Football player real-time physical value prediction method and system based on multi-source perception and individual correction

Abstract

The application discloses a football player real-time physical value prediction method and system based on multi-source perception and individual correction, which comprises the following steps: acquiring multi-source real-time data of a target football player and performing time synchronization, player identity matching and abnormal data detection to generate continuous time sequence state data; extracting physical state related features based on a preset sliding time window and constructing individual physical benchmark according to historical physical data, on-field position attributes and historical game stage physical change curve and performing normalization correction; dynamically weighting and fusing according to the quality confidence of different data sources and inputting the fused individual physical state features into a time sequence fusion prediction model to output the real-time physical value, future physical value change trend, physical decline rate and prediction confidence of the target football player. The application can realize continuous, dynamic, individual and real-time prediction of the physical state of the football player and reduce the misjudgment caused by unified threshold and single movement index.

Description

Technical Field

[0001] This application relates to the field of sports data analysis technology, specifically to a method and system for predicting the real-time physical fitness of football players based on multi-source perception and individualized correction. Background Technology

[0002] Football matches are characterized by high intensity, strong physical contact, and rapid transitions between offense and defense. Athletes' physical condition fluctuates constantly throughout the match, depending on running distance, number of sprints, intensity of contact, match time, and tactical objectives. An athlete's physical condition not only affects their running ability, the quality of their technical execution, and their participation in the game, but also influences the coach's judgment regarding substitutions, training load control, and tactical adjustments.

[0003] Current methods for monitoring the physical fitness of football players mainly rely on coaches' experience and observation, post-match running data statistics, GPS location data, heart rate monitoring devices, or wearable sensors. While these methods can provide some information on exercise load, they still have the following shortcomings: First, existing methods mostly use single or a few indicators such as running distance, speed, and heart rate as the basis for judging physical fitness, which makes it difficult to fully reflect the comprehensive physical condition of athletes in real competitions. Second, existing methods usually use a uniform threshold or a uniform evaluation standard to judge the physical condition of different athletes, ignoring the individual athlete's physical foundation, position responsibilities, historical athletic ability and differences in the stage of the competition, which can easily lead to inaccurate estimation of physical strength. Third, during the competition, data sources such as video, location, and physiological sensors may be affected by factors such as obstruction, signal drift, sensor malfunction, or data loss. Directly fusing multi-source data can easily lead to unstable physical fitness prediction results. Fourth, existing monitoring methods are mostly limited to post-competition statistics or current status display, making it difficult to predict athletes' physical fitness trends in real time over a future period.

[0004] Therefore, there is an urgent need for a method that can integrate multi-source real-time data, combine individualized physical fitness benchmarks of athletes, and utilize time-series prediction models to predict the real-time physical fitness of football players. Summary of the Invention

[0005] Purpose of the invention: In order to overcome the above shortcomings, the purpose of this application is to provide a method and system for predicting the real-time physical fitness of football players based on multi-source perception and individualized correction, which solves the problems of single judgment of physical fitness status, insufficient consideration of individual differences, poor stability of multi-source data fusion, and difficulty in predicting future physical fitness trends in real time in existing football player physical fitness monitoring methods.

[0006] To address the aforementioned technical problems, this application provides a method for predicting the real-time physical condition of football players based on multi-source sensing and individualized correction, including: S1: Acquire multi-source real-time data of the target football player during the game or training process, and perform time synchronization, player identity matching and abnormal data detection on the multi-source real-time data to generate continuous time-series state data matching the target football player. S2: Based on a preset sliding time window, extract physical state-related features from the continuous time-series state data and construct an individualized physical benchmark according to the target football player's historical physical data, on-field position attributes, and physical change curves during historical matches. S3: Based on the individualized physical fitness benchmark, the physical fitness-related features are normalized and corrected to generate individualized physical fitness features that characterize the degree of deviation of the target football player's current physical fitness status relative to his / her normal range of physical fitness changes. S4: Based on the quality confidence of data from different sources, the individualized physical condition features are dynamically weighted and fused, and the fused individualized physical condition features are input into the time series fusion prediction model to predict and output the real-time physical condition value of the target football player and the trend of physical condition value change within a preset time period in the future.

