent node vector. Obtain a 3x3 joint rotation matrix by solving the linear transformation relationship between the parent node vector and the child node vector. Map the values ​​in the rotation matrix as data points to the three-dimensional special orthogonal group manifold space, and output the joint rotation matrix data representing the joint orientation state.

[0072] S13. Receive joint rotation matrix data, perform trace operation on each 3x3 rotation matrix, that is, extract the three elements on the diagonal of the matrix and add the three to obtain the trace of the matrix, use the trace value to calculate the four component parameters of the corresponding unit quaternion, and concatenate the calculated unit quaternion parameters with the original joint rotation matrix data in the channel dimension to construct a multi-channel data structure containing 9 rotation matrix values ​​and 4 quaternion values, and generate a multi-channel human pose tensor manifold.

[0073] In this embodiment, S2 specifically includes:

[0074] S21. Read the multi-channel human posture tensor manifold data, select the first frame manifold data point of the action sequence as the initial manifold trajectory reference center point, and use the Riemann logarithmic mapping operator to project the manifold data points of each time step in the sequence onto the tangent plane coordinate system of the reference center point. In the tangent plane coordinate system, sum the coordinate values ​​of all projected vectors and divide them by the total number of data points to obtain the tangent space average vector. Use the Riemann exponential mapping operator to map the tangent space average vector back to the manifold space along the geodesic direction to obtain the updated reference center point. Calculate the Frobenius norm value between the updated reference center point and the reference center point at the previous time step as the change value. When the change value is less than 0.001, it is determined that the preset convergence condition is met and the update is stopped. If it is not met, the updated reference center point is set as the new reference and the above projection and mapping process is repeated until the condition is met and the Karcher mean data is output.

[0075] S22. Receive Karcher mean data and set it as the geometric reference point of the manifold trajectory. For each manifold data point in the action trajectory, use the Riemann logarithmic mapping operator to calculate the tangent space vector coordinates of the data point relative to the geometric reference point. Use the Pythagorean theorem to calculate the sum of squares of each component of the tangent space vector coordinates and take the square root. Use the calculation result as the geodesic distance from the manifold data point to the geometric reference point. Arrange and combine the calculated geodesic distances according to the chronological order of the actions to construct the distribution curve of the geodesic distance values ​​changing with time.

[0076] S23. Receive the geodesic distance distribution curve, uniformly divide the total length interval of the geodesic distance distribution curve according to the target sampling rate of 100 Hz, determine 64 target interpolation positions, find the interval between two adjacent original manifold data points on the original curve for each target interpolation position, calculate the tangent space direction vector between the manifold data points at the interval endpoints using the Riemann logarithmic mapping, scale the coordinate values ​​of the tangent space direction vector proportionally according to the distance the target interpolation position moves on the curve, map the scaled tangent space vector back to the manifold space using the Riemann exponential mapping to obtain the manifold coordinates of the interpolation point, and stitch all the generated interpolation points together in chronological order to generate an aligned motion trajectory feature where the geodesic arc length values ​​between adjacent data points are equal and consistent.

[0077] The Riemann logarithmic mapping operator involves reading the input human posture tensor manifold data, locating the corresponding rotation matrix channel from the tensor data, and extracting the three-dimensional numerical matrix corresponding to a specific joint from the data block according to a preset joint index. For the reference point, the manifold center point data generated in the previous iteration is read, and the 3x3 numerical matrix contained in the center point data is directly extracted and loaded into the current processing memory as the reference point rotation matrix. For the data point to be mapped, the posture frame corresponding to the current time is located from the action sequence based on the current processing timestamp, and the rotation parameters of the current joint are extracted from the posture frame to construct a 3x3 matrix containing 9 floating-point numbers. The matrix operation is used to verify whether the matrix satisfies the orthogonality constraint condition. If it does, it is confirmed as the rotation matrix of the data point to be mapped. By calculating the product of the transpose of the reference point rotation matrix and the data point rotation matrix, the relative rotation matrix describing the relative transformation relationship between the two postures is obtained. The diagonal elements of the relative rotation matrix are extracted, and the trace of the matrix is ​​obtained by adding these three values. The rotation angle value corresponding to the relative rotation is calculated using inverse trigonometric functions based on the trace value of the matrix. By subtracting the values ​​of the off-diagonal elements in the relative rotation matrix pairwise, the vector components representing the rotation axis direction are calculated. A scaling factor is then constructed based on the calculated rotation angle values. This scaling factor is used to numerically scale the vector components along the rotation axis direction, generating a three-dimensional coordinate vector located in the tangent plane of the reference point.

[0078] In this embodiment, S3 specifically includes:

[0079] S31. Read the aligned motion trajectory features from the input. Sum the values ​​of all manifold data points in the feature sequence and divide by the total number of frames to calculate the arithmetic mean vector of the motion trajectory features. Set the arithmetic mean vector as the reference base point of the manifold tangent space. For each frame of manifold data points in the motion trajectory features, use the Riemann logarithmic mapping operator to calculate the tangent space projection vector coordinates of the manifold data point relative to the reference base point. Linearly arrange all the calculated tangent space projection vectors according to the original time sequence of the action occurrence to generate a set of tangent space vector sequences containing time dimension information.

