A sports exercise AI standing long jump evaluation data management system

Abstract

The present invention relates to the field of sports artificial intelligence and data management technology, and in particular to a physical exercise AI standing long jump evaluation data management system, which includes modules such as dynamic environment calibration, multimodal identity recognition, and three-dimensional posture analysis; the dynamic environment calibration module utilizes a multispectral sensor array and a deep learning algorithm to establish a dynamic coordinate system without the need for fixed markers; the multimodal identity recognition module integrates gait and facial features to achieve contactless identity authentication; the three-dimensional posture analysis module uses a multi-view camera and advanced algorithms to perform high-precision modeling and analysis of long jump movements; the intelligent violation judgment engine combines rules with AI models to automatically identify violations and dynamically adjust judgment criteria. The present invention breaks away from the limitations of traditional evaluation methods, achieves high-precision, intelligent evaluation, and effectively solves problems such as poor environmental adaptability, inaccurate identity authentication, and lack of data security. The evaluation results are scientific and reliable, and data management is safe and efficient.

Description

Technical Field

[0001] The present invention relates to the field of sports artificial intelligence and data management technology, and specifically to a sports exercise AI standing long jump evaluation data management system. Background Art

[0002] As a fundamental and important sport in physical education, competitive training, and mass fitness, the accuracy and efficiency of standing long jump assessment have always been a focus. With the development of sports and the rise of the national fitness craze, the limitations of traditional standing long jump assessment methods have become increasingly apparent.

[0003] Currently, standing long jump evaluations mostly rely on manual judging or simple sensor devices. Manual judging is subject to significant subjective factors, with varying standards among referees. It also struggles to accurately capture detailed movements, leading to misjudgments and missed calls, resulting in a lack of fairness and accuracy in the evaluation results. Simple sensor devices, such as standard rangefinders, can only capture distance data and are unable to analyze key performance indicators such as takeoff angle, aerial posture, and landing stability. This makes it difficult to meet the comprehensive and in-depth data requirements of professional training and scientific fitness.

[0004] Amidst rapid technological advancements, while some intelligent assessment systems are beginning to be applied in the sports sector, numerous challenges remain. Existing systems often rely on fixed markers for site calibration. This can significantly reduce calibration accuracy when the environment changes, such as changes in lighting intensity, complex ground textures, or the presence of obstacles, affecting overall assessment results. Identity authentication often relies on simple methods like card swiping or code scanning, which carries the risk of identity theft and data mismatches, making it impossible to guarantee the relevance and security of assessment data. Regarding data management, traditional systems lack effective data encryption and privacy protection measures, resulting in centralized data storage, making them vulnerable to attacks and data leaks. Furthermore, collaborative optimization of data from multiple sites and model upgrades are difficult.

[0005] Therefore, to address the above problems, a physical exercise AI standing long jump evaluation data management system is proposed. Summary of the Invention

[0006] The purpose of the present invention is to provide a physical exercise AI standing long jump evaluation data management system to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A physical exercise AI standing long jump evaluation data management system, comprising:

[0009] Dynamic Environmental Calibration Module: This module collects real-time environmental data from the field using a multispectral sensor array deployed around the long jump area. A deep learning-based adaptive calibration algorithm is then used to establish a dynamic coordinate system. This algorithm integrates ambient light intensity, ground texture characteristics, and obstacle distribution to generate a virtual calibration grid without relying on fixed markers.

[0010] Multimodal identity recognition module: By integrating gait recognition and facial feature extraction technology with user biometric information in the cloud database, it achieves contactless identity authentication;

[0011] 3D Posture Analysis Module: This module uses multi-view cameras to synchronously capture long jump action videos, builds a motion trajectory model using a 3D human posture estimation algorithm, and detects take-off angles, mid-air postures, and landing stability in real time.

[0012] Intelligent Violation Detection Engine: Based on a rule engine and AI time series analysis model, combined with kinematic parameters and spatial trajectory characteristics, it automatically identifies violations including stepping, crossing the line, and body imbalance.

[0013] Distributed data management unit: uses blockchain technology to store assessment data, supports multi-node synchronous verification and tamper-proof performance records, and realizes localized data processing through edge computing modules to reduce cloud dependence.

[0014] As a preferred solution, the implementation method of the dynamic environment calibration module includes:

[0015] A multispectral sensor array is used to obtain real-time environmental information about the long jump area, including light intensity, ground texture, and obstacle distribution.

[0016] Based on the improved convolutional neural network (CNN), environmental features are extracted and virtual calibration grids that adapt to different site conditions are dynamically generated. The loss function of the improved CNN consists of two parts:

[0017] Feature reconstruction loss: Calculates the mean square error between the input environment features and the reconstructed features, and the weight coefficient α is used to control its importance;

[0018] Environmental robustness loss: By maximizing the tolerance of feature differences under light interference, the weight coefficient β is used to balance the anti-interference ability, and and The sum is 1;

[0019] Calibration parameters are optimized through reinforcement learning algorithms to ensure the stability of the coordinate system in complex environments such as shadows and reflections.

[0020] As a preferred solution, the three-dimensional posture analysis module further includes:

[0021] Multi-view cameras are used to synchronously capture stereo images of the long jump action, and the 3D motion trajectory is reconstructed using an improved stereo matching algorithm. The energy function of the algorithm is composed of the following three weighted components:

[0022] Data item: Evaluate the consistency of 3D pose through projection error of multi-view images;

[0023] Smoothness term: constrains the continuity of adjacent joint points in space;

[0024] Timing term: ensures the physical rationality of the posture changes between adjacent frames. The weight coefficients λ1, λ2, and λ3 control the contribution ratio of the three terms respectively;

[0025] Based on the temporal convolutional network (TCN), the joint motion sequence is analyzed to predict the center of mass offset at the moment of landing.

[0026] Combined with the inertial measurement unit (IMU) data, the acceleration error of the aerial attitude is corrected through a filtering algorithm to achieve sub-centimeter precision measurement.

[0027] As a preferred solution, the determination method of the intelligent violation determination engine includes:

[0028] The continuous frame action sequence is modeled through an improved LSTM network, whose forget gate output is integrated with predefined kinematic constraints, and the constraint weight coefficient γ controls the contribution of physical priors;

[0029] Define a dynamic threshold rule base and adjust the judgment criteria based on user age, gender, and historical data. The threshold is the weighted average of the baseline threshold and the adaptive offset, and the balance coefficient μ controls the weight of the two.

[0030] Generates explainable violation reports, including action breakdown diagrams and improvement suggestions, and provides real-time feedback to users via AR devices.

[0031] As a preferred solution, the distributed data management unit further includes:

[0032] The blockchain-based smart contract mechanism ensures the integrity and traceability of score data during transmission and storage. Data hashing is achieved by concatenating the current data, a random number, and the hash value of the previous block, and using the SHA-3 algorithm to generate a unique identifier.

[0033] A federated learning framework is used to aggregate anonymized training data from multiple sites to optimize global AI model performance. The objective function is the weighted sum of the local losses of each site, and a regularization term is added to prevent overfitting.

[0034] Support users to independently manage personal sports data through encrypted tokens NFT and authorize third-party access.

[0035] It can be seen from the technical solutions provided by the present invention that the physical exercise AI standing long jump evaluation data management system provided by the present invention has the following beneficial effects:

[0036] High-precision evaluation and reliable judgment: The dynamic environment calibration module eliminates reliance on fixed markers. Through a multispectral sensor array and deep learning algorithms, it accurately constructs a dynamic coordinate system in complex environments, providing an accurate spatial reference for 3D posture analysis. The 3D posture analysis module utilizes multi-view cameras and advanced algorithms to achieve sub-centimeter-level precision measurement of long jump movements, detecting key indicators during takeoff, mid-air, and landing in real time. The intelligent violation judgment engine combines a rule engine with an AI timing analysis model to automatically and accurately identify violations based on kinematic parameters and spatial trajectory characteristics. It can also dynamically adjust judgment criteria based on individual user characteristics to ensure scientific, fair, and reliable evaluation results. Compared with traditional evaluation methods, this greatly improves the accuracy and credibility of the evaluation.

