A high-level football referee intelligent training system based on computer vision

Abstract

The application discloses a high-level football referee intelligent training system based on computer vision and relates to the technical fields of sports training and artificial intelligence.The application comprises a hardware subsystem and a software subsystem which are connected with each other in communication.The hardware subsystem adopts a five-layer architecture and comprises a perception layer, a collection layer, a processing layer, an interaction layer and a transmission layer.The software subsystem adopts a four-layer architecture and comprises a perception layer, a platform layer, an application layer and a display layer.Through deep cooperation between software and hardware, multi-modal physiological signal fusion, generative AI virtual counter-training and blockchain storage authentication, a full-process digital referee training system covering theoretical teaching, video decision-making, gesture training, psychological assessment, outdoor collection and cross-institution authentication is constructed, systematic, intelligent and closed-loop training and quantitative evaluation of the four-dimensional abilities of referees, i.e., knowledge, decision-making, behavior and psychology, are realized, and the objectivity, individualization and public credibility of football referee training are significantly improved.

Description

Technical Field

[0001] This invention relates to the fields of sports training and artificial intelligence technology, specifically to a high-level intelligent training system for football referees based on computer vision. Background Technology

[0002] The training of football referees relies heavily on experience transfer and subjective evaluation. Currently, referee training both domestically and internationally primarily employs the traditional model of "theoretical instruction + video analysis + field practice." This existing technology has the following significant shortcomings: 1. The assessment is limited in scope and highly subjective: Existing systems primarily focus on online examinations of theoretical knowledge, lacking automated and quantitative assessments of referees' core officiating behaviors (such as the standardization of penalty gestures and running posture). Assessment results heavily rely on the examiners' subjective experience, resulting in insufficient consistency and fairness.

[0003] 2. The separation of "knowledge, decision-making, and action": Theoretical learning, video decision-making, and gesture practice are independent of each other, failing to form a closed loop of "knowledge acquisition - decision-making training - behavioral feedback", resulting in low training effectiveness and difficulty in transferring to actual combat.

[0004] 3. Lack of data collection on outdoor officiating behavior: The lack of non-intrusive, real-time data collection methods for referees' behavior in real or simulated match scenarios makes it impossible to form objective and traceable data archives of the officiating process, resulting in a lack of quantitative basis for promotion assessment.

[0005] 4. Low level of intelligence and limited training materials: There is a lack of a system that integrates large-scale intelligent question answering and automated short-answer question scoring. At the same time, the training relies on a limited library of real competition videos, which is prone to "memory effect" and affects the objectivity of the evaluation.

[0006] 5. Gap in Psychological and Physiological Assessment: Current technology completely ignores the impact of referees' psychological load, attention distribution, and stress level on the quality of their decisions, and cannot assess their psychological resilience.

[0007] 6. Insufficient data credibility and cross-institutional mutual recognition: Training data is stored on local servers, which poses a risk of tampering. Data silos exist between different institutions, making it impossible to form a nationally unified referee competence certification network.

[0008] 7. If the existing system lacks real-time performance and cannot meet the requirements of high frame rate attitude detection, a closed loop of problem and effect will be formed.

[0009] In summary, existing technologies cannot achieve systematic, intelligent, and closed-loop training and quantitative assessment of referees' four-dimensional abilities: knowledge, decision-making, behavior, and psychology. Summary of the Invention

[0010] The purpose of this invention is to provide a high-level intelligent training system for football referees based on computer vision, in order to solve the problems raised in the background art, such as how to objectively and quantitatively evaluate the referee's officiating behavior, how to construct a high-ecological-validity closed-loop training environment that integrates theoretical learning, high-pressure decision-making training, and real-time behavioral feedback, how to achieve distributed, non-intrusive collection and three-dimensional quantitative analysis of real outdoor officiating behavior, how to achieve intelligent interaction, personalized feedback, and data-driven management throughout the training process, how to assess the impact of the referee's psychological and physiological state on the quality of judgment, and how to ensure the immutability, traceability, and cross-institutional recognition of training and assessment data.

[0011] To this end, the present invention provides a high-level intelligent training system for football referees based on computer vision, comprising a hardware subsystem and a software subsystem that are interconnected.

