Large Model-based Ball Game Action Evaluation Method, Device, Equipment and Medium

Abstract

The present invention relates to a ball motion evaluation method, device, equipment and medium based on a large model. The method includes: collecting the three-dimensional coordinates of the joints of the trainee and the trajectory data of the sports equipment, and generating a standardized motion sequence through spatio-temporal compensation and noise suppression; separating and extracting the human joint rotation features and the equipment motion parameters based on inverse kinematics and differential kinematics technologies to construct a multi-dimensional motion feature vector; calculating the dynamic Mahalanobis distance and the contribution degree of key joints by using a contrast learning mechanism and a standard motion template to generate a quantitative evaluation result; and finally, outputting a motion correction scheme through inverse kinematics optimization and three-dimensional visualization rendering. This method breaks through the limitations of traditional single-modal evaluation, realizes the in-depth analysis of the collaborative motion characteristics of the equipment-human body, effectively solves the technical problems of inaccurate capture of motion details, single evaluation dimension and insufficient feedback visualization in high-speed scenarios, and significantly improves the scientificity and real-time performance of training guidance.

Description

Technical Field

[0001] The present invention relates to the technical field of sports teaching evaluation, and specifically relates to a ball action evaluation method, device, equipment and medium based on a large model. Background Art

[0002] In the field of sports science, the standardization of ball movement actions is crucial for athletes' skill improvement and sports injury prevention. Traditional action evaluation methods rely on coaches' empirical observations or manual analysis of basic sensor data, and judge the action quality through simple parameters such as speed and angle. However, ball actions are highly dynamic and complex, especially in high-speed movement scenarios, and subtle joint linkage differences or equipment trajectory deviations are often difficult to be effectively captured. In recent years, with the development of artificial intelligence technology, intelligent evaluation methods based on visual sensors and big data analysis have gradually become a research hotspot, aiming to provide objective and quantitative action improvement guidance for athletes through multi-dimensional data fusion and algorithm modeling.

[0003] In the prior art, evaluation systems based on single sensors such as monocular cameras or inertial measurement units have obvious limitations: on the one hand, visual data is easily affected by environmental light and occlusion interference, resulting in large joint positioning errors; on the other hand, traditional algorithms are difficult to extract deep-level correlation features of actions from high-dimensional time-series data, the evaluation dimension is single and lacks dynamic adaptability. In addition, most systems can only provide static scores and cannot generate targeted correction suggestions in combination with individual physiological differences, resulting in the disconnection between training guidance and actual needs. This technical defect makes the existing methods difficult to meet the requirements of accurately depicting action details and providing real-time feedback in high-level competitive training, especially unable to solve the dynamic analysis problem of the coordinated movement of equipment trajectories and human joints. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a ball action evaluation method, device, equipment and medium based on a large model that can fuse multi-source data, achieve in-depth analysis of action features, and dynamically generate visual correction schemes.

[0005] The purpose of the present invention is achieved by the following solutions:

[0006] In the first aspect, the present invention provides a ball action evaluation method based on a large model, including the following steps:

[0007] S1: Collect the original action data of the trainee based on multi-source sensors, and perform time and space synchronization processing and noise data suppression processing on the collected original action data to generate a trainee action data set;

[0008] S2: Separate and extract the human joint motion features and the motion equipment trajectory features from the trainee action dataset, and generate an action feature vector set containing spatio-temporal correlation features;

[0009] S3: Perform action quality deviation analysis on the action feature vector set based on a large model with a contrastive learning mechanism, and generate an action evaluation result containing a comprehensive score and key action deviation terms;

[0010] S4: Perform inverse kinematics optimization based on the action evaluation result, and generate a visualization correction scheme containing three-dimensional action comparison information.

[0011] In one embodiment, S1 of a ball game action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0012] S11: Collect the trainee and the motion equipment through a multi-view optical sensor to generate three-dimensional position data of the trainee's joint points and equipment usage data containing the rotation speed and acceleration data of the motion equipment;

[0013] S12: Perform spatio-temporal compensation on the missing parts of the three-dimensional position data and the equipment usage data based on the rotation group interpolation compensation algorithm to generate a time-space synchronized action sequence;

[0014] S13: Perform smoothing filtering on the time-space synchronized action sequence based on polynomial fitting to generate a trainee action dataset with high-frequency noise removed.

[0015] In one embodiment, S2 of a ball game action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0016] S21: Calculate the joint rotation angles of the trainee action dataset based on inverse kinematics technology to generate a kinematic feature matrix composed of relative rotation angles of each joint;

[0017] S22: Extract the motion state parameters of the motion equipment trajectory data in the trainee action dataset based on differential kinematics to generate an equipment trajectory feature set containing linear velocity, angular velocity, and acceleration;

[0018] S23: Perform multi-modal feature fusion on the kinematic feature matrix and the equipment trajectory feature set based on the dynamic weighted fusion algorithm to generate a fused action feature vector set. The calculation formula of the action feature vector set is:

[0019] ;

[0020] Among them, is the action feature vector set, is thek The rotation feature vector of a joint is the joint contribution weight coefficient is the set of instrument trajectory features, including linear velocity, angular velocity, and acceleration is the instrument feature weighting coefficient is the feature dimension concatenation operation

[0021] In one embodiment, S3 of the ball game action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0022] S31: Perform contrastive learning model construction processing based on the action feature vector set to construct a neural network model including a similarity contrast mechanism. The expression of the training objective function of the neural network model is:

[0023] ;

[0024] Among them, is the training objective function is the i th sample in the action feature vector set is the pre-stored standard action template is the deviation amount between the relative rotation angle of the joint and the standard angle is the temperature coefficient for controlling the similarity distribution range is the hyperparameter for adjusting the sparse constraint intensity is the set of trainable parameters of the neural network model is the number of batch samples for model training;

[0025] S32: Input the action feature vector into the neural network model for gradient backpropagation processing to generate the contribution degree parameters of the influence of each joint on the action deviation;

[0026] S33: Based on the contribution degree parameters and the pre-stored standard action template, perform similarity matching processing on the action feature vector in the neural network model to generate an action evaluation result including a comprehensive score and a key action deviation term

[0027] In one embodiment, S32 of the ball game action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0028] S321: Perform gradient backpropagation calculation processing on the action feature vector at the output layer of the neural network model based on the chain derivative algorithm to generate the original gradient matrix of each joint;

[0029] S322: Perform temporal smoothing processing on the original gradient matrix based on the sliding window average algorithm to generate denoised gradient data. The calculation formula of the denoised gradient data is:

[0030] ;

[0031] Among them, is the denoised gradient data, is the window width, is the k th joint's original gradient at the i th moment;

[0032] S323: Normalize the denoised gradient data based on the Z-score normalization algorithm to generate a contribution parameter, and the calculation formula of the contribution parameter is:

[0033] ;

[0034] Among them, is the contribution parameter, is the mean of all joint gradients, is the standard deviation.

[0035] In one embodiment, S33 of a ball action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0036] S331: Calculate the matching degree between the action feature vector and the standard action template based on the cosine similarity algorithm to generate a comprehensive score;

[0037] S332: Perform anomaly detection on the comprehensive score based on the dynamic threshold determination technology to generate a dynamic alarm threshold;

[0038] S333: Identify whether the comprehensive score is lower than the dynamic alarm threshold. If it is lower, perform anomaly joint positioning on the contribution parameter to generate a key action deviation term;

[0039] S334: Integrate the comprehensive score and the key action deviation term to generate an action evaluation result including the comprehensive score and the key action deviation term.