[0037] Efficient identity authentication and data security: The multimodal identity recognition module integrates gait and facial features to achieve contactless identity authentication, reducing user operation processes and improving ease of use; the distributed data management unit uses blockchain technology to store data, ensuring that the data cannot be tampered with and is traceable; edge computing enables local processing, reducing cloud dependence and improving data processing efficiency; the federated learning framework aggregates data from multiple sites to optimize AI models while protecting data privacy; users can also independently manage data permissions through NFTs, comprehensively ensuring data security and preventing data leakage and abuse. Compared with similar systems, it has obvious advantages in data security and privacy protection.

[0038] Intelligent Analysis and Personalized Services: The system's modules collaborate to generate multi-dimensional data. Through in-depth analysis of long jump movements, it provides users with detailed action breakdowns and improvement suggestions, helping them improve their athletic performance in a targeted manner. Furthermore, based on users' historical data and individual characteristics, it dynamically adjusts evaluation standards and recommends personalized training plans, meeting the diverse needs of different users, enhancing user stickiness, and promoting scientific exercise. Compared with traditional sports evaluation systems, it is more intelligent and personalized.

[0039] Wide applicability and technical scalability: The system's technical solution can be quickly expanded to the evaluation of other sports such as high jump and shot put by adjusting sensor layout, defining key points for posture analysis, and loading pre-trained models, effectively reducing development costs and improving system versatility. Its integration of advanced technologies such as multispectral sensing, deep learning, and blockchain provides innovative ideas and practical experience for the technological development in the field of sports evaluation, promotes the application and expansion of related technologies in more fields, and promotes the intelligent upgrading of the sports industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1This is a schematic diagram of the overall structure of a physical exercise AI standing long jump evaluation data management system of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0042] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0043] like Figure 1 As shown, an embodiment of the present invention provides a physical exercise AI standing long jump evaluation data management system, comprising:

[0044] Dynamic Environmental Calibration Module: This module collects real-time environmental data from the field using a multispectral sensor array deployed around the long jump area. A deep learning-based adaptive calibration algorithm is then used to establish a dynamic coordinate system. This algorithm integrates ambient light intensity, ground texture characteristics, and obstacle distribution to generate a virtual calibration grid without relying on fixed markers.

[0045] Multimodal identity recognition module: By integrating gait recognition and facial feature extraction technology with user biometric information in the cloud database, it achieves contactless identity authentication;

[0046] 3D Posture Analysis Module: This module uses multi-view cameras to synchronously capture long jump action videos, builds a motion trajectory model using a 3D human posture estimation algorithm, and detects take-off angles, mid-air postures, and landing stability in real time.

[0047] Intelligent Violation Detection Engine: Based on a rule engine and AI time series analysis model, combined with kinematic parameters and spatial trajectory characteristics, it automatically identifies violations including stepping, crossing the line, and body imbalance.

[0048] Distributed data management unit: uses blockchain technology to store assessment data, supports multi-node synchronous verification and tamper-proof performance records, and realizes localized data processing through edge computing modules to reduce cloud dependence.

[0049] In this embodiment, the implementation method of the dynamic environment calibration module includes:

[0050] A multispectral sensor array is used to obtain real-time environmental information about the long jump area, including light intensity, ground texture, and obstacle distribution.

[0051] Based on the improved convolutional neural network (CNN), environmental features are extracted and virtual calibration grids that adapt to different site conditions are dynamically generated. The loss function of the improved CNN consists of two parts:

[0052] Feature reconstruction loss: Calculate the mean square error between the input environment features and the reconstructed features, weight coefficient Used to control its importance;

[0053] Environmental robustness loss: By maximizing the tolerance of feature differences under light interference, the weight coefficient Used to balance anti-interference capabilities, and and The sum is 1;

[0054] Optimize calibration parameters through reinforcement learning algorithms to ensure the stability of the coordinate system in complex environments such as shadows and reflections;

[0055] Furthermore, the dynamic environment calibration module serves as the "spatial positioning center" of the AI ​​standing long jump evaluation data management system for physical exercise. Through the deep integration of a multispectral sensor array and a deep learning algorithm, it achieves dynamic perception of the long jump venue environment and establishes a precise coordinate system, ensuring the accuracy and stability of evaluation data collection. The following will elaborate on this module in terms of its overall function, sub-module composition, key technical principles, workflow, and application value:

[0056] 1. Overview of overall functions:

[0057] The dynamic environmental calibration module uses a multispectral sensor array deployed around the long jump area to collect real-time field environmental data, including information such as light intensity, ground texture, and obstacle distribution. Based on a deep learning-based adaptive calibration algorithm, the module dynamically generates a virtual calibration grid and constructs a dynamic coordinate system suitable for complex environments, providing a precise spatial reference for subsequent functions such as motion capture and posture analysis. This module eliminates the traditional calibration's reliance on fixed markers and effectively addresses the impact of environmental changes (such as shadows and reflections) on calibration accuracy, providing reliable basic data support for standing long jump evaluation.

[0058] 2. Submodule composition and functions:

[0059] (1) Environmental data collection unit:

[0060] Sensor Deployment: A multispectral sensor array is strategically placed around and above the long jump area. This array integrates visible light cameras and infrared cameras. The visible light cameras capture visual information such as the field's texture and color. The infrared cameras acquire thermal imaging data in low-light or complex lighting conditions, complementing the visible light cameras and enabling comprehensive, all-weather monitoring of the field environment.

[0061] Real-time data acquisition: The sensor array continuously collects real-time environmental information of the long jump area at a high frame rate, preliminarily encodes the collected image data, light intensity data, temperature data, etc., and sends it to the data processing unit through a high-speed data transmission channel;

[0062] (2) Feature extraction and model building unit:

[0063] Environmental feature extraction: This algorithm extracts features from collected environmental data using an improved convolutional neural network (CNN). The network structure utilizes a combination of multiple convolutional layers, pooling layers, and fully connected layers. Through convolution operations, it automatically extracts features such as light intensity variations, ground texture features, and obstacle outlines. The improved CNN incorporates an attention mechanism, allowing the network to focus on key regional features that have a significant impact on coordinate system calibration, improving the efficiency and accuracy of feature extraction.

[0064] Loss function optimization: The loss function of the improved CNN is defined as (in, is the feature reconstruction loss weight coefficient, which is used to control the influence of the mean square error between the input environment features and the reconstructed features; is the environmental robustness loss weight coefficient, and ; , is the sample size, is the input environment feature, To reconstruct features; , is the number of test samples, is the tolerance threshold, which is used to enhance the model’s ability to resist interference from illumination changes. is the environmental feature after adding noise, is the original environmental feature); by optimizing the loss function, accurate extraction of environmental features and improvement of model robustness are achieved;

[0065] Virtual calibration grid generation: Based on the extracted environmental features, a virtual calibration grid is dynamically generated to adapt to different site conditions. The density and accuracy of the grid are automatically adjusted according to the complexity of the site. In areas with many obstacles or drastic texture changes, the grid division is finer to ensure the accuracy of the coordinate system construction.

[0066] (III) Parameter optimization and coordinate system establishment unit:

[0067] Reinforcement learning optimization: A reinforcement learning algorithm is used to optimize the parameters of the virtual calibration grid. A reward function is defined to provide positive rewards when the generated coordinate system accurately matches the actual spatial position in complex environments, and negative rewards when it does not. Through continuous interactive learning between the agent and the environment, the calibration parameters are adjusted to ensure that the coordinate system remains stable in complex environments such as shadows and reflections.

