Intelligent correction method for sports teaching action based on spatio-temporal graph convolution

Abstract

The application provides a sports teaching action intelligent correction method based on space-time graph convolution, comprising the following steps: acquiring original action sequence data from a preset multi-project sports action video database, performing initial segmentation on action characteristics of different projects such as basketball and football, and obtaining preliminary space-time sequence segments; adopting a convolutional neural network to extract features of the obtained space-time sequence segments, capturing dynamic changes of key joint nodes in the action, and determining a universal space-time relationship vector across projects; through an obtained adjusted mixed feature model, performing real-time matching on newly input sports action data, determining a matching degree, and outputting action rule deviation to obtain an analysis result after deviation correction; and according to a generated final universal analysis template, performing batch application on various project data of a comprehensive sports teaching platform to obtain a unified action evaluation report.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an intelligent correction method for physical education movements based on spatiotemporal graph convolution. Background Technology

[0002] Sports motion analysis, as a crucial field in intelligent sports teaching and exercise assessment, plays an indispensable role in improving training efficiency and providing personalized guidance. With the integration of technology, building a comprehensive analysis platform capable of covering multiple sports has become an urgent need for industry development. However, the significant differences in motion characteristics and analytical requirements among different sports present serious challenges to technological implementation.

[0003] Current technical solutions often struggle to adapt to the diversity of sports, and generally suffer from insufficient adaptability in motion feature extraction.

[0004] Many methods lack the ability to dynamically capture the spatiotemporal relationships behind movements when dealing with different projects, resulting in analysis results that cannot accurately reflect the unique movement patterns of the project.

[0005] This limitation means that when the system is used across projects, it often needs to be designed from scratch, making it impossible to achieve modularity and universality.

[0006] Against this backdrop, the heterogeneous nature of sports has become a core challenge for technological breakthroughs.

[0007] The movements of different sports vary greatly in terms of time and space. For example, basketball emphasizes the coordination between the arms and the ball, while football emphasizes the power transmission between the legs and the torso. This difference requires the system to be able to handle the movement characteristics of different sports in a targeted manner.

[0008] A deeper challenge lies in how to find common movement characteristics that can be shared across projects, while acknowledging the differences between projects, such as key nodes of the human skeleton and the changing patterns of movement speed.

[0009] If it is not possible to effectively distinguish which features are universal and which require specific design, the system will be difficult to reuse efficiently across multiple projects.

[0010] Therefore, how to balance the personalized needs of motion analysis for different sports with the universal requirements of underlying characteristics in a comprehensive sports teaching platform has become a key issue that urgently needs to be addressed.

[0011] Especially in practical applications, such as analyzing basketball dribbling and soccer shooting, ensuring that the extracted features, such as the rate of change of joint angles, have consistent physical meaning between the two projects, and determining which features can be shared and which need to be customized, has become a problem that must be overcome in the process of technology implementation.

[0012] Furthermore, it is important to emphasize that basketball and football, as mainstream mass sports, serve as the optimal benchmarks for extracting universal movement features across sports (compared to specialized sports like gymnastics and rhythmic gymnastics). Basketball and football movements do not require overly specialized body postures, and their movement characteristics are more aligned with the basic laws of human movement. Compared to single-dimensional sports like volleyball and badminton, basketball primarily involves coordinated movements of the upper limbs and core, while football focuses on explosive movements of the lower limbs and core. Combining these two approaches comprehensively covers the spatiotemporal movement characteristics of core joints such as the shoulder, elbow, wrist, hip, knee, and ankle. Simultaneously, basketball and football have high educational reach, resulting in a much larger accumulation of movement video data than other sports, providing ample and high-quality sample support for universal feature extraction. If universal features are extracted solely from other sports without considering the benchmark features of basketball and football, problems such as feature specialization, poor transferability, and insufficient sample size can easily arise, failing to meet the universal analysis needs of a comprehensive sports teaching platform. Therefore, the raw data used in this invention will include relevant data from basketball and football.

[0013] Terminology Explanation 1. Spatiotemporal Graph Convolutional Network (ST-GCN): The convolutional neural network mentioned in this application specifically refers to the spatiotemporal graph convolutional network, which is a convolutional neural network based on the topological graph of the human skeleton that can simultaneously capture spatial structural features and temporal evolution features. It is the core network model for sports motion feature extraction.

[0014] 2. Spatial Adjacency Strength: A feature value used to quantify the degree of spatial correlation between different key joint nodes in the human body. The larger the value, the stronger the spatial cooperation between joints. The value range is [0,1].

[0015] 3. Temporal Dependency Strength: A feature value used to quantify the degree of temporal correlation between different key joint nodes in the human body. The larger the value, the higher the temporal coordination between joints. The value range is [0,1].

[0016] 4. General Feature Set: A feature set formed by filtering and integrating spatiotemporal relationship vectors shared across projects. The shared feature library is the underlying data storage carrier of the general feature set. The two have a one-to-one data storage and logical set relationship.

[0017] 5. Movement pattern deviation: Abbreviated as deviation, it refers to the quantitative value of the characteristic difference between actual sports movement data and standard movement template. The larger the deviation value, the higher the degree of non-standard movement.

[0018] 7. Modular requirements: These refer to the reusability and standardization conditions that a general analysis template must meet. They are the core criteria for judging whether a template can be applied in batches across projects.

[0019] 8. Hybrid Feature Model: A feature model that integrates general features across projects with project-specific parameters, enabling both general motion analysis and personalized correction. Summary of the Invention

[0020] This invention provides an intelligent correction method for physical education movements based on spatiotemporal graph convolution, mainly comprising: Raw motion sequence data is obtained from a pre-set multi-sports motion video database. Initial segmentation is performed based on the motion characteristics of different sports, such as basketball and football, to obtain preliminary spatiotemporal sequence fragments.

[0021] Convolutional neural networks are used to extract features from the obtained spatiotemporal sequence segments, capture dynamic changes in key joint nodes of the action, and determine a general spatiotemporal relationship vector across projects.

[0022] Based on the determined spatiotemporal relationship vectors, the personalized action patterns of different items are classified by support vector machine. If the classification results show that the feature similarity is higher than the preset threshold, these vectors are included in the shared feature library to obtain the general feature set after classification.

[0023] After obtaining the set of general features after classification, the remaining personalized needs are adjusted in the multi-project analysis platform to determine whether project-specific parameters need to be added. If so, specific indicators of basketball arm coordination or football leg strength are integrated to obtain the adjusted hybrid feature model.

[0024] The adjusted hybrid feature model is used to match the newly input sports motion data in real time, determine the matching degree and output the deviation of the motion pattern, and obtain the analysis results after deviation correction.

[0025] Extract the shared parts across projects from the analysis results after bias correction, and use a convolutional neural network to optimize them again to enhance the accuracy of dynamic capture. Determine whether the optimized results meet the modular requirements. If they do, generate the final general analysis template.

[0026] Based on the generated final universal analysis template, data from various projects on the integrated physical education teaching platform are applied in batches to obtain a unified motion evaluation report.

[0027] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an intelligent correction method for sports teaching movements based on spatiotemporal graph convolution. Addressing the unique business scenario where commonalities and individual characteristics of movements coexist across multiple sports, making it difficult to construct a unified intelligent analysis model, this invention achieves unified modeling and personalized correction of movements in various sports such as basketball and football by extracting a universal spatiotemporal relationship vector across sports and dynamically fusing project-specific parameters. Specifically, initial spatiotemporal segmentation is performed from the original multi-sports movement video sequences. A convolutional neural network is used to capture dynamic changes in key joints to construct a universal spatiotemporal relationship vector. Then, a support vector machine is used for classification and selection of shared feature sets. Subsequently, project-specific indicators, such as arm coordination in basketball or leg strength in football, are adaptively fused according to individual needs to construct an adjusted hybrid feature model. This enables real-time matching and deviation correction of new input movement data. Finally, secondary spatiotemporal graph convolution is used to optimize and extract the shared parts, generating a universal analysis template that meets modular requirements. This allows for batch application of data from multiple sports on a comprehensive sports teaching platform, outputting a unified movement evaluation report. This invention significantly improves the universality, accuracy, and teaching correction efficiency of cross-sports movement analysis, possessing strong practical value and promising prospects for widespread application. Attached Figure Description

[0028] Figure 1 This is a flowchart of a method for intelligent correction of physical education movements based on spatiotemporal graph convolution, according to the present invention.

[0029] Figure 2 This is a schematic diagram of an intelligent correction method for physical education teaching movements based on spatiotemporal graph convolution, according to the present invention.

[0030] Figure 3 This is another schematic diagram of the intelligent correction method for physical education teaching movements based on spatiotemporal graph convolution according to the present invention.

[0031] Figure 4 This is a visualization of the cross-project sharing relationship mechanism based on spatiotemporal graph convolution of this invention.

[0032] Figure 5 This is a visualization of the core mechanism for action deviation correction driven by hybrid features in this invention.

[0033] Figure 6 This is a schematic diagram of the deployment of the sports teaching motion collection site according to the present invention.

[0034] Figure 7 This is a cross-sectional view of the wearable IMU sensor node of the present invention.

[0035] Figure 8 This is a convergence curve of the motion correction accuracy of the present invention.

[0036] Figure 9 This invention provides a heatmap for improving the efficiency of cross-project batch analysis.

[0037] Figure 10 The band diagram shows the contribution of the core mechanism of this invention to performance indicators. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, this method consists of a closed-loop technical process from S101 to S107. First, in S101, raw sequences are obtained from a multi-sports movement video database and initial segmentation is performed. Then, in S102, key joint dynamic features are extracted through spatiotemporal graph convolution to form a cross-sports universal spatiotemporal relationship vector. S103 uses a support vector machine to complete project-specific pattern classification and shared feature selection. S104 combines unique parameters such as basketball arm coordination and soccer leg strength for fusion modeling, resulting in a hybrid feature model. S105 performs real-time matching and deviation correction based on the hybrid feature model. S106 performs secondary optimization of the correction results and determines modularity conditions. Finally, in S107, multi-sports batch application is completed and a unified evaluation report is output, achieving integrated processing of cross-sports movement analysis and teaching correction.