[0040] In one embodiment, S4 of a ball action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0041] S41: Process the key action deviation term in the action evaluation result based on the inverse kinematics optimization algorithm to generate a joint angle correction amount;

[0042] S42: Perform kinematic interpolation on the joint angle correction amount based on the trajectory smoothing technology to generate a continuous executable optimized action sequence;

[0043] S43: Based on the double-buffered rendering technology, perform three-dimensional visualization processing on the optimized action sequence and the original action sequence to generate a visualization correction plan. The visualization correction plan includes a comparison dual-view of the action skeleton models before and after optimization, a heat map of joint angle differences, and a dynamic marking box for key deviation parts.

[0044] In a second aspect, the present invention provides a ball game action evaluation device based on a large model. The device is configured with the following modules:

[0045] A data acquisition and processing module, which is used to collect the original action data of the trainee based on multi-source sensors, and perform time and space synchronization processing and noise data suppression processing on the collected original action data to generate a trainee action dataset;

[0046] A spatio-temporal correlation feature module, which is used to perform separation and extraction processing on the human joint motion features and the motion trajectory features of the sports equipment in the trainee action dataset to generate an action feature vector set containing spatio-temporal correlation features;

[0047] An action evaluation module, which is used to perform action quality deviation analysis processing on the action feature vector set based on a large model containing a contrastive learning mechanism to generate an action evaluation result containing a comprehensive score and key action deviation items;

[0048] A correction plan generation module, which is used to perform inverse kinematics optimization processing based on the action evaluation result to generate a visualization correction plan containing three-dimensional action comparison information.

[0049] In a third aspect, the present application provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements any one of the above-mentioned ball game action evaluation methods based on a large model.

[0050] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements any one of the above-mentioned ball game action evaluation methods based on a large model.

[0051] In summary, the ball action evaluation method based on the large model provided by the present invention collects and preprocesses multi-source sensor data, extracts spatio-temporal correlation features from human joint movement features and instrument trajectory features, and performs noise suppression and synchronization calibration to generate a high-precision student action data set; constructs an action feature vector set based on these data, uses the large model of the contrast learning mechanism to analyze the action quality deviation, and generates an evaluation result including a comprehensive score and key deviation items; further uses the inverse kinematics optimization algorithm to deeply process the evaluation result to generate a visualization correction plan including three-dimensional comparison information. This process can effectively solve the limitations of traditional action evaluation methods in terms of dynamics, complexity, and individual differences, and make full use of the feature extraction capabilities of multi-source sensor data and deep learning models. Through the analysis of the large model of the contrast learning mechanism, the deep-level correlation between action features is mined, thereby improving the feature expression ability of the evaluation model. At the same time, the inverse kinematics optimization algorithm can enhance the model's ability to generate action correction plans, thereby improving the pertinence and effectiveness of correction suggestions. The visualization correction plan further improves the interpretability of the evaluation results, provides intuitive action improvement guidance for athletes and coaches, helps to accurately depict action details in sports training, optimize training plans, reduce the risk of sports injuries, and improve training efficiency and competitive level.

[0052] For better understanding and implementation, the present invention will be described in detail below with reference to the accompanying drawings. Brief Description of the Drawings

[0053] Figure 1 It is a schematic flow chart of the ball action evaluation method based on the large model provided by the embodiment of the present application;

[0054] Figure 2 It is a schematic flow chart of generating an action evaluation result provided by the embodiment of the present application;

[0055] Figure 3 It is a schematic flow chart of generating a contribution degree parameter provided by the embodiment of the present application;

[0056] Figure 4 It is a schematic structural diagram of the ball action evaluation device based on the large model provided by another embodiment of the present application. Detailed Embodiments

[0057] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the understanding of the disclosure of the invention more thorough and comprehensive.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the description of the invention herein are for the purpose of describing specific embodiments only and are not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0059] In one embodiment, as Figure 1 shown, a ball game action evaluation method based on a large model is provided. In this embodiment, taking the application of this method to a terminal as an example, it can be understood that this method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0060] S1: Collect the original action data of the trainee based on multi-source sensors, and perform time and space synchronization processing and noise data suppression processing on the collected original action data to generate a trainee action dataset.

[0061] Specifically, the system non-invasively collects the actions of the trainee through a multi-source sensor network deployed outside the body to ensure the convenience and efficiency of data acquisition. This network consists of a high-resolution binocular vision system, a depth sensor, and an inertial measurement unit (IMU). All sensors are fixed around the training ground to avoid any interference with the trainee's actions. The binocular vision system uses the principle of stereoscopic imaging to capture the three-dimensional coordinates of human joints and achieves sub-millimeter spatial accuracy through parallax calculation. The depth sensor measures the distance information of the human body and the equipment in the three-dimensional space in real time through infrared structured light technology. The inertial measurement unit is deployed near the equipment and records its six-degree-of-freedom dynamic parameters, including three-axis acceleration and three-axis angular velocity.

[0062] Preferably, the system can adopt a timestamp-based synchronization algorithm to align different modality data in the time dimension to ensure the continuity and consistency of the action sequence. To cope with environmental light changes and occasional occlusions, the visual data undergoes adaptive histogram equalization processing to enhance edge features while suppressing overexposed areas. The depth data removes noise through bilateral filtering and retains edge details. The IMU data is drift-corrected using a Kalman filter, and the acceleration and angular velocity information are fused using a complementary filtering algorithm to output stable Euler angles and quaternion representations.

[0063] The system can further perform spatio-temporal domain interpolation on the occluded parts through the U-Net architecture in deep learning, utilize context information to reconstruct missing data, and ensure the integrity of the action sequence. The preprocessed data is integrated into the trainee action dataset and stored in a four-dimensional tensor structure of sample × time × space × modality. The spatial dimension includes X, Y, and Z coordinates, and the modality dimension differentiates visual, depth, and IMU data, providing a structured input for subsequent feature extraction.

[0064] S2: Perform separate extraction processing on the trainee action dataset for human joint motion features and exercise equipment trajectory features to generate an action feature vector set containing spatio-temporal correlation features.

[0065] Specifically, the system deeply analyzes the trainee action dataset and separates and extracts human joint motion features and exercise equipment trajectory features. The extraction of human joint motion features is based on the three-dimensional coordinate sequence of joint points. The system first calculates kinematic parameters such as the angle, angular velocity, and angular acceleration of joint points. These parameters are obtained from the coordinate sequence through numerical differentiation and integration methods, reflecting the dynamic changes of the human body during movement. To reduce the data dimension and highlight key features. Preferably, the system can use principal component analysis (PCA) to perform dimensionality reduction processing on the kinematic parameters and extract the most representative feature vectors.

[0066] The extraction of exercise equipment trajectory features focuses on the movement path and dynamic characteristics of the equipment. The system extracts the three-dimensional coordinate sequence of key points of the equipment from the dataset, such as extracting the data of the center of the racket face, and calculates parameters such as the speed, acceleration, and curvature of the trajectory. These parameters can be feature fused through a Gaussian mixture model to generate a feature vector that can characterize the movement pattern of the equipment. The Gaussian mixture model fits the data distribution through a probability density function and can effectively capture the complex features of the trajectory.