[0080] S32. Receive the sequence of tangent space vectors. Calculate the sum of squares of the coordinate components of each tangent space projection vector in the sequence using the Pythagorean theorem and take the square root. Use the calculated value as the geodesic distance parameter. The specific steps for constructing a Gaussian radial basis function kernel model with the geodesic distance parameter as the independent variable include obtaining the calculated geodesic distance parameter, setting the bandwidth coefficient of the kernel function (which is twice the median of all geodesic distance parameters), calculating the square of the geodesic distance parameter using the geodesic distance parameter as the independent variable, performing normalized division on the square of the geodesic distance parameter using the square of the bandwidth coefficient to obtain the normalized distance ratio, substituting the normalized distance ratio into the negative exponential function (i.e., using the negative normalized distance ratio as the exponent of the natural constant e), and calculating the real number result between 0 and 1, which is the weight coefficient of each tangent space projection vector, thus completing the construction of the Gaussian radial basis function kernel model. Substitute the geodesic distance parameter into the Gaussian radial basis function kernel model to calculate the weight coefficient value corresponding to each tangent space projection vector. The arithmetic mean of all vectors in the tangent space vector sequence is calculated as the tangent space center vector. The tangent space center vector is then subtracted from each tangent space projection vector in the sequence to obtain a centered bias vector. The calculated weight coefficients are used to perform a weighted summation operation on the outer product matrix of each bias vector, outputting a weighted covariance matrix that considers the geometric characteristics of the manifold.

[0081] S33. Receive the weighted covariance matrix and perform eigenvalue decomposition on it using the Jacobi iterative algorithm to extract the eigenvalues ​​and corresponding eigenvectors. Construct a diagonal matrix composed of all eigenvalues ​​arranged in ascending order. Detect each negative eigenvalue in the diagonal matrix; if a negative eigenvalue is detected, correct it by replacing it with a positive threshold of 0.001. Extract the diagonal elements of the corrected eigenvalue diagonal matrix as micro-curvature features representing the local geometry. Reconstruct the covariance matrix using matrix multiplication with the extracted eigenvector matrix, outputting a symmetric positive definite manifold feature that eliminates numerical singularities and satisfies the positive definiteness condition.

[0082] The manifold feature extraction process proposed in this step is similar to the traditional Riemannian geometric analysis method in that it is based on manifold learning theory. That is, it linearizes the nonlinear data on the manifold by mapping it to the tangent space, uses geodesic distance to measure the geometric difference between data points, and uses Riemannian logarithmic mapping and exponential mapping to realize the transformation between the manifold and the tangent space.

[0083] The difference lies in that this invention breaks away from the limitations of traditional methods that only perform simple alignment or clustering analysis on manifold data. Instead of simply calculating the average trajectory in traditional models, this invention adds a weighted covariance calculation step. It dynamically calculates sample weights using a Gaussian radial basis function with geodesic distance as the independent variable, rather than applying equal weights to all samples. In the eigenvalue decomposition layer, micro-curvature features are extracted using the eigenvalue diagonal matrix, and a negative eigenvalue truncation correction step is introduced to address numerical calculation errors, rather than directly outputting the original covariance matrix. Finally, in the symmetric positive definite manifold feature output step, a matrix satisfying positive definiteness constraints is reconstructed based on the corrected eigenvalues ​​and eigenvectors.

[0084] The beneficial effect of the improvements lies in the fact that by introducing a geodesic distance-weighted kernel function and a positive definite eigenvalue correction step, this invention can transform the local geometric deformation of the motion trajectory on the manifold into high-dimensional statistical distribution features. This breaks through the limitations of traditional methods in capturing the microscopic bending structure and dynamic topological changes of the manifold, and realizes a deep mapping from the physical posture space to the Riemannian manifold space. This design significantly enhances the model's ability to perceive subtle posture differences and nonlinear dynamic features in human motion, and can more accurately represent the intrinsic geometric properties of complex motion trajectories. The positive definite processing based on eigenvalue correction effectively solves the singularity problem in numerical calculation and enhances the robustness and mathematical stability of the feature vector in subsequent similarity measurement and classification tasks.

[0085] In this embodiment, S4 specifically includes:

[0086] S41. Obtain multiple symmetric positive definite manifold feature data from the input, and select any feature data point as the initial center point for iteration, setting the preset convergence threshold to 0.01%. In each iteration, for each feature data point, calculate the tangent space vector of that data point relative to the current center point using the Riemann logarithmic mapping operator, and arithmetically average all calculated tangent space vectors to obtain the center point update gradient vector. Map the center point update gradient vector back to the manifold space along the geodesic direction using the Riemann exponential mapping operator to obtain the temporary update point for this iteration. Calculate the geodesic distance between the temporary update point and the center point at the previous time step as the change, and determine if the change value is less than 0.01%. If it is less than the threshold, the algorithm is considered converged, and the current center point is determined as the final geometric center; otherwise, the temporary update point is set as the new center point, and the iteration is repeated. Set the final geometric center as the reference point for the tangent space mapping, and project each symmetric positive definite manifold feature in the manifold space onto the tangent plane of the reference point using the Riemann logarithmic mapping operator to generate the tangent space feature matrix. Construct a trainable linear transformation weight matrix to generate query vectors and key vectors respectively, and calculate the similarity value using the Riemann metric tensor at the reference point.

[0087] S42. Receive the tangent space feature matrix and similarity values. Perform Softmax normalization on the similarity values, mapping the value range to between 0 and 1, to obtain the manifold inner product weights. Use the calculated manifold inner product weights to perform a weighted summation and aggregation operation on all column vectors in the tangent space feature matrix, and calculate the weighted aggregation mean vector as the reconstructed tangent space vector. Calculate the autocovariance of the reconstructed tangent space vector based on the Riemann metric tensor at the reference point and take its square root to obtain the magnitude of the reconstructed tangent space vector.

