Tennis ball real-time AI evaluation system based on multi-view visual feature fusion

Abstract

The invention relates to the technical field of sports match analysis, and discloses a tennis real-time AI evaluation system based on multi-view visual feature fusion, which comprises the steps of acquiring multi-angle video stream data of a tennis match, fusing multi-view data by using a time sequence diagram convolutional network to generate spatial-temporal feature representation, and constructing a time sequence cause and effect diagram model based on the fused spatial-temporal features. Key turning points of a match are identified by utilizing graph structure analysis and event influence evaluation, the influence of the turning points is verified through an anti-factual reasoning technology, and personalized tactical suggestions are generated in combination with a player feature model. According to the method, mining and understanding of a deep causal mechanism of the competition are realized by constructing the time sequence causal graph model and identifying the key turning points, and more accurate technical and tactical analysis and decision support are provided for coaches and players.

Description

Technical Field

[0001] The present invention relates to the technical field of sports game analysis, and more specifically, it relates to a real-time AI evaluation system for tennis based on the fusion of multi-perspective visual features. Background Art

[0002] With the development of data science and computer vision technology, sports game analysis is gradually shifting from traditional experience-based subjective analysis to data-based scientific analysis. In high-speed and complex adversarial sports such as tennis, accurately capturing and analyzing key events and turning points in the game is of great significance for improving athlete performance and game strategies.

[0003] Existing tennis game analysis technologies mainly focus on basic data statistics and single-perspective video analysis. These methods have the following obvious deficiencies: First, there is a lack of effective fusion of multi-perspective data, resulting in an incomplete understanding of the on-field situation; second, only single-event recognition and basic action analysis are concerned, lacking in-depth understanding of the temporal causal relationships during the game; third, key turning points that affect the game trend cannot be effectively identified and explained; fourth, in a complex and changing game environment, the dynamic causal chain between events cannot be accurately captured; fifth, it cannot provide in-depth tactical insights and targeted improvement suggestions for coaches and players.

[0004] Currently, although there are studies attempting to analyze key events in sports games through statistical models or rule systems, these methods usually have difficulty capturing complex temporal relationships and causal mechanisms, with limited recognition accuracy and lack of interpretability, and cannot provide in-depth tactical guidance. Traditional video analysis tools are often limited to a single perspective and are difficult to comprehensively grasp the game scenario, resulting in one-sided or inaccurate analysis results. In addition, existing technologies generally lack the ability of in-depth causal analysis and are difficult to reveal the underlying reasons for changes in the game trend, thus limiting their application value in high-level competitive games.

[0005] Therefore, there is an urgent need for a real-time evaluation method for tennis that can fuse multi-perspective visual features, deeply explore the causal mechanism of the game, accurately identify key turning points, and provide personalized tactical suggestions to meet the needs of modern competitive sports for scientific training and game analysis. Summary of the Invention

[0006] The present invention provides a real-time AI evaluation method for tennis based on the fusion of multi-perspective visual features, which solves the technical problems in related technologies such as the difficulty in deeply understanding the causal relationship of the game, the inability to accurately identify key turning points, and the lack of targeted tactical suggestions.

[0007] The present invention discloses a real-time AI evaluation method for tennis based on multi-view visual feature fusion, comprising the following steps: collecting multi-angle video stream data of a tennis match; using a temporal graph convolutional network to fuse multi-view data to generate spatio-temporal feature representations; constructing a temporal causal graph model based on the fused spatio-temporal features, where nodes represent match events and edges represent potential causal relationships between events; using graph structure analysis and event influence evaluation to identify key turning points in the match; verifying the influence of turning points through counterfactual reasoning technology to generate a quantitative influence evaluation; combining with a player feature model to generate personalized tactical suggestions; wherein, the temporal causal graph model uses a dynamic causal strength function to quantify the causal relationship between events and forms a sparse causal relationship network through adaptive threshold filtering.

[0008] Further, the temporal graph convolutional network includes a spatial graph convolutional layer and a temporal convolutional layer, wherein: the spatial graph convolutional layer is used to capture the spatial relationship between nodes; the temporal convolutional layer is used to capture the temporal dependence relationship; and the residual connection and layer normalization are used to improve the training stability of the network.