[0067] Through the dynamic time warping (DTW) algorithm, the system aligns the human joint motion features and the equipment trajectory features in the time dimension and extracts the correlation between the two in the time series. In the spatial dimension, the system calculates parameters such as the relative position, distance, and included angle between the human joint points and the equipment trajectory through a geometric transformation algorithm to generate spatial correlation features. These spatio-temporal correlation features are integrated into the action feature vector set, providing a comprehensive feature input for subsequent action quality assessment.

[0068] S3: Perform action quality deviation analysis processing on the action feature vector set based on a large model containing a contrast learning mechanism to generate an action evaluation result containing a comprehensive score and key action deviation terms.

[0069] Specifically, the system can use a large model incorporating a contrastive learning mechanism to deeply analyze the set of action feature vectors. Contrastive learning enhances the discriminative ability of the model by constructing positive and negative sample pairs. Positive samples are extracted from a standard action database and are similar to the trainee's actions; negative samples are generated through data augmentation techniques and are significantly different from the trainee's actions. The model optimizes its parameters through a contrastive loss function (such as the InfoNCE loss), making positive samples closer and negative samples farther apart in the feature space, thereby improving the ability to distinguish action features.

[0070] The large model architecture is based on Transformer and can effectively process high-dimensional time-series data. Transformer captures the global dependencies between feature vectors through self-attention mechanisms, avoiding the vanishing gradient problem of traditional recurrent neural networks. The encoder part of the model performs multi-layer feature extraction on the input features, and the decoder part generates the action quality assessment results. During the training process, the system uses the labeled standard action dataset and the trainee action dataset to optimize the model parameters through supervised learning to ensure the accuracy and reliability of the assessment results.

[0071] The action quality assessment results include a comprehensive score and key action deviation terms. The comprehensive score is calculated based on multi-dimensional indicators such as action normativity, instrument trajectory accuracy, and human joint coordination through a weighted average method. The key action deviation terms are identified through the attention mechanism. The model automatically learns the feature positions that have the greatest impact on the assessment results during the feature extraction process, and these positions correspond to significant deviations in the trainee's actions. The system further conducts a quantitative analysis of the deviation terms to generate a detailed deviation report, which includes the position, degree, and improvement direction of the deviation.

[0072] S4: Perform inverse kinematics optimization based on the action assessment results to generate a visual correction plan containing three-dimensional action comparison information.

[0073] Specifically, the system performs inverse kinematics optimization according to the action assessment results to generate a corrected action posture. Inverse kinematics optimization adjusts the motion parameters of the human joint points to make the trainee's actions as close as possible to the standard actions. Specifically, the system first defines the target postures, which are extracted from the standard action database and contain standard joint angles and instrument trajectories. During the optimization process, the system uses an algorithm based on gradient descent to iteratively adjust the joint point positions to minimize the difference between the trainee's actions and the target postures. To ensure that the optimization results conform to human kinematics laws, the system introduces human physiological constraints, such as joint range of motion and muscle strength limitations, which are obtained through previous biomechanics experiments.

[0074] Preferably, the system can generate intermediate transitional postures between the original actions and the corrected actions of the trainees through an interpolation algorithm, and construct a dynamic three-dimensional comparison model. During the generation of transitional postures, the system ensures that the posture changes at each step conform to physical laws and avoids unnatural actions. The key deviation points are automatically marked in the comparison model, and the positions and improvement directions of the deviations are intuitively displayed through color coding and arrow indication. The system further converts the comparison model into a dynamic three-dimensional display format, supporting frame-by-frame playback and real-time interaction, to help trainees intuitively understand the action deviations and correction methods.

[0075] The output forms of the visualized correction scheme are diverse, including videos, animations, and interactive 3D models. The scheme can not only include three-dimensional action comparison information, but also generate personalized correction suggestions in combination with physiological characteristics such as the height, weight, and strength level of the trainees. These suggestions are presented in the form of text descriptions and graphical annotations to ensure that trainees can clearly understand and apply them to actual training.

[0076] In summary, the ball game action evaluation method based on a large model provided by the present invention collects and preprocesses multi-source sensor data, extracts spatio-temporal correlation features from human joint movement features and instrument trajectory features, and performs noise suppression and synchronization calibration to generate a high-precision trainee action data set; constructs an action feature vector set based on these data, uses a large model with a contrastive learning mechanism to analyze the action quality deviation, and generates an evaluation result including a comprehensive score and key deviation items; further uses an inverse kinematics optimization algorithm to deeply process the evaluation result and generate a visualized correction scheme including three-dimensional comparison information. This process can effectively solve the limitations of traditional action evaluation methods in terms of dynamics, complexity, and individual differences, and make full use of the feature extraction capabilities of multi-source sensor data and deep learning models. Through the analysis of the large model with a contrastive learning mechanism, the deep-level correlations between action features are mined, thereby enhancing the feature expression ability of the evaluation model. At the same time, the inverse kinematics optimization algorithm can enhance the model's ability to generate action correction schemes, thereby improving the pertinence and effectiveness of correction suggestions. The visualized correction scheme further improves the interpretability of the evaluation results, provides intuitive action improvement guidance for athletes and coaches, helps to accurately depict action details in sports training, optimize training plans, reduce the risk of sports injuries, and improve training efficiency and competitive level.

[0077] In one embodiment, S1 of the ball game action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0078] S11: Collect data on trainees and sports equipment through multi-view optical sensors to generate three-dimensional position data of trainee joint points and equipment usage data containing the rotational speed and acceleration data of the sports equipment.

[0079] Specifically, the system performs non-contact data acquisition on trainees and sports equipment through a multi-view optical sensor network deployed outside the body, ensuring the convenience and efficiency of data acquisition. This network consists of a high-resolution binocular vision system and depth sensors, and all sensors are fixed around the training venue to avoid any interference with the trainees' movements. The binocular vision system uses the principle of stereoscopic imaging to capture the three-dimensional position data of human joint points, achieving sub-millimeter spatial accuracy through parallax calculation. The depth sensor measures the rotational speed and acceleration data of sports equipment in real time through infrared structured light technology.

[0080] The binocular vision system captures the same scene from different angles through two cameras and calculates depth information using the parallax principle. Parallax refers to the position difference of the same object on the imaging planes of the two cameras. The greater the parallax, the closer the object. The system finds corresponding points in the left and right images through a stereo matching algorithm, calculates the parallax map, and then generates a depth map. The infrared structured light technology projects infrared light with a known pattern and measures the degree of pattern deformation to calculate depth information, which has the advantages of high precision and anti-environmental light interference. The timestamp synchronization algorithm embeds accurate timestamps in each data packet to ensure the temporal alignment of data from different sensors, usually implemented using the Network Time Protocol (NTP) or dedicated synchronization hardware.

[0081] Preferably, the system can adopt a timestamp-based synchronization algorithm to align data of different modalities in the time dimension, ensuring the continuity and consistency of the action sequence. To cope with environmental light changes and occasional occlusions, the visual data is processed by adaptive histogram equalization to enhance edge features while suppressing overexposed areas. The depth data is denoised by bilateral filtering to retain edge details. The preprocessed data is integrated into the three-dimensional position data of the trainee's joint points and the equipment usage data containing the rotational speed and acceleration data of the sports equipment, providing a structured input for subsequent processing.