[0088] S43. Receive the reconstructed tangent space vector and the reference point of the tangent space mapping. Using the reference point of the tangent space mapping as the origin of the exponential mapping, use the Riemann exponential mapping operator to map the reconstructed tangent space vector back from the tangent space to the manifold space along the geodesic direction, generating a reconstruction instance. Read all manifold data points contained in the reconstruction instance, calculate the arithmetic mean of the data points as the distribution mean, and calculate the discrete covariance matrix of the data points relative to the distribution mean. Combine the mean and covariance matrix to output the deep manifold distribution parameters describing the distribution characteristics of features in the manifold space.

[0089] The deep manifold distribution parameter extraction process proposed in this step is similar to the traditional Riemannian attention mechanism in that it is based on the Riemannian geometric manifold learning theory. That is, it linearizes the nonlinear data on the manifold by projecting it onto the tangent plane, uses the Riemannian logarithmic mapping and exponential mapping to achieve spatial transformation, and uses inner product operation to measure the similarity relationship between features.

[0090] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on Euclidean inner product for attention calculation or ignore manifold metric structures. Building upon the traditional model's direct use of dot products to calculate similarity, this invention adds a Riemannian geometric center localization step. This step calculates the geometric center of the manifold data as the reference point for tangent space mapping, rather than arbitrarily selecting a reference point. In the similarity calculation step, the Riemannian metric tensor at the reference point replaces the Euclidean metric to perform induced inner product operations on the query vector and key vector, instead of a simple vector dot product. Finally, in the distribution parameter output step, based on the magnitude calculation of the reconstructed tangent space vector and the statistical analysis of reconstructed manifold instances, the mean and covariance matrices are jointly output as deep manifold distribution parameters, rather than a single feature vector.

[0091] The beneficial effects of this improvement are that, through Riemannian geometric center localization and Riemannian metric tensor-induced inner product operation, the curvature features and local geometric structure on the manifold can be integrated into the attention weight calculation. This breaks the limitation of traditional methods that ignore metric variation in curved space, which leads to feature inaccuracies, and achieves an accurate mapping from Euclidean similarity to Riemannian geometric distance. This design significantly enhances the model's ability to perceive the inherent geometric topology of complex manifold data and can more accurately capture the nonlinear dependencies of features in tangent space. The reconstruction instance based on exponential mapping and parameterization effectively improve the completeness of feature representation and enhance the robustness and discrimination accuracy of the system in complex human pose estimation and action recognition tasks.

[0092] In this embodiment, S5 specifically includes:

[0093] S51. Construct a big data motion sample library. Use high-precision motion capture equipment to collect original coordinate data of joints of multi-dimensional standard athletes in various sports scenarios at a frequency of 60 frames per second. Use the Riemann manifold mapping algorithm to transform the original coordinate data into manifold space, extract posture manifold feature vectors and standardize and store them to form a benchmark motion sample library containing 100,000 frames of standard posture manifold features.

[0094] S52. Obtain standard motion distribution parameters. Read the standard posture manifold features from 100,000 frames in the motion sample library, and set the bandwidth parameter of the Gaussian kernel function to 0.5. Using the kernel density estimation algorithm, calculate the Gaussian kernel function value between each data point in the manifold space and all other data points. Summate all kernel function values ​​and divide by 100,000 to obtain the probability density value at each data point, thus constructing a probability distribution model for standard motions. Extract the first-order mean vector and the second-order covariance matrix from the probability distribution model, and combine them to generate standard motion distribution parameters that characterize the regularity of standard motions.

[0095] S53. Obtain the manifold features of the athlete's current movement. Collect the athlete's posture data during training in real time and map it to the manifold space. Select continuous manifold feature data within a 2-second time window. Calculate the probability density distribution model of the current movement in the manifold space using the same kernel density estimation algorithm as the standard movement. Define the number of data points in the current movement probability distribution as the number of supply points, and the number of data points in the standard movement probability distribution as the number of demand points. For each supply point in the current movement distribution, calculate the manifold geodesic distance between the supply point and each demand point in the standard movement distribution, and set the square of the geodesic distance as the basic transmission cost between the two points. Arrange all pairwise corresponding basic transmission cost values ​​in a matrix according to the order of rows corresponding to supply points and columns corresponding to demand points to construct an initial cost matrix. The Sinkhorn algorithm is used to perform logarithmic-domain iterative operations on the initial cost matrix. An entropy regularization coefficient of 0.1 is introduced to optimize the solution speed. Under the constraints that the total output of supply points equals the total probability of the current action distribution and the total input of demand points equals the total probability of the standard action distribution, the joint probability transition matrix is ​​calculated to minimize the total expected transmission cost, thus obtaining the constructed joint probability cost matrix. The constructed joint probability cost matrix is ​​set as the optimization variable in linear programming. Each element in the matrix represents the probability of transmitting the probability quality of supply points in the current action distribution to demand points in the standard action distribution. An objective function is constructed by multiplying each element in the joint probability cost matrix with the corresponding element of the initial cost matrix, summing all the product results, and setting the sum value as the optimization objective. Constraints are set as follows: the first constraint is that the sum of the elements in each row of the joint probability cost matrix equals the probability value of the corresponding supply point in the current action distribution; the second constraint is that the sum of the elements in each column of the joint probability cost matrix equals the probability value of the corresponding demand point in the standard action distribution; and the third constraint is that all element values ​​in the joint probability cost matrix must be greater than or equal to zero. The objective function and constraints are input into the simplex method solver for iterative calculation. The optimal joint probability transition matrix that satisfies all constraints and minimizes the objective function is calculated. The value obtained by substituting the optimal joint probability transition matrix into the objective function is the minimum expected transmission cost. The minimum expected transmission cost is used as the Wasserstein distance between the two probability distributions.