[0040] like Figure 2 As shown, the system consists of a data acquisition layer, a model analysis layer, and an application output layer. The data acquisition layer includes a multi-project video database and edge acquisition devices, responsible for providing multi-source spatiotemporal motion data. The model analysis layer extracts spatiotemporal features of key joints using a spatiotemporal graph convolutional network, then uses a support vector machine to perform project difference classification and shared feature filtering, and combines project-specific indicators to perform weighted fusion, forming a hybrid feature model that can be used for real-time discrimination. The application output layer performs motion matching, deviation correction, and template updating based on this model, and outputs the results as a unified motion evaluation report. This structure achieves synergy between the reuse of common features across projects and personalized correction for specific projects.

[0041] like Figure 6 As shown, in the comprehensive sports teaching venue, high-definition sports cameras, wearable IMU sensors, and edge acquisition gateways are respectively deployed in the basketball teaching area and the football teaching area. After the cameras and sensors synchronously collect the trainees' movements, they are aggregated to the gateway through a wireless link, and then viewed and verified on-site by the teacher's tablet terminal, thus forming a physical collection and deployment foundation for movement data of multiple sports.

[0042] like Figure 3As shown, the deployment of this invention in a comprehensive physical education teaching platform consists of three parts: an edge acquisition side, a cloud analysis platform, and teaching terminals. The edge side collects joint movement and video data via wearable IMU sensors and high-definition cameras, and uploads it to the cloud via a wireless network. The cloud then sequentially performs data cleaning, spatiotemporal graph convolutional analysis, movement evaluation standard comparison, and evaluation report generation. The final results are displayed on both teacher and student terminals, achieving a closed loop of movement anomaly alerts, correction suggestions, and teaching feedback. This deployment method supports concurrent access for multiple projects and unified standard output.

[0043] Specifically, an embodiment of the present invention provides a method for intelligent correction of physical education movements based on spatiotemporal graph convolution, which may include: S101. Obtain raw motion sequence data from a pre-set multi-sports motion video database, perform initial segmentation based on the motion characteristics of different sports such as basketball and football, and obtain preliminary spatiotemporal sequence fragments.

[0044] Based on the aforementioned business content and extracted relevant attributes, the following business solution is generated. The technical process revolves around attributes such as sports movements, video databases, raw sequences, movement characteristics, preliminary segmentation, spatiotemporal segments, basketball, football, data extraction, project differentiation, sequence processing, and initial acquisition. Attributes that cannot form a logical connection are discarded. The solution is implemented using the following steps: From a pre-defined video database, raw sequence data related to sports movements is extracted through a specific data interface. This data is then categorized and stored according to different projects, resulting in a pre-organized set of movement data. For this pre-organized set, pre-defined rules are used to differentiate the movement characteristics of basketball and football. If a specific movement pattern is detected in the data, it is categorized into the corresponding project group, thus determining the categorized movement data groups. Based on these categorized movement data groups, the raw sequence data within each group is obtained. Preliminary segmentation is performed using timestamps and spatial location information, resulting in a set of segmented sequence segments. For these segmented sequence segments, a pre-defined movement characteristic template is compared. If a sequence segment matches the template with a pre-defined threshold, it is marked as a valid segment, thus identifying a valid set of spatiotemporal segments. Based on a valid set of spatiotemporal segments, the sequence processing results for each segment are obtained, and a unified formatted storage method is used to determine the processed spatiotemporal segment data. The integrity of the processed spatiotemporal segment data is checked using data verification tools. If missing or abnormal data is found, a supplementary extraction process is triggered to obtain a complete and standardized spatiotemporal segment dataset.

[0045] The specific data interface described in this application is a custom video frame extraction interface developed based on OpenCV. It supports mainstream video formats such as AVI, MP4, and MOV, and can extract data such as continuous frame images, frame timestamps, and joint coordinates from videos. The data interface adopts a RESTful API architecture and supports both batch data extraction and real-time data extraction modes.

[0046] The data validation and data cleaning tools described in this application are all developed based on Python Pandas and NumPy, and are custom automation scripts. Missing value imputation in data cleaning adopts the linear interpolation method between consecutive frames to impute continuous data such as timestamps and joint coordinates. Outlier judgment in data validation adopts the 3σ principle, and data exceeding 3 times the standard deviation is judged as outlier data. Data formatting and storage adopts JSON structure, and each record contains core fields such as fragment ID, time range, feature vector, and project identifier.

[0047] In this application, the preset threshold for the matching degree between action segments and templates is set to 0.8. This threshold is set based on the feature similarity statistics of 1,000 sets of standard action samples for basketball and football, which can effectively filter out noisy segments. The preset threshold for data integrity is set to 95%. That is, when the integrity rate of spatiotemporal segment data is lower than 95%, the supplementary extraction process is triggered. This threshold is set in combination with the actual situation of sports action video data collection, which can take into account both data integrity and processing efficiency.

[0048] For example, in the process of extracting raw sequence data related to sports movements from a pre-set video database.

[0049] Understandably, this video database is typically a large storage system containing recordings of various sporting events, such as basketball games or the World Cup football matches. These files are saved in high-resolution format and contain time-series frames and motion capture data. The extraction is performed through a specific data interface, such as an API-based query interface. Users input keywords like "basketball shot" or "football shot," and the system automatically scans the database, filtering out matching video clips to form raw sequence data. This sequence data includes consecutive image frame sequences, each frame containing a timestamp and spatial coordinates, thus initially organizing it into a motion data set. This data is then stored according to sport categories; for example, basketball data is stored in one subdirectory, and football data in another, ensuring efficient subsequent processing.

[0050] In one possible implementation, when differentiating the movement characteristics of basketball and football based on a pre-compiled set of movement data.

[0051] It should be noted that the preset rules can be based on action pattern recognition. For example, basketball action characteristics often involve high jumps and arm swings, while soccer action characteristics emphasize foot kicks and fast running.

[0052] Specifically, if the detected data contains a pattern of arm raising combined with the ball's trajectory, it is classified into the basketball group; conversely, if the sequence shows a vector of lower limb swing and goal direction, it is classified into the soccer group. This differentiation process is achieved through machine learning algorithms, such as using convolutional neural networks to analyze sequence frames, calculate pattern matching scores, and determine the classified motion data groups, thus providing an accurate basis for subsequent segmentation.

[0053] For example, when obtaining the original sequence data within each group based on the classified action data, the process involves using timestamps and spatial location information for initial segmentation.

[0054] Understandably, a timestamp refers to the precise moment of a video frame, such as a marker for 30 frames per second, while spatial location information includes the coordinates of the subject in the frame, such as the position of a player's joints.

[0055] In one possible implementation, the segmentation method divides the sequence into small segments, such as from the start time stamp to the end time stamp of a shooting action, and detects boundary points based on spatial position changes to obtain a set of segmented sequence fragments, which helps to isolate independent action units.

[0056] In one possible implementation, when matching a set of segmented sequence fragments by comparing them with a preset action feature template.

[0057] Specifically, the template is a standardized library of motion models, such as a basketball "dunk template" which includes jump height and speed thresholds. If the similarity between the feature vector of a sequence segment and the template reaches a preset threshold, such as 0.8, it is marked as a valid segment, thus identifying a valid set of spatiotemporal segments. This matching uses cosine similarity calculation, and the process involves extracting keyframe features from the segment, such as pose estimation points, and then comparing them with the template to filter out noisy segments and improve the purity of the data.

[0058] For example, when obtaining the sequence processing results of each segment based on a valid set of spatiotemporal segments, a uniform formatted storage method is used to determine the processed spatiotemporal segment data.

[0059] Understandably, formatting involves converting data into a JSON structure, with each record containing a fragment ID, time range, and feature vector to ensure compatibility.

[0060] In one possible implementation, when the processed spatiotemporal segment data is checked for completeness using a data validation tool, if any missing information, such as a missing timestamp, is detected, a supplementary extraction process is triggered to retrieve data from the original database, resulting in a complete and standardized spatiotemporal segment dataset. This validation tool is an automated script that scans the dataset for completeness; if it falls below 95%, the supplementary mechanism is activated, ultimately improving the reliability of the dataset and providing high-quality input for sports analytics applications.

[0061] S102. Use a convolutional neural network to extract features from the obtained spatiotemporal sequence segments, capture dynamic changes for key joint nodes in the action, and determine the general spatiotemporal relationship vector across projects.

[0062] A convolutional neural network (CNN) is used to process spatiotemporal sequence segments to obtain multi-channel feature maps. The positions of key joint nodes are analyzed using these feature maps to obtain the coordinate sequences of each node in the time dimension. The displacement vectors of joint nodes between adjacent frames are calculated based on the coordinate sequences to obtain the joint motion trajectory sequences. Temporal features are extracted from the joint motion trajectory sequences to determine the motion velocity sequence of each key joint node. Based on the correspondence between the motion velocity sequences and spatial positions, the relative distance sequences between key joint nodes are calculated. Spatial dimension features between joints are constructed using the relative distance sequences to obtain the spatial adjacency strength between joints. A CNN is then used to process the joint input of the spatial adjacency strength between joints and the motion velocity sequences to obtain intermediate features that fuse spatiotemporal information. The temporal dependency strength between each joint node is calculated based on the intermediate features to determine a spatiotemporal relationship vector applicable across the project.

[0063] In this application, the spatial adjacency strength between joints is quantified by the mean and variance of the relative distance sequence. First, the mean and variance of the relative distance sequence between key joint nodes are calculated, and then normalized to obtain a spatial adjacency strength value with a value range of [0,1]. The temporal dependency strength is quantified by an attention mechanism. The intermediate features that fuse spatiotemporal information are used as attention input, and after the attention weight is calculated, a temporal dependency strength value with a value range of [0,1] is obtained. The spatiotemporal relationship vector that is universal across projects is a 128-dimensional feature vector, which contains 64-dimensional spatial features and 64-dimensional temporal features, and is a feature embedding in a high-dimensional feature space.

[0064] Furthermore, to implement spatiotemporal graph convolution operations, the spatiotemporal sequence segments are constructed internally as a human skeleton topological graph structure. In this structure, key joint nodes of the human skeleton (such as shoulders, elbows, and knees) are defined as nodes in the graph, while skeletal connections within a single frame that conform to human physiological connections are defined as spatial edges, and connections between the same joint nodes in consecutive frames are defined as temporal edges. Based on this, the convolutional neural network is specifically configured as a spatiotemporal graph convolutional network, where the convolutional kernels no longer slide on a regular image grid, but rather aggregate features within the neighborhood of nodes according to the connection relationships of the aforementioned topological graph structure. This processing method can accurately capture the evolution of actions in the temporal dimension while preserving the spatial structural information of the human body.