[0012] The hardware subsystem includes a hardware sensing layer, a data acquisition layer, a processing layer, an interaction layer, and a transmission layer. Preferably, the hardware sensing layer includes: Distributed visual acquisition network: Composed of multiple intelligent visual nodes based on ESP32-CAM, each node adopts a modular stacked structure, including, from top to bottom, an upper protective shell, an ESP32 main control board, a camera module, optical lenses, a lower protective shell, and a lithium battery; the ESP32 main control board integrates a lightweight AI inference engine, used to run the MediaPipe model at the edge to extract 33 skeletal key points and 21 hand key points of the human body; multiple nodes are distributed and deployed at the four corners and both sides of the center line of the football field, communicating with the central processing unit via Wi-Fi / 4G / 5G to form a 360° blind-spot-free multi-view coverage network, used to capture the human skeletal key points and hand key points of the referee in real time in outdoor scenes; Referee physiological state acquisition module: includes a wearable heart rate monitoring device, eye-tracking glasses, and a skin conductance response sensor; the wearable heart rate monitoring device uses a MAX30102 heart rate sensor to collect heart rate variability (HRV), instantaneous heart rate, and heart rate recovery rate in real time, and communicates with edge computing nodes via Bluetooth Low Energy protocol; the eye-tracking glasses integrate an eye-tracking device to collect gaze point trajectory, gaze duration, pupil diameter changes, and saccade speed in real time; the skin conductance response sensor is integrated into a palm patch to measure skin conductance level (SCL) and skin conductance response (SCR). Intelligent goal sensing device: includes an array of infrared beam sensors densely arranged inside the goal frame to form a full-coverage infrared light curtain, used to capture football crossing events, timestamps and crossing directions in milliseconds.

[0013] Preferably, the hardware acquisition layer includes a multi-view high frame rate camera array, consisting of 8-12 synchronously triggered high frame rate cameras, evenly deployed behind the goal, along the diagonal of the field, and around the perimeter of the training field. It achieves microsecond-level synchronous triggering through the NTP network time protocol and is used to acquire real-time video streams of the entire training process.

[0014] Preferably, the hardware processing layer includes multiple distributed edge computing nodes and a central industrial server: The distributed edge computing nodes are evenly deployed around the training field and equipped with a lightweight MediaPipe posture estimation model for real-time extraction and data compression of key points of the front-end referees, with a processing latency of ≤30ms. The central industrial server is deployed in the local control room and is equipped with a posture compliance (PCI) quantitative assessment model, a psycho-physiological fusion assessment model, and a blockchain evidence storage unit. It is used to receive data from edge nodes and complete multi-source data fusion, PCI score calculation, psychological load index calculation, training data storage, and blockchain evidence storage.

[0015] Preferably, the hardware interaction layer includes a student terminal, an administrator terminal, an examiner evaluation terminal, and a VR / AR immersive training terminal; the VR / AR immersive training terminal is used to access a generative AI virtual adversarial training environment, enabling students to immerse themselves in judging virtual matches from a first-person perspective.

[0016] Preferably, the hardware transmission layer includes a Wi-Fi / 5G wireless communication unit and a gigabit Ethernet wired communication unit, used to realize low-latency data transmission and synchronization between hardware layers, with an end-to-end wireless transmission latency of ≤20ms.

[0017] The software subsystem includes a software perception layer, a platform layer, an application layer, and a presentation layer. Preferably, the software perception layer is used to interface with the hardware subsystem to complete data reception, format conversion, and validity verification.

[0018] Preferably, the software platform layer includes a user authentication module, a training engine module, an AI intelligent judgment engine module, a video analysis engine module, a generative AI training environment module, a data management module, and a cross-module communication unit. The user authentication module adopts a role-based access control (RBAC) model for user identity verification and hierarchical management of permissions. The training engine module is used to implement training process control, competition status management, precise countdown control, and command issuance. The AI ​​intelligent judgment engine module includes: Gesture Recognition and PCI Quantitative Assessment Submodule: Based on key points extracted by MediaPipe, a gesture compliance PCI quantitative assessment model is constructed. PCI is a quantitative score of 0-100 points calculated based on multiple sub-indicators such as joint angle error, movement range, and stability. By collecting and annotating more than 3,000 standard referee gesture sample images to train the classifier, the key point features of trainees are compared with the standard model to calculate the PCI score. The psycho-physiological fusion assessment submodule establishes a multimodal physiological characteristic-decision quality correlation model and generates the psychological load index (MWI), calculated using the following formula: In the formula , and The weighting coefficients are determined through principal component analysis. When the MWI exceeds a preset threshold, a psychological adjustment prompt is automatically pushed.