[0009] Further, the temporal causal graph model has a hierarchical structure, including: the bottom layer is composed of basic event nodes, representing hitting actions, player movements, ball landing points, etc.; the middle layer is composed of tactical combination nodes, representing tactical sequences formed by consecutive basic events; the top layer is composed of tactical effect nodes, representing the results of tactical executions.

[0010] Further, the steps of identifying key turning points in the match include: calculating the in-degree, out-degree and various centrality metrics of nodes; calculating the change value of graph structure centrality after removing nodes; calculating the event influence metric based on conditional probability and importance weights; sorting comprehensively according to node importance and event influence, and identifying events exceeding the threshold as key turning points.

[0011] Further, the counterfactual reasoning technology includes: constructing a counterfactual scenario where the turning point does not occur or occurs in a different way; removing or modifying the corresponding nodes in the temporal causal graph; using a probability propagation algorithm to predict the probability distribution of events occurring in the counterfactual scenario; calculating the difference between the actual result and the counterfactual result, which is defined as the causal effect of the turning point.

[0012] Further, the player feature model includes: technical features, representing serving position distribution, hitting force distribution, landing point distribution, etc.; tactical features, representing common tactical combinations and strategy adjustments for different opponents, etc.; psychological features, representing performance in key points and stress coping abilities, etc.

[0013] Furthermore, the steps for generating personalized tactical suggestions include: identifying effective tactical sequences leading to scores through causal analysis; identifying key technical or tactical problems leading to points conceded; analyzing the opponent's technical and tactical patterns and response strategies; and using a multi-objective optimization algorithm to comprehensively consider success rate, risk level, and player execution ability to generate the most suitable tactical suggestions.

[0014] Furthermore, the personalized tactical suggestions include: suggestions for adjusting the distribution of hitting locations; suggestions for optimizing serving strategies; suggestions for adjusting technical movements; suggestions for controlling the game rhythm.

[0015] Furthermore, it also includes the steps of forming visual tactical decision support information, and the visual tactical decision support information includes: tactical effect evaluation charts; heat maps of advantageous areas; sequential tactical recommendations; real-time decision prompts.

[0016] The beneficial effects of the present invention are as follows: Through the innovative combination of multi-perspective visual feature fusion, sequential causal graph construction, and turning point recognition algorithms, it solves the technical problems that traditional tennis analysis systems cannot deeply understand causal relationships and identify key turning points, and achieves the following remarkable technical effects: 1. It realizes the excavation and understanding of the deep causal mechanism in the game process, enabling coaches and players to understand the essential reasons behind the game trend; 2. Through the innovative turning point recognition algorithm and counterfactual reasoning technology, it accurately identifies and quantifies the impact of key turning points, providing interpretable causal analysis; 3. It supports real-time event recognition and analysis at the millisecond level, and can provide instant tactical adjustment suggestions during the game; 4. It generates highly personalized tactical suggestions, directly serving the actual combat needs of coaches and players; 5. Through sequential graph convolutional technology, it effectively fuses multi-perspective video data, providing more comprehensive and accurate basic data for analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is the overall flowchart of the present invention;

[0018] Figure 2 is the flowchart of the multi-perspective visual feature acquisition and fusion steps of the present invention;

[0019] Figure 3 is the flowchart of the sequential causal graph construction steps of the present invention;

[0020] Figure 4 is the flowchart of the turning point recognition algorithm steps of the present invention;

[0021] Figure 5 is the flowchart of the counterfactual reasoning and verification steps of the present invention;

[0022] Figure 6 is the flowchart of the personalized tactical decision support steps of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0023] This embodiment is applied to the field of tennis match analysis, especially for scenarios that require in-depth understanding of the causal relationships between various events during the match and identification of key turning points. The existing tennis analysis systems mainly have the following technical problems:

[0024] Only focus on single event recognition and basic action analysis, lacking in-depth understanding of the temporal causal relationships during the match;

[0025] Unable to effectively identify and explain the key turning points that affect the trend of the match;

[0026] In a complex and ever-changing match environment, unable to accurately capture the dynamic causal chain between events;

[0027] Unable to provide in-depth tactical insights and targeted improvement suggestions for coaches and players.