[0065] The specific configuration of the spatiotemporal graph convolutional network in this application is as follows: the network contains 4 spatiotemporal convolutional layers, 2 pooling layers, and 1 fully connected layer; the kernel size of the spatiotemporal convolutional layers is 3×3, and the number of convolutional kernels is 64, 128, 256, and 256 respectively; the pooling layer uses max pooling with a kernel size of 2×2; the activation function is the ReLU function; the learning rate of the network training is set to 0.001, the number of iterations is set to 500 rounds, and the batch size is set to 32; the key joint nodes of the human skeleton topology graph are set to 17, which are the core joints of the human body such as the shoulder, elbow, wrist, hip, knee, ankle, and trunk.

[0066] For example, when processing spatiotemporal sequence segments using convolutional neural networks (CNNs), we can first use spatiotemporal sequence segments extracted from basketball shooting videos as input. These segments contain consecutive frames of the player's arms and legs. CNNs capture edge and texture features in the images through multiple convolutional operations. For instance, the first layer of the network uses 3x3 convolutional kernels to scan the sequence segments, progressively extracting low-level features such as joint contours. Then, pooling layers reduce dimensionality, ultimately generating multi-channel feature maps. These feature maps represent the spatial distribution of the action at different time points, thus providing fundamental data support for subsequent analysis.

[0067] like Figure 4 As shown, the left side constructs a skeleton topological coupling graph with 17 key joint nodes. The grayscale and line width of the edges represent the spatial adjacency strength, and the node size represents the temporal dependency strength. The right side presents the corresponding shared spatiotemporal relationship vector spectrum matrix, which is used to demonstrate the reusable relationship of basketball and football actions in the key joint coordination mode. This figure intuitively reflects the extraction results of spatiotemporal graph convolution on cross-project general mechanisms, providing a structured foundation for subsequent feature fusion and action matching.

[0068] In one possible implementation, when analyzing the location of key joint nodes through feature map analysis, the locations of key nodes such as the knee, ankle, and hip joints can be identified based on the feature map of a soccer shooting motion.

[0069] Specifically, the system uses pre-trained pose estimation models such as OpenPose to assist in analysis, calculating the pixel coordinates of each node from the feature map. For example, in frame 1, the knee node is located at coordinates (120, 200), while in frame 2 it moves to (130, 210). This frame-by-frame acquisition of coordinate sequences over time ensures that the sequence captures the dynamic changes of the movement. This method helps to accurately track the evolution of joints in space and time, providing reliable coordinate data for motion analysis.

[0070] For example, when calculating the displacement vector of joint nodes between adjacent frames based on a coordinate sequence, in a basketball dribbling sequence, the vector can be obtained by taking the coordinate difference between two adjacent frames. For instance, the displacement vector of the knee node from (100, 150) to (105, 155) is (5, 5). Repeating this process forms the entire joint motion trajectory sequence. This calculation reveals the direction and amplitude of joint movement, which helps in understanding the smoothness of the movement.

[0071] In one possible implementation, when extracting time-dimensional features from the joint motion trajectory sequence, for a soccer passing action, the magnitude of the displacement vector can be calculated and divided by the inter-frame time interval to obtain the motion velocity sequence of each key joint node. For example, the velocity sequence of the ankle joint shows the process of accelerating from 0.5 m / s to 2 m / s, thereby quantifying the rhythmic changes of the action.

[0072] For example, when calculating the relative distance sequence between key joint nodes based on the correspondence between motion speed sequence and spatial position, in a basketball dunking action, the Euclidean distance formula can be used to calculate the distance between the wrist and shoulder. For example, the distance is 0.8 meters in the peak speed frame and 1.2 meters in the low speed frame, forming a sequence to reflect the dynamic interaction between joints.

[0073] In one possible implementation, when constructing the spatial dimensional features between joints through a relative distance sequence, for a soccer heading action, the statistical values ​​of the distance sequence, such as the mean and variance, can be used to obtain the spatial adjacency strength between joints. For example, the strength value of the knee and hip joints is 0.6, indicating close cooperation, thereby quantifying the spatial relationship.

[0074] For example, when using a convolutional neural network to process the joint input of inter-joint spatial adjacency strength and motion velocity sequence, in the analysis of basketball defensive actions, these data are concatenated into an input tensor, and intermediate features are extracted through the network's fusion layer. These features integrate spatiotemporal information and improve the accuracy of action recognition.

[0075] In one possible implementation, when calculating the temporal dependency strength between joint nodes based on intermediate features, for general analysis across sports such as basketball and football, an attention mechanism can be used to evaluate dependencies. For example, the dependency strength between the wrist and ankle is calculated to be 0.7, forming a spatiotemporal relationship vector. This vector supports unified modeling of movements across multiple sports, enabling more efficient sports data processing.

[0076] Specifically, the cross-sports universal spatiotemporal relationship vector is essentially an embedding of basic kinematic features after removing the specific movement amplitudes of each sport (such as the large leg swing in soccer or the high arm raise in basketball). By setting shared weight layers in the network layers, joint movements of different sports are mapped to the same high-dimensional feature space, extracting features reflecting the commonalities of human biomechanics, such as the temporal transmission rate of core muscle groups driving limb force, the relative coordination stability between joints, and the rate of change of acceleration at the extremities. These features do not depend on specific ball game rules but on human anatomical constraints, thus possessing cross-sports universality.

[0077] S103. Based on the determined spatiotemporal relationship vectors, the personalized action patterns of different items are classified using a support vector machine. If the classification results show that the feature similarity is higher than a preset threshold, these vectors are included in the shared feature library to obtain a general feature set after classification.

[0078] Step 1: Obtain spatiotemporal relationship vectors from the raw data, and perform preliminary organization of action patterns for different items to obtain an initial vector set. Step 2: Use a support vector machine to classify the initial vector set, and compare the feature similarity in the classification results to determine the vector groups with high similarity. Step 3: If the number of vector groups with high similarity exceeds a preset threshold, these vectors are added to a shared feature library to obtain a preliminary shared feature set. Step 4: Perform secondary analysis on the preliminary shared feature set to extract feature elements with strong generality and determine whether they meet the criteria of a general set. Step 5: Integrate the extracted feature elements according to the criteria of a general set to obtain the final general feature set. Step 6: For the final general feature set, record the corresponding item differentiation information and determine the feature mapping relationship after classification. Step 7: Apply the general feature set to subsequent action pattern analysis through the feature mapping relationship to obtain more accurate personalized classification basis.

[0079] In this application, the support vector machine used for action pattern classification is a classification SVM. The kernel function is a radial basis function (RBF) kernel, the penalty factor C is set to 1.0, and the kernel function parameter γ is set to 0.1. The feature similarity is calculated using cosine similarity, and the degree of feature similarity is quantified by the cosine angle between vectors.

[0080] When extracting spatiotemporal relationship vectors from raw motion data, preliminary sorting is usually performed for different sports such as basketball, football, and gymnastics.

[0081] In one embodiment, for a layup in basketball, the relative positions of the wrist and shoulder at the moment of release can be separated from the three-dimensional coordinate sequence of the joints to form an initial vector set; for a free kick in soccer, the focus is on organizing the chain motion vectors from the hip joint to the ankle; and for the landing motion in gymnastics, the focus is on organizing the coordination vectors of the torso and limbs. In this way, the initial vector sets retain the movement pattern information unique to each sport. Next, a support vector machine is used to classify these initial vector sets.

[0082] Specifically, the vectors of actions such as basketball layups, soccer shots, and gymnastic somersaults are labeled with different categories and input into a support vector machine (SVM) model for training and prediction. In the classification results, feature similarity is compared by calculating the cosine similarity or Euclidean distance between vectors. For example, the similarity between the shoulder-to-hand motion trajectory vectors of a basketball dunk and a volleyball spike might reach 0.87, while the similarity with the leg vector of a soccer shot is only 0.32, thus identifying vector groups with higher similarity. When the similarity of vector groups exceeds a preset threshold of 0.80, these vectors are included in a shared feature library, forming a preliminary shared feature set. For example, the "shoulder-elbow-wrist" rapid extension pattern common in basketball jump shots and volleyball spikes, as well as the "hip-knee-ankle" explosive swing pattern in soccer volleys and taekwondo side kicks, are included in this set for subsequent feature reuse across different sports. In the secondary analysis of the preliminary shared feature set, feature elements that appear frequently across multiple sports and have low variability are mainly extracted. For example, the temporal pattern of "core trunk exerting force before limbs" is observed in landings in basketball, volleyball, and gymnastics. Statistical analysis determined that this pattern conforms to the criteria of a general set, while the "toe-pointing" detail specific to certain sports was removed. Based on the criteria of the general set, the extracted feature elements are integrated to obtain a cross-sports-compatible feature set. For instance, three high-frequency common elements—"trunk rotation leading to limb swing," "rapid transfer of weight to the supporting leg," and "delayed force exertion at distal joints"—are integrated into a unified descriptive framework, forming a transferable general feature set. For the final general feature set, it is necessary to record the sports-specific information corresponding to each feature to determine the feature mapping relationship after classification.

[0083] One possible implementation involves creating a mapping table where "torso rotation angle exceeding 45 degrees" is mapped to offensive movements in basketball and volleyball, while "hip joint twisting exceeding 60 degrees" is more often mapped to kicking movements in soccer and martial arts. Through this mapping relationship, subsequent motion pattern analysis can quickly distinguish between different categories while retaining the recognition basis provided by common features, thus obtaining more accurate personalized classification criteria.

[0084] S104. Obtain the set of general features after classification, and adjust them in the multi-project analysis platform according to the remaining personalized needs. Determine whether to add project-specific parameters. If so, integrate specific indicators of basketball arm coordination or football leg strength to obtain the adjusted hybrid feature model.