[0096] S54. Based on the Wasserstein distance value, substitute the calculated distance value into the preset Sigmoid nonlinear mapping function. The scaling factor of the function is set to 10, and the offset is set to -5. Input the distance value to calculate the function value, and obtain the distribution deviation index with a value between 0 and 1. Use the distribution deviation index as a weighting factor to perform weighted bias processing on each feature component of the standard movement distribution parameters. Map the processed result to a 128-dimensional high-dimensional feature space to generate a global distribution deviation feature that characterizes the degree of difference between the athlete's current movement and the standard movement in the overall distribution pattern.

[0097] In this embodiment, the improved AGCN network model includes a data-dependent topology construction layer, a multi-scale spatiotemporal convolutional layer, a manifold feature enhancement layer, an adaptive adjacency aggregation layer, and a physical error parsing layer.

[0098] The algorithm receives the deep manifold distribution parameters and the standard motion distribution parameters. The standard motion distribution parameters are used as the reference point on the Riemannian manifold, and the deep manifold distribution parameters are used as the target point to be aligned. Both are confirmed to be symmetric positive definite matrices of the same dimension. The standard motion distribution parameters used as the reference point are inverted to obtain the corresponding inverse matrix. The calculated inverse matrix is ​​used as a left multiplication factor and multiplied by the deep manifold distribution parameters used as the target point to obtain a product matrix. This product matrix is ​​defined as the initial deviation matrix of the deep manifold distribution parameters relative to the standard motion distribution parameters. The initial deviation matrix is ​​transposed, and the transposed matrix is ​​summed with the initial deviation matrix to generate a symmetric deviation matrix. Eigenvalue decomposition is performed based on the symmetric deviation matrix to obtain an orthogonal eigenvector matrix and a diagonal matrix composed of eigenvalues. The diagonal elements of the diagonal matrix are replaced with their natural logarithmic values ​​to construct a logarithmic eigenvalue matrix. The orthogonal eigenvector matrix is ​​used to map the logarithmic eigenvalue matrix back to the original space to obtain the residual vector of the deep manifold distribution parameters relative to the standard motion distribution parameters.

[0099] The data-dependent topology construction layer receives the residual vector and initializes it into a physical adjacency matrix according to the human skeleton connection relationship. Based on the node features, it calculates the cosine similarity values ​​between each pair to construct the original similarity matrix. It performs a Top-k filtering operation on each row of the original similarity matrix, retains strong connections and sparsifies the matrix. It uses learnable normalization coefficients to perform a weighted transformation on the filtered matrix to generate an implicit adjacency matrix. The physical adjacency matrix and the implicit adjacency matrix are weighted and fused with a preset weight of 0.5 to obtain an adaptive topological adjacency matrix.

[0100] The multi-scale spatiotemporal convolutional layer inputs the residual vector and the adaptive topological adjacency matrix into the temporal and spatial graph convolutional units. It uses a temporal convolutional kernel with a kernel size of 3 to extract temporal dynamic features and uses a spatial convolutional kernel with Chebyshev polynomial expansion to aggregate spatial semantic features. The temporal dynamic features and spatial semantic features are fused to generate a multi-scale spatiotemporal feature map.

[0101] The manifold feature enhancement layer receives a multi-scale spatiotemporal feature map and maps it to the Lie algebra tangent space using a manifold projection operator to obtain a tangent space feature manifold. The number of features to be extracted is set to 64. The eigenvalue decomposition algorithm is used to extract 64 eigenvalues ​​and their corresponding eigenvectors of the tangent space feature manifold, and a diagonal spectral matrix composed of the 64 eigenvalues ​​is constructed. The diagonal spectral matrix is ​​used as the geometric constraint feature of the manifold. The geometric constraint feature of the manifold is concatenated with the multi-scale spatiotemporal feature map in the channel dimension to obtain the enhanced manifold feature tensor.

[0102] The adaptive adjacency aggregation layer expands the enhanced manifold feature tensor along the node dimension, calculates the global attention coefficient, performs weighted pooling on the enhanced manifold feature tensor, and obtains the aggregated feature tensor through residual connection and nonlinear mapping with a multilayer perceptron containing two hidden layers.

[0103] The physical error analysis layer maps the aggregated feature tensor into a three-dimensional rotation vector through the axis-angle transformation layer. It uses the Euclidean norm formula to calculate the magnitude of the three-dimensional rotation vector as the joint rotation angle deviation. It extracts the unit vector direction of the three-dimensional rotation vector as the joint deviation orientation. The joint rotation angle deviation and the joint deviation orientation are combined to generate the joint-level physical angle error.

[0104] The improved AGCN network proposed in this invention is similar to the traditional AGCN network in that it retains the core architecture of adaptive graph convolution, that is, it constructs a graph topology structure through a data-dependent adaptive mechanism, and uses graph convolution operators to aggregate the spatiotemporal features of nodes on the constructed topology to capture the dynamic correlation of human motion. Both networks use multilayer perceptrons for nonlinear feature mapping.