[0028] This embodiment provides a real-time AI evaluation method for tennis based on multi-perspective visual feature fusion. By constructing a temporal causal graph model and implementing a turning point recognition algorithm, the above technical problems are solved, realizing the mining and understanding of the deep causal mechanism of the match, and providing more in-depth technical and tactical analysis and targeted improvement suggestions for coaches and players.

[0029] The method of this embodiment includes:

[0030] Step 1: Multi-perspective visual feature collection and fusion

[0031] Use a temporal graph convolutional network to process multi-angle video streams and generate a fused spatio-temporal feature representation.

[0032] Specifically, it includes the following sub-steps:

[0033] 1.1 Install high-frame-rate cameras at both ends of the tennis court to collect multi-angle video stream data of the match. Each camera records at a rate of 60 frames per second to ensure that the picture clarity is not less than 1080p;

[0034] 1.2 Apply an object detection algorithm to process the video stream, extract the position information and motion trajectory data of players, balls and court elements, and form a preliminary spatio-temporal coordinate sequence;

[0035] 1.3 Construct a temporal graph structure G st =(V st , E st , F node ), where the node V st represents the positions of players and balls, the edge E st represents the interaction relationship between nodes, and F node is the node feature vector. The node feature vector includes position coordinates, velocity vectors, acceleration vectors and local visual features;

[0036] 1.4 Apply the spatio-temporal graph convolutional network (ST-GCN) to fuse multi-view data. This network aggregates and updates the node features at different time points through spatio-temporal graph convolutional layers. The convolutional operation is defined as:

[0037]

[0038] where represents the feature vector of node v at the l-th layer, N(v) is the neighbor set of node v, and c v,u is the normalization constant, W (l) and b (l) are learnable parameters, and σ is the activation function.

[0039] The spatio-temporal graph convolutional network in this embodiment specifically consists of the following structures:

[0040] Spatial graph convolutional layer: Captures the spatial relationships between nodes and performs convolutional operations on the graph at each time step

[0041] Temporal convolutional layer: Uses one-dimensional convolution to capture temporal dependencies, with a window size of 3

[0042] Residual connection: Adds residual connections every two layers to improve the training effect of the deep network

[0043] Layer normalization: Applies layer normalization after each convolutional layer to improve training stability

[0044] The network contains a total of 9 graph convolutional layers. The input is the original node features, and the output is the fused spatio-temporal feature representation. In the real-time analysis scenario of tennis matches, this network processes the input data at a rate of 30 frames per second and can accurately capture the movement trajectories and mutual relationships during complex actions such as rapid movement and turning of players.

[0045] Through the above sub-steps, the multi-view video data is converted into a unified spatio-temporal graph feature representation, providing a basis for subsequent temporal causal analysis. The output result of this step is a spatio-temporal feature tensor that fuses multi-angle visual information where N represents the number of nodes, T len represents the time step, and D represents the feature dimension.

[0046] Step 2: Temporal causal graph construction

[0047] Based on the fused spatio-temporal feature tensor, apply the temporal causal inference algorithm to generate the temporal causal graph model of the tennis match.

[0048] Specifically, it includes the following sub-steps:

[0049] 2.1 Based on the spatio-temporal feature tensor F output in Step 1st , use the event recognition model to detect various key events in a tennis match, such as serving, returning, scoring points, etc., and form an event sequence E seq = {e1, e2,..., e n}, and each event e i contains the event type, occurrence time, and related features;

[0050] 2.2 Construct the temporal causal graph model G tc = (V e , E c , T dim ), where:

[0051] The node set V e represents various events detected from the video, and each node contains the event type, timestamp, and feature vector. The edge set E c represents potential causal relationships, and a directed edge (e i , e j ) indicates that event e i may cause event e j

[0052] T dim represents the time dimension and is used to record the temporal information of event occurrences

[0053] The implementation of the temporal causal graph model in an actual tennis match adopts a hierarchical structure:

[0054] Bottom layer: Consists of basic event nodes, such as hitting actions, player movements, ball landing points, etc.