[0085] The process begins by acquiring a set of general features after classification. Data is read from a pre-established feature library, and format matching is performed on the data structure within the multi-project analysis platform to obtain a preliminary feature dataset. Based on this dataset, the matching degree of personalized needs is analyzed. If the matching degree between a specific project's personalized needs and the general features is below a preset threshold, a project parameter filtering process is triggered to determine the specific parameter types to be incorporated. Using the filtered specific parameter types, specific indicator data related to basketball arm coordination or soccer leg strength are obtained, and corresponding quantified values ​​are extracted from a pre-established indicator library to obtain a specific indicator dataset. For the specific indicator dataset and the general feature dataset, a support vector machine algorithm is used for feature fusion processing. Vector mapping and weight allocation are performed on the two types of data to construct a preliminary hybrid feature model. Based on this preliminary hybrid feature model, the distribution characteristics of multi-project data in the platform's analysis module are analyzed. If a deviation is found in the fusion results of specific indicators and general features, weight adjustments are made to obtain an optimized feature combination. Finally, the optimized feature combination is adapted to the application scenarios of the multi-project analysis platform, embedding the feature combination into the platform's data processing flow to determine if the final hybrid feature model meets the application requirements. Based on the final hybrid feature model, feature application templates suitable for multi-project analysis are generated. Key parameters are extracted from the templates to determine cross-project feature application schemes.

[0086] In this application, the support vector machine used for feature fusion is a binary classification SVM, the kernel function is a radial basis function (RBF) kernel, the penalty factor C is set to 0.8, and the kernel function parameter γ is set to 0.05; the standard / non-standard of the action is used as the classification label, and the feature importance is evaluated through the hyperplane normal vector obtained by training.

[0087] The feature fusion process here does not directly use Support Vector Machines (SVMs) for simple classification. Instead, it uses the hyperplane normal vectors obtained from SVM training as the evaluation criterion for feature importance. Specifically, a specific indicator dataset and a general feature dataset are concatenated as input to train a binary classification SVM model to distinguish between standard and non-standard actions. After training, the weight coefficients (i.e., the absolute values ​​of the normal vector components) corresponding to each feature dimension are extracted. Using these weight coefficients, the original specific indicator features and general features are weighted and summed or recombined to construct a hybrid feature model that includes both commonalities and highlights project characteristics. This approach effectively solves the feature dimension redundancy problem caused by direct concatenation.

[0088] In this application, the weight allocation method for feature fusion is as follows: first, the absolute values ​​of the components of the SVM hyperplane normal vector are extracted as the weight coefficients of each feature dimension; then, each feature dimension of the general feature dataset and the specific index dataset is multiplied by the corresponding weight coefficients; finally, the weighted two types of features are summed dimension by dimension to obtain the fused hybrid features; among them, the arm coordination-specific parameter weight ratio of basketball is 0.6, and the general feature weight ratio is 0.4; the leg strength-specific parameter weight ratio of soccer is 0.6, and the general feature weight ratio is 0.4. The weight ratios can be adaptively adjusted according to the personalized needs of different sports.

[0089] After obtaining the general feature set after classification, the data reading operation is first completed through the pre-established feature library.

[0090] Specifically, the system will perform a format matching check on each record in the common feature set according to the data structure definition of the multi-project analysis platform.

[0091] For example, in sports training scenarios, the platform may manage data for basketball, football and swimming at the same time. The feature library will pre-store common indicators across sports, such as heart rate change curves and peak acceleration. When reading, these common features will be automatically mapped to a unified time series format to obtain a preliminary feature dataset.

[0092] In one possible implementation, based on a preliminary feature dataset, the system calculates the degree of matching between the personalized requirements of each specific project and the general features.

[0093] It should be noted that the degree of matching is usually quantified by the cosine of the vector angle or the normalized form of the Euclidean distance. For example, for basketball, general features may include upper limb explosive power and core stability indicators, while the personalized needs of basketball players focus more on arm coordination and quick change of direction. If the calculated matching degree is lower than the preset threshold of 0.75, the project parameter filtering process will be automatically triggered.

[0094] Preferably, the screening process prioritizes identifying unique parameter types that differ significantly from general characteristics. For example, in a basketball scenario, the system might screen for arm coordination indicators, including the range of shoulder joint rotation angles, peak wrist rotation speed, and micro-vibration frequency when the hand is controlling the ball; in a soccer scenario, it might focus on leg strength-related indicators, such as the quadriceps activation timing, ankle extension angle, and peak ground reaction force. These quantified values ​​are then extracted from a pre-defined indicator library to form a specific indicator dataset. For instance, basketball arm coordination might extract an average phase difference of 35 degrees between arm swing and torso twist, while soccer leg strength might extract specific values ​​such as the maximum extension force reaching 3.2 times body weight.

[0095] Understandably, after obtaining the specific indicator dataset, the system uses the support vector machine algorithm for feature fusion processing. Specifically, the general feature vector is concatenated with the specific indicator vector, mapped to a high-dimensional space, and a classification hyperplane with the largest margin is calculated using a kernel function. Simultaneously, different weights are assigned based on the contribution of each feature to the classification boundary, constructing a preliminary hybrid feature model. For example, in basketball, the weight of the general feature of trunk stability might be 0.4, while the weight of the arm coordination-specific parameter is adjusted to 0.6 to highlight individual differences.

[0096] In one embodiment, based on the preliminary hybrid feature model, the platform analysis module further examines the distribution characteristics of the multi-project data after fusion. If it is found that the leg strength index of the football project deviates significantly from the general features after fusion, for example, the fusion vector of some forward samples is significantly biased towards the high explosive power area, the system will make targeted adjustments to the weights of such samples, usually by increasing the penalty parameter C or optimizing the selection range of support vectors, to obtain an optimized feature combination.

[0097] For example, after weight adjustment, the optimized feature combination can better adapt to the application scenarios of multi-project analysis platforms. For instance, the fused features can be embedded into the real-time training monitoring process to determine whether an athlete's current movement deviates from the optimal pattern. Finally, based on the optimization results, the system generates feature application templates suitable for multi-project analysis. Key parameters extracted from the templates include the dimension of the fused vector, the principal component contribution rate threshold, and the weight allocation ratio for each project, thereby determining the cross-project feature application scheme.

[0098] S105. Using the adjusted hybrid feature model, the newly input sports motion data is matched in real time to determine the matching degree and output the deviation of the motion pattern, thus obtaining the analysis results after deviation correction.

[0099] Acquire new input sports motion data. Perform real-time matching of the sports motion data using an adjusted hybrid feature model to obtain a matching degree value. Compare the matching degree value with a preset threshold; if the matching degree is lower than the preset threshold, a motion pattern deviation is identified, and the motion pattern deviation value is obtained. Perform deviation correction processing on the original matching sequence using the motion pattern deviation value to obtain a corrected matching sequence. Recalculate the corrected matching degree for the corrected matching sequence using the hybrid feature model to obtain a corrected matching degree value. Determine whether the current motion meets the analysis conditions based on the corrected matching degree value; if the corrected matching degree reaches the preset standard, output the final analysis result.

[0100] In this application, the preset threshold for the matching degree between sports movement data and the hybrid feature model is set to 0.9, and the preset standard for correcting the matching degree is not less than 0.9. This threshold is set based on the evaluation criteria for movement standardization in sports teaching, which can effectively distinguish between standard and non-standard movements.

[0101] For example, in a sports training system, the first step is to acquire new input sports movement data, which can be understood as collecting the athlete's limb movement information in real time through sensor devices.

[0102] Specifically, when a basketball player is practicing shooting, the system uses wearable sensors installed on the arm and torso to capture data, including quantitative indicators such as arm swing angle, speed, and acceleration. This data is wirelessly transmitted to the analysis platform to ensure the real-time nature and accuracy of the data, thus providing a basis for subsequent matching.

[0103] like Figure 7 As shown, the wearable IMU node consists of an elastic strap, a shell cover, an IMU inertial measurement chip, a microcontroller, a battery unit, a Bluetooth / 5G antenna, a charging interface, and a skin-friendly cushioning layer. It can stably fit the limbs and continuously output joint motion data during sports teaching and training, providing high-quality raw input for subsequent spatiotemporal graph convolutional analysis.

[0104] like Figure 5 As shown, the left side uses matching degree terrain and correction trajectory to illustrate the process of motion state converging from a low matching region to a high matching threshold region, while the right side uses contribution migration phase diagram to illustrate the collaborative mechanism of gradually increasing contributions from general features and gradually converging contributions from project-specific parameters. This figure demonstrates that the present invention can achieve personalized motion correction under the premise of unified modeling across projects and stably output high matching degree analysis results.

[0105] In one possible implementation, these sports motion data are matched in real time using an adjusted hybrid feature model to obtain a matching degree value. Here, the hybrid feature model is a computational framework that integrates general motion features and project-specific indicators.

[0106] For example, it combines parameters such as arm coordination and leg strength to assess the standard of a movement. In the basketball shooting example, the model compares the collected arm angle data with a preset standard shooting template, and generates a matching score by calculating vector similarity. For example, if the standard template requires an arm angle of 45 degrees, while the actual data is 40 degrees, the matching score may be 0.85. This value reflects how close the movement is to the ideal state.

[0107] For example, the matching score is then compared to a preset threshold. If the matching score is lower than the preset threshold, a deviation in the movement pattern is identified, and the movement pattern deviation value is calculated. The preset threshold can be set to 0.9, meaning that the movement needs to achieve 90% similarity to be considered acceptable. In soccer kicking training, if the matching score is 0.75, which is lower than the threshold, the system will analyze the source of the deviation, such as trajectory deviation caused by insufficient leg strength, and calculate a deviation value of 0.15. This value quantifies the difference between the movement and the standard pattern, helping coaches identify problems.

[0108] In one possible implementation, the original matching sequence is corrected using the deviation value of the movement pattern to obtain a corrected matching sequence. This correction involves adjusting the data sequence, for example, using the deviation value as a correction factor to weight and compensate each data point in the original sequence. In the basketball example, if the deviation value is 0.1, the system will adjust the original arm angle sequence from [40,42,45] to [41,43,46], thereby generating a sequence closer to the standard and improving the reliability of the analysis.

[0109] For example, the corrected matching sequence is recalculated using a hybrid feature model to obtain the corrected matching degree value. This step ensures the effectiveness of the correction. In the football scenario, after the corrected sequence is input into the model, a corrected matching degree of 0.92 may be obtained, indicating that the deviation has been effectively compensated.

[0110] In one possible implementation, the current action is judged to meet the analysis conditions based on the corrected matching degree value. If the corrected matching degree reaches the preset standard, the final analysis result is output.

[0111] For example, if the standard is 0.9 and the correction value is 0.92, the system output will be something like "The movement is qualified, it is recommended to strengthen strength training", which can improve training efficiency in multi-project platforms.