[0105] The difference lies in that this invention breaks away from the limitations of traditional AGCN networks, which rely solely on Euclidean coordinate differences or fixed physical connections to construct features. It introduces Riemannian geometric operators and a dynamic topology generation step. Building upon the traditional model's simple extraction of statistical features, this invention adds Riemannian geometric operators, utilizing translation operators and matrix logarithmic mapping to calculate residual vectors between distributions, rather than directly aligning coordinates. In the data-dependent topology construction layer, an implicit adjacency matrix is ​​generated through the cosine similarity of node features and then weighted and fused with the physical adjacency matrix, instead of using a static graph structure. Finally, in the manifold feature enhancement layer, eigenvalue decomposition is used to construct manifold geometric constraint features and concatenate them with spatiotemporal features, rather than directly outputting the original feature map.

[0106] The beneficial effects of the improvements are that by introducing Riemannian geometric operators and adaptive topology steps, the improved AGCN network can accurately map the distribution deviation of motion data in curved manifold space to flat tangent space, breaking the limitation of traditional methods that rely solely on Euclidean distance to measure posture differences, and realizing feature decoupling from high-dimensional nonlinear geometric space to linear analyzable space. This design significantly enhances the model's ability to perceive the intrinsic geometric structure and physical errors of motion, and can more accurately analyze the motion coupling relationship between joints. The steps based on manifold spectrum constraints and attention aggregation effectively improve the completeness of feature representation and enhance the robustness and error calculation accuracy of the system in complex motion posture analysis scenarios.

[0107] In this embodiment, S7 specifically includes:

[0108] S71. Obtain micro-curvature features, global distribution deviation features, and joint-level physical angle errors. Analyze the first feature dimension of the micro-curvature features, the second feature dimension of the global distribution deviation features, and the third feature dimension of the joint-level physical angle errors respectively. Determine the target splicing dimension based on the sum of the first feature dimension, the second feature dimension, and the third feature dimension. Through tensor cascading operations, fuse the micro-curvature features, global distribution deviation features, and joint-level physical angle errors along the target splicing dimension to generate an initial high-dimensional feature tensor containing local geometric shape, overall probability distribution, and physical motion constraint information.

[0109] S72. Receive the initial high-dimensional feature tensor, reshape the initial high-dimensional feature tensor into a two-dimensional data matrix, calculate the Euclidean distance between each row vector in the data matrix and the preset manifold center vector, use the Gaussian radial basis kernel function to perform a nonlinear transformation on the Euclidean distance to calculate the scalar potential energy value of each data point on the manifold surface, calculate the corresponding manifold topological constraint force matrix based on the negative gradient direction of the scalar potential energy value on the manifold surface, and use the manifold topological constraint force matrix to perform a channel dimension weighted modulation operation on the initial high-dimensional feature tensor to generate an energy field feature vector carrying manifold geometric structure information;

[0110] S73. Receive the energy field feature vector, construct a self-attention mechanism weighted network, and use a linear transformation layer to map the energy field feature vector into a query matrix, a key matrix, and a value matrix, respectively. Calculate the product of the query matrix and the transpose of the key matrix and divide by a scaling factor to obtain the original attention score matrix. Perform Softmax normalization operation on each row of the original attention score matrix to generate attention weight coefficients. Use the attention weight coefficients to perform a weighted summation on the value matrix and output a score vector containing the comprehensive evaluation result.

[0111] refer to Figure 3A motion sports scoring device based on big data analysis specifically includes the following modules:

[0112] The Human Pose Tensor Manifold Construction Module is used to acquire motion videos and extract key points of the skeleton, model joint orientation as points on a special orthogonal group in 3D, and construct the human pose tensor manifold using rotation matrices.

[0113] The nonlinear resampling alignment module is used to calculate the Karcher mean of the motion trajectory on the human posture tensor manifold. It uses geodesics to perform nonlinear resampling on the motion sequence and outputs the aligned motion trajectory features.

[0114] The micro curvature feature extraction module is used to project the motion trajectory features onto the tangent space via Riemann exponential mapping, calculate the weighted covariance matrix using the manifold geodesic distance kernel function, extract the matrix eigenvalues ​​as micro curvature features, and output symmetric positive definite manifold features after positive definite processing.

[0115] The manifold feature reconstruction module is used to map the features of a symmetric positive definite manifold to the tangent space to construct query and key vectors, use the manifold inner product for weighted aggregation, and map back to the manifold via the Riemann index to generate reconstruction instances, outputting the deep manifold distribution parameters;

[0116] The global distribution deviation measurement module is used to calculate the Wasserstein distance between the probability distribution of an athlete's movement and that of a standard movement based on the deep manifold distribution parameters, and output the global distribution deviation feature.

[0117] The physical angle error analysis module is used to calculate the residual vector based on the deep manifold distribution parameters and the standard motion distribution parameters, and input it into the improved AGCN model for processing. Based on the residual vector, an adaptive topology is constructed, and after multi-scale spatiotemporal convolution and manifold enhancement, it is aggregated and analyzed into joint-level physical angle error.

[0118] The manifold energy field scoring module is used to cascade tensors based on micro curvature features, global distribution deviation features, and joint-level physical angle errors to generate an initial high-dimensional feature tensor. It then uses the manifold topological constraint force matrix to generate an energy field feature vector through weighted modulation, and finally processes it through a self-attention mechanism to output a comprehensive scoring vector.