[0055] Middle layer: Consists of tactical combination nodes, representing tactical sequences formed by consecutive basic events

[0056] Top layer: Consists of tactical effect nodes, representing the results of tactical executions, such as scoring, losing points, advantage conversion, etc.

[0057] In practical applications, this model can capture tactical causal chains such as "serving landing point close to the sideline → opponent's return quality decreases → attacking side obtains an offensive opportunity → scoring".

[0058] 2.3 Apply the dynamic causal intensity function to calculate the causal relationship intensity between events. The function is defined as:

[0059]

[0060] Where:

[0061] I c (e i , e j , Δt) represents the influence of event e i on event ej Causal influence strength

[0062] P(e j |e i , Δt) represents the conditional probability that event e occurs within Δt time after event e i occurs j conditional probability

[0063] exp(-β·Δt) is the time decay factor, and β is the decay coefficient

[0064] α is the causal weight coefficient

[0065] In the specific implementation, the conditional probability P(e j |e i , Δt) is calculated by a temporal conditional model based on the attention mechanism. This model considers event types, time intervals, and context features, and is trained using historical match data. The time decay coefficient β is dynamically adjusted according to different event types. A larger β value (fast decay) is used for short-term tactical events, while a smaller β value (slow decay) is used for long-term strategic events.

[0066] 2.4 Based on the above causal strength function, a complete temporal causal graph is constructed. Edges with causal strength greater than the threshold θ are retained in the graph to form a sparse causal relationship network. In practical applications, the threshold θ is set as an adaptive value and dynamically adjusted according to the match stage and importance.

[0067] The output result of this step is a directed graph structure that can represent the temporal causal relationship between events in a tennis match. Each edge has a corresponding causal strength value, which is used to quantify the causal influence degree between events.

[0068] Step 3: Turning point identification algorithm

[0069] Based on the constructed temporal causal graph, apply graph structure analysis and event influence evaluation algorithms to identify and quantify the key turning points in the match.

[0070] Specifically, it includes the following sub-steps:

[0071] 3.1 Analyze the topological structure of the temporal causal graph, paying particular attention to the following features:

[0072] In-degree and out-degree of nodes: Reflect the influence range and degree of being influenced of events

[0073] Node centrality: Including degree centrality, betweenness centrality, eigenvector centrality, etc.

[0074] Community structure: Identify densely associated groups of events

[0075] In a tennis match scenario, this step particularly focuses on potential high - impact events such as consecutive winning points, saving break points in service games, and key points in tie - break games, and determines the positions and importance of these events in the causal network through graph - structure analysis.

[0076] 3.2 Calculate the importance index of key nodes, defined as the change in graph - structure centrality after removing this node:

[0077] ΔC(G tc ,e)=|C(G tc ) - C(G tc \e)|

[0078] Where:

[0079] C(G tc ) represents the centrality metric value of the original graph

[0080] C(G tc \e) represents the centrality metric value of the graph after removing node e

[0081] Based on the above calculations, identify the most important nodes:

[0082] e * =argmax e ΔC(G tc ,e)

[0083] The implementation of this algorithm uses a combination of multiple centrality metrics, including eigenvector centrality (evaluating the global influence of nodes), betweenness centrality (evaluating the importance of nodes as "bridges"), and PageRank value (evaluating the importance of nodes in a directed graph). In the actual analysis of tennis matches, this algorithm can accurately identify turning - point events such as "key breakbacks" and "consecutive winning points forming a psychological advantage".

[0084] 3.3 Combine the event - influence index to evaluate the degree of influence of events on the game trend, and define the event influence as:

[0085]

[0086] Where:

[0087] P(e j |e) represents the conditional probability that event e j occurs after event e occurs

[0088] w j is the importance weight of event e j

[0089] α j is the adjustment coefficient

[0090] In practical applications, the importance weight w j is predefined according to the event type. For example, the weight of a scoring point is higher than that of an ordinary round, and the weight of a key ball is higher than that of an ordinary ball. The adjustment coefficient α j is dynamically adjusted according to the game stage and score situation. For example, the influence of an event in the deciding set will be given a higher weight.