[0112] like Figure 8As shown, with the increase in training rounds, the accuracy of motion correction in basketball and soccer scenarios using the method of this invention continuously improves and tends to stabilize, ultimately reaching approximately 92.5% and 93.2%, respectively. Compared with traditional single-item analysis methods, there is a significant difference in accuracy after convergence, demonstrating the effectiveness of the fusion strategy of cross-item general features and item-specific parameters in improving recognition accuracy. This result verifies the high-precision correction capability of this method on complex teaching motion data.

[0113] S106. Extract the cross-project shared parts from the analysis results after deviation correction, and use a convolutional neural network to optimize again to enhance the dynamic capture accuracy. Determine whether the optimized results meet the modular requirements. If they do, generate the final general analysis template.

[0114] Obtain the analysis results after deviation correction. Perform secondary optimization processing on the deviation-corrected analysis results using a convolutional neural network to obtain the optimized results. Extract cross-project shared content from the optimized results to obtain cross-project shared content. Compare the cross-project shared content with preset modularization conditions. If the cross-project shared content meets the preset modularization conditions, the current optimized results are determined to meet the modularization requirements. Generate a general analysis template using the optimized results that meet the modularization requirements to obtain the initial draft of the general analysis template. Perform structural adjustment processing on the initial draft of the general analysis template using a convolutional neural network to obtain the structurally adjusted general analysis template. Perform a final confirmation operation on the structurally adjusted general analysis template to obtain the final general analysis template.

[0115] In one possible implementation, the bias-corrected analysis results can be obtained by extracting the processed motion data from the sports training system.

[0116] For example, in basketball shooting training, the system first collects data on the athlete's shooting posture. After deviation correction, standardized shooting trajectory analysis results are obtained. These results include corrected values ​​for arm angle, ball speed, and landing point, ensuring that the data reflects the true state of optimized movement. In this way, the analysis results become a reliable foundation for subsequent processing, connecting the entire optimization process.

[0117] For example, when performing secondary optimization on the analysis results after bias correction, convolutional neural networks (CNNs) can be used to capture spatial features in the data. A CNN is a deep learning model that extracts local patterns from image or sequence data through multiple convolutional and pooling layers. In sports motion analysis, CNNs can treat the corrected motion sequence as a temporal image and perform layer-by-layer convolution operations to identify hidden pattern biases.

[0118] Specifically, in the context of soccer shooting training, the input is corrected leg swing data, which is then scanned by the sliding kernel of a CNN to calculate feature maps. The nonlinear representation is then enhanced using activation functions such as ReLU, ultimately outputting optimized results, such as a more accurate shot force distribution map. This process helps improve the accuracy of the analysis.

[0119] The subsequent optimization and template generation process is essentially a reverse inductive process from specific samples to abstract standards. The convolutional neural network here acts as a discriminator or the encoder end of a generative adversarial network (GAN) to evaluate whether the features extracted from the bias correction results are statistically representative. If the optimized results exhibit Gaussian clustering characteristics in the feature space and the dispersion is below a preset threshold, then the modularity requirement is met. The resulting "final general analysis template" is not an image, but a set of standardized parameter matrices containing the ideal spatial coordinate range of each key joint within a standard motion cycle, the velocity threshold range, and the standard phase difference data between each joint.

[0120] In this application, the pre-defined modular conditions include three core requirements: 1. Feature reusability is no less than 90%, meaning that the template features can be used for motion analysis in at least 90% of sports; 2. Module interface is standardized, using a unified JSON data format for feature input and output; 3. Feature dispersion is less than 0.15, meaning that the features exhibit good clustering characteristics in high-dimensional space. The parameter matrix of the final general analysis template is a 17×6 two-dimensional matrix. The 17 rows correspond to the 17 key joint nodes of the human body, and the 6 columns are, in order, the minimum and maximum values ​​of the ideal space coordinates of the joint, the minimum and maximum values ​​of the movement velocity, the minimum and maximum values ​​of the phase difference between the joints, respectively. The values ​​of the parameter matrix are set based on the statistical results of 2000 sets of standard motion samples each for basketball and football, serving as standard reference values ​​for sports motion analysis.

[0121] In this application, the preset threshold for the dispersion of the feature space distribution is set to 0.15. This threshold is set based on the statistical distribution characteristics of cross-project general features. A value lower than this indicates that the features have good clustering and generality.

[0122] In one possible implementation, the process of extracting cross-project shared content from the optimized results involves identifying common elements between different sports.

[0123] For example, shared vertical take-off height and landing cushioning patterns can be extracted from the optimization results of jumping movements in basketball and volleyball. These can be obtained through feature vector comparison algorithms, ensuring that shared content across sports, such as joint pressure distribution, is uniformly summarized, thus providing a reference for multi-sports training.

[0124] For example, when comparing shared content across projects with preset modular conditions, if the shared content, such as the motion coordination index, exceeds the threshold of 0.8, it is determined that it meets the modular requirements.

[0125] Specifically, the pre-defined modular conditions may include the reusability and standardization of the content. In tennis and badminton swing analysis, if the shared swing speed patterns meet the conditions, it can be confirmed that the optimization results are applicable to the modular framework, which lays the foundation for subsequent template generation.

[0126] In one possible implementation, a draft of a general analysis template is generated from the optimized results that meet the modular requirements, and the shared content can be integrated through a template building tool.

[0127] For example, in cross-training for track and field and swimming events, optimized data on long jump take-off and breaststroke stroke are integrated into a template draft, including standard parameters such as the timing of force exertion and energy consumption models, to ensure that the initial draft is universal.

[0128] For example, when performing structural adjustments on a general analysis template draft using a convolutional neural network (CNN), the CNN intervenes again to optimize the template layout. The principle is to adjust the structural weights using the network's feedback mechanism. In a gymnastics tumbling training template, the initial draft data is input, and the CNN refines the module connections through backpropagation, resulting in a structurally adjusted template, such as a more compact motion sequence diagram. This adjustment enhances the template's practicality and leads to more efficient cross-project applications in business.

[0129] In one possible implementation, a final confirmation operation is performed on the general analysis template after structural adjustments, and consistency can be checked through a verification algorithm.

[0130] For example, in mixed martial arts training, it is confirmed that the striking and defensive modules of the template are conflict-free, and a final universal analysis template is output for practical guidance of training for athletes in multiple sports.

[0131] like Figure 10 As shown in the figure, the contribution distribution of the four core mechanisms—"spatiotemporal relationship extraction, personalized parameter fusion, deviation correction, and templated output"—to the three performance indicators—"accuracy improvement," "cross-project efficiency improvement," and "template reusability improvement"—is depicted in a stream-like manner. It is evident that spatiotemporal relationship extraction and deviation correction contribute significantly to accuracy improvement, deviation correction and parameter fusion contribute significantly to efficiency improvement, while templated output makes the most prominent contribution to reusability improvement, reflecting the inherent coupling path of this invention from mechanism design to performance achievement.

[0132] S107. Based on the generated final general analysis template, apply the data from various projects on the comprehensive sports teaching platform in batches to obtain a unified motion evaluation report.

[0133] Raw data from various projects is obtained from a comprehensive platform using a pre-established general template. Data cleaning tools are used to standardize the format and imput missing values, resulting in a processed basic dataset. For this processed dataset, a classification method is applied to group the data according to different projects in physical education. A simple grouping logic is used: if a data item's project identifier matches a preset classification condition, it is assigned to the corresponding project group, thus determining the grouped project dataset. Based on the grouped project datasets, a preset motion evaluation standard library is invoked to extract evaluation indicators relevant to each project. For each project group, the corresponding standard parameters are obtained, resulting in a matched evaluation basis set. Using this matched evaluation basis set, the data for each project group is compared and analyzed item by item. Logical judgment is used: if a data value deviates from the standard parameter range, it is marked as an outlier, resulting in an analysis result set containing anomaly markers. Based on the analysis result set containing anomaly markers, motion evaluation details for each project are generated. A data aggregation tool is used to categorize outliers and normal data separately, obtaining a categorized evaluation detail set. Using the categorized evaluation detail set and adhering to the format requirements of a unified report, the data content is automatically populated. A template rendering method is used to map assessment details into the report structure, determining the final action assessment document. Based on the final action assessment document, a data storage and distribution process is executed. The document content is saved to the database of the integrated platform and simultaneously pushed to relevant teaching modules via an interface, completing the data flow.

[0134] Raw data from various projects can be obtained from the integrated platform using pre-established generic templates.

[0135] In one embodiment, the general template predefines field mapping rules, such as mapping basketball shooting video data, badminton hitting trajectory data, and football running path data to the same structured table, so that the original records of different sports can be read by the platform in a consistent manner.

[0136] Specifically, the integrated platform periodically pulls the latest data packages from the folders uploaded by each teaching terminal. The template is responsible for verifying the data source identifier and initially archiving the data. Data cleaning tools are used to standardize the format of the acquired data and impute missing values, resulting in the processed basic dataset.

[0137] For example, when the timestamp format of a badminton swing data is "2025-10-15 14:30" while that of a basketball shooting data is "20251015143000", the cleaning tool will automatically convert it to the unified ISO 8601 standard format. For missing frames caused by device disconnection for some students, the tool preferentially uses linear interpolation between preceding and following frames to fill in the swing angle or running speed values, thus forming a continuous and analyzable basic dataset. For the processed basic dataset, a classification method is applied to group the data according to different sports in physical education teaching.

[0138] In one possible implementation, the system reads the item identifier field of each record. If the identifier is "BASKETBALL_SHOT", it is classified into the basketball group; "BADMINTON_SMASH" into the badminton group; and "SOCCER_DRIBBLE" into the soccer group. After this simple logic, independent datasets for basketball shooting, badminton smashing, and soccer dribbling can be obtained. Based on the grouped item datasets, a preset motion evaluation standard library is called to extract evaluation indicators related to each item.

[0139] For example, the basketball group extracts three core indicators—release angle, release speed, and trajectory height—from the standard library; the badminton group extracts racket face angle, instantaneous speed of impact, and height of impact; and the soccer group extracts contact point, dribbling frequency, and change of direction angle. These indicators all come from a pre-defined evaluation standard library in the physical education syllabus, ensuring that different sports use specific and scientific parameter sets. For each sports group's data, the corresponding standard parameters are obtained, resulting in a matched set of evaluation criteria.

[0140] Specifically, for the basketball shooting group, standard parameters might be set at a release angle range of 45° to 60° and a release speed of no less than 6.5 meters per second; while the badminton smash group requires a racket face angle close to vertical and a hitting speed higher than 25 meters per second. Through the matching process, each set of data carries a project-specific acceptable range as the basis for subsequent comparisons. Using the matched set of evaluation criteria, the data for each project group is compared and analyzed item by item.