[0119] To verify the feasibility of this invention in intelligent human-computer interaction and sports health assessment, the method of this invention was applied to the motion capture and posture assessment system of a leading motion-sensing game platform (hereinafter referred to as "Platform G"). In traditional motion-sensing game interaction systems, algorithms based on Euclidean coordinate differences or simple manifold statistics methods are typically used for motion matching and scoring. These methods not only have limited computational accuracy but also cannot effectively handle the nonlinear geometric structure and dynamic topological relationships of human movements in curved manifold space, easily leading to problems such as inaccurate scoring, feedback delay, and poor immersion. To solve the above problems, Platform G decided to adopt the motion sports scoring method and device based on big data analysis proposed in this invention.

[0120] During implementation, platform G first utilizes its high-precision visual sensors to acquire multi-channel human posture manifold data of players in real time, including skeletal node coordinates, joint rotation quaternions, and velocity acceleration information. After data cleaning, timestamp alignment, and noise filtering, a standardized motion trajectory feature flow is formed. Simultaneously, platform G integrates a pre-set standard motion template library to accurately express the deep manifold distribution parameters and physical constraint characteristics of standard skill movements in different game levels.

[0121] Platform G introduces Riemannian geometric operators and utilizes techniques such as translation operators, matrix logarithmic mapping, and eigenvalue decomposition to map players' real-time motion data to the tangent space of the SPD manifold. It calculates the residual vector relative to the standard motion, accurately characterizing the initial deviation of the motion in the manifold's geometric space. Through an improved data-dependent topology construction layer and a multi-scale spatiotemporal convolutional layer, it adaptively constructs an adaptive topological adjacency matrix that conforms to the current motion state by weighted fusion of the physical adjacency matrix and the implicit adjacency matrix based on cosine similarity. Furthermore, it aggregates spatiotemporal semantic features using a spatial convolution kernel expanded by Chebyshev polynomials. Subsequently, by combining a manifold feature enhancement layer and a physical error analysis layer, it maps multi-scale spatiotemporal features into joint-level physical angle errors, achieving deep analysis and precise quantification of the microscopic morphology and overall geometric structure of the motion.

[0122] During implementation, the technical team at Platform G discovered that, compared to traditional scoring methods based on coordinate difference and Euclidean distance, the method of this invention significantly improves the accuracy and robustness of motion assessment. Traditional methods cannot precisely characterize the complex intrinsic coupling relationships and non-rigid body deformations between human joints, while the method of this invention, through Riemannian geometric operators and adaptive topology mechanisms, effectively achieves deep insight into complex nonlinear motion characteristics and precise analysis of physical-level errors.

[0123] To further verify the actual performance of the method of the present invention, platform G conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1:

[0124] Table 1. Performance Comparison of Motion Evaluation Methods for Platform G-Sensing Games

[0125] index Traditional methods Method of the present invention Increase Motion assessment accuracy rate (%) 82.5 94.8 +14.9% Joint angle error (degrees) 5.2 1.8 -65.4% Action rating correlation 0.76 0.93 +22.4% Manifold feature matching time (milliseconds) 35 12 -65.7% Complex pose recognition rate 78.0 92.5 +18.6% Real-time interactive response latency (milliseconds) 150 48 -68.0% User immersion rating 7.2 9.1 +26.4% Abnormal movement missed rate (%) 8.5 1.2 -85.9% System resource utilization (CPU%) 45 28 -37.8%

[0126] As shown in Table 1, the performance of the motion assessment system for motion-sensing games was comprehensively improved after applying the method of this invention. The motion assessment accuracy increased from 82.5% of the traditional method to 94.8%, and the joint angle error decreased from 5.2 degrees to 1.8 degrees, significantly improving the accuracy of the assessment and enhancing the competitive fairness of the game. The correlation of motion scores increased from 0.76 to 0.93, and the time consumption for manifold feature matching and the latency of real-time interaction response decreased by more than 60%, effectively solving the problems of lag and delayed feedback in game interaction. Furthermore, the rate of missed detection of motion anomalies decreased significantly from 8.5% to 1.2%, and the user immersion score was significantly improved, effectively enhancing the player's gaming experience. The reduction in system resource consumption also means that this method can run smoothly on a wider variety of terminal devices.

[0127] Through the method of this invention, platform G successfully achieved high-precision motion assessment for motion-sensing games, effectively improving the realism of game interaction and player participation, ensuring user stickiness and activity in game operation, significantly improving the intelligence level of the motion-sensing interaction system, significantly reducing the consumption of computing resources by the algorithm, enhancing the robustness of the system in complex scenarios, and providing strong technical support for the intelligent development of the next generation of motion-sensing games.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method and apparatus for scoring motion sports based on big data analysis, characterized in that, Includes the following steps: S1. Collect motion video and extract key points of bones, model joint orientation as points on a special orthogonal group in three dimensions, and construct human posture tensor manifold using rotation matrix; S2. Calculate the Karcher mean of the motion trajectory on the human posture tensor manifold, use geodesics to perform nonlinear resampling of the motion sequence, and output the aligned motion trajectory features. S3. Project the motion trajectory features onto the tangent space using the Riemann exponential mapping, calculate the weighted covariance matrix using the manifold geodesic distance kernel function, extract the matrix eigenvalues ​​as micro curvature features, and output symmetric positive definite manifold features after positive definite processing. S4. Map the features of the symmetric positive definite manifold to the tangent space to construct the query and key vectors. Use the manifold inner product for weighted aggregation, and map back to the manifold via the Riemann index to generate reconstructed instances. Output the deep manifold distribution parameters. S5. Obtain standard motion distribution parameters based on a large-scale motion sample library, calculate the Wasserstein distance between the athlete's motion and the probability distribution corresponding to the standard motion distribution parameters, and quantify the global distribution deviation characteristics. S6. Calculate the residual vector based on the deep manifold distribution parameters and standard motion distribution parameters, and input it into the improved AGCN model for processing. Construct an adaptive topology based on the residual vector, and after multi-scale spatiotemporal convolution and manifold enhancement, aggregate and analyze it into joint-level physical angle error. S7. Tensor concatenation is performed based on micro curvature features, global distribution deviation features, and joint-level physical angle errors to generate an initial high-dimensional feature tensor. The energy field feature vector is generated by weighted modulation using the manifold topological constraint force matrix and then processed by a self-attention mechanism to output a comprehensive score vector.