[0019] Three-level review management unit: The review level is divided into three levels according to the confidence level of the penalty - Level 1 review (confidence level ≥ 0.9) The system automatically completes the penalty scoring; Level 2 review (confidence level 0.7-0.9) is pushed to the examiner's terminal for manual confirmation; Level 3 review (confidence level < 0.7) triggers the final manual decision; The video analysis engine module is used to realize multi-view video synchronization, slow motion, frame-by-frame viewing and 3D trajectory reconstruction, and generate eagle eye 3D playback view; The generative AI training environment module includes: Virtual adversarial scenario generator: Based on the diffusion model or generative adversarial network, using real match video clips as training material, it automatically generates an unlimited number of dynamically changing virtual match adversarial scenarios, and can adjust parameters such as match situation, score pressure, time pressure, foul type, difficulty of judgment, and environmental interference. Adaptive Difficulty Adjuster: Based on the student's historical ACC and RT data, the difficulty parameters of the virtual scene are dynamically adjusted using reinforcement learning algorithms. The difficulty is automatically increased when the student makes 5 consecutive correct judgments and the RT is below the threshold, and the difficulty is automatically decreased when the student makes 3 consecutive incorrect judgments. Multi-person collaborative training sub-module: Supports multiple trainees to access the same virtual competition scene at the same time, each playing a different referee role, and generates a team collaboration effectiveness evaluation report; The data management module includes a blockchain evidence storage unit: The key data of each training session and assessment (theoretical exam scores, ACC / RT, PCI scores, MWI index, virtual training performance) of trainees are generated into encrypted hash values ​​and uploaded to the consortium blockchain for notarization, ensuring the authenticity, integrity and immutability of the data; The system automatically generates verifiable digital certificates of referee competence based on full-cycle training data. These certificates include trainee digital identity identifiers, quantitative competence scores for various dimensions, blockchain-based evidence record summaries, and digital signatures from the issuing authority. The referee promotion rules are encoded into smart contracts, and the promotion review process is automatically triggered when the trainees' various ability indicators meet the preset conditions. It supports automatically retrieving blockchain-stored evidence data to generate a complete and verifiable evidence package during the appeal process, and the entire review process is stored on the blockchain.

[0020] Preferably, the software application layer includes: Student training console: integrates functions such as theoretical exams, video judgment practice, gesture and posture simulation training, VR / AR immersive training, error notebook, and personal learning profile; Administrator backend: integrates question bank management, test paper management, student management, exam record and dashboard statistics functions; Examiner Evaluation Workbench: Integrates functions such as viewing AI-generated penalty suggestions, retrieving Eagle Eye replays, confirming and modifying scores, and three-level review processing.

[0021] Preferably, the system can be configured as a competition assessment mode and a daily training mode; in the training mode, the system's automatic scoring function can be configured to be used only for recording and feedback, and all detection events and AI analysis results serve the trainees' skill improvement and training review.

[0022] The present invention proposes a high-level intelligent training system for football referees based on computer vision, the advantages of which are as follows: (1) Objectification and standardization of evaluation: This invention pioneered the application of computer vision systems to the evaluation of football referee gestures and defined a quantitative index for posture compliance (PCI). Tests show that the intragroup correlation coefficient (ICC) between the AI ​​score and expert review in this invention reaches 0.87, achieving "excellent" consistency and completely changing the subjective evaluation model.

[0023] (2) "Knowledge-Decision-Action-Mind" four-dimensional integrated closed loop: It integrates theoretical learning, video decision-making, real-time behavioral feedback and psychological state assessment. Through instant linkage and physiological-psychological fusion assessment, it solves the problem of the disconnect between theory and practice and the neglect of psychological factors in traditional training, and the training ecology has extremely high validity.

[0024] (3) Intelligent and personalized: It integrates a large-scale intelligent assistant to realize 24 / 7 rule-based question and answer and short-answer intelligent evaluation. Through the generative AI-driven virtual adversarial training environment, it can generate a variety of training scenarios and adaptively adjust the difficulty to achieve personalized and precise training for each person, effectively avoiding the "memory effect".

[0025] (4) Traceability of outdoor adjudication behavior: Based on the distributed vision network of ESP32-CAM, the system realizes seamless, full-process, traceable collection and quantitative analysis of real outdoor adjudication behavior. Hardware prototype testing shows that the continuous working time of the nodes is ≥8 hours, the key point recognition frame rate is ≥15fps, and the end-to-end latency is <500ms.

[0026] (5) Superior performance: On a standard client, the system can stably maintain real-time attitude detection and rendering at 25 frames per second, with a total end-to-end latency of less than 200 milliseconds. The average accuracy (ACC) of the video penalty module is 89.2%, and the average reaction time (RT) is 1.8 seconds.

[0027] (6) Quantitative assessment of psychological stress resistance: For the first time, multimodal physiological signals such as heart rate, eye movement, and skin conductance are introduced into the referee training assessment system. The psychological stress resistance and decision fatigue of referees are quantitatively assessed through the MWI index, providing data support for targeted psychological training.