[0082] S12: Perform spatio-temporal compensation processing on the missing parts of the three-dimensional position data and the equipment usage data based on the rotation group interpolation compensation algorithm to generate a time-space synchronized action sequence.

[0083] Specifically, the rotation group interpolation compensation algorithm is an interpolation method based on Lie group theory for handling missing and compensating rotation data. The three-dimensional rotation group (Special Orthogonal Group, SO(3)) represents all rotations in three-dimensional space, parameterized by quaternions or Euler angles. Interpolation is performed by exponential mapping on the rotation group, ensuring smooth and continuous interpolation results in the rotation space. This algorithm can handle the periodic and non-linear characteristics of rotation data, avoiding errors caused by traditional linear interpolation. Spatiotemporal compensation processing combines time series interpolation and spatial interpolation to ensure data consistency in the time and space dimensions, usually implemented using cubic spline interpolation or Bezier curves.

[0084] Specifically, the system performs spatiotemporal compensation processing on the missing parts of the three-dimensional position data and instrument usage data based on the rotation group interpolation compensation algorithm. The rotation group interpolation compensation algorithm minimizes the marker point residuals by optimizing the objective function, accurately calculating the three-dimensional attitude angles and displacement vectors of each joint point. For missing data, the algorithm interpolates a set of uniformly spaced points, whose variations are between the maximum and minimum noise levels, to generate a noise profile. Finally, the system selects the subbands with an energy proportion exceeding 95% for reconstruction, generating the trajectory features after noise reduction. Through this algorithm, the system can effectively handle data missing and noise problems, generating a time-space synchronized action sequence, providing complete and accurate data support for subsequent evaluation.

[0085] S13: Perform smoothing filtering on the time-space synchronized action sequence based on polynomial fitting to generate a dataset of trainee actions with high-frequency noise removed.

[0086] Specifically, the system performs smoothing filtering on the time-space synchronized action sequence based on polynomial fitting to generate a dataset of trainee actions with high-frequency noise removed. Polynomial fitting fits the data with a polynomial function of an appropriate order to remove high-frequency noise and retain the main trend. The system further optimizes the data quality using a time series denoising model based on Transformer. The Transformer encoder contains global self-attention layers and feed-forward networks, capable of processing the entire input sequence simultaneously, thus learning global dependencies. The model is trained on synthetic clean signal / noise signal pairs, using the Adam optimizer, mean squared error (MSE) loss function, and monitoring the mean absolute error (MAE) during training. The trained model can effectively remove noise, generating a smooth and accurate dataset of trainee actions, providing high-quality input data for subsequent evaluation.

[0087] In one embodiment, step S2 of a ball game action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0088] S21: Calculate the joint rotation angles of the trainee's motion data set based on inverse kinematics technology, and generate a kinematic feature matrix composed of the relative rotation angles of each joint.

[0089] Specifically, the system uses inverse kinematics technology to calculate the joint rotation angles of the trainee's motion data set. Inverse kinematics is a technology that solves joint variables through known end-effector positions and postures. In this embodiment, the end-effector can be a sports equipment in the trainee's hand, such as a racket or a club. The system calculates the rotation angles of each joint of the human body through an inverse kinematics algorithm according to the end position and posture of the sports equipment.

[0090] Inverse kinematics algorithms usually include closed-form solutions, numerical solutions, and screw algebra methods, etc. This system uses a numerical solution method to gradually approximate the true value of the joint variables through iterative calculations. The numerical solution method has strong adaptability and can handle complex kinematic models. The system first establishes a kinematic model of the human joints, representing the relationship between human joint points as a series of transformation matrices. Then, through iterative calculations, the joint variables are adjusted to make the end position and posture of the sports equipment as close as possible to the actual measured values.

[0091] Through inverse kinematics technology, the system can accurately calculate the relative rotation angles of each joint of the trainee in ball sports, and integrate these angle data into a kinematic feature matrix. This matrix contains the rotation characteristics of each joint point of the human body during the movement, providing key input data for subsequent motion evaluation.

[0092] S22: Extract the motion state parameters of the sports equipment trajectory data in the trainee's motion data set based on differential kinematics, and generate an equipment trajectory feature set containing linear velocity, angular velocity, and acceleration.

[0093] Specifically, the system uses differential kinematics to extract the motion state parameters of the sports equipment trajectory data in the trainee's motion data set. Differential kinematics is a discipline that studies the relationship between joint velocities and end-effector velocities. In this system, through a differential kinematics algorithm, motion state parameters such as the linear velocity, angular velocity, and acceleration of the sports equipment can be extracted.

[0094] The differential kinematics algorithm is based on the Jacobian matrix of the kinematic model, which describes the linear relationship between joint velocities and end-effector velocities. The system calculates the Jacobian matrix through numerical differentiation methods to obtain the linear velocity and angular velocity of the sports equipment. The acceleration is obtained from the velocity data through second-order differentiation or numerical integration methods.

[0095] The system processes the trajectory data of the sports equipment, extracts its motion state parameters, and integrates these parameters into the equipment trajectory feature set. This feature set contains information such as linear velocity, angular velocity, and acceleration during the movement of the sports equipment, providing comprehensive input for subsequent action evaluation.

[0096] S23: Perform multi-modal feature fusion processing on the kinematic feature matrix and the equipment trajectory feature set based on the dynamic weighted fusion algorithm to generate a fused action feature vector set.

[0097] Specifically, the system uses the dynamic weighted fusion algorithm to perform multi-modal feature fusion processing on the kinematic feature matrix and the equipment trajectory feature set. The dynamic weighted fusion algorithm generates a fused action feature vector set by assigning different weights to different features and dynamically adjusting the importance of the features. The core of the dynamic weighted fusion algorithm lies in the dynamic adjustment of weights. The system automatically adjusts the weight coefficients according to the correlation and contribution degree of the features. For example, for the rotation angles of some key joints, the system will assign higher weights; for the trajectory parameters of the sports equipment, the system will also assign corresponding weights according to their impact on the action quality. The adjustment of weights can be achieved through machine learning algorithms or expert systems. The calculation formula for the action feature vector set is:

[0098] ;

[0099] where is the action feature vector set, is the rotation feature vector of the k th joint in the kinematic feature matrix, is the joint contribution weight coefficient, is the equipment trajectory feature set, including linear velocity, angular velocity, and acceleration, is the equipment feature weighting coefficient, is the feature dimension splicing operation. Through the dynamic weighted fusion algorithm, the system can comprehensively consider the human joint motion features and the equipment trajectory features, generate a comprehensive action feature vector set, and provide reliable data support for subsequent action evaluation and optimization.

[0100] In one embodiment, as Figure 2 shown, S3 of a ball action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0101] S31: Perform contrastive learning model construction processing based on the action feature vector set to construct a neural network model including a similarity comparison mechanism.