[0091] 3.4 Based on the comprehensive score of node importance and event influence, sort the events, identify the events with a score exceeding the threshold τ as the key turning points of the game, and form the turning point set E key ={e k |score(e k )>τ}.

[0092] In actual cases, the algorithm can identify the following turning points:

[0093] Getting two break points in a row when the score is close

[0094] Using a rare but effective tactical combination on a key ball

[0095] Successfully saving a ball and reversing the situation in a high-pressure situation

[0096] The output result of this step is a set of quantitatively identified key turning points of the game, and each turning point contains event information, occurrence time, importance score, and a description of the potential impact on the game trend.

[0097] Step 4: Counterfactual reasoning and verification

[0098] Using counterfactual reasoning technology, verify and explain the identified turning points, and generate causal explanations and quantitative impact assessments.

[0099] Specifically, it includes the following sub-steps:

[0100] 4.1 For each key turning point e k ∈E key , construct a counterfactual scenario S cf (e k ), that is, assume a scenario where the event e k does not occur or occurs in a different way;

[0101] In the analysis of tennis matches, constructing counterfactual scenarios considers multiple possibilities. For example, for a key serve and score point, the following counterfactual scenarios can be constructed:

[0102] What if the serve lands in a different area?

[0103] What if the serve speed or spin is different?

[0104] What if different tactical combinations are used?

[0105] 4.2 Apply the counterfactual reasoning algorithm to simulate the development path of the game in counterfactual situations. This algorithm is based on the temporal causal graph model and is implemented through the following process:

[0106] Remove or modify node e in the temporal causal graph k

[0107] Based on the modified graph structure, use the probability propagation algorithm to predict the probability distribution of subsequent events

[0108] Calculate the game state transition path in the counterfactual situation

[0109] The mathematical expression of counterfactual reasoning is:

[0110] P(Y cf |do(X=x′),G tc )

[0111] Where:

[0112] Y cf Represents the result variable in the counterfactual situation

[0113] do(X=x′) represents the intervention operation, setting the variable X to the counterfactual value x′

[0114] G tc Is the temporal causal graph model

[0115] The counterfactual reasoning algorithm in this embodiment is implemented based on the structural causal model and combines the Monte Carlo tree search method for path exploration. The algorithm maintains the consistency of time and event logic during execution to ensure that the generated counterfactual situations conform to the rules and real-world constraints of tennis matches.

[0116] For example, when analyzing a crucial volley point, the algorithm will simulate various possible results if the player chooses a groundstroke instead of a volley and calculate the probability of the impact of this decision change on the game result.

[0117] 4.3 Calculate the difference between the actual result and the counterfactual result, defined as the causal effect of the turning point:

[0118] CE(e k ) = M(Y actual ) - M(Y cf )

[0119] Where:

[0120] M(Y actual ) is the actually observed result metric

[0121] M(Ycf ) is the result metric predicted in the counterfactual scenario

[0122] In a specific implementation, the result metric M includes multiple dimensions, such as scoring probability, change in game advantage, tactical effect, etc. These metrics are calculated by a prediction model trained with historical game data.

[0123] 4.4 Generate a causal explanation and a quantitative impact assessment report for each turning point, including:

[0124] A detailed description of the turning point (event type, time, relevant players)

[0125] Causal path analysis (how this turning point affects the subsequent development of the game)

[0126] Quantitative impact assessment (the degree of impact on the game result)

[0127] Visualization of the causal chain display

[0128] In an actual application scenario, this technology can generate the following analysis:

[0129] "Saving the break point in the 7th game of the third set increased the player's winning probability by about 35%, which is the most crucial turning point in this game."

[0130] "Three consecutive precise serve landing point selections forced the opponent to change their stance in the return shot, creating conditions for subsequent winning points."

[0131] "The tactical adjustment of changing the hitting rhythm at a critical moment disrupted the opponent's defensive rhythm and was the key factor in reversing the game."

[0132] The output result of this step is a set of causal explanation reports for turning points. Each report contains a detailed causal analysis and a quantitative impact assessment, enabling coaches and players to clearly understand the impact mechanism of key events in the game.