[0028] (7) Data credibility and cross-institutional mutual recognition: The blockchain-based evidence storage and authentication mechanism ensures the immutability and traceability of training and assessment data, providing "ironclad proof" for referee promotion. The smart contract-driven automatic promotion review and cross-institutional data mutual recognition have built a nationally unified referee competency certification network. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a block diagram of the overall system architecture of the present invention; Figure 2 This is an exploded view of the intelligent vision node (ESP32-CAM) of the present invention; Figure 3 This is a flowchart of the AI ​​intelligent judgment and three-level review process of the present invention; Figure 4 This is a flowchart illustrating the multimodal physiological signal fusion evaluation process of the present invention. Figure 5 This is a system interaction diagram of the generative AI-driven virtual adversarial training environment of the present invention; Figure 6 A business process diagram for blockchain-enabled adjudication evidence storage and authentication in this invention; Figure 7 This is a schematic diagram of the student training console interface of the present invention (theoretical exam / video judgment / posture detection); Figure 8 This is a schematic diagram of the attitude detection and PCI scoring real-time feedback interface of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described in this specification are merely for explaining the invention and are not intended to limit the invention.

[0032] Example: Please see Figure 1-8 This invention provides a high-level intelligent training system for football referees based on computer vision. Taking a standard indoor training field of a football training base as an example, the system is deployed in its entirety.

[0033] Hardware awareness layer deployment: See Figure 2 Distributed visual acquisition network deployment: Six intelligent visual nodes are deployed at the four corners and both sides of the center line of the training field. Each node's camera module 3 and optical lens 4 are responsible for capturing video of the referee's gestures. The ESP32 main control board 2 runs a lightweight MediaPipe model at the edge, extracting 33 key points of the referee's skeletal structure and 21 key points of their hands in real time, and transmitting the key point data to the central processing unit via a Wi-Fi module. The upper protective shell 1 and lower protective shell 5 form an IP67-rated waterproof sealed cavity, with a built-in lithium battery 6 providing independent power to the nodes, supporting continuous operation for more than 8 hours.

[0034] Referee physiological status acquisition module deployment: Each trainee is equipped with a wearable heart rate monitoring wristband (integrating a MAX30102 sensor), eye-tracking glasses (integrating a Tobii eye-tracking device), and a palmar electrodermal response patch. All devices are paired with edge computing nodes via Bluetooth Low Energy protocol to collect physiological indicators such as heart rate variability (HRV), gaze trajectory, pupil diameter, and skin conductance response (SCL) in real time, with a sampling frequency of 10Hz.

[0035] Deployment of intelligent goal sensing device: On the inner side of the two goal frames, multiple rows of densely packed infrared photoelectric sensors (model E3F-20C1-L) are densely arranged along the horizontal and vertical sides of the goal frames to form a full-coverage infrared light curtain. A total of 100 pairs of infrared photoelectric tubes are arranged, with a response time of ≤1ms. It communicates with the edge computing node through RS485 bus.

[0036] For the hardware acquisition layer, eight synchronously triggered high-frame-rate network cameras (4-megapixel resolution, 60FPS) are deployed around the training field: one behind each goal, two diagonally across the field, and two on each side of the long side, achieving 360° coverage of the training field. All cameras are directly connected to the edge computing nodes via gigabit Ethernet cables, and microsecond-level synchronous triggering is achieved through the NTP network time protocol.

[0037] Hardware processing layer deployment: Four edge computing boxes (Jetson XavierNX model) are evenly deployed around the training field as distributed edge computing nodes. Each node connects to two cameras, runs on the Ubuntu 20.04 operating system, and comes pre-installed with a lightweight MediaPipe pose estimation model to complete real-time extraction of referee key points from the front-end video stream, with edge processing latency ≤30ms. The pre-processed data is transmitted to the central industrial server via a gigabit wired network.

[0038] The central industrial server is deployed in the local control room and is equipped with a PCI quantitative assessment model for attitude compliance, a psycho-physiological fusion assessment model, and a blockchain evidence storage unit. It is responsible for receiving preprocessed data from edge nodes, completing PCI score calculation, MWI index calculation, AI penalty suggestion generation, full system data storage, and blockchain on-chain evidence storage.

[0039] Hardware interaction layer deployment: 20 ordinary laptops are provided to trainees as training terminals, which can access the system's web interface via Wi-Fi 6 network; one ruggedized tri-proof tablet is provided to each administrator and examiner, which can communicate with the central server in real time via a dedicated 5G network; 5 HTC VIVE headsets are provided for VR / AR training mode, which can be used to access the generative AI virtual adversarial training environment.

[0040] Hardware transmission layer deployment: A dedicated Wi-Fi 6 wireless network is built, using 3 enterprise-grade wireless APs, evenly deployed on the top of the training field to achieve full coverage, with end-to-end wireless transmission latency ≤20ms; a gigabit wired local area network is built, with all edge computing nodes, central server, and cameras directly connected via the wired network; a 5G private network is deployed for remote examiner access and cross-institutional data mutual recognition.