[0102] Specifically, the system deeply analyzes the action feature vector set through a contrastive learning model and constructs a neural network model with a similarity contrast mechanism. Contrastive learning is a self-supervised learning method that learns feature representations by maximizing the similarity of positive sample pairs and the dissimilarity of negative sample pairs. In this embodiment, the positive sample pair consists of the trainee's action feature vector and the standard action template, while the negative sample pair consists of the trainee's action feature vector and other trainees' action feature vectors. This design enables the model to effectively distinguish different action features, thereby improving the accuracy of action evaluation. The expression of the training objective function of the neural network model is:

[0103] ;

[0104] where, is the training objective function, where, is the i th sample in the action feature vector set, is the pre-stored standard action template, is the deviation amount of the relative joint rotation angle from the standard angle, is the temperature coefficient that controls the similarity distribution range, is the hyperparameter that adjusts the sparse constraint strength, is the set of trainable parameters of the neural network model, is the number of batch samples for model training.

[0105] During the training process, the system optimizes the model parameters through the similarity contrast mechanism, making the positive sample pairs closer in the feature space and the negative sample pairs farther away. The similarity function usually uses cosine similarity or Euclidean distance to measure the similarity between two feature vectors. The temperature coefficient controls the range of the similarity distribution. A smaller value makes the similarity distribution more concentrated, while a larger value makes the distribution smoother. The sparse constraint term is used to prevent the model from overfitting and ensure the generalization ability of the model. In this way, the system can effectively learn the deep representation of action features, providing a solid foundation for subsequent action evaluation.

[0106] S32: Input the action feature vector into the neural network model for gradient backpropagation processing to generate the contribution degree parameters of the influence of each joint pair on the action deviation.

[0107] Specifically, backpropagation of gradients is the core algorithm for neural network training. By calculating the gradients of the loss function with respect to the model parameters, the model parameters are updated to minimize the prediction error. In this system, backpropagation of gradients is not only used to update the model parameters, but also to generate contribution parameters, which reflect the degree of influence of each joint on the action deviation. Specifically, as Figure 3 shown, the contribution parameters are obtained through the following steps:

[0108] S321: Perform backpropagation calculation on the action feature vector at the output layer of the neural network model based on the chain rule for differentiation, and generate the original gradient matrix of each joint.

[0109] Specifically, the chain rule for differentiation calculates the gradients of the loss function with respect to the network parameters and updates the network parameters layer by layer to make the output of the network closer to the true value. Specifically, the system first passes the input samples to the neural network through forward propagation to obtain the output value of the network. Then, by calculating the loss function, the difference between the output value of the network and the true value is used as the input of the loss function to obtain the error between the predicted value and the true value. According to the value of the loss function, the gradients of each parameter with respect to the loss function are calculated layer by layer. At the same time, using the chain rule, the gradient of the previous layer is multiplied by the derivative of the activation function of the current layer with respect to the input to obtain the gradient of the current layer, and it is passed to the previous layer, and the calculation is performed layer by layer until the input layer. Finally, the system uses the gradient descent algorithm to adjust the value of each parameter in the opposite direction of the gradient, so that the loss function gradually decreases. In this way, the system can generate the original gradient matrix of each joint, providing a basis for subsequent temporal smoothing processing.

[0110] S322: Perform temporal smoothing processing on the original gradient matrix based on the moving window average algorithm to generate denoised gradient data.

[0111] Specifically, the moving window average algorithm slides a window with a fixed width over the time series data and calculates the average value of the data within the window, thereby smoothing the data and removing high-frequency noise. The calculation formula for the denoised gradient data is:

[0112] ;

[0113] where is the denoised gradient data, is the window width, is the k th joint at the i th moment of the original gradient.

[0114] S323: Perform normalization processing on the denoised gradient data based on the Z-score normalization algorithm to generate contribution parameters.

[0115] Specifically, Z-score normalization makes data comparable by converting it into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula for the contribution parameter is as follows:

[0116] ;

[0117] where is the contribution parameter, is the mean of all joint gradients, is the standard deviation. The Z-score normalization algorithm makes data comparable by converting it into a standard normal distribution with a mean of 0 and a standard deviation of 1. This method can effectively eliminate the dimensional difference of data and improve the comparability of data. In this system, the Z-score normalization algorithm is not only used for normalization processing but also for improving the interpretability of the contribution parameter, making the influence degree of each joint on the action deviation more intuitive. In this way, the system can generate contribution parameters, which reflect the influence degree of each joint on the action deviation and provide a scientific basis for subsequent action evaluation and optimization.

[0118] S33: Based on the contribution parameter and the pre-stored standard action template, perform similarity matching processing on the action feature vector in the neural network model to generate an action evaluation result including a comprehensive score and key action deviation terms.

[0119] Specifically, similarity matching evaluates the normality of the trainee's action by calculating the similarity between the trainee's action feature vector and the standard action template. Cosine similarity or Euclidean distance is usually used for similarity calculation, and these methods can effectively measure the similarity between two feature vectors; the comprehensive score is calculated by the weighted average method, and the weights are determined by the contribution parameter. Specifically, as Figure 2 shown, the action evaluation result is generated through the following steps:

[0120] S331: Based on the cosine similarity algorithm, perform matching degree calculation processing on the action feature vector and the standard action template to generate a comprehensive score.

[0121] Specifically, cosine similarity is used to measure the similarity in direction between two vectors. The system first compares the trainee's action feature vector with the pre-stored standard action template and calculates the cosine similarity between them. The calculation formula for cosine similarity is as follows:

[0122] ;

[0123] where is the trainee's action feature vector, It is the standard action template vector. The similarity value ranges from 0 to 1. The closer the value is to 1, the closer the action is to the standard template. The system maps the similarity value to the scoring range of 0 to 100 to generate a comprehensive score. The mapping formula is:

[0124] ;

[0125] Through this method, the system can quantify the similarity between the trainee's action and the standard action, providing an objective scoring basis for subsequent evaluation. The cosine similarity algorithm performs well in high-dimensional spaces and is especially suitable for measuring the similarity of feature vectors. Different from the Euclidean distance, the cosine similarity is not affected by the vector length and is more suitable for measuring the similarity of feature vectors. The similarity value ranges from -1 to 1. The closer the value is to 1, the more similar the directions are; the closer the value is to -1, the more opposite the directions are. The system converts the similarity value to a score from 0 to 100 through linear mapping to ensure the intuitiveness and interpretability of the score.

[0126] S332: Perform anomaly detection processing on the comprehensive score based on the dynamic threshold determination technology to generate a dynamic alarm threshold.

[0127] Specifically, the dynamic threshold determination technology dynamically adjusts the alarm threshold by analyzing historical scoring data to ensure the accuracy and adaptability of anomaly detection. The system first calculates the mean of the historical scoring data and the standard deviation , and then generates the dynamic alarm threshold according to the following formula:

[0128] ;

[0129] where is the dynamic adjustment coefficient, which is usually determined according to the distribution of historical data and the application scenario. The system monitors the comprehensive score in real time to determine whether it is lower than the dynamic alarm threshold. If it is lower than the threshold, the anomaly detection mechanism is triggered, providing a basis for subsequent action deviation analysis.

[0130] The dynamic threshold determination technology dynamically adjusts the threshold through statistical analysis of historical data to ensure the accuracy and adaptability of anomaly detection. The mean and the standard deviation are the core statistical quantities describing the data distribution, and the dynamic adjustment coefficient controls the strictness of the threshold. This method can effectively cope with the changes in data distribution and improve the robustness of anomaly detection.