[0068] Dynamic coordinate system establishment: Based on the optimized virtual calibration grid, a dynamic coordinate system is constructed. This coordinate system uses a fixed reference point in the long jump area as its origin and updates the coordinate axis direction and scale in real time based on environmental changes, providing a unified and accurate spatial coordinate reference for subsequent 3D posture analysis, violation determination, and other functions.

[0069] 3. Key technical principles:

[0070] (1) Principle of multispectral data fusion:

[0071] The visible light and infrared data collected by the multispectral sensor array have different characteristics and advantages. Through data fusion technology, the two types of data can be fused at the pixel level, feature level, or decision level. In pixel-level fusion, the corresponding pixels of the visible light image and infrared image are directly weighted and combined. In feature-level fusion, the features of the two images are extracted and then fused. Decision-level fusion makes a comprehensive decision based on the analysis results of the two images. Through data fusion, more comprehensive and accurate environmental information is obtained, providing a high-quality data foundation for subsequent calibration algorithms.

[0072] (2) Principle of improved CNN feature extraction:

[0073] The improved CNN is based on the basic architecture of convolutional neural networks. By introducing an attention mechanism and a special network layer design, it enhances its ability to extract environmental features. The attention mechanism enables the network to automatically learn the importance of features in different regions, focusing on areas that are critical to coordinate system calibration, such as the long jump take-off line and the landing area boundary. Special network layer designs, such as residual connections and dilated convolutions, effectively solve the vanishing gradient problem caused by increased network depth and improve the network's ability to express complex environmental features.

[0074] (III) Principles of reinforcement learning parameter optimization:

[0075] Reinforcement learning uses continuous trial and error learning within the environment to adjust strategies with the goal of maximizing cumulative rewards. In the dynamic environment calibration module, the agent adjusts the calibration parameter strategy, using the complex environment of a long jump field. The agent adjusts the calibration parameters based on the current environmental state (such as light intensity and obstacle distribution) and receives corresponding reward feedback. Through continuous learning, the agent gradually finds the optimal calibration parameter strategy, ensuring that the dynamic coordinate system remains stable in various environments.

[0076] 4. Module workflow:

[0077] (1) Initialization phase:

[0078] After the dynamic environmental calibration module is started, it loads the initial parameters of the improved CNN model, the initial strategy of the reinforcement learning algorithm, and the preset virtual calibration grid template; establishes data communication connections with the multispectral sensor array and other modules of the system, and prepares to receive environmental data and control instructions;

[0079] (2) Data collection stage:

[0080] The multispectral sensor array continuously collects environmental data from the long jump area at a set frequency and transmits the data in real time to the environmental data acquisition unit. This unit converts the data format and performs preliminary verification to ensure its integrity and accuracy before sending it to the feature extraction and model building unit.

[0081] (3) Feature extraction and model calculation stage:

[0082] The feature extraction and model building unit uses an improved CNN to extract features from environmental data, calculates the loss function and performs backpropagation to optimize network parameters and improve the accuracy of feature extraction. Based on the extracted features, a virtual calibration grid is dynamically generated and the grid parameters are optimized through a reinforcement learning algorithm.

[0083] (IV) Coordinate system establishment stage:

[0084] The parameter optimization and coordinate system establishment unit constructs a dynamic coordinate system based on the optimized virtual calibration grid. The coordinate system information is synchronized to other modules of the system, such as the 3D posture analysis module and the intelligent violation determination engine, providing a spatial coordinate benchmark for subsequent evaluation functions.

[0085] (V) Dynamic update stage:

[0086] During system operation, the system continuously monitors environmental changes. When environmental changes such as sudden changes in light intensity or movement of obstacles are detected, the system restarts the feature extraction, model calculation, and parameter optimization processes, and updates the dynamic coordinate system in real time to ensure that the coordinate system always adapts to environmental changes.

[0087] (6) Ending stage:

[0088] When the system stops running or receives a stop command, the dynamic environment calibration module stops data collection and calculation operations, saves the current calibration parameters and model status, closes the communication connection with other modules, and releases system resources;

[0089] 5. Application value of the module:

[0090] 1. Improving evaluation accuracy:

[0091] Freed from the limitations of fixed markers, the dynamic environmental calibration module can accurately construct a coordinate system in complex and changing environments, providing a precise spatial reference for functions such as 3D posture analysis and violation determination, ensuring the accuracy of standing long jump evaluation data and reducing evaluation errors caused by environmental factors.

[0092] (2) Enhance system adaptability:

[0093] By dynamically sensing and adapting to different lighting conditions, ground textures, and obstacle distribution, the module enables the entire evaluation system to operate stably in a variety of field environments, expanding the system's application scenarios and improving its versatility and practicality.

[0094] (3) Reduce maintenance costs:

[0095] Traditional calibration methods rely on fixed markers, which are easily damaged and lost, and require frequent maintenance and calibration. The dynamic environmental calibration module does not require fixed markers, reducing the maintenance cost and calibration workload of the markers, improving the system's maintenance efficiency and reducing long-term use costs.

[0096] In this embodiment, the multimodal identity recognition module serves as the "digital gatekeeper" of the AI ​​standing long jump assessment data management system for physical exercise. By integrating multiple biometric recognition technologies and interacting with cloud data, it achieves contactless, high-precision user identity authentication, ensuring a one-to-one correspondence between assessment data and user information, and improving system security and data reliability. The following describes this module in detail from the aspects of overall functionality, submodule composition, key technical principles, workflow, and application value:

[0097] 1. Overview of overall functions:

[0098] The multimodal identity recognition module utilizes a combination of gait recognition and facial feature extraction technologies, combined with user biometric information from a cloud database, to automatically complete identity authentication when a user enters the long jump area. This module requires no active user interaction and uses dynamic behavioral data and static biometric data collected by multiple sensors to quickly and accurately identify the user, eliminating the cumbersome process and errors associated with manual registration. Furthermore, the module is capable of dynamically updating user biometric data to adapt to changes in individual characteristics over time, laying the foundation for accurate recording, personalized analysis, and user profile management of subsequent evaluation data.

[0099] 2. Submodule composition and functions:

[0100] (1) Biometric collection unit:

[0101] Sensor deployment: Multi-view cameras and depth sensors are deployed at the entrance and surrounding areas of the long jump area. The multi-view cameras capture the user's facial images and full-body movement postures from different angles. The depth sensors obtain three-dimensional spatial information of various parts of the user's body, providing depth data support for gait recognition and ensuring the integrity and accuracy of biometric feature collection.

[0102] Real-time data acquisition: When the user enters the acquisition range, the sensor array synchronously collects multimodal data at a high frame rate, including color facial image sequences, human body contour dynamic video, joint motion trajectories during the gait cycle, and other data. The raw data is encoded and transmitted to the feature extraction and analysis unit;

[0103] (2) Feature extraction and analysis unit:

[0104] Facial feature extraction: A deep learning-based convolutional neural network (CNN) is used to extract features from collected facial images. Through multiple convolutional and pooling layers, key facial features (such as the shape and relative position of eyes, nose, and mouth) and texture information (such as wrinkles and skin color) are extracted to generate high-dimensional facial feature vectors. Pre-trained facial recognition models (such as FaceNet) are used in conjunction with transfer learning techniques to fine-tune the model in the system application scenario to improve the accuracy and robustness of facial feature extraction.