[0041] Software System Deployment: The software perception layer deploys a data access gateway to connect with all sensing and acquisition devices in the hardware subsystem, completing data reception, format conversion, and validity verification. The platform layer adopts a microservice architecture, integrating a user authentication module (RBAC model), a training engine module, an AI intelligent judgment engine module (including PCI quantitative assessment, psycho-physiological fusion assessment, and a three-level review unit), a video analysis engine module, a generative AI training environment module (including a diffusion model scene generator and a reinforcement learning difficulty adjuster), and a data management module (including a blockchain evidence storage unit). The application layer develops four dedicated function entry points: a student training console, an administrator management backend, an examiner evaluation workbench, and a multi-user collaborative training interface. The presentation layer uses the Vue3 + ElementPlus front-end framework to provide a visual operation interface for all users. The blockchain evidence storage unit connects to the Hyperledger Fabric consortium blockchain, interoperating with provincial football association training center nodes to achieve cross-institutional data mutual recognition.

[0042] Example 2: Indoor intelligent training and assessment (including psycho-physiological integrated assessment), based on the system deployed in Implementation Example 1, taking the standard process of trainees using the system of this invention for indoor training as an example: 1. Login and Registration: Students log in by entering their name and ID through the front-end interface. The system verifies the user through the ` / api / user / login` interface; new users are automatically registered.

[0043] 2. Theoretical Learning: Students click the "Knowledge Explanation" button to enter the knowledge explanation module. Based on the pre-generated `knowledge_content.json` index, the system displays digital human explanation videos and PPT / PDF courseware for students to access and learn from.

[0044] 3. Video Judgment Training (Decision-Making Training): Trainees enter the [Video Posture Judgment] module. The system randomly selects 5 video clips from the video library, and trainees watch them and make judgment choices. The system records the trainee's judgment options and automatically calculates the matching degree with the standard answer, providing real-time feedback on judgment accuracy (ACC) and reaction time (RT). Simultaneously, the system uses wearable devices to synchronously collect the trainee's heart rate variability (HRV), pupil diameter changes, and skin conductance response levels during this task, which are used for subsequent calculation of the psychological load index.

[0045] 4. Officiating Behavior Training and PCI Quantitative Assessment: The trainee clicks on [Posture Practice] and turns on the camera. The system calls the ` / api / pose / detect` interface to send video frames to the backend. The backend uses the MediaPipe model to extract key points, performs recognition through a proprietary gesture classifier, and calculates the PCI score. The frontend interface draws skeletal lines in real time, displays the recognition results and PCI score, and provides textual improvement suggestions such as "Arm needs to be raised higher."

[0046] 5. Psychophysiological Integration Assessment: Based on the collected physiological data, the system automatically calculates the Psychological Load Index (MWI). When a trainee's MWI exceeds the preset threshold of 0.75 in a high-pressure scenario (such as penalty kick judging), the system sends a notification: "Your current psychological load is high. It is recommended to pause training and take deep breaths to relax." After the assessment, the system generates a psychological stress resistance score, which is included in the trainee's comprehensive ability assessment report.

[0047] 6. Comprehensive Assessment and AI-Powered Intelligent Review (Three-Tier Review): Administrators distribute comprehensive exam papers through the exam paper management system in the management backend. After students answer the questions, objective questions are automatically scored, while short-answer questions are reviewed intelligently using a large model. The system divides the review into three levels based on confidence level: judgments with a confidence level ≥ 0.9 are automatically archived (Level 1 Review); judgments with a confidence level between 0.7 and 0.9 are pushed to the examiner's workbench for manual confirmation (Level 2 Review); judgments with a confidence level < 0.7 trigger Eagle Eye 3D playback and final manual decision-making (Level 3 Review).

[0048] 7. Blockchain Evidence Storage and Digital Certificates: Each training session and assessment (ACC, RT, PCI, MWI, etc.) of the trainees generates an encrypted hash value and uploads it to the consortium blockchain for evidence storage. After completing all training courses and meeting all indicators, the system automatically generates a digital certificate of referee competence, which includes the trainee's DID identity identifier, quantitative scores for each dimension, and a summary of the blockchain evidence storage record. The certificate becomes effective after being digitally signed by the issuing authority.

[0049] Example 3: Generative AI-driven virtual adversarial training, based on the system deployed in Implementation Example 1, takes virtual adversarial training using VR / AR immersive mode as an example: 1. Scene Generation: Trainees wear HTC VIVE headsets and enter the VR immersive training module. The system calls the virtual adversarial scene generator of the generative AI training environment module to automatically generate a brand-new virtual match scene based on the diffusion model: the second half of the final, stoppage time, score 2-2, home crowd booing loudly, away team player goes down in the penalty area. The system sets the parameter configuration of this scene (score pressure, time pressure, environmental interference) to a high difficulty level.