[0133] Step 5: Personalized tactical decision support

[0134] Based on the analysis results of the foregoing steps and combined with historical data, generate targeted tactical suggestions and decision support information.

[0135] Specifically, it includes the following sub-steps:

[0136] 5.1 Build a player's technical and tactical feature model, which includes the following dimensions:

[0137] Technical features: serving position distribution, hitting force distribution, landing point distribution, etc.

[0138] Tactical features: common tactical combinations, strategy adjustments for different opponents, etc.

[0139] Psychological characteristics: key score performance, stress coping ability, etc.

[0140] In this embodiment, the player's technical and tactical feature model adopts a hybrid feature representation method, combining statistical features and deep learning features. The model automatically extracts technical patterns and habits by analyzing the player's historical game data, and constructs a personalized player profile. This model can identify the strengths and weaknesses of players, such as features like "serving preference in the T area" and "unforced errors are likely to occur in backhand returns".

[0141] 5.2 Based on the time-series causal analysis and the turning point recognition results, combined with the player feature model, identify the technical and tactical advantages and disadvantages in the current game, specifically including:

[0142] Effective tactical combinations: Identify the effective tactical sequences that lead to scores through causal analysis

[0143] Weak links: Identify the key technical or tactical problems that lead to points being lost

[0144] Opponent's pattern: Analyze the opponent's technical and tactical patterns and coping strategies

[0145] In practical applications, the system can real-time identify tactical information such as "the opponent is likely to send high-arc balls to the backhand area under pressure" and "is likely to make mistakes after more than 5 consecutive baseline rallies", and form targeted suggestions.

[0146] 5.3 Generate targeted tactical suggestions, including:

[0147] Suggestions for adjusting the distribution of hitting points: Based on heat map analysis, recommend more effective hitting point areas

[0148] Optimization of serving strategy: According to the opponent's receiving service characteristics, provide optimization suggestions for serving position, speed and spin

[0149] Adjustment of technical movements: Specific improvement suggestions for key technical links

[0150] Control of the game rhythm: Based on time-series causal analysis, provide strategic suggestions for controlling the game rhythm

[0151] The generation of targeted tactical suggestions adopts a multi-objective optimization algorithm, comprehensively considering success rate, risk degree and player execution ability. This algorithm is trained based on a large amount of historical game data and can generate the most suitable tactical suggestions according to the current game state and player characteristics.

[0152] In specific game scenarios, the system will generate the following suggestions:

[0153] "It is recommended to increase the deep balls to the opponent's backhand area to reduce the opponent's offensive conversion probability"

[0154] "It is recommended to adopt a high-risk serving strategy during key time points and serve to the T area."

[0155] "For the second serve, it is recommended to increase spin rather than pursue speed to improve the serving success rate."

[0156] 5.4 Generate visual tactical decision support information, including:

[0157] Tactical effect evaluation chart: Quantitatively display the success rate and influence of different tactics

[0158] Heat map of advantageous areas: Highlight the advantageous areas on the court and the areas that should be focused on attacking

[0159] Tactical recommendations by time sequence: Tactical adjustment suggestions for different stages of the game

[0160] Real-time decision-making prompts: Provide concise tactical prompts during the game

[0161] The visual decision support system adopts a multi-screen interactive design and is suitable for viewing on the sidelines coach's tablet and when players are resting. Through color coding, dynamic charts, and concise text, the system realizes the intuitive presentation of complex tactical information, ensuring that coaches and players can understand and apply tactical suggestions in a short time.

[0162] The output result of this step is a set of personalized tactical decision support information, including specific tactical suggestions, technical adjustment directions, and visual decision-making auxiliary materials, directly serving the actual combat needs of coaches and players.

[0163] This implementation method realizes the following remarkable technical effects through a series of innovative technical steps such as multi-perspective visual feature fusion, time-sequence causal graph construction, turning point recognition algorithm, counterfactual reasoning and verification, and personalized tactical decision support:

[0164] Deep causal mechanism mining: Breaking through the limitation of traditional tennis analysis systems that only focus on surface phenomena, through constructing a time-sequence causal graph model, it realizes the mining and understanding of the deep causal mechanism during the game, enabling coaches and players to understand the essential reasons behind the game trend.