[0105] Gait feature extraction: A spatiotemporal graph convolutional network (ST-GCN) is used to process human motion videos to capture the spatiotemporal characteristics of gait. The spatial positions and time series information of human joints in consecutive frames are constructed into a spatiotemporal graph. Dynamic features such as gait period, stride length, and swing angle are extracted through graph convolution operations to generate a gait feature vector. At the same time, an attention mechanism is introduced to enable the network to focus on gait features that are critical for identity recognition, such as unique walking posture and body swing amplitude.

[0106] Feature fusion: A weighted fusion strategy is used to fuse the extracted facial feature vector with the gait feature vector. The fusion weight is dynamically adjusted based on the recognition accuracy of the two features in different scenarios. For example, in an environment with sufficient light and clear faces, the weight of the facial feature is appropriately increased; in cases of long distance or face occlusion, the weight of the gait feature is increased to generate the final multimodal feature vector.

[0107] (3) Identity authentication and data management unit:

[0108] Cloud data comparison: The fused multimodal feature vector is sent to the cloud database for comparison with the stored user biometric template. The similarity between the feature vectors is evaluated using methods such as cosine similarity calculation or Euclidean distance measurement. An identity authentication threshold is set. When the similarity exceeds the threshold, the user identity is determined to be a successful match. If the threshold is not reached, a secondary authentication process is triggered, such as prompting the user to enter a registered mobile phone number or verify other backup biometrics.

[0109] User data management: After successful authentication, the module binds the user's identity information with the current evaluation data and updates the relevant data in the user's sports profile. At the same time, if the user's biometric characteristics (such as slight changes in facial features caused by changes in hairstyle, changes in gait caused by long-term exercise) are detected to be different from the original template and meet the update conditions, the new feature data will be automatically uploaded to the cloud database, replacing the old template, realizing dynamic update of the user's biometric characteristics.

[0110] Security protection: Encryption algorithms (such as AES-256) are used to encrypt transmitted biometric data to prevent data theft or tampering during transmission. In the cloud database storage process, blockchain technology is used to distribute user identity information to ensure data immutability and traceability, thereby improving the security of the identity authentication system.

[0111] 3. Key technical principles:

[0112] (1) Principle of multimodal biometric fusion:

[0113] Multimodal biometric fusion is based on the concept of information complementarity, combining the advantages of facial and gait biometrics. Facial features are stable and unique, making them suitable for close-range, clear image recognition. Gait features can be collected at a distance and in a non-cooperative state, and contain information about an individual's behavioral habits. Through feature-level fusion methods, the two features are integrated at the data level to generate a feature vector containing more identity information, compensating for the recognition defects of single-modal features in complex scenarios and improving the accuracy and reliability of identity authentication.

[0114] (2) Principles of deep learning feature extraction:

[0115] Deep learning models (such as CNN and ST-GCN) automatically learn complex patterns and feature representations in data through multi-layer neural networks. In facial feature extraction, CNN extracts semantic features at different levels by sliding convolution kernels across the image. In gait feature extraction, ST-GCN treats human motion as a spatiotemporal graph and uses graph convolution to capture the dynamic relationship between joints in space and time. The model is trained on a large amount of annotated data to optimize network parameters, enabling the network to accurately extract features that are discriminative for identity recognition.

[0116] (3) Principle of cloud-based collaborative authentication:

[0117] The cloud database stores a large number of user biometric templates. During the identity authentication process, the local module uploads the extracted feature vectors to the cloud for comparison. The cloud server uses distributed computing and parallel processing technology to quickly retrieve matching user templates and return the authentication results. At the same time, a two-way data synchronization mechanism is established between the cloud and local modules to ensure the real-time update and consistency of user biometric data, realizing seamless identity authentication across devices and scenarios.

[0118] 4. Module workflow:

[0119] (1) Initialization phase:

[0120] After the multimodal identity recognition module is started, it loads the pre-trained face and gait recognition model parameters, establishes a secure communication connection with the cloud database, initializes the feature fusion weight parameters and identity authentication threshold, and completes the self-test and calibration of the sensor array to ensure the normal operation of the acquisition equipment.

[0121] (2) Data collection stage:

[0122] When a user is detected entering the long jump area, the multi-view camera and depth sensor of the biometric acquisition unit begin to synchronously collect facial images and body motion data, and transmit the collected raw data to the feature extraction and analysis unit at a fixed frequency;

[0123] (III) Feature extraction and fusion stage:

[0124] The feature extraction and analysis unit processes the facial image and human motion video respectively to extract the facial feature vector and gait feature vector. It uses a dynamic weighted fusion strategy to merge the two feature vectors into a multimodal feature vector to provide data support for subsequent identity authentication.

[0125] (IV) Identity Authentication Stage:

[0126] The multimodal feature vector is sent to the cloud database and similarity is calculated with the stored user biometric template. If the similarity is higher than the set threshold, the identity authentication is determined to be successful and the user's identity information is obtained. If the similarity does not meet the threshold, a secondary authentication process is triggered, requiring the user to provide additional identity verification information until the authentication is successful or the authentication operation is terminated.

[0127] (V) Data binding and update stage:

[0128] After successful authentication, the user's identity information is bound to the standing long jump test data and recorded in the user's sports profile. At the same time, the difference between the newly collected biometrics and the original template is analyzed. If the difference meets the update standard, the new feature data is uploaded to the cloud database and the user's biometric template is updated.

[0129] (6) Ending stage:

[0130] When the system stops running or receives a stop command, the multimodal identity recognition module stops data collection and authentication operations, disconnects from the cloud database, releases system resources, and saves the current model parameters and configuration information;

[0131] 5. Application value of the module:

[0132] 1. Improving user experience:

[0133] Achieve contactless and automated identity authentication, eliminating the need for users to carry identity credentials or perform manual operations, reducing waiting time, improving the convenience and smoothness of using the system, and providing users with more intelligent and efficient sports evaluation services;

[0134] (2) Ensuring data accuracy:

[0135] Through multimodal biometric recognition, we ensure that assessment data accurately matches the user's identity, avoiding data confusion or erroneous recording. This provides a reliable data foundation for user performance analysis and training effect evaluation, making exercise guidance and suggestions more targeted.

[0136] (3) Enhance system security:

[0137] Encrypted transmission and blockchain storage technologies are used to protect user biometric data and prevent identity information leakage and fraudulent use. Multimodal feature fusion and dynamic update mechanisms improve the anti-counterfeiting capability of identity authentication, effectively resist forgery attacks, and ensure the security and stability of system operation.

[0138] (IV) Support personalized services:

[0139] Based on accurate identity authentication and user sports profile management, the system can provide different users with personalized sports plans, training suggestions and performance analysis reports to meet users' diverse sports needs, promote users to continue using the system for physical exercise, and improve user stickiness and satisfaction.

[0140] In this embodiment, the three-dimensional posture analysis module further includes:

[0141] Multi-view cameras are used to synchronously capture stereo images of the long jump action, and the 3D motion trajectory is reconstructed using an improved stereo matching algorithm. The energy function of the algorithm is composed of the following three weighted components:

[0142] Data item: Evaluate the consistency of 3D pose through projection error of multi-view images;

[0143] Smoothness term: constrains the continuity of adjacent joint points in space;

[0144] Timing term: ensure the physical rationality of the posture changes between adjacent frames, weight coefficient 、 、 Control the contribution ratio of the three items separately;

[0145] Based on the temporal convolutional network (TCN), the joint motion sequence is analyzed to predict the center of mass offset at the moment of landing.

[0146] Combined with the inertial measurement unit (IMU) data, the acceleration error of the aerial attitude is corrected through a filtering algorithm to achieve sub-centimeter precision measurement;

[0147] Furthermore, the 3D Posture Analysis Module, serving as the "sports digital mirror" of the AI ​​standing long jump evaluation data management system for physical exercise, integrates multi-perspective visual perception with advanced algorithms to perform high-precision 3D modeling and dynamic analysis of long jump movements, providing key motion data support for the evaluation system. The following details its overall functionality, sub-module composition, key technical principles, workflow, and application value.