[0131] S333: Identify whether the comprehensive score is lower than the dynamic alarm threshold. If it is lower, perform anomaly joint positioning processing on the contribution degree parameter to generate a key action deviation item.

[0132] Specifically, the system identifies whether the comprehensive score is lower than the dynamic alarm threshold. If it is lower, the system performs abnormal joint positioning on the contribution parameters to generate key action deviation terms. The contribution parameters reflect the influence degree of each joint on the action deviation. By analyzing these parameters, the system can accurately locate the key joints causing abnormal actions. Specifically, the system first compares the comprehensive score with the dynamic alarm threshold. If the comprehensive score is lower than the dynamic alarm threshold, the system will trigger an abnormal handling process. At this time, the system analyzes the contribution parameters and identifies the joint features with larger values. The joints corresponding to these features have the greatest influence on the action deviation and need to be focused on. The system sets a contribution threshold, screens out the joints whose contribution parameters exceed this threshold, and determines them as key action deviation terms.

[0133] To further improve the accuracy of positioning, the system can combine domain knowledge and expert experience to prioritize the key joints. For example, the movement of certain joints has a decisive impact on the success rate of specific actions, and the deviations of these joints should be considered first. In this way, the system can generate key action deviation terms to provide specific improvement directions for the trainees.

[0134] S334: Integrate the comprehensive score and the key action deviation terms to generate an action evaluation result that includes the comprehensive score and the key action deviation terms.

[0135] Specifically, the comprehensive score provides an overall quality assessment of the trainee's actions, while the key action deviation terms point out specific improvement directions. By integrating these two parts of information, the system can provide a comprehensive action evaluation result for the trainees. Specifically, the system first integrates the comprehensive score and the key action deviation terms into a structured evaluation report. The evaluation report contains two main parts: one is the comprehensive score, which intuitively shows the overall quality of the trainee's actions in numerical form; the other is the key action deviation terms, which list in detail the joints that need to be improved and their deviation degrees in a list form. The system can also use visualization technology to mark the key action deviation terms on the three-dimensional action model to help the trainees more intuitively understand the position and degree of the deviations.

[0136] In addition, the system can generate personalized training suggestions according to the severity of the key action deviation terms. For example, for joints with larger deviations, the system can recommend specific training exercises or adjustment methods. In this way, the system not only provides the evaluation result but also provides scientific guidance for the trainees' subsequent training.

[0137] A ball action evaluation method based on a large model provided above constructs a neural network model containing a similarity comparison mechanism, performs deep feature learning using an action feature vector set, and optimizes the training objective function of the model to generate an action quality evaluation result. By constructing and optimizing the contrastive learning model, this method can effectively solve the limitations of traditional action evaluation methods in feature expression and deviation analysis. Through gradient backpropagation processing, contribution parameters reflecting the influence degree of each joint on the action deviation are generated, thereby enhancing the model's ability to identify key action deviations. The similarity matching process combines with a standard action template to generate an evaluation result containing a comprehensive score and key action deviation items, significantly improving the objectivity and pertinence of the evaluation. Through the deep feature extraction ability of the contrastive learning mechanism, this process explores the complex relationships between action features, thereby enhancing the feature expression ability of the evaluation model. At the same time, the gradient backpropagation algorithm can enhance the model's sensitivity to action deviations, thereby improving the reliability and stability of the evaluation results. The generation of contribution parameters further improves the interpretability of the evaluation results, provides intuitive action improvement guidance for athletes and coaches, helps to accurately depict action details in sports training, optimize training plans, reduce the risk of sports injuries, and improve training efficiency and competitive level.

[0138] In one embodiment, step S4 of the ball action evaluation method based on a large model provided by the present invention specifically includes the following steps:

[0139] S41: Process the key action deviation items in the action evaluation result based on the inverse kinematics optimization algorithm to generate joint angle correction amounts.

[0140] Specifically, the system uses the inverse kinematics optimization algorithm to process the key action deviation items in the action evaluation result to generate joint angle correction amounts. The inverse kinematics optimization algorithm is a technique for solving joint variables based on the known position and posture of the end effector. In this embodiment, the end effector can be a sports instrument in the trainee's hand, such as a racket or a club. The system calculates the rotation angles of each joint of the human body according to the end position and posture of the sports instrument through the inverse kinematics algorithm, and compares them with the standard action template to identify the key action deviation items.

[0141] Inverse kinematics optimization algorithms usually include closed - form solutions, numerical solutions, screw algebra methods, etc. In this embodiment, the system adopts a numerical solution and gradually approaches the true values of joint variables through iterative calculations. The numerical solution has strong adaptability and can handle complex kinematic models. The system first establishes a kinematic model of the human joints and represents the relationships between human joint points as a series of transformation matrices. Then, through iterative calculations, the joint variables are adjusted to make the end - position and attitude of the exercise equipment as close as possible to the standard action template. This adjustment process not only optimizes the joint angles but also ensures the coherence and naturalness of the actions.

[0142] S42: Perform kinematic interpolation processing on the joint angle correction amount based on the trajectory smoothing technology to generate a continuous and executable optimized action sequence.

[0143] Specifically, the trajectory smoothing technology slides a window with a fixed width over the time - series data and calculates the average value of the data within the window, thereby smoothing the data and removing high - frequency noise. In this embodiment, the system generates a smooth action trajectory on the time - series through a kinematic interpolation algorithm. The kinematic interpolation algorithm adjusts the joint angles step by step by calculating the differences between adjacent joint angles, making the action sequence more coherent. The system first performs preliminary processing on the input trajectory data, such as denoising and outlier processing, to create a good starting point for interpolation. Then, according to the complexity of the trajectory and the requirement for smoothness, methods such as linear interpolation and cubic spline interpolation are selected. By optimizing the distribution of interpolation nodes and dynamically adjusting the positions of control points, the system can generate a more natural and smooth curve. Finally, through local error analysis and global error minimization, the system balances the smoothness of the curve and the faithfulness to the original data to generate a continuous and executable optimized action sequence.

[0144] S43: Perform three - dimensional visualization processing on the optimized action sequence and the original action sequence based on the double - buffer rendering technology to generate a visualization correction plan. The visualization correction plan includes a comparison dual - view of the action skeleton models before and after optimization, a heat map of joint angle differences, and a dynamic marking box for key deviation parts.

[0145] Specifically, the double-buffer rendering technology is a technique used to reduce screen tearing and improve rendering efficiency. By alternately rendering and displaying between two buffers, it ensures the smoothness and stability of the screen. The system first converts the optimized action sequence and the original action sequence into the data format of a three-dimensional skeletal model. Then, it renders the optimized action sequence and the original action sequence in the double buffers respectively to generate a comparison double view. The use of the double buffers ensures the efficiency of the rendering process and the stability of the screen. The system calculates the joint angle differences to generate a heat map of joint angle differences. The heat map visually shows the deviation degree of joint angles through color coding, helping trainees quickly identify key deviation areas. In addition, the system marks the key deviation areas in the three-dimensional model and highlights the joints that need improvement through dynamic marking frames. The dynamic marking frames change with the change of the action sequence, providing real-time visual feedback. Finally, the generated visualization correction plan includes the comparison double view of the action skeletal models before and after optimization, the heat map of joint angle differences, and the dynamic marking frames of key deviation areas, providing intuitive action deviation analysis and correction guidance for trainees.