[0165] High-precision turning point recognition: Through the innovative turning point recognition algorithm, the system can accurately identify the key turning points that affect the game trend and provide interpretable causal analysis through counterfactual reasoning technology, enabling coaches and players to clearly understand the influence mechanism of key events.

[0166] Real-time analysis and decision support: The system supports real-time event recognition and analysis at the millisecond level, can provide instant tactical adjustment suggestions during the game, meet the needs of instant tactical adjustment during the game, and significantly improve the scientific level of training and competition.

[0167] Personalized Tactical Optimization: Based on in-depth analysis of player characteristics and understanding of the game's causal mechanisms, the system can generate highly personalized tactical suggestions, including adjustments to the distribution of hitting points, optimization of serving strategies, and adjustments to technical movements, directly serving the actual combat needs of coaches and players.

[0168] Multi-dimensional Data Fusion: Through temporal graph convolutional technology, the system realizes the effective fusion of multi-perspective video data, overcomes the limitations of single-perspective data, and provides more comprehensive and accurate basic data for subsequent analysis.

[0169] Generally speaking, the real-time AI evaluation method for tennis based on multi-perspective visual feature fusion provided by this embodiment has achieved a technological breakthrough in the field of tennis game analysis, and can provide more in-depth, comprehensive, and real-time valuable technical and tactical analysis and decision-making support for coaches and players.

[0170] Through an actual case analysis of a high-level tennis match, demonstrate the application process and effect of this embodiment. The analysis object is the men's quarter-final of a certain international tennis tour, involving the match between player A and player B, both of whom are ranked among the top 20 in the world.

[0171] This application is carried out on a standard indoor hard court tennis court. Four high-definition cameras are installed at both ends of the court, with a collection frequency of 60 frames per second and a resolution of 1920×1080 pixels. The camera layout is as follows:

[0172] Table 1 Multi-perspective Camera Layout and Functions

[0173]

[0174]

[0175] Multi-perspective Visual Feature Fusion

[0176] The system first collects video streams from 8 cameras, and extracts and fuses multi-perspective visual features in real time. Taking a typical service-winning rally as an example, the system processing flow is shown in Table 2:

[0177] Table 2 Example of the Processing Flow for Multi-perspective Visual Feature Fusion

[0178]

[0179] The graph structure contains 9 nodes (each part of the two players and the ball), each node has a 15-dimensional feature vector, and the dimension of the fused feature tensor is 9×120×64 (number of nodes × time steps × feature dimension).

[0180] Temporal Causal Graph Construction

[0181] Based on the fused spatio-temporal features, the system identified 28 key events and constructed a temporal causal graph. Table 3 shows some of the key events and their causal relationships:

[0182] Table 3 Examples of Key Events and Their Causal Relationships in the Game

[0183]

[0184] Turning Point Identification and Counterfactual Reasoning

[0185] The system analyzed the temporal causal graph and identified 5 key turning points. The two most critical turning points and their counterfactual reasoning results are shown in Table 4:

[0186] Table 4 Results of Key Turning Point Identification and Counterfactual Reasoning

[0187]

[0188] The details of the counterfactual reasoning performed by the system on TP2 are as follows: Assuming that player B used a different serving strategy at this break point (serving to the backhand area instead of the forehand area as actually used), the simulation results show that the probability of player B saving the break point in this game would be reduced from the actual 78% to approximately 31%, and the impact on the final winning rate would be approximately 41%.

[0189] This embodiment is applied to the analysis of 10 tennis matches at different levels. The main technical effect verification results are as follows:

[0190] Verification of Turning Point Identification Accuracy

[0191] Table 5 Comparison of Turning Point Identification Accuracy with the Existing Technology

[0192]

[0193] 4.3.2 Real-time Performance Verification

[0194] Table 6 System Real-time Performance Metrics

[0195]

[0196] The experimental results show that this embodiment can complete the entire process from video input to tactical advice generation within 200 milliseconds, meeting the requirements of real-time analysis of tennis matches. Compared with the existing technology, the accuracy of turning point identification of this system has increased by 21.8% - 25.6%, and the ability to identify key tactical decision points has increased by 26.1% - 31.7%, providing more timely and accurate tactical analysis and decision support for coaches and players.