[0148] 1. Overview of overall functions:

[0149] The 3D Posture Analysis module uses multi-view cameras to synchronously capture long jump action videos and constructs a motion trajectory model using a 3D human posture estimation algorithm. This module can accurately detect key indicators such as take-off angle, aerial posture, and landing stability in real time. This module not only captures the motion trajectory of each human joint in three-dimensional space, but also quantitatively evaluates movement quality based on kinematic principles, providing users with detailed feedback on the movements and providing core data for intelligent violation determination, ensuring the scientific and fair evaluation results.

[0150] 2. Submodule composition and functions:

[0151] (1) Multi-view image acquisition unit:

[0152] Equipment deployment: Multi-view cameras are distributed in a circular or multi-angle pattern around the long jump area to ensure that the take-off point, flight path, and landing area are all within the camera's field of view. Camera parameters, including frame rate, resolution, and focal length, are uniformly calibrated to ensure temporal synchronization and spatial consistency of multi-view images.

[0153] Data acquisition: When the user begins a long jump, the multi-view camera synchronously captures the action video at a high frame rate (e.g., 60 fps) and sends the raw image data to the data processing unit in real time via a high-speed data transmission link, providing raw material for subsequent posture analysis.

[0154] (2) 3D motion trajectory reconstruction unit:

[0155] Stereo matching processing: An improved stereo matching algorithm is used to process multi-view images, and temporal consistency constraints are introduced to optimize matching accuracy; its energy function is

[0156] (in, Indicates the The three-dimensional posture parameters of the frame, including the coordinates of each joint of the human body in three-dimensional space; The image consistency is measured by calculating the multi-view image projection error. The smaller the error, the more accurate the matching. It is a smoothing term used to constrain the continuity of the spatial positions of adjacent joints to avoid sudden changes in the trajectory; It is a timing item that ensures that the posture changes between adjacent frames conform to physical laws and prevents jumps that do not conform to motion logic; 、 、 is a weight coefficient used to adjust the relative importance of each energy term and is determined through optimization of a large amount of experimental data); by minimizing the energy function, accurate matching of corresponding points in multi-view images is achieved, thereby reconstructing the motion trajectory of the human body in three-dimensional space;

[0157] Trajectory optimization: Filter the reconstructed 3D trajectory to remove abnormal data points caused by noise or matching errors, and use Kalman filtering or median filtering algorithms to smooth the trajectory curve to make the motion trajectory more consistent.

[0158] (3) Key indicator analysis unit:

[0159] Take-off angle calculation: Based on the reconstructed 3D posture data at the moment of take-off, the spatial vector angle between the center of gravity and the take-off and landing points is calculated, and the take-off angle is determined in conjunction with the ground coordinate system. The rationality of the take-off action is evaluated by analyzing the deviation between the take-off angle and the ideal angle.

[0160] Aerial posture assessment: Using a temporal convolutional network (TCN), the joint motion sequence is analyzed to predict the center of mass offset at the moment of landing. The TCN effectively captures the temporal dependencies of motion sequences, learns the patterns of human posture changes in the air, and determines whether there is any posture imbalance or unreasonable movement.

[0161] Landing stability testing: Combining inertial measurement unit (IMU) data with 3D trajectory information, the Kalman filter algorithm is used to correct for acceleration errors in mid-air posture, achieving sub-centimeter-level precision measurement. Landing stability is assessed by analyzing joint angles, body inclination, and impact force distribution at the moment of landing to determine whether there is a risk of falling or slipping.

[0162] (4) Data output unit:

[0163] The analyzed data, such as take-off angle, aerial posture parameters, and landing stability indicators, are formatted and output in the form of numbers and charts. At the same time, key data is transmitted to the intelligent violation determination engine and result display module to provide a basis for violation determination and provide users with an intuitive action analysis report.

[0164] 3. Key technical principles:

[0165] (1) Principle of multi-view stereo matching:

[0166] Multi-view stereo matching is based on the principle of triangulation. By finding the corresponding pixels of the same object point in multiple images, the camera's internal and external parameters are used to calculate its 3D coordinates. The improved stereo matching algorithm introduces temporal consistency constraints, considers the motion continuity between adjacent frames, and constructs an energy function combining data terms, smoothing terms, and timing terms. The optimal matching result is solved through an optimization algorithm, thereby improving the accuracy and stability of 3D reconstruction.

[0167] (2) Principle of temporal convolutional network analysis:

[0168] The Temporal Convolutional Network (TCN) is a deep learning model specifically designed to process sequential data. Through causal and dilated convolution operations, it can effectively capture long-range temporal dependencies without losing time series information. In 3D posture analysis, TCN uses joint motion sequences as input, learning the patterns and regularities of motion changes over time, enabling prediction of future postures and evaluation of current postures.

[0169] (3) Kalman filter correction principle:

[0170] The Kalman filter is an optimal estimation algorithm based on a linear system state space model. It estimates the system state through two steps: prediction and update. In 3D attitude analysis, the Kalman filter is used to fuse IMU data with visual measurement data to correct the acceleration error of the aerial attitude. Based on the system's state equation and observation equation, the state estimate is continuously updated to improve the accuracy of attitude measurement.

[0171] 4. Module workflow:

[0172] (1) Initialization phase:

[0173] After the 3D posture analysis module is started, it completes the parameter configuration and calibration of the multi-view camera, loads the initial parameters of the stereo matching algorithm, TCN model, and Kalman filter algorithm, establishes a data communication channel with other modules, and prepares to receive image data and control instructions;

[0174] (2) Data collection stage:

[0175] The multi-view image acquisition unit monitors the long jump area in real time. When the user's jump action is detected, the multi-view cameras start collecting data synchronously and quickly transmit the video data to the 3D motion trajectory reconstruction unit.

[0176] (3) Trajectory reconstruction stage:

[0177] The 3D motion trajectory reconstruction unit performs stereo matching processing on multi-view images, calculates 3D posture parameters by optimizing the energy function, reconstructs the human motion trajectory, and performs filtering optimization to remove data noise and outliers.

[0178] (IV) Key indicator analysis stage:

[0179] The key indicator analysis unit calculates the take-off angle, analyzes the aerial posture, and detects landing stability based on the reconstructed trajectory data. It uses the TCN network to predict the center of mass offset and combines IMU data with the Kalman filter algorithm to improve measurement accuracy.

[0180] (V) Data output stage:

[0181] The data output unit organizes and formats the analysis results and outputs them to the intelligent violation determination engine and result display module, while also storing key data for subsequent query and analysis;

[0182] (6) Ending stage:

[0183] When the system stops running or receives a stop command, the 3D posture analysis module stops data collection and processing, saves algorithm parameters and intermediate results, closes communication connections with other modules, and releases system resources;

[0184] 5. Application value of the module:

[0185] 1. Improving the professionalism of assessments:

[0186] Through high-precision three-dimensional posture analysis, it can quantify various technical indicators of long jump movements and provide users with professional and detailed movement analysis reports, helping users understand their own movement defects, improve training methods in a targeted manner, and enhance sports performance;

[0187] (2) Enhanced judgment accuracy:

[0188] Providing reliable data support for intelligent violation determination. Based on accurate analysis of take-off angle, aerial posture, and landing stability, it can accurately identify violations such as step jumps and line crossings, ensuring the fairness and authority of the evaluation results.

[0189] 3. Supporting personalized training:

[0190] Based on each user's movement characteristics and evaluation data, the system can develop a personalized training plan for the user, recommend appropriate training intensity and methods, meet the training needs of different users, and improve training effects and user satisfaction;

[0191] (IV) Promoting technological innovation:

[0192] This module integrates advanced technologies such as multi-perspective visual perception, deep learning and sensor fusion, providing practical experience and innovative ideas for the technological development in the field of sports evaluation, and promoting the application and expansion of related technologies in other sports.