[0141] For example, in the soccer dribbling dataset, the system reads the point of contact between the athlete and the ball frame by frame. If three consecutive touches are made using the outside of the foot instead of the standard inside, it is marked as an abnormal movement. Similarly, if the arc height of a basketball shot is lower than a preset minimum value, it is also marked as an anomaly, thus forming an analysis result set with anomaly markers. Based on the analysis result set containing anomaly markers, a detailed movement evaluation for each item is generated.

[0142] In one embodiment, for ten badminton smashes performed by a student, the system generates a detailed report listing the instantaneous speed, racket face angle, and whether each hit deviates from the standard range, and marks the third and seventh hits as "errors caused by premature racket face opening." A data aggregation tool is used to categorize outliers and normal data separately, obtaining a categorized set of evaluation details.

[0143] For example, the aggregated data provides statistics such as a 20% abnormal percentage of a student's badminton smashes and an average normal shot speed of 28.4 meters per second, allowing teachers to quickly grasp the overall performance. The categorized assessment details are automatically populated with data based on the standardized report format. A template rendering method maps these assessment details to a pre-defined report structure; for example, abnormal details are placed in the "Technical Problem Analysis" section, and statistical data in the "Overall Performance" section, ultimately creating a visually appealing action assessment document. Based on the final action assessment document, a data storage and distribution process is executed. The document content is saved in PDF format to the integrated platform's database and simultaneously pushed in real-time to the teacher's teaching module and the student's personal account via a messaging interface, completing the entire data flow from analysis to feedback.

[0144] like Figure 9 As shown, in sports such as basketball, football, volleyball, and badminton, this invention demonstrates stable efficiency improvements under different batch processing scales. Thermal distribution data shows that the processing efficiency improvement remains at approximately 40% or higher in most scenarios, with a more significant improvement observed with larger batch inputs. This indicates that the method possesses excellent parallel processing capabilities and cross-sport scalability, effectively supporting the high-concurrency evaluation needs of a comprehensive sports teaching platform.

[0145] Additional details on core technologies: The core technical solution of the intelligent correction method for sports teaching movements based on spatiotemporal graph convolution described in this application is further explained below. The scope of protection of this method covers the entire process of intelligent movement correction, construction of hybrid feature models, generation of general analysis templates, and real-time movement correction, among other core technical points. Specific details are as follows: 1. This method targets not only basketball and football, but also extends to mainstream sports such as volleyball, badminton, and table tennis. Basketball and football serve as benchmark sports for extracting cross-sport common features. The overall technical solution is adaptable to motion analysis of all sports based on human core joint movements, and can adaptively adjust specific parameters and weight allocation ratios according to the motion characteristics of different sports.

[0146] 2. Regarding the technical details of batch application of multi-project data and generation of unified evaluation reports, the data cleaning tools used in this method are developed based on Python Pandas and NumPy. Missing value imputation adopts the linear interpolation method between the preceding and following frames, and outlier judgment adopts the 3σ principle. The supporting motion evaluation standard library contains the core evaluation indicators and standard parameter ranges of each sports project, which can be dynamically updated according to the sports teaching syllabus. The final generated motion evaluation report adopts the standardized PDF format and includes core contents such as motion anomaly point marking, targeted motion correction suggestions, and motion standardization score. The evaluation report can be pushed to the teacher's computer / tablet and the student's mobile phone / tablet in real time through the RESTful API interface.

[0147] 3. The construction of the hybrid feature model is one of the core technical points of this method. Specifically, after obtaining the set of general features across projects, the model is first matched with the data structure of the multi-project analysis platform to obtain a preliminary feature dataset. If the matching degree between the personalized needs of the project and the general features is less than 0.75, specific parameter types such as basketball arm coordination and football leg strength are selected, and quantitative values ​​are extracted from the preset indicator library to form a specific indicator dataset. A binary classification support vector machine with a radial basis function kernel, a penalty factor C of 0.8, and a kernel parameter γ of 0.05 is used to complete feature fusion. The absolute value of the hyperplane normal vector component is extracted as the weight coefficient, and the general features and specific parameters are weighted and summed (the weight of basketball and football project specific parameters accounts for 0.6, and the weight of general features accounts for 0.4). The weights are then adjusted in combination with the distribution characteristics of multi-project data. After adapting to the platform application scenario, the final hybrid feature model is obtained, and a feature application template is generated.

[0148] 4. The method for generating the general analysis template in this approach is as follows: First, obtain the analysis results after motion deviation correction. Then, perform secondary optimization through a spatiotemporal graph convolutional network to enhance the accuracy of motion capture. Next, extract the cross-project shared content from the optimization results. If the shared content meets the modularization conditions of feature reusability not less than 90%, module interface standardization, and feature dispersion less than 0.15, a draft of the general analysis template is generated. After the spatiotemporal graph convolutional network structure is adjusted and finally confirmed, the final general analysis template is obtained. The template is a 17×6 standardized parameter matrix. The 17 rows correspond to the 17 key joint nodes of the human body, and the 6 columns are, in order, the minimum and maximum values ​​of the ideal space coordinates of the joints, the minimum and maximum values ​​of the motion velocity, the minimum and maximum values ​​of the inter-joint phase difference, and the maximum and maximum values ​​of the inter-joint phase difference. All values ​​in the matrix are set based on the statistical results of 2000 sets of standard motion samples each for basketball and football.

[0149] 5. The specific implementation method of real-time correction of physical education teaching movements in this method is as follows: Newly inputted physical education movement data is acquired in real time using edge acquisition devices such as wearable sensors equipped with inertial measurement units and high-definition action cameras, and transmitted to the cloud server via Bluetooth / 5G; the hybrid feature model constructed above is used to perform real-time matching of the movement data and obtain a matching degree value. If the matching degree is lower than 0.9, the movement pattern deviation value is calculated, and the original matching sequence is weighted and compensated using the deviation value to obtain a corrected matching sequence. The corrected matching degree is then recalculated; if the corrected matching degree is not lower than 0.9, the analysis results after deviation correction and personalized movement correction suggestions are immediately output and pushed to each teaching terminal of the integrated physical education teaching platform; if the corrected matching degree is still lower than 0.9, the deviation correction process is repeated until the corrected matching degree reaches the preset standard.

[0150] 6. The specific implementation of each technical point covered by this method is based on the above details, and together with the contents of the technical field, background technology, invention content, and specific implementation methods in the specification, they form a complete technical solution. The technical points are interconnected and mutually supportive, and together realize the intelligent correction and unified evaluation of sports teaching movements in multiple sports.

[0151] Implementation Environment The intelligent correction method for physical education teaching movements based on spatiotemporal graph convolution described in this application can be implemented based on a hardware architecture of "cloud server + edge acquisition device + teaching terminal": 1. Edge acquisition device: It adopts wearable sensors with inertial measurement units (IMU) and high-definition action cameras to collect human joint motion data and sports action video data in real time, and supports wireless Bluetooth / 5G data transmission.

[0152] 2. Cloud Server: Servers using NVIDIA A100 GPUs are used for model training and inference of spatiotemporal graph convolutional networks and support vector machines, as well as batch processing and storage of multi-sport motion data.

[0153] 3. Teaching Terminals: These include teacher-side computers / tablets and student-side mobile phones / tablets, used for displaying movement assessment reports, pushing movement correction suggestions, and providing physical education teaching guidance. The software environment for this method is Ubuntu 20.04 operating system, and the algorithm development is based on Python 3.8, PyTorch 1.13, OpenCV 4.7, and Scikit-learn 1.2. The integrated physical education teaching platform adopts a B / S architecture and supports multi-terminal access via web and mobile app.

[0154] Experimental results verification This application uses basketball shooting and soccer shooting as experimental subjects, selecting movement data from 100 physical education participants as experimental samples. The samples consist of 2000 sets of standard movements and 1000 sets of non-standard movements for each activity. The method described in this application is used to conduct a movement intelligent correction experiment. The experimental results are as follows: 1. Accuracy of motion matching and correction: The accuracy of correction for basketball shooting motion is 92.5%, and the accuracy of correction for soccer shooting motion is 93.2%, which is more than 25% higher than the traditional single-item motion analysis method.

[0155] 2. Cross-project analysis efficiency: It can process motion data from multiple projects such as basketball, football, and volleyball in batches, improving data processing efficiency by more than 40% compared to traditional methods.

[0156] 3. Template Reusability: The final generated general analysis template can be reused in eight mainstream sports, including basketball, football, volleyball, and badminton, with a feature reusability of 95%, meeting the general analysis needs of cross-sports teaching. Experimental results show that this method can effectively achieve unified modeling and personalized correction of movements in multiple sports, significantly improving the universality, accuracy, and efficiency of cross-sports movement analysis and sports teaching correction.

[0157] Supplementary Instruction Manual I. Specific Implementation of Spatiotemporal Graph Convolutional Network (ST-GCN) The Spatiotemporal Graph Convolutional Network (ST-GCN) used in this invention is a deep learning model based on the topology of the human skeleton, and its specific implementation method is as follows: Network architecture design: This invention employs a network structure consisting of 4 spatiotemporal convolutional layers, 2 pooling layers, and 1 fully connected layer.

[0158] The first spatiotemporal convolutional layer has 64 kernels, the second layer has 128, the third layer has 256, and the fourth layer has 256.

[0159] The pooling layer uses max pooling, and the pooling kernel size is 2×2.

[0160] All activation functions use the ReLU function.

[0161] The network input consists of spatiotemporal sequence data of 17 key nodes, and the output is a 128-dimensional feature vector.

[0162] Human skeleton topology construction: The key joints of the human skeleton are defined as 17: top of head, neck, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, right ankle, center of trunk, left foot, and right foot.

[0163] Spatial edges are defined as skeletal connections within a single frame of an image that conform to the physiological connections of the human body, such as shoulder-elbow, elbow-wrist, etc.

[0164] A temporal edge is defined as a connection between the same joint node between consecutive frames.

[0165] Training methods: Training was conducted using 2000 sets of standard movement samples each from basketball and soccer.

[0166] The Adam optimizer was used for training, with an initial learning rate of 0.001 and a decay rate of 0.95.

[0167] Set the batch size to 32 and the number of training rounds to 500.