Claims

1. A real-time AI evaluation method for tennis based on multi-view visual feature fusion, characterized in that, It includes the following steps: Collect multi-angle video stream data of tennis matches; Use a temporal graph convolutional network to fuse multi-view data and generate spatio-temporal feature representations; Construct a temporal causal graph model based on the fused spatio-temporal features, where nodes represent match events and edges represent potential causal relationships between events; Identify key turning points in the match using graph structure analysis and event influence assessment; Verify the influence of turning points through counterfactual reasoning technology and generate a quantitative influence assessment; Combine the player feature model to generate personalized tactical suggestions; Among them, the temporal causal graph model uses a dynamic causal strength function to quantify the causal relationship between events, and forms a sparse causal relationship network through adaptive threshold filtering.

2. The method according to claim 1, characterized in that, The temporal graph convolutional network includes a spatial graph convolutional layer and a temporal convolutional layer, where: The spatial graph convolutional layer is used to capture the spatial relationship between nodes; The temporal convolutional layer is used to capture temporal dependence relationships; And improve the network training stability through residual connections and layer normalization.

3. The method according to claim 1, characterized in that, The temporal causal graph model has a hierarchical structure, including: The bottom layer consists of basic event nodes, representing hitting actions, player movements, and ball landing points; The middle layer consists of tactical combination nodes, representing tactical sequences formed by multiple consecutive basic events; The top layer consists of tactical effect nodes, representing the results of tactical execution.

4. The method according to claim 1, wherein The steps for identifying key turning points in the match include: Calculate the in-degree, out-degree, and various centrality metrics of nodes; Calculate the change value of the graph structure centrality after removing nodes; Calculate the event influence metric based on conditional probability and importance weights; Rank based on node importance and event influence, and identify events exceeding the threshold as key turning points.

5. The method according to claim 1, characterized in that, The counterfactual reasoning technology includes: Construct a counterfactual scenario where the turning point does not occur or occurs in a different way; Remove or modify the corresponding nodes in the temporal causal graph; Use the probability propagation algorithm to predict the probability distribution of events occurring in the counterfactual scenario; Calculate the difference between the actual result and the counterfactual result, defined as the causal effect of the turning point.

6. The method according to claim 1, wherein The player feature model includes: Technical features, representing serving position distribution, hitting power distribution, and landing point distribution; Tactical features, representing common tactical combinations and strategy adjustments for dealing with different opponents; Psychological features, representing performance in key points and stress coping abilities.

7. The method according to claim 1, characterized in that The steps for generating personalized tactical suggestions include: Identify effective tactical sequences leading to scores through causal analysis; Identify key technical or tactical problems leading to points lost; Analyze the opponent's technical and tactical patterns and coping strategies; Use a multi-objective optimization algorithm to comprehensively consider success rate, risk degree, and player execution ability to generate the most suitable tactical suggestions.

8. The method according to claim 1, wherein The personalized tactical suggestions include: Suggestions for adjusting the hitting landing point distribution; Suggestions for optimizing serving strategies; Suggestions for adjusting technical movements; Suggestions for controlling the match rhythm.

9. The method according to claim 1, characterized in that, It also includes the step of forming visual tactical decision support information, and the visual tactical decision support information includes: Tactical effect evaluation charts; Heat maps of advantageous areas; Temporal tactical recommendations; Real-time decision-making tips.

10. A real-time AI evaluation system for tennis based on the fusion of multi-view visual features, characterized in that, It includes: A multi-view video acquisition module for collecting multi-angle video stream data of tennis matches; A spatio-temporal feature fusion module for using a temporal graph convolutional network to fuse multi-view data and generate spatio-temporal feature representations; A temporal causal graph construction module for constructing a temporal causal graph model based on the fused spatio-temporal features; A turning point recognition module for identifying the key turning points of the game by using graph structure analysis and event influence evaluation; A counterfactual reasoning module for verifying the influence of turning points and generating a quantitative influence evaluation; A tactical decision support module for generating personalized tactical suggestions and visual decision support information in combination with the player feature model.

Images