[0193] In this embodiment, the determination method of the intelligent violation determination engine includes:

[0194] The continuous frame action sequence is modeled through an improved LSTM network, whose forget gate output is integrated with predefined kinematic constraints, and the constraint weight coefficient γ controls the contribution of physical priors;

[0195] Define a dynamic threshold rule base and adjust the judgment criteria based on user age, gender, and historical data. The threshold is the weighted average of the baseline threshold and the adaptive offset, and the balance coefficient μ controls the weight of the two.

[0196] Generates explainable violation reports, including action breakdowns and improvement suggestions, and provides real-time feedback to users via AR devices;

[0197] Furthermore, the intelligent violation determination engine, serving as the "rule guardian" of the AI ​​standing long jump evaluation data management system for physical exercise, deeply integrates the rule engine with the AI ​​time series analysis model. Based on kinematic parameters and spatial trajectory characteristics, it automatically and intelligently determines violations during the long jump. The following will elaborate on this module in terms of its overall function, sub-module composition, key technical principles, workflow, and application value:

[0198] 1. Overview of overall functions:

[0199] The intelligent violation determination engine automatically identifies violations such as step jumps, line crossings, and body imbalances through real-time analysis of human motion data and spatial trajectory information during the long jump, combined with a predefined rule library and AI algorithm model. The engine not only quickly and accurately determines the type of violation but also dynamically adjusts the determination criteria based on individual user characteristics (age, gender, historical data, etc.) to ensure the scientific and fair nature of the results. It also generates visual and interpretable violation reports and provides real-time feedback to users via AR devices, helping them to promptly understand the cause of the violation and improve their movements.

[0200] 2. Submodule composition and functions:

[0201] (1) Data receiving and preprocessing unit:

[0202] Data access: Establish data communication interfaces with other modules in the system, such as the 3D posture analysis module and the dynamic environment calibration module, to receive real-time 3D posture data (joint angles, center of mass position, etc.) of the human body during long jump, motion trajectory data, take-off and landing point coordinate data, and environmental calibration information;

[0203] Data cleaning and conversion: Clean the received data to remove noise and outliers, such as unreasonable data points caused by sensor errors or matching errors. At the same time, data in different formats are uniformly converted into a standard format that meets the engine's processing requirements, providing a high-quality data foundation for subsequent analysis.

[0204] (2) Violation Analysis Unit:

[0205] Action sequence modeling: An improved LSTM network is used to model continuous frame action sequences. Its gating mechanism introduces kinematic prior knowledge, and the update formula is: (in, It is the output of the forget gate, which is used to control the degree of retention of the information at the previous moment; is the weight matrix of the forget gate, which determines the influence weight of the input information on the output of the forget gate; is the hidden state at the previous moment, containing information about the previous action sequence; is the input at the current moment, i.e. the human motion data of the current frame; The bias vector of the forget gate is used to adjust the benchmark of the forget gate output; It is a predefined kinematic constraint matrix that stores kinematic rule information such as joint angle range, center of mass speed limit, etc. To constrain the weight coefficient, control the influence of kinematic constraints on the output of the forget gate; The activation function, usually a sigmoid function, is used to map the output value to between 0 and 1); the model is used to learn the time series characteristics of the long jump action and capture the continuity and abnormal changes of the action;

[0206] Dynamic threshold calculation: Define a dynamic threshold rule base and adjust the judgment criteria based on user age, gender and historical data. The threshold calculation formula is: (in, is the final judgment threshold; The baseline threshold is a universal standard derived from a large amount of normal long jump action data. It is an adaptive offset based on the user's historical data, reflecting the user's individual action characteristics and ability differences; is the balance coefficient, and its value range is , used to adjust the weight of the baseline threshold and the adaptive offset, which can be dynamically adjusted according to actual needs); through the dynamic threshold, the judgment standard is more in line with the individual situation of the user, reducing misjudgment;

[0207] Violation identification: Processed data is compared and analyzed with dynamic thresholds and kinematic rules. For example, abnormal movement of the center of mass before takeoff is detected to determine whether a step jump has occurred. The positional relationship between the takeoff point coordinates and the takeoff line is used to determine whether the line has been crossed. Changes in body joint angles and center of mass stability are used to assess whether there is body imbalance. If the data exceeds the threshold range or does not comply with the kinematic rules, a violation is determined and the type and time of the violation are recorded.

[0208] (3) Report generation and feedback unit:

[0209] Violation report generation: When a violation is detected, a detailed violation report is generated. The report includes a breakdown of the violation, noting the specific time point, body part, and movement deviation. In addition, based on kinematic principles and rules, the report provides an analysis of the cause of the violation and improvement suggestions, helping users understand the problem and conduct targeted training.

[0210] Real-time feedback: Violation reports are displayed to users in real time via AR devices. Users can intuitively see virtual annotations and prompts of their own violations, achieving "instant feedback" and allowing them to correct incorrect actions in a timely manner. In addition, violation results are synchronously transmitted to the system's data management module for subsequent performance recording and statistical analysis.

[0211] 3. Key technical principles:

[0212] (1) Principle of improved LSTM network:

[0213] The improved LSTM network integrates kinematic prior knowledge into the gating mechanism based on the traditional LSTM structure. The kinematic constraint matrix intervenes in the calculation of the forget gate, input gate, and output gate, enabling the network to follow the physical laws of human motion when learning action sequences. For example, when joint angles exceed the normal range, the network adjusts information transmission based on the constraint matrix, more accurately identifying abnormal action patterns and improving the ability to detect violations.

[0214] (2) Principle of dynamic threshold adjustment:

[0215] Dynamic threshold adjustment is based on statistics and personalized learning concepts; the baseline threshold provides a general judgment standard, and the adaptive offset analyzes the user's historical data (such as movement parameters and performance of multiple long jumps) to explore the user's individual movement characteristics and ability level; the balance coefficient It is used to balance the relationship between the two, so that the threshold can reflect the general rule while adapting to individual differences to achieve accurate judgment;

[0216] (3) Principles of integration of rule engine and AI:

[0217] The rules engine pre-defines the basic rules and kinematic constraints of the long jump competition, such as take-off rules and landing requirements. The AI ​​time series analysis model (such as an improved LSTM network) learns complex movement patterns and abnormal characteristics from large amounts of movement data. After the two are integrated, the rules engine provides clear judgment boundaries for the AI ​​model, and the AI ​​model compensates for the rules engine's shortcomings in recognizing complex movement patterns. The two complement each other and achieve comprehensive and accurate judgment of violations.

[0218] 4. Module workflow:

[0219] (1) Initialization phase:

[0220] After the intelligent violation determination engine is started, it loads pre-trained improved LSTM network model parameters, dynamic threshold rule library, kinematic constraint matrix and other configuration information; establishes data communication connections with other modules in the system, and prepares to receive relevant data during the long jump;

[0221] (2) Data reception and processing stage:

[0222] The data receiving and pre-processing unit receives data from the 3D posture analysis module in real time, performs cleaning and format conversion, and transmits the processed data to the violation analysis unit;

[0223] (III) Violation determination stage:

[0224] The violation analysis unit uses an improved LSTM network to model and analyze action sequences, combining dynamically calculated thresholds with kinematic rules to identify violations. Once a violation is determined, it records information such as the violation type and time, and passes the results to the report generation and feedback unit.