2. The motion sports scoring method and device based on big data analysis according to claim 1, characterized in that, S1 includes: S11. Acquire motion video frames and extract the three-dimensional coordinates of skeletal key points, and output a sequence of skeletal key points containing spatial location information; S12. Receive the sequence of key points of the skeleton, construct a local orthogonal coordinate system based on the vectors of the parent node and child node of the joint, obtain the joint rotation matrix by solving the coordinate transformation matrix between the basis vectors, represent the joint orientation state as manifold data points on a three-dimensional special orthogonal group, and output the joint rotation matrix data. S13. Receive joint rotation matrix data, extract the corresponding unit quaternion parameters using matrix trace operation, and perform joint tensor quantization of the rotation matrix and unit quaternion in the channel dimension to construct a multi-channel human pose tensor manifold.

3. The motion sports scoring method and device based on big data analysis according to claim 1, characterized in that, S2 specifically includes: S21. Receive the multi-channel human posture tensor manifold, and use the Riemann logarithm and exponential mapping operators to perform iterative operations in the manifold space to calculate the Karcher mean data of the manifold trajectory. S22. Using the Karcher mean data as the geometric benchmark, the geodesic distance from the data points to the benchmark at each time point is calculated using the Riemann logarithmic mapping operator and a distribution curve is constructed. The interpolation interval is determined based on the target sampling rate. S23. Within the interpolation interval, the Riemann logarithm and exponential mapping operators are used to project and reconstruct between the tangent space and the manifold space to generate aligned motion trajectory features with equal geodesic arc lengths for adjacent data points.

4. The motion sports scoring method and device based on big data analysis according to claim 1, characterized in that, S3 specifically includes: S31. Receive motion trajectory features, calculate the arithmetic mean vector of the motion trajectory features and set it as the reference base point of the manifold tangent space. For each manifold data point in the motion trajectory features, use the Riemann logarithmic mapping operator to calculate the tangent space projection vector of the manifold data point relative to the reference base point. Arrange all tangent space projection vectors in time order to obtain a tangent space vector sequence that characterizes the attitude change trend. S32. Receive the tangent space vector sequence, construct a Gaussian radial basis kernel function based on geodesic distance to calculate the weight coefficients, and combine it with the centered deviation vector to perform a weighted outer product summation, outputting a weighted covariance matrix that considers the characteristics of the manifold geometry. S33. Perform eigenvalue decomposition on the weighted covariance matrix, correct negative eigenvalues ​​through positive thresholding to extract micro curvature features, and reconstruct manifold features that eliminate numerical singularity and satisfy positive definiteness.

5. The motion sports scoring method and device based on big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Receive the symmetric positive definite manifold features, calculate the Riemann mean of the symmetric positive definite manifold features as the reference point for tangent space mapping, use the reference point to project the symmetric positive definite manifold features to the tangent space through the Riemann logarithmic mapping operator to obtain the tangent space feature matrix, process the tangent space feature matrix through linear transformation to generate the corresponding query vector and key vector, and use the query vector and key vector to calculate the similarity coefficient in the inner product space induced by the Riemann metric. S42. Receive the tangent space feature matrix and similarity coefficients, perform normalization on the similarity coefficients to obtain the manifold inner product weights, use the manifold inner product weights to perform weighted summation and aggregation operations on the tangent space feature matrix, use the weighted aggregation mean vector as the reconstructed tangent space vector, and calculate the magnitude of the reconstructed tangent space vector based on the Riemann metric tensor. S43. Receive the reconstructed tangent space vector and the reference point of the tangent space mapping. Using the reference point of the tangent space mapping as the origin, use the Riemann exponential mapping operator to map the reconstructed tangent space vector from the tangent space back to the manifold space, generate reconstructed instances, calculate the mean and covariance matrix of the reconstructed instances, and combine the mean and covariance matrix as deep manifold distribution parameters describing the characteristics of the feature distribution.

6. The motion sports scoring method and apparatus based on big data analysis according to claim 1, characterized in that, S5 specifically includes: S51. Construct a big data motion sample library, collect multi-dimensional standard athlete motion data and map it to the manifold space to form a motion sample library containing massive standard posture manifold features. S52. Obtain standard motion distribution parameters, calculate the probability distribution model of standard motions using statistical analysis methods based on the motion sample library, and extract the statistical feature vector of the probability distribution model as standard motion distribution parameters. S53. Obtain the manifold features of the athlete's current movement and calculate the Wasserstein distance between the probability distribution of the athlete's movement and the probability distribution corresponding to the standard movement distribution parameters. S54. Based on the Wasserstein distance values, generate global distribution bias features by mapping.