[0050] 2. Adaptive Difficulty Adjustment: The system retrieved the student's historical data and found that the student's accuracy rate in the "penalty kick judgment" scenario was only 65%, below the average level. The adaptive difficulty adjuster automatically increases the frequency of penalty kick controversy scenarios and reinforces the key judgment points in the current scenario.

[0051] 3. Immersive Officiating and Real-Time Feedback: Trainees officiate in a virtual stadium from a first-person perspective, making rulings using gesture recognition devices. The system recognizes the gestures in real time and calculates the PCI score, while simultaneously monitoring the trainee's physiological indicators through wearable devices. When a trainee makes a correct ruling, the system provides positive reinforcement; when a misjudgment occurs, the system automatically triggers Hawk-Eye replay, showing the controversial moment from multiple angles and providing rule hints.

[0052] 4. Multi-person collaborative training: Three trainees simultaneously access the system, playing the roles of head referee, assistant referee, and fourth official, respectively. The system records indicators such as the frequency of communication among the three, consistency of rulings, and role cooperation, generates a team collaboration effectiveness evaluation report, and stores it on the blockchain.

[0053] Example 4: Distributed data collection and evaluation of outdoor officiating behavior, based on the system deployed in Implementation Example 1, taking the data collection and evaluation of referee performance in outdoor standard football fields as an example: 1. Deployment and Data Acquisition: Six intelligent vision nodes are fixed to the four corner flag areas and the light poles on both sides of the center line of the football field (3-4 meters high) using brackets. After the system is started, the camera modules and optical lenses of each node acquire video of the referee. The ESP32 main control board extracts 33 key points of the referee's human skeleton and 21 key points of his hands in real time at the edge end, and transmits the key point data to the central processing unit on the sidelines via Wi-Fi / 5G.

[0054] 2. Fusion and Analysis: The central processing unit receives multi-view data, performs time synchronization and three-dimensional motion trajectory reconstruction. It calls upon the referee behavior model to automatically calculate quantitative indicators such as the referee's PCI score, movement efficiency, and average reaction time (RT) during the entire field movement.

[0055] 3. Assessment and Filing: All data is encrypted and uploaded to the consortium blockchain for evidence storage, generating an immutable electronic adjudication file for the referee. Examiners can view these quantitative indicators in the background. If the referee's "offside flag signal" PCI score is found to be below 80 points (the passing score), targeted indoor training courses can be pushed to them, forming a precise closed loop of "practice-data-supplementary training".

[0056] Example 5: Cross-institutional mutual recognition and smart contract advancement, based on the consortium blockchain network deployed in Implementation Example 1, taking the certification of a student in training institutions in different provinces as an example: 1. Cross-institutional data recognition: Trainee A completed the intermediate referee course at the Football Association Training Center of Province A, and their training data (theoretical scores, PCI ratings, MWI index, etc.) has been stored on the blockchain. Trainee A then transferred to the Football Association Training Center of Province B to participate in the advanced referee course. Province B's system automatically retrieved Trainee A's historical stored data from the consortium blockchain, eliminating the need for repeated assessments of the intermediate course content, allowing direct entry into the advanced course.

[0057] 2. Automated Smart Contract Promotion: After completing the advanced course, Student A's various performance indicators (PCI ≥ 88 points, ACC ≥ 92%, RT ≤ 1.6 seconds, MWI ≤ 0.65) meet the promotion requirements. The smart contract automatically detects that the conditions have been met, triggering the promotion review process and sending a notification to the examiner stating "Student A has met the national-level referee promotion standards" along with a complete evidence package. After the examiner's review and confirmation, the system automatically generates a national-level referee digital certificate and stores it on the blockchain. The entire process requires no manual application or paper approval.

[0058] Example 6: Verifiable appeals and reviews, based on the system deployed in Implementation Example 1, taking a student's appeal against assessment results as an example: 1. Appeal Initiation: Student B disagrees with the "offside penalty error" result given by the system in a certain video penalty assessment and initiates an online appeal through the student training console.

[0059] 2. Evidence Package Generation: The system automatically retrieves the blockchain-stored evidence data for this assessment, including the original answer records, the video clip of the judgment, the confidence level of the AI ​​judgment, PCI score details, and the time series of the trainee's physiological indicators (heart rate, pupil data) at that moment, to generate a complete and verifiable evidence package.