[0146] Preferably, as Figure 4 shown, the present invention provides a ball game action evaluation device 500 based on a large model, and this device is configured with the following modules:

[0147] A data acquisition and processing module 510, which is used to collect the original action data of trainees based on multi-source sensors, and perform time and space synchronization processing and noise data suppression processing on the collected original action data to generate a trainee action data set;

[0148] A spatio-temporal correlation feature module 520, which is used to perform separation and extraction processing on the human joint movement features and the movement equipment trajectory features of the trainee action data set to generate an action feature vector set containing spatio-temporal correlation features;

[0149] An action evaluation module 530, which is used to perform action quality deviation analysis processing on the action feature vector set based on a large model containing a contrastive learning mechanism to generate an action evaluation result containing a comprehensive score and key action deviation items;

[0150] A correction plan generation module 540, which is used to perform inverse kinematics optimization processing based on the action evaluation result to generate a visualization correction plan containing three-dimensional action comparison information.

[0151] In summary, the ball action evaluation device based on the large model provided by the present invention extracts spatio-temporal correlation features from human joint movement features and instrument trajectory features through multi-source sensor data acquisition and preprocessing, and performs noise suppression and synchronization calibration to generate a high-precision student action data set; constructs an action feature vector set based on these data, uses the large model of the contrast learning mechanism to analyze the action quality deviation, and generates an evaluation result including a comprehensive score and key deviation items; further uses the inverse kinematics optimization algorithm to deeply process the evaluation result to generate a visualization correction plan including three-dimensional comparison information. This process can effectively solve the limitations of traditional action evaluation methods in terms of dynamics, complexity, and individual differences, and make full use of the feature extraction capabilities of multi-source sensor data and deep learning models. Through the analysis of the large model of the contrast learning mechanism, the deep-level correlation between action features is mined, thereby enhancing the feature expression ability of the evaluation model. At the same time, the inverse kinematics optimization algorithm can enhance the model's ability to generate action correction plans, thereby improving the pertinence and effectiveness of the correction suggestions. The visualization correction plan further improves the interpretability of the evaluation results, provides intuitive action improvement guidance for athletes and coaches, helps to accurately depict action details in sports training, optimize training plans, reduce the risk of sports injuries, and improve training efficiency and competitive level.

[0152] Preferably, the data acquisition and processing module 510 is configured with the following units:

[0153] The motion data acquisition unit is used to collect the student and the sports equipment through a multi-view optical sensor, and generate three-dimensional position data of the student's joint points and equipment usage data containing the rotation speed and acceleration data of the sports equipment;

[0154] The spatio-temporal compensation unit is used to perform spatio-temporal compensation processing on the missing parts of the three-dimensional position data and the equipment usage data based on the rotation group interpolation compensation algorithm, and generate a time-space synchronous action sequence;

[0155] The data smoothing processing unit is used to perform smoothing filtering processing based on polynomial fitting on the time-space synchronous action sequence, and generate a student action data set that eliminates high-frequency noise.

[0156] Preferably, the spatio-temporal correlation feature module 520 is configured with the following units:

[0157] The kinematic feature matrix generation unit is used to perform joint rotation angle calculation processing on the student action data set based on inverse kinematics technology, and generate a kinematic feature matrix composed of relative rotation angles of each joint;

[0158] An instrument trajectory feature set generation unit, which is used to extract and process motion state parameters from the motion instrument trajectory data in the trainee action dataset based on differential kinematics, and generate an instrument trajectory feature set including linear velocity, angular velocity, and acceleration;

[0159] An action feature vector set generation unit, which is used to perform multi-modal feature fusion processing on the kinematic feature matrix and the instrument trajectory feature set based on the dynamic weighted fusion algorithm, and generate a fused action feature vector set.

[0160] Preferably, the action evaluation module 530 is configured with the following units:

[0161] A contrast learning model construction unit, which is used to perform contrast learning model construction processing based on the action feature vector set, and construct a neural network model including a similarity comparison mechanism;

[0162] A contribution parameter generation unit, which is used to input the action feature vector into the neural network model for gradient backpropagation processing, and generate contribution parameters for the influence degree of each joint on the action deviation;

[0163] An action evaluation result generation unit, which is used to perform similarity matching processing on the action feature vector in the neural network model based on the contribution parameter and the pre-stored standard action template, and generate an action evaluation result including a comprehensive score and key action deviation items.

[0164] Preferably, the contribution parameter generation unit includes an original gradient matrix generation subunit, a denoised gradient data generation subunit, and a contribution parameter generation subunit. Among them, the original gradient matrix generation subunit is used to perform gradient backpropagation calculation processing on the action feature vector at the output layer of the neural network model based on the chain rule for differentiation, and generate the original gradient matrix of each joint; the denoised gradient data generation unit is used to perform temporal smoothing processing on the original gradient matrix based on the sliding window average algorithm, and generate denoised gradient data; the contribution parameter generation subunit is used to perform normalization processing on the denoised gradient data based on the Z-score normalization algorithm, and generate contribution parameters.

[0165] Preferably, the action evaluation result generation unit includes a comprehensive score generation subunit, a dynamic threshold generation subunit, a deviation term generation subunit, and an evaluation result generation subunit. Among them, the comprehensive score generation unit is used to calculate the matching degree between the action feature vector and the standard action template based on the cosine similarity algorithm, and generate a comprehensive score; the dynamic threshold generation unit is used to perform anomaly detection processing on the comprehensive score based on the dynamic threshold determination technology, and generate a dynamic alarm threshold; the deviation term generation unit is used to identify whether the comprehensive score is lower than the dynamic alarm threshold. If it is lower, it performs abnormal joint positioning processing on the contribution parameter to generate a key action deviation term; the evaluation result generation unit is used to integrate the comprehensive score and the key action deviation term to generate an action evaluation result including the comprehensive score and the key action deviation term.

[0166] Preferably, the correction scheme generation module 540 is configured with the following units:

[0167] The joint angle correction amount generation unit is used to process the key action deviation term in the action evaluation result based on the inverse kinematics optimization algorithm to generate a joint angle correction amount;

[0168] The optimized action sequence generation unit is used to perform kinematic interpolation processing on the joint angle correction amount based on the trajectory smoothing technology to generate a continuous executable optimized action sequence;

[0169] The visualization scheme generation unit is used to perform three-dimensional visualization processing on the optimized action sequence and the original action sequence based on the double-buffer rendering technology to generate a visualization correction scheme. The visualization correction scheme includes a comparison dual view of the action skeleton models before and after optimization, a heat map of joint angle differences, and a dynamic marking box for key deviation parts.

[0170] In one embodiment, the present application also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the above-mentioned ball action evaluation method based on a large model.

[0171] In one embodiment, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the above-mentioned ball action evaluation method based on a large model.