[0168] An early stopping mechanism is adopted, and training is stopped when the validation set loss does not decrease for 5 consecutive rounds.

[0169] Use cross-validation to evaluate model performance and ensure model generalization ability.

[0170] Feature extraction process: Input the original motion sequence data and extract the coordinate sequence of key joint nodes.

[0171] The coordinate sequence is processed using the ST-GCN network to generate a multi-channel feature map.

[0172] Extract the time-dimensional coordinate sequence of key joint nodes from the feature map.

[0173] Calculate the displacement vectors of joint nodes between adjacent frames to obtain the joint motion trajectory sequence.

[0174] By extracting time-dimensional features from joint motion trajectory sequences, motion velocity sequences can be determined.

[0175] Calculate the relative distance sequence between key joint nodes based on the relationship between the motion velocity sequence and spatial position.

[0176] Construct the spatial adjacency strength between joints to form a feature vector that integrates spatiotemporal information.

[0177] II. Specific Implementation of Support Vector Machine (SVM) This invention uses two support vector machine models, one for feature classification and the other for feature fusion: Support vector machines for feature classification: A typed SVM was adopted, and the radial basis function (RBF) kernel was selected.

[0178] The penalty factor C is set to 1.0, and the kernel function parameter γ is set to 0.1.

[0179] The training data includes spatiotemporal relationship vectors of movements from different sports such as basketball, football, and volleyball.

[0180] The training process uses cross-validation, dividing the dataset into 5 parts, using 4 parts for training and 1 part for validation each time.

[0181] The optimal parameter combination is determined by grid search.

[0182] The classification results are used to calculate feature similarity using cosine similarity. Vectors with a similarity greater than 0.8 are included in the shared feature library.

[0183] Support vector machines for feature fusion: A binary classification SVM was used, and the radial basis function (RBF) kernel was selected.

[0184] The penalty factor C is set to 0.8, and the kernel function parameter γ is set to 0.05.

[0185] The training labels are standard / non-standard actions, and the hyperplane normal vector is obtained through training.

[0186] The absolute values ​​of the components of the hyperplane normal vector are used as weight coefficients for each feature dimension.

[0187] Feature fusion is achieved through weighted summation. In basketball, the weight of arm coordination-specific parameters accounts for 0.6, and the weight of general features accounts for 0.4. In soccer, the weight of leg strength-specific parameters accounts for 0.6, and the weight of general features accounts for 0.4.

[0188] III. Basis and Adjustment Methods for Threshold Determination Action segment matching template threshold (0.8): Basis: Based on the feature similarity statistics of 1000 sets of standard motion samples each for basketball and football.

[0189] Determination method: Calculate the cosine similarity between each action segment and the standard template. Statistically, 90% of the samples have a similarity between 0.75 and 0.95. Take the median value of 0.8 as the threshold.

[0190] Adjustment method: When the amount of action data for a new project reaches more than 500 sets, the similarity distribution can be recalculated based on the new data, and the threshold can be fine-tuned within a range of ±0.05.

[0191] Data integrity threshold (95%): Basis: Based on the actual situation of sports motion video data collection, considering common problems such as device disconnection and obstruction.

[0192] Methodology: The completeness of 1000 sports action video clips was statistically analyzed, and it was found that 95% of the clips had a completeness rate between 90% and 100%.

[0193] Adjustment method: When the stability of the motion data acquisition equipment for a specific sport is improved, the threshold can be appropriately increased to 97%, and vice versa.

[0194] Feature similarity threshold (0.8): Basis: Similarity statistics based on the movement characteristics of sports such as basketball, football, and volleyball.

[0195] Method for determination: The cosine similarity of the motion features of key joints among the projects was calculated, and it was found that the cross-project common feature similarity was between 0.7 and 0.9.

[0196] Adjustment method: When adding a new sports item, first collect 100 sets of standard movement samples for the new item, calculate the similarity with existing common features, and redetermine the threshold.

[0197] Personalized requirement matching threshold (0.75): Basis: Based on the degree of difference between project-specific parameters and general characteristics.

[0198] Methodology: The matching degree between project-specific parameters and general features in sports such as basketball and football was statistically analyzed, and it was found that the matching degree of project-specific parameters was between 0.65 and 0.85 for 90% of the projects.

[0199] Adjustment method: When project-specific parameters change, recalculate the matching degree, and the threshold can be adjusted by ±0.05.

[0200] Matching threshold (0.9) and modified matching threshold (0.9): Basis: Based on the evaluation criteria for movement standardization in physical education teaching, and with reference to the scoring criteria for standard movements by physical education teachers.

[0201] Methodology: Ten professional physical education teachers were invited to score 1,000 standard movement samples. Statistical analysis showed that 90% of the standard movements had a matching degree between 0.85 and 0.95.

[0202] Adjustment method: When the teaching standards are updated, the threshold is re-determined based on teacher scores, with an adjustment range of ±0.05.

[0203] Feature dispersion threshold (0.15): Basis: Based on the statistical distribution characteristics of common features across projects.

[0204] Method for determination: The feature dispersion of 2000 sets of standard motion samples for basketball and football was calculated, and it was found that the dispersion of 90% of the samples was between 0.05 and 0.20.

[0205] Adjustment method: When the number of new sports samples reaches more than 1,000, the feature dispersion is recalculated and the threshold is fine-tuned.

[0206] IV. Explanation of Key Terms Key joint nodes: These refer to 17 joint positions in human movement that have a significant impact on motion analysis, including the top of the head, neck, shoulders, elbows, wrists, hips, knees, ankles, and the center of the trunk. These nodes are connected by bones to form the topology of the human skeleton.

[0207] Spatial adjacency strength: a feature value that quantifies the degree of spatial correlation between different key joint nodes in the human body. It is obtained by calculating the average and variance of the relative distance sequence between key joint nodes and then normalizing the result to obtain a value in the range of [0,1]. The larger the value, the stronger the spatial cooperation between joints.

[0208] Temporal dependency strength: a feature value that quantifies the degree of temporal correlation between different key joint nodes in the human body. The intermediate features of the fused spatiotemporal information are processed through an attention mechanism to obtain a value in the range of [0,1]. The larger the value, the higher the temporal coordination between joints.

[0209] Feature fusion: The process of combining cross-project common features with project-specific parameters. The absolute values ​​of the components of the hyperplane normal vector are extracted by support vector machine as weight coefficients. The common features and specific parameters are weighted and summed to form a hybrid feature model.

[0210] Modular requirements: These refer to the reusability and standardization conditions that general analysis templates must meet, including three core requirements: feature reusability of no less than 90%, standardized module interfaces (using a unified JSON data format), and feature dispersion of less than 0.15.

[0211] Deviation correction: When the matching degree is lower than the preset threshold, the original matching sequence is weighted and compensated using the deviation value of the action pattern, so that the corrected matching sequence is more consistent with the standard action template.

[0212] V. Detailed Explanation of the Implementation Process Data preprocessing workflow: The raw motion sequence data is extracted from the video database, and continuous frame images, frame timestamps, joint coordinates and other data are obtained through a specific data interface (a custom video frame extraction interface developed based on OpenCV).

[0213] Different categories of projects are stored separately. Basketball and soccer projects are distinguished by preset rules. For example, basketball actions include a pattern of arm raising combined with the trajectory of the ball, while soccer actions include a vector of lower limb swing and the direction of the goal.

[0214] The sequence is initially segmented into independent action units using timestamps and spatial location information.

[0215] The segment is marked as a valid segment by comparing it with a preset action feature template (such as a basketball "dunk template"). The matching degree reaches 0.8.

[0216] The processed spatiotemporal fragment data is stored using a uniform formatted storage method (JSON structure).

[0217] The integrity of the data is checked using a data validation tool (developed using Python Pandas and Numpy). If the integrity rate is lower than 95%, a supplementary extraction process is triggered.

[0218] Feature extraction and fusion process: ST-GCN was used to process the spatiotemporal sequence fragments to obtain multi-channel feature maps.

[0219] By analyzing the location of key joint nodes through feature maps, the coordinate sequence of each node in the time dimension is obtained.

[0220] Calculate the displacement vectors of joint nodes between adjacent frames to obtain the joint motion trajectory sequence.

[0221] Extract time-dimensional features to determine the motion velocity sequence.

[0222] Calculate the relative distance sequence between key joint nodes to construct the spatial adjacency strength between joints.

[0223] ST-GCN is used to process the joint input of spatial adjacency strength and motion velocity sequence to obtain intermediate features that fuse spatiotemporal information.

[0224] Calculate the temporal dependency strength based on intermediate features to determine the cross-project general spatiotemporal relationship vector.

[0225] The general spatiotemporal relation vectors are classified using SVM, and vectors with feature similarity higher than 0.8 are included in the shared feature library.

[0226] Obtain the set of general features after classification, adjust them according to individual needs, and determine whether to add project-specific parameters.

[0227] Feature fusion is performed using SVM to construct a hybrid feature model.

[0228] A hybrid feature model is used to perform real-time matching of newly input sports motion data.

[0229] Deviation correction and template generation process: The new input action data is matched in real time using a hybrid feature model to obtain a matching degree value.

[0230] When the matching degree is less than 0.9, the deviation value of the action pattern is calculated, and deviation correction is performed on the original matching sequence.

[0231] Recalculate the corrected match score. If the corrected match score is not less than 0.9, output the final analysis results.

[0232] Extract shared content across projects from the analysis results after bias correction.

[0233] The shared content is further optimized using ST-GCN to enhance the accuracy of motion capture.

[0234] Determine whether the requirements are met based on the preset modularization conditions (feature reusability not less than 90%, module interface standardization, feature dispersion less than 0.15).

[0235] Once the requirements are met, a general analysis template is generated, which is represented by a 17×6 standardized parameter matrix.

[0236] Based on a general analysis template, data from various projects on the integrated physical education teaching platform are applied in batches to generate a unified motion assessment report.

[0237] This supplementary description details the specific implementation methods and determination basis of key components such as algorithms, models, and thresholds involved in this invention, ensuring that those skilled in the art can implement this invention based on the description and meet the requirements of Article 26, Paragraph 3 of the Patent Law.

[0238] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. The present invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of protection claimed by the present invention.