[0225] (IV) Report generation and feedback stage:

[0226] The report generation and feedback unit generates detailed reports based on violation information, feeds the reports back to users in real time through AR devices, and synchronizes the judgment results to other modules of the system;

[0227] (V) Ending stage:

[0228] When the system stops running or receives a stop instruction, the intelligent violation determination engine stops data processing and determination operations, saves model parameters and determination records, closes communication connections with other modules, and releases system resources;

[0229] 5. Application value of the module:

[0230] 1. Ensuring the fairness of the evaluation:

[0231] Through an automated and standardized violation determination process, human interference is eliminated, ensuring that all users are evaluated under the same rules, ensuring the fairness and impartiality of evaluation results, and enhancing the credibility of the evaluation system;

[0232] (2) Improving user training effects:

[0233] Real-time, intuitive violation feedback helps users identify and correct movement errors in a timely manner, deepens their understanding and memory of correct movements, accelerates the transformation of training results, and effectively improves users' long jump technical level and athletic performance;

[0234] (3) Reduce operating costs:

[0235] This reduces the need for manual refereeing, lowering both labor and time costs. Furthermore, the efficiency and accuracy of intelligent judgments reduce the need for repeated evaluations or dispute resolution due to judgment disputes, improving system operational efficiency.

[0236] (IV) Promoting the scientific development of sports:

[0237] The accumulation of a large amount of violation data and motion analysis results provides data support for sports training research and optimization of sports rules, helps to deeply understand the technical points and causes of violations in long jump, and promotes the scientific development of sports training and competition rules.

[0238] In this embodiment, the distributed data management unit further includes:

[0239] The blockchain-based smart contract mechanism ensures the integrity and traceability of score data during transmission and storage. Data hashing is achieved by concatenating the current data, a random number, and the hash value of the previous block, and using the SHA-3 algorithm to generate a unique identifier.

[0240] A federated learning framework is used to aggregate anonymized training data from multiple sites to optimize global AI model performance. The objective function is the weighted sum of the local losses of each site, and a regularization term is added to prevent overfitting.

[0241] Support users to independently manage their personal sports data through encrypted tokens NFT and authorize third parties to access;

[0242] Furthermore, the distributed data management unit, serving as the "data hub guardian" of the AI ​​standing long jump assessment data management system for physical exercise, comprehensively utilizes technologies such as blockchain and federated learning to achieve secure storage, efficient processing, and collaborative optimization of assessment data, ensuring data integrity, traceability, and privacy. The following will elaborate on this module in terms of its overall functionality, sub-module composition, key technical principles, workflow, and application value:

[0243] 1. Overview of overall functions:

[0244] The distributed data management unit uses blockchain technology to store standing long jump evaluation data, ensuring that the data cannot be tampered with and is traceable during transmission and storage. A multi-node synchronous verification mechanism also ensures data accuracy and consistency. The edge computing module enables localized data processing, reducing reliance on the cloud and improving data processing efficiency. Furthermore, a federated learning framework aggregates anonymized training data from multiple sites to optimize global AI model performance. Users can manage their personal sports data independently through non-cryptographic tokens (NFTs), enabling secure data sharing and authorized access, providing solid support for stable system operation and data value mining.

[0245] 2. Submodule composition and functions:

[0246] (1) Blockchain data storage unit:

[0247] Data on-chain: After the standing long jump test is completed, the test data (including user identity information, performance data, movement analysis data, violation determination results, etc.) is packaged and processed to generate a data block; the data block contains information such as data content, timestamp, and hash value of the previous block;

[0248] Hash generation and verification: using hash generation algorithm (in, Represents a data concatenation operation, which is used to combine different types of data together; It is a random number, and the hash value is adjusted by constantly changing its value to ensure that the hash result meets specific requirements; It is the hash value of the previous block, which is used to build the chain structure of the blockchain to ensure the continuity and traceability of data; Evaluation data stored for the current block, covering all kinds of user-related information; The hash value of the current block is generated, which is the unique identifier of the block data. The data block is hashed to generate a unique hash value. Through the multi-node synchronous verification mechanism, each node verifies the hash value of the data block to ensure that the data has not been tampered with.

[0249] Smart contract execution: Based on the blockchain-based smart contract mechanism, rules for data storage, access, and modification are pre-set. When specific conditions are met (such as data on-chain, user authorized access, etc.), the corresponding contract terms are automatically executed to ensure the integrity and traceability of performance data during transmission and storage, and prevent illegal operations.

[0250] (2) Federated Learning Optimization Unit:

[0251] Data collection and anonymization: Establish connections with the evaluation systems of multiple venues to collect anonymized training data generated by each venue, including user long jump data, environmental data, and violation data. The collected data is anonymized to remove sensitive user information, retaining only feature data related to model training to ensure data privacy and security.

[0252] Model collaborative training: Using the federated learning framework, each site trains AI models based on its own data locally (such as models for dynamic environment calibration, 3D posture analysis, etc.), and only uploads model parameters (not raw data) to the central server; the central server uses the objective function to train AI models. (in, are model parameters, which are the variables that need to be optimized in the training model of each site; For the The amount of data for each venue reflects the contribution of the venue data to the overall training; is the total amount of data for all sites, which is used to calculate the proportion of data for each site; For the The local loss function of each site measures the training error of the model on the data of that site; (where is the regularization coefficient, used to control model complexity and prevent overfitting) aggregates and updates the model parameters uploaded by each site, generates globally optimized model parameters, and distributes them to each site. Each site continues training using the updated parameters, optimizing the global AI model performance through multiple rounds of iterations.

[0253] Model evaluation and feedback: Regularly evaluate the optimized global model and judge its performance based on its performance on test datasets (e.g., evaluation accuracy, violation determination accuracy, and other indicators). Based on the evaluation results, provide feedback on model training results to each site, guiding them to adjust their training strategies or data processing methods to further improve model performance.

[0254] (3) Data rights management unit:

[0255] NFT token generation: Generates a cryptographic token (NFT) for each user, which serves as a unique identifier and management credential for the user's personal sports data. The NFT token is bound to the user's biometric information or account information to ensure uniqueness and non-replicability.

[0256] Permission management and authorization: Users can independently manage access rights to their personal sports data through NFT tokens and authorize third parties (such as sports coaches, scientific research institutions, etc.) to access part or all of the data. During the authorization process, the system records authorization information, including the authorized object, authorized data scope, authorization time, etc., to ensure the legitimacy and controllability of data access.

[0257] Data access control: When a third party applies to access user data, the system verifies its access rights. Only with valid authorization from the user can the third party access the corresponding data, and the access process is strictly monitored to prevent data leakage or abuse.

[0258] (4) Edge computing processing unit:

[0259] Local data processing: Deploy edge computing modules at the data generation end (such as the equipment end at the long jump venue) to perform preliminary processing on the collected raw data (such as multispectral sensor data and camera video data), including data cleaning, feature extraction, format conversion, and other operations, reducing data transmission volume and alleviating cloud computing pressure.

[0260] Real-time response support: For tasks with high real-time requirements (such as partial data processing for real-time violation determination and preliminary analysis of movement posture), the edge computing module quickly processes the data locally and promptly feeds back the results to the user or related modules, improving the system's response speed and user experience.

[0261] Data caching and synchronization: The edge computing module sets up a data caching mechanism to temporarily store some processed data locally; when network conditions permit, the cached data is synchronized to the cloud blockchain storage unit to ensure data integrity and consistency;

[0262] 3. Key technical principles:

[0263] (1) Principles of blockchain technology:

[0264] Blockchain uses distributed ledger technology to store data on multiple nodes, with each node maintaining a complete copy of the ledger. Through hashing algorithms and a chain structure, data immutability is ensured. Once data is recorded on the chain, any modification will cause the hash value to change, making it detectable by other nodes. Based on the decentralized nature of blockchain, smart contracts automatically execute pre-set rules, automating and standardizing data operations and ensuring data security and credibility.