[0172] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0173] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial descriptions of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present disclosure solution. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0174] As described above, the above are only the specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of various changes or substitutions, and these should all be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A ball game action evaluation method based on a large model, characterized in that, It includes the following steps: S1: Collect the original action data of the trainee based on multi-source sensors, and perform time and space synchronization processing and noise data suppression processing on the collected original action data to generate a trainee action dataset; S2: Perform separation and extraction processing on the human joint motion characteristics and the motion equipment trajectory characteristics of the trainee action dataset to generate an action feature vector set containing spatio-temporal correlation characteristics; S3: Perform action quality deviation analysis processing on the action feature vector set based on a large model containing a contrast learning mechanism to generate an action evaluation result containing a comprehensive score and key action deviation items; S4: Perform inverse kinematics optimization processing based on the action evaluation result to generate a visualization correction plan containing three-dimensional action comparison information; Among them, S3 includes: S31: Perform contrast learning model construction processing based on the action feature vector set to construct a neural network model containing a similarity comparison mechanism. The expression of the training objective function of the neural network model is: ； Among them ， is the training objective function, is the i -th sample in the action feature vector set, is the pre-stored standard action template, is the deviation amount between the relative joint rotation angle and the standard angle, is the temperature coefficient for controlling the similarity distribution range, is the hyperparameter for adjusting the sparse constraint intensity, is the set of trainable parameters of the neural network model, is the number of batch samples for model training; S32: Input the action feature vector into the neural network model for gradient backpropagation processing to generate contribution degree parameters of the influence degree of each joint on the action deviation; S33: Based on the contribution degree parameters and a pre-stored standard action template, perform similarity matching processing on the action feature vector in the neural network model to generate an action evaluation result containing a comprehensive score and key action deviation items; Among them, S32 includes: S321: Perform gradient backpropagation calculation processing on the action feature vector at the output layer of the neural network model based on the chain derivative algorithm to generate an original gradient matrix of each joint; S322: Perform temporal smoothing processing on the original gradient matrix based on the sliding window average algorithm to generate denoised gradient data. The calculation formula of the denoised gradient data is: ； Among them, is the denoised gradient data, is the window width, is the k th joint's original gradient at the i th moment; S323: Perform normalization processing on the denoised gradient data based on the Z-score normalization algorithm to generate contribution degree parameters. The calculation formula of the contribution degree parameters is: ； Among them, is the contribution parameter, is the mean of all joint gradients, is the standard deviation.

2. The method according to claim 1, wherein S1 includes: S11: Collect the trainee and the exercise equipment through a multi-view optical sensor to generate three-dimensional position data of the trainee's joint points and equipment usage data containing the rotational speed and acceleration data of the exercise equipment; S12: Perform spatio-temporal compensation processing on the missing parts of the three-dimensional position data and the equipment usage data based on the rotation group interpolation compensation algorithm to generate a time-space synchronized action sequence; S13: Perform smoothing filtering processing based on polynomial fitting on the time-space synchronized action sequence to generate a trainee action dataset that eliminates high-frequency noise.

3. The method according to claim 1, wherein S2 includes: S21: Perform joint rotation angle calculation processing on the trainee action dataset based on inverse kinematics technology to generate a kinematic feature matrix composed of relative rotation angles of each joint; S22: Perform motion state parameter extraction processing on the motion equipment trajectory data in the trainee action dataset based on differential kinematics to generate an equipment trajectory feature set containing linear velocity, angular velocity, and acceleration; S23: Perform multimodal feature fusion processing on the kinematic feature matrix and the instrument trajectory feature set based on the dynamic weighted fusion algorithm to generate a fused action feature vector set. The calculation formula for the action feature vector set is as follows: ； Among them, is the set of action feature vectors, is the rotation feature vector of the k -th joint in the kinematic feature matrix, is the joint contribution weight coefficient, is the set of instrument trajectory features, including linear velocity, angular velocity and acceleration, is the instrument feature weighting coefficient, is the feature dimension splicing operation.

4. The method according to claim 1, wherein The S33 includes: S331: Calculate the matching degree between the action feature vector and the standard action template based on the cosine similarity algorithm to generate a comprehensive score; S332: Perform anomaly detection processing on the comprehensive score based on the dynamic threshold determination technology to generate a dynamic alarm threshold; S333: Identify whether the comprehensive score is lower than the dynamic alarm threshold. If it is lower, perform anomaly joint positioning processing on the contribution parameter to generate a key action deviation term; S334: Integrate the comprehensive score and the key action deviation term to generate an action evaluation result including the comprehensive score and the key action deviation term.

5. The method according to any one of claims 1 to 4, characterized in that, The S4 includes: S41: Process the key action deviation term in the action evaluation result based on the inverse kinematics optimization algorithm to generate a joint angle correction amount; S42: Perform kinematic interpolation processing on the joint angle correction amount based on the trajectory smoothing technology to generate a continuous executable optimized action sequence; S43: Perform 3D visualization processing on the optimized action sequence and the original action sequence based on the double-buffer rendering technology to generate a visualization correction scheme. The visualization correction scheme includes a comparison dual-view of the action skeleton models before and after optimization, a heat map of joint angle differences, and a dynamic marking box for key deviation parts.

6. A ball motion evaluation device based on a large model, characterized in that, The device includes: A data acquisition and processing module, configured to collect the original action data of the trainee based on multi-source sensors, and perform time and space synchronization processing and noise data suppression processing on the collected original action data to generate a trainee action data set; A spatio-temporal correlation feature module, configured to perform separation and extraction processing on the human joint motion features and the motion instrument trajectory features in the trainee action data set to generate an action feature vector set including spatio-temporal correlation features; An action evaluation module, configured to perform action quality deviation analysis processing on the action feature vector set based on a large model including a contrastive learning mechanism to generate an action evaluation result including a comprehensive score and a key action deviation term; A correction scheme generation module, configured to perform inverse kinematics optimization processing based on the action evaluation result to generate a visualization correction scheme including 3D action comparison information; Among them, the action evaluation module includes: A contrastive learning model construction unit, configured to perform contrastive learning model construction processing based on the action feature vector set to construct a neural network model including a similarity comparison mechanism. The expression of the training objective function of the neural network model is: ； Among them, is the training objective function, is the i-th sample in the action feature vector set, is the pre-stored standard action template, is the deviation between the relative joint rotation angle and the standard angle, is the temperature coefficient for controlling the similarity distribution range, is the hyperparameter for adjusting the sparse constraint intensity, is the set of trainable parameters of the neural network model, is the number of batch samples for model training; A contribution parameter generation unit, configured to input the action feature vector into the neural network model for gradient backpropagation processing to generate a contribution parameter of the influence degree of each joint on the action deviation; An action evaluation result generation unit, configured to perform similarity matching processing on the action feature vector in the neural network model based on the contribution parameter and a pre-stored standard action template to generate an action evaluation result including a comprehensive score and a key action deviation term. Among them, the contribution parameter generation unit includes: An original gradient matrix generation subunit, configured to perform gradient backpropagation calculation processing on the action feature vector at the output layer of the neural network model based on the chain rule of differentiation, and generate an original gradient matrix for each joint; A denoised gradient data generation subunit, configured to perform temporal smoothing processing on the original gradient matrix based on a sliding window average algorithm to generate denoised gradient data, and the calculation formula for the denoised gradient data is: ； Among them, is the denoised gradient data, is the window width, is the original gradient of the k-th joint at the i-th moment; A contribution parameter generation subunit, configured to perform normalization processing on the denoised gradient data based on the Z-score normalization algorithm to generate contribution parameters, and the calculation formula for the contribution parameters is: ； Among them, is the contribution parameter, is the mean of all joint gradients, is the standard deviation.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the method described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the method described in any one of claims 1 to 5 is implemented.

Images