7. The motion sports scoring method and apparatus based on big data analysis according to claim 1, characterized in that, The improved AGCN network model includes a data-dependent topology construction layer, a multi-scale spatiotemporal convolutional layer, a manifold feature enhancement layer, an adaptive adjacency aggregation layer, and a physical error parsing layer. The algorithm receives the deep manifold distribution parameters and the standard motion distribution parameters, calculates the initial deviation matrix of the deep manifold distribution parameters relative to the standard motion distribution parameters using the translation operator under the Riemannian geometric framework, performs a transpose operation on the initial deviation matrix, and adds the transpose matrix to the initial deviation matrix to generate a symmetric deviation matrix. The symmetric deviation matrix is ​​then mapped to the corresponding tangent space vector using the matrix logarithm operator to obtain the residual vector of the deep manifold distribution parameters relative to the standard motion distribution parameters. The data-dependent topology construction layer is used to receive the residual vector and initialize it into a physical adjacency matrix according to the human skeleton connection relationship. It calculates the data adjacency matrix based on the cosine similarity of node features, generates an implicit adjacency matrix, and weightedly fuses it with the physical adjacency matrix to obtain an adaptive topology adjacency matrix. The multi-scale spatiotemporal convolutional layer is used to input the residual vector and the adaptive topological adjacency matrix into the temporal and spatial graph convolutional units, respectively extracting temporal dynamic features and aggregating spatial semantic features, and fusing them to generate a multi-scale spatiotemporal feature map. The manifold feature enhancement layer receives multi-scale spatiotemporal feature maps and maps them to the Lie algebra tangent space using a manifold projection operator to obtain the tangent space feature manifold. Eigenvalues ​​and eigenvectors of the tangent space feature manifold are extracted using an eigenvalue decomposition algorithm, and a diagonal spectral matrix composed of eigenvalues ​​is constructed. This diagonal spectral matrix is ​​used as the geometric constraint feature of the manifold. The geometric constraint feature of the manifold is then concatenated with the multi-scale spatiotemporal feature map to obtain the enhanced manifold feature tensor. The adaptive adjacency aggregation layer is used to expand the enhanced manifold feature tensor along the node dimension, calculate the global attention coefficient, perform weighted pooling on the enhanced manifold feature tensor, and obtain the aggregated feature tensor through residual connection and nonlinear mapping of multilayer perceptron. The physical error analysis layer is used to map the aggregated feature tensor into a three-dimensional rotation vector through the axis-angle transformation layer, calculate the magnitude of the three-dimensional rotation vector as the joint rotation angle deviation, extract the unit vector direction of the three-dimensional rotation vector as the joint deviation orientation, and combine the joint rotation angle deviation and the joint deviation orientation to generate the joint-level physical angle error.

8. The motion sports scoring method and apparatus based on big data analysis according to claim 1, characterized in that, Specifically, S7 includes: S71. Obtain micro curvature features, global distribution deviation features and joint-level physical angle errors. Determine the splicing dimension based on each feature dimension and perform tensor concatenation to generate an initial high-dimensional feature tensor containing local geometric shape, overall probability distribution and physical motion constraint information. S72. Receive the initial high-dimensional feature tensor, calculate the scalar potential energy value of each data point relative to the center of the manifold using the Gaussian radial basis kernel function, generate the manifold topological constraint force matrix based on the negative gradient direction of the potential energy, perform channel dimension weighted modulation on the initial high-dimensional feature tensor, and output the energy field feature vector carrying the geometric structure information of the manifold. S73. The energy field feature vector is mapped to a query, key and value matrix through a self-attention weighting mechanism. The attention weight coefficients are calculated and the value matrix is ​​weighted and summed to output a score vector containing the comprehensive evaluation results.

9. A motion sports scoring device based on big data analysis, executing the motion sports scoring method based on big data analysis as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The Human Pose Tensor Manifold Construction Module is used to acquire motion videos and extract key points of the skeleton, model joint orientation as points on a special orthogonal group in 3D, and construct the human pose tensor manifold using rotation matrices. The nonlinear resampling alignment module is used to calculate the Karcher mean of the motion trajectory on the human posture tensor manifold. It uses geodesics to perform nonlinear resampling on the motion sequence and outputs the aligned motion trajectory features. The micro curvature feature extraction module is used to project the motion trajectory features onto the tangent space via Riemann exponential mapping, calculate the weighted covariance matrix using the manifold geodesic distance kernel function, extract the matrix eigenvalues ​​as micro curvature features, and output symmetric positive definite manifold features after positive definite processing. The manifold feature reconstruction module is used to map the features of a symmetric positive definite manifold to the tangent space to construct query and key vectors, use the manifold inner product for weighted aggregation, and map back to the manifold via the Riemann index to generate reconstruction instances, outputting the deep manifold distribution parameters; The global distribution deviation measurement module is used to calculate the Wasserstein distance between the probability distribution of an athlete's movement and that of a standard movement based on the deep manifold distribution parameters, and output the global distribution deviation feature. The physical angle error analysis module is used to calculate the residual vector based on the deep manifold distribution parameters and the standard motion distribution parameters, and input it into the improved AGCN model for processing. Based on the residual vector, an adaptive topology is constructed, and after multi-scale spatiotemporal convolution and manifold enhancement, it is aggregated and analyzed into joint-level physical angle error. The manifold energy field scoring module is used to cascade tensors based on micro curvature features, global distribution deviation features, and joint-level physical angle errors to generate an initial high-dimensional feature tensor. It then uses the manifold topological constraint force matrix to generate an energy field feature vector through weighted modulation, and finally processes it through a self-attention mechanism to output a comprehensive scoring vector.

Images