[0060] 3. Review and Adjudication: The examiner reviews the evidence package through the examiner evaluation workbench and retrieves the Eagle Eye 3D playback for review. If the review confirms an error in the system's judgment, the examiner modifies the judgment and fills in the review comments. All review operations (examiner identity, operation time, viewed records, modified content) are recorded on the blockchain for evidence preservation.

[0061] 4. Result Synchronization: The corrected score is automatically synchronized to the student's training console, and student B can view the complete chain of verification evidence. The system also updates the relevant ability scores in student B's digital certificate.

[0062] Example 7: In daily training mode, the system can be configured as a training mode. In this mode, the automatic penalty function and the three-level review process are disabled. All detection events and AI analysis results (ACC, RT, PCI, MWI) are used only for recording and feedback and are not included in the formal assessment score. Coaches can view trainees' training data in real time through the system, retrieve video replays and 3D trajectories for any time period, and provide real-time explanations and guidance for trainees' operational errors. After training, the system automatically generates a personal skills assessment report and psychological state analysis report for each trainee, supporting data export in multiple formats such as JSON, Excel, and PDF for post-competition review and targeted training plan development.

[0063] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A high-level intelligent training system for football referees based on computer vision, characterized in that: This includes hardware and software subsystems that are interconnected. The hardware subsystem includes a hardware perception layer, a data acquisition layer, a processing layer, an interaction layer, and a transmission layer. The hardware perception layer is used to acquire core physical data and physiological state data of the referee's officiating behavior. The data acquisition layer is used to acquire video streams and environmental data throughout the training process. The processing layer is used to preprocess, fuse multiple sources, and perform intelligent analysis on the acquired data. The interaction layer is used to provide human-computer interaction interfaces for different users. The transport layer is used to achieve low-latency data transmission and synchronization between different hardware layers; The software subsystem includes a software perception layer, a platform layer, an application layer, and a presentation layer. The software perception layer is used to interface with the hardware subsystem to receive, convert, and validate data. The platform layer is used to implement the core training logic processing and intelligent analysis. The application layer is used to provide dedicated function interfaces for different user roles. The presentation layer is used to provide users with a visual operation and viewing interface.

2. The intelligent training system for high-level football referees based on computer vision according to claim 1, characterized in that: The hardware perception layer includes a distributed visual acquisition network, a referee physiological state acquisition module, and an intelligent goal perception device. The distributed visual acquisition network consists of multiple intelligent visual nodes based on ESP32-CAM. Each node adopts a modular stacked structure, which includes, from top to bottom, an upper protective shell, an ESP32 main control board, a camera module, an optical lens, a lower protective shell, and a lithium battery. The ESP32 main control board integrates a lightweight AI inference engine, which is used to run the MediaPipe model at the edge to extract 33 skeletal key points and 21 hand key points of the human body; multiple nodes are distributed and deployed at the four corners and both sides of the center line of the football field, and communicate with the central processing unit through Wi-Fi / 4G / 5G to form a 360° blind spot-free multi-view coverage network, which is used to capture the human skeletal key points and hand key points of the referee in real time in outdoor scenes; The referee's physiological state acquisition module includes a wearable heart rate monitoring device, eye-tracking glasses, and a skin conductance response sensor. The wearable heart rate monitoring device uses a MAX30102 heart rate sensor to collect heart rate variability, instantaneous heart rate, and heart rate recovery rate in real time, and communicates with edge computing nodes via Bluetooth Low Energy. The eye-tracking glasses integrate an eye-tracking device to collect gaze point trajectory, gaze duration, pupil diameter changes, and saccade speed in real time. The skin conductance response sensor is integrated into a palm patch to measure skin conductance levels and skin conductance response. The intelligent goal sensing device includes an array of infrared beam sensors densely arranged inside the goal frame, forming a full-coverage infrared light curtain for capturing football crossing events, timestamps, and crossing directions in milliseconds.

3. The intelligent training system for high-level football referees based on computer vision according to claim 1, characterized in that: The acquisition layer includes a multi-view high frame rate camera array and training environment acquisition equipment; The multi-view high frame rate camera array includes 8-12 synchronously triggered high frame rate cameras, which are evenly deployed behind the goal, diagonally across the field, and around the perimeter of the training field. The microsecond-level synchronous triggering is achieved through the NTP network time protocol, which is used to collect real-time video streams of the entire training process. The training environment acquisition device is used to collect environmental data on lighting, temperature, and humidity in the training field, providing data support for adaptive optimization of the system algorithm environment.