Claims

1. A method for intelligent correction of physical education movements based on spatiotemporal graph convolution, characterized in that, The method includes: Raw motion sequence data is obtained from a pre-set multi-sports motion video database. Initial segmentation is performed based on the motion characteristics of different sports, including basketball and football, to obtain preliminary spatiotemporal sequence fragments. Convolutional neural networks are used to extract features from the obtained spatiotemporal sequence segments, capture dynamic changes for key joint nodes in the action, and determine a general spatiotemporal relationship vector across projects; Based on the determined spatiotemporal relationship vectors, the personalized action patterns of different items are classified by support vector machine. If the classification results show that the feature similarity is higher than the preset threshold, these vectors are included in the shared feature library to obtain the general feature set after classification. After obtaining the set of general features after classification, the remaining personalized needs are adjusted in the multi-project analysis platform to determine whether project-specific parameters need to be added. If so, specific indicators of basketball arm coordination or football leg strength are integrated to obtain the adjusted hybrid feature model. The adjusted hybrid feature model is used to match the newly input sports motion data in real time, determine the matching degree and output the deviation of the motion pattern, and obtain the analysis results after deviation correction. Extract the cross-project shared parts from the analysis results after deviation correction, and use a convolutional neural network to optimize it again to enhance the dynamic capture accuracy. Determine whether the optimized results meet the modular requirements. If they do, generate the final general analysis template. Based on the generated final universal analysis template, data from various projects on the integrated physical education teaching platform are applied in batches to obtain a unified motion evaluation report.

2. The method according to claim 1, characterized in that, The process involves acquiring raw motion sequence data from a pre-set multi-sports motion video database, where the multi-sports include basketball and football. Initial segmentation is performed based on the motion features of basketball and football to obtain preliminary spatiotemporal sequence fragments, including: From a pre-set video database, raw sequence data related to sports movements is extracted through a specific data interface, and stored according to different sports to obtain a preliminary set of movement data. For the initially compiled set of motion data, preset rules are used to distinguish the motion characteristics of basketball and football. If a specific motion pattern is detected in the data, it is classified into the corresponding project group, and the classified motion data grouping is determined. Based on the classified action data, the original sequence data in each group is obtained, and preliminary segmentation is performed using timestamps and spatial location information to obtain a set of segmented sequence fragments. For the segmented set of sequence fragments, by comparing with the preset action feature template, if a sequence fragment matches the template to a preset threshold, it is marked as a valid fragment, and a valid set of spatiotemporal fragments is determined. Based on the valid set of spatiotemporal segments, the sequence processing results of each segment are obtained, and a unified formatted storage method is used to determine the processed spatiotemporal segment data. The processed spatiotemporal fragment data is checked for completeness using a data verification tool. If data is missing or abnormal, a supplementary extraction process is triggered to obtain a complete and standardized spatiotemporal fragment dataset.

3. The method according to claim 1, characterized in that, The method employs a convolutional neural network to extract features from the obtained spatiotemporal sequence segments, captures dynamic changes at key joint nodes in the action, and determines a universal spatiotemporal relationship vector across projects, including: A convolutional neural network is used to process spatiotemporal sequence segments to obtain multi-channel feature maps; By analyzing the location of key joint nodes through feature maps, the coordinate sequence of each node in the time dimension is obtained. The displacement vectors of joint nodes between adjacent frames are calculated based on the coordinate sequence to obtain the joint motion trajectory sequence; Temporal features are extracted from the joint motion trajectory sequence to determine the motion velocity sequence of each key joint node; Calculate the relative distance sequence between key joint nodes based on the correspondence between motion velocity sequence and spatial position; Spatial dimensional features between joints are constructed by using relative distance sequences to obtain the spatial adjacency strength between joints; A convolutional neural network is used to process the joint input of inter-joint spatial adjacency strength and motion velocity sequence to obtain intermediate features that fuse spatiotemporal information. Calculate the temporal dependency strength between each joint node based on intermediate features, and determine the spatiotemporal relationship vector that is applicable across projects.

4. The method according to claim 1, characterized in that, The process involves classifying personalized action patterns for different items using a support vector machine based on determined spatiotemporal relationship vectors. If the classification results show a feature similarity higher than a preset threshold, these vectors are then incorporated into a shared feature library to obtain a general feature set after classification, including: Step 1: Obtain the spatiotemporal relationship vector from the raw data, and perform preliminary organization based on the action patterns of different projects to obtain an initial vector set; Step 2: Use a support vector machine to classify the initial vector set, compare the feature similarity in the classification results, and determine the vector group with high similarity. Step 3: If the number of vector groups with high similarity exceeds a preset threshold, these vectors are included in the shared feature library to obtain a preliminary shared feature set; Step 4: By performing a secondary analysis on the initial shared feature set, extract feature elements with strong generality and determine whether they meet the criteria of a general set; Step 5: Based on the standard of the general set, integrate the extracted feature elements to obtain the final general feature set; Step Six: For the final set of general features, record the corresponding item differentiation information and determine the feature mapping relationship after classification; Step 7: Apply the general feature set to subsequent action pattern analysis through feature mapping relationships.

5. The method according to claim 1, characterized in that, The obtained general feature set after classification is adjusted in the multi-project analysis platform for the remaining personalized needs. It is determined whether project-specific parameters need to be added. If so, specific indicators of basketball arm coordination or soccer leg strength are fused to obtain an adjusted hybrid feature model, including: Obtain the general feature set after classification, read the data through the pre-established feature library, and perform format matching on the data structure in the multi-project analysis platform to obtain the preliminary feature dataset. Based on the preliminary feature dataset, the degree of matching of personalized needs is analyzed. If the degree of matching between the personalized needs of a specific project and the general features is lower than the preset threshold, the project parameter screening process is triggered to determine the type of unique parameters to be incorporated. By selecting specific parameter types, specific indicator data related to basketball arm coordination or soccer leg strength are obtained, and corresponding quantitative values ​​are extracted from a preset indicator library to obtain a specific indicator dataset. For specific indicator datasets and general feature datasets, the support vector machine algorithm is used for feature fusion processing. The two types of data are vector-mapped and weighted to construct a preliminary hybrid feature model. Based on the preliminary hybrid feature model, the distribution characteristics of multi-item data in the platform's analysis module are analyzed. If a deviation is found in the fusion results of specific indicators and general features, the weights are adjusted to obtain an optimized feature combination. By optimizing the feature combination, we adapt it to the application scenarios of the multi-project analysis platform, embed the feature combination into the platform's data processing flow, and determine whether the final hybrid feature model meets the application requirements. Based on the final hybrid feature model, feature application templates suitable for multi-project analysis are generated. Key parameters are extracted from the templates to determine cross-project feature application schemes.

6. The method according to claim 1, characterized in that, The adjusted hybrid feature model is used to perform real-time matching on newly input sports motion data, determine the matching degree, output the deviation of motion patterns, and obtain the analysis results after deviation correction, including: Acquire newly input sports motion data; The sports motion data is matched in real time using the adjusted hybrid feature model to obtain the matching degree value. The matching degree value is compared with a preset threshold. If the matching degree is lower than the preset threshold, it is determined that there is a deviation in the movement pattern, and the movement pattern deviation value is obtained. The original matching sequence is subjected to deviation correction processing using the deviation value of the action pattern to obtain the corrected matching sequence; The corrected matching sequence is recalculated using the hybrid feature model to obtain the corrected matching degree value; The analysis results are output based on the corrected matching degree value to determine whether the current action meets the analysis conditions. If the corrected matching degree reaches the preset standard, the final analysis result is output.

7. The method according to claim 1, characterized in that, The process involves extracting the cross-project shared portions from the bias-corrected analysis results, further optimizing them using a convolutional neural network to enhance dynamic capture accuracy, and determining whether the optimized results meet modularity requirements. If so, a final universal analysis template is generated, including: Obtain the analysis results after deviation correction; The analysis results after bias correction are subjected to secondary optimization processing through a convolutional neural network to obtain the optimized results. Extract cross-project shared content from the optimized results to obtain cross-project shared content; Based on the comparison between the cross-project shared content and the preset modularization conditions, if the cross-project shared content meets the preset modularization conditions, then the current optimized result is determined to meet the modularization requirements. A general analysis template is generated from the optimized results that meet the modular requirements, resulting in a draft of the general analysis template. The initial draft of the general analysis template is structurally adjusted using a convolutional neural network to obtain the structurally adjusted general analysis template. Perform a final confirmation operation on the general analysis template after structural adjustment to obtain the final general analysis template.

8. The method according to claim 3, characterized in that, The step of constructing spatial dimension features between joints through relative distance sequences to obtain spatial adjacency strength between joints includes: calculating the mean and variance of the relative distance sequences between key joint nodes, normalizing the mean and variance to obtain spatial adjacency strength values ​​with a range of [0,1]; the step of calculating temporal dependency strength between joint nodes based on intermediate features includes: using intermediate features that fuse spatiotemporal information as input to the attention mechanism, calculating attention weights to obtain temporal dependency strength values ​​with a range of [0,1].

9. The method according to claim 5, characterized in that, The method involves using a support vector machine (SVM) algorithm to perform feature fusion processing on a specific indicator dataset and a general feature dataset. This process maps and assigns weights to the two types of data. The steps include: constructing a binary classification SVM model labeled with whether an action is standard or not; training the model using the specific indicator dataset and the general feature dataset; extracting the absolute values ​​of the components of the hyperplane normal vector of the trained model and using them as weight coefficients for the corresponding feature dimensions; and using these weight coefficients to perform a weighted summation of each feature dimension of the specific indicator dataset and the general feature dataset to obtain a hybrid feature model.

10. The method according to claim 7, characterized in that, The process of generating the final general analysis template includes: generating a parameter matrix, wherein the number of rows in the parameter matrix corresponds to the number of key joint nodes in the human body, and the number of columns corresponds to the dimensions of the motion feature parameters; the dimensions of the motion feature parameters include the minimum value of the ideal space coordinates of the joint, the maximum value of the ideal space coordinates of the joint, the minimum value of the movement speed, the maximum value of the movement speed, the minimum value of the inter-joint phase difference, and the maximum value of the inter-joint phase difference; the values ​​in the parameter matrix are set based on the statistical results of standard motion samples.

11. The method according to claim 2, characterized in that, The integrity of the processed spatiotemporal segment data is checked using a data verification tool, including: checking whether there are missing values ​​in the spatiotemporal segment data; if so, filling in the timestamps and joint coordinates using a linear interpolation method between the preceding and following frames; checking whether there are outliers in the spatiotemporal segment data; using the three-standard-deviation principle to determine outliers; and removing or correcting data that exceeds three standard deviations.

Images