[0265] (2) Principles of Federated Learning:

[0266] Federated learning is a distributed machine learning framework that allows participants to collaboratively train machine learning models without sharing original data. By training the model locally and uploading parameters, a central server aggregates the parameters, enabling joint model optimization. This approach protects data privacy while fully leveraging the value of data from multiple parties, improving the model's generalization and performance.

[0267] (3) Edge computing principles:

[0268] Edge computing moves computing and data processing tasks from the cloud to network edge devices (such as sensors, cameras, and local servers). By processing data at the source, it reduces data transmission latency, lowers cloud load, and improves system real-time performance and reliability. Furthermore, edge computing is combined with cloud computing to form a cloud-edge collaborative computing model, optimizing resource utilization efficiency.

[0269] 4. Module workflow:

[0270] (1) Initialization phase:

[0271] After the distributed data management unit is started, it loads the blockchain network configuration information, federated learning framework parameters, data permission management rules, and initialization settings of the edge computing module; establishes data communication connections with other modules of the system (such as the evaluation module and user management module), and prepares to receive and process data;

[0272] (2) Data generation and processing stage:

[0273] After the standing long jump test is completed, the test data is first transmitted to the edge computing processing unit. The edge computing module performs preliminary processing such as data cleaning and feature extraction. For tasks with high real-time requirements, local calculations are performed and the processed data is cached.

[0274] (3) Data storage stage:

[0275] The cleaned evaluation data is packaged to generate data blocks. The blockchain data storage unit calculates the block hash value through a hash generation algorithm and performs multi-node synchronous verification. After verification, the data block is stored according to the chain structure of the blockchain, and the smart contract is executed at the same time to record data storage related operations.

[0276] (IV) Model optimization stage:

[0277] The federated learning optimization unit collects anonymized training data from each site. Each site trains the AI ​​model locally and uploads the parameters. The central server aggregates and updates the model parameters based on the objective function, generates a globally optimized model, and sends it to each site for a new round of training to continuously optimize model performance.

[0278] (V) Data rights management stage:

[0279] The data rights management unit generates NFT tokens for users, and users use tokens to manage personal data access rights. When a third party applies to access data, the system verifies the permissions, controls data access based on the authorization status, and records the access operations.

[0280] (6) Data synchronization and termination phase:

[0281] The edge computing processing unit synchronizes the cached data to the blockchain storage unit at the appropriate time to ensure data integrity. When the system stops running or receives a stop command, the distributed data management unit stops data processing, saves relevant configurations and data, closes the communication connection, and releases system resources.

[0282] 5. Application value of the module:

[0283] (1) Ensuring data security and reliability:

[0284] Blockchain technology ensures the immutability and traceability of evaluation data, and the multi-node verification mechanism prevents data errors or malicious modifications, providing solid protection for the authenticity and integrity of the data and improving its credibility and availability.

[0285] (2) Improving model performance:

[0286] The federated learning framework breaks down data silos, aggregates data from multiple sites, and optimizes AI models, enabling them to learn richer features and patterns, improving their accuracy and generalization capabilities in different scenarios and enhancing the intelligence of the system.

[0287] (3) Protecting user privacy:

[0288] Through anonymization and strict data permission management, users can independently control access to their personal data and protect privacy during data sharing, meeting users' needs for data privacy protection and enhancing their trust in the system.

[0289] (IV) Improving system efficiency:

[0290] Edge computing reduces data transmission volume and cloud computing pressure, enables rapid data processing and real-time response, improves the overall system operation efficiency and user experience, reduces system operating costs, and enables the system to better cope with application scenarios with high concurrency and high real-time requirements.

[0291] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A physical exercise AI standing long jump evaluation data management system, characterized by: include: Dynamic Environmental Calibration Module: This module collects real-time field environmental data using a multispectral sensor array deployed around the long jump area. A deep learning-based adaptive calibration algorithm is used to establish a dynamic coordinate system. This algorithm integrates ambient light intensity, ground texture characteristics, and obstacle distribution to generate a virtual calibration grid without relying on fixed markers. The implementation method of the dynamic environment calibration module includes: A multispectral sensor array is used to obtain real-time environmental information about the long jump area, including light intensity, ground texture, and obstacle distribution. Based on the improved convolutional neural network (CNN), environmental features are extracted and virtual calibration grids that adapt to different site conditions are dynamically generated. The loss function of the improved CNN consists of two parts: Feature reconstruction loss: Calculates the mean square error between the input environment features and the reconstructed features, and the weight coefficient α is used to control its importance; Environmental robustness loss: By maximizing the tolerance of feature differences under light interference, the weight coefficient β is used to balance the anti-interference ability, and the sum of α and β is 1; Optimize calibration parameters through reinforcement learning algorithms to ensure the stability of the coordinate system in complex environments such as shadows and reflections; Multimodal identity recognition module: By integrating gait recognition and facial feature extraction technology with user biometric information in the cloud database, it achieves contactless identity authentication; 3D Posture Analysis Module: This module uses multi-view cameras to synchronously capture long jump action videos, builds a motion trajectory model using a 3D human posture estimation algorithm, and detects take-off angles, mid-air postures, and landing stability in real time. Intelligent Violation Detection Engine: Based on a rule engine and AI time series analysis model, combined with kinematic parameters and spatial trajectory characteristics, it automatically identifies violations including stepping, crossing the line, and body imbalance. Distributed data management unit: uses blockchain technology to store assessment data, supports multi-node synchronous verification and tamper-proof performance records, and realizes localized data processing through edge computing modules to reduce cloud dependence.

2. A physical exercise AI standing long jump evaluation data management system according to claim 1, characterized in that: The three-dimensional posture analysis module further includes: Multi-view cameras are used to synchronously capture stereo images of the long jump action, and the 3D motion trajectory is reconstructed using an improved stereo matching algorithm. The energy function of the algorithm is composed of the following three weighted components: Data item: Evaluate the consistency of 3D pose through projection error of multi-view images; Smoothness term: constrains the continuity of adjacent joint points in space; Timing term: ensures the physical rationality of the posture changes between adjacent frames. The weight coefficients λ1, λ2, and λ3 control the contribution ratio of the three terms respectively; Based on the temporal convolutional network (TCN), the joint motion sequence is analyzed to predict the center of mass offset at the moment of landing. Combined with the inertial measurement unit (IMU) data, the acceleration error of the aerial attitude is corrected through a filtering algorithm to achieve sub-centimeter precision measurement.

3. A physical exercise AI standing long jump evaluation data management system according to claim 1, characterized in that: The determination method of the intelligent violation determination engine includes: The continuous frame action sequence is modeled through an improved LSTM network, whose forget gate output is integrated with predefined kinematic constraints, and the constraint weight coefficient γ controls the contribution of physical priors; Define a dynamic threshold rule base and adjust the judgment criteria based on the user's age, gender and historical data. The threshold is the weighted average of the baseline threshold and the adaptive offset, and the balance coefficient μ controls the weight of the two. Generates explainable violation reports, including action breakdown diagrams and improvement suggestions, and provides real-time feedback to users via AR devices.

4. A physical exercise AI standing long jump evaluation data management system according to claim 1, characterized in that: The distributed data management unit further comprises: The blockchain-based smart contract mechanism ensures the integrity and traceability of score data during transmission and storage. Data hashing is achieved by concatenating the current data, a random number, and the hash value of the previous block, and using the SHA-3 algorithm to generate a unique identifier. A federated learning framework is used to aggregate anonymized training data from multiple sites to optimize global AI model performance. The objective function is the weighted sum of the local losses of each site, and a regularization term is added to prevent overfitting. Support users to independently manage personal sports data through encrypted tokens NFT and authorize third-party access.

Images