4. The intelligent training system for high-level football referees based on computer vision according to claim 1, characterized in that: The hardware processing layer includes multiple distributed edge computing nodes and a central industrial server. The distributed edge computing nodes are evenly deployed around the training field. Each node connects to 2-3 cameras and is equipped with a lightweight MediaPipe pose estimation model for real-time extraction and data compression of key points of the front-end referees. The central industrial server is deployed in the local control room and is equipped with a posture compliance quantification assessment model, a psycho-physiological fusion assessment model, and a blockchain storage unit. It is used to receive preprocessed data from edge nodes, complete multi-source data fusion, posture compliance PCI score calculation, psychological load index MWI calculation, training data storage and blockchain storage. The storage adopts a combination architecture of PostgreSQL relational database and Redis cache database.

5. The intelligent training system for high-level football referees based on computer vision according to claim 4, characterized in that: The hardware interaction layer includes a student terminal, an administrator terminal, an examiner evaluation terminal, and a VR / AR immersive training terminal. The student terminal uses a general-purpose computer, smartphone, or tablet computer, and accesses the corresponding functional modules of the system through a browser. The administrator terminal and examiner evaluation terminal are ruggedized tri-proof tablets or laptops that communicate with the central industrial server in real time. The VR / AR immersive training terminal is used to access a generative AI virtual adversarial training environment, enabling trainees to immerse themselves in virtual matches from a first-person perspective.

6. The intelligent training system for high-level football referees based on computer vision according to claim 4, characterized in that: The hardware transmission layer includes a dedicated wireless communication unit, a wired communication unit, and a data synchronization module; The dedicated wireless communication unit uses Wi-Fi 6 or 5G private network to cover the entire training ground and is used for wireless communication connection between the perception layer, acquisition layer, interaction layer and processing layer. The wired communication unit uses a gigabit Ethernet line for wired direct connection between distributed edge computing nodes and the central industrial server. The data synchronization module is used to achieve real-time timeline synchronization of the entire system's video stream, sensor data, training control commands, and judgment results.

7. The intelligent training system for high-level football referees based on computer vision according to claim 1, characterized in that: The software platform layer includes a user authentication module, a training engine module, an AI intelligent judgment engine module, a video analysis engine module, a generative AI training environment module, a data management module, and a cross-module communication unit. The user authentication module adopts a role-based access control (RBAC) model for user identity verification and hierarchical management of permissions. The training engine module is used to implement training process control, competition status management, precise countdown control, and command issuance. The AI ​​intelligent judgment engine module includes: a gesture recognition and PCI quantitative assessment submodule, a psychological-physiological fusion assessment submodule, and a three-level review management unit; The video analysis engine module is used to realize multi-view video synchronization, slow motion, frame-by-frame viewing and 3D trajectory reconstruction, and generate eagle eye 3D playback view; The generative AI training environment module includes a virtual adversarial scenario generator, an adaptive difficulty adjuster, and a multi-person collaborative training sub-module. The data management module includes a blockchain evidence storage unit, which is used to generate encrypted hash values ​​for key data of each training and assessment of trainees and upload them to the consortium blockchain for evidence storage. The cross-module communication unit is used to realize real-time data interaction between modules of the platform layer and between the platform layer and the upper and lower layers.

8. The intelligent training system for high-level football referees based on computer vision according to claim 7, characterized in that: The specific configuration of the three-level review management unit of the AI ​​intelligent judgment engine module is as follows: Level 1 review: The confidence level of the event judgment is ≥0.

9. The system automatically completes the judgment scoring and synchronizes the results, and automatically confirms and archives the results without manual intervention. Level 2 review: If the confidence level of the event judgment is in the range of 0.7-0.9, the system pushes an AI judgment suggestion containing a complete chain of evidence to the examiner's terminal, which is then confirmed or modified by the examiner to complete the decision. Level 3 review: If the confidence level of the event judgment is <0.7, the system pushes the original event data to the examiner's terminal, triggers the Eagle Eye 3D playback retrieval entry, and the examiner makes the final decision entirely manually.

9. The intelligent training system for high-level football referees based on computer vision according to claim 1, characterized in that: The software application layer includes a student training console, an administrator management backend, and an examiner evaluation workbench. The student training console integrates functions such as theoretical exams, video judgment practice, gesture and posture simulation training, VR / AR immersive training, error notebook, and personal learning profile. The administrator management backend integrates question bank management, test paper management, student management, examination records and dashboard statistics functions; The examiner evaluation workbench integrates functions such as viewing AI-generated penalty suggestions, retrieving Eagle Eye replays, confirming and modifying scores, and three-level review processing.

10. The intelligent training system for high-level football referees based on computer vision according to claim 1, characterized in that: The system can be configured as a competition assessment mode and a daily training mode; In the competition assessment mode, the system enables full-function judgment and three-level review process, and the results are officially recorded in the student's file; In daily training mode, the system's automatic scoring function is configured to be used only for recording and feedback. All detection events and AI analysis results are used only for recording and feedback to serve the trainees' skill improvement and training review.

Images