[0007] As a preferred embodiment of this application, in S1, the method includes the following steps: S11: Collect multi-source real-time data of the target football player using one or more of the following: video acquisition equipment, positioning equipment, inertial sensing equipment, physiological monitoring equipment, and match status recording equipment. ,in For video data, For motion trajectory data, For action posture data, For physiological state data, This refers to match status data; S12: Based on the multi-source real-time data, detect speed anomalies, heart rate anomalies, trajectory anomalies, and posture anomalies, and process the anomalies using methods such as linear interpolation, previous frame filling, sliding window smoothing, or weight reduction. S13: Using video frame timestamps or a unified competition clock as a reference, all data in the multi-source real-time data are unified to the same time axis through linear interpolation, nearest neighbor interpolation, or time window averaging. S14: Identify player numbers, team uniform colors, or initial positioning information, bind video trajectory IDs, establish a correspondence between video trajectory IDs and various device numbers, and construct a unique player ID mapping table to ensure that data from different sources all belong to the same target football player; S15: Based on the multi-source real-time data and the unique player ID mapping table, generate continuous time-series state data matching the target football player.

[0008] As a preferred embodiment of this application, in S2, the method includes the following steps: S21: Set the sliding time window W with the current time t as the endpoint; S22: Calculate the number of sprints, sudden stops, changes of direction, acceleration, and deceleration within a short time window to generate explosive exercise consumption characteristics; S23: Calculate running distance, high-speed running distance, average speed, average heart rate, and heart rate recovery rate within a medium time window to generate continuous exercise load characteristics; S24: Calculate the cumulative load, duration of continuous high-intensity exercise, trend of decline in movement stability, and historical trend of decline in physical strength within a long time window to generate overall fatigue accumulation characteristics; S25: Based on the explosive exercise consumption characteristics, continuous exercise load characteristics, and overall fatigue accumulation characteristics, generate physical state-related features of multi-source feature vectors: ,in This is characterized by explosive energy expenditure within a short time window. Characteristics of continuous motion load within a medium time window. This represents the overall fatigue accumulation characteristics within a long time window. Adjust features for game status.

[0009] As a preferred embodiment of this application, in step S2, the method further includes the following steps: S26: Extract historical physical fitness data from the target football player's preset number of matches, preset number of special training sessions, and pre-match physical fitness tests. The historical physical fitness data includes one or more of the following: maximum speed, average running intensity, high-speed running distance, sprint recovery time, heart rate recovery ability, and movement stability variation patterns. S27: Statistics are segmented according to competition stage, training, or competition type; S28: Calculate the typical physical fatigue decay curve for each position according to the on-field position attributes, wherein the on-field position attributes include at least forwards, midfielders, defenders and goalkeepers, and different positions correspond to different running patterns and physical fatigue benchmarks; S29: Combining correction factors including score difference, offensive and defensive status, game rhythm, weather conditions and field conditions, generate individual normal physical strength ranges, mean and standard deviations for different times, positions, intensities and game conditions to form individualized physical strength benchmarks.

[0010] As a preferred embodiment of this application, in step S3, the method further includes the following steps: Based on the individualized physical fitness benchmark, the physical fitness-related features are normalized and corrected to generate individualized physical fitness features: ,in, This serves as a baseline for the normal physical condition of the target football player under the current stage of the match and their position attributes. To correspond to the historical fluctuation scale, the standard deviation of historical characteristics under the same competition stage and the same position task is taken; ε is a constant to prevent the denominator from being zero. Used to characterize the degree of deviation of a target football player's current physical condition from their historical normal level; where, when When the absolute value of increases, it indicates that the current state deviates more from the individual's normal range of physical strength variation.

[0011] As a preferred embodiment of this application, in step S4, the method includes the following steps: S41: Set quality confidence level: ,in, For the reliability of the k-th type of data source, For the integrity of the k-th type of data source, For the time continuity of the k-th type of data source, The degree of occlusion or anomaly of the k-th type of data source. These are weight parameters; S42: Calculate the fusion weights based on the quality confidence level: Generate the fused individual physical condition characteristics: , For the feature encoding function corresponding to the k-th type of data source, For the k-th type of data source, the features are after individualized correction. The characteristics of the individualized physical condition after integration; Let be the quality confidence score of the j-th type of data source at time t.

[0012] As a preferred embodiment of this application, in step S4, the method further includes the following steps: S43: Input the fused features within the continuous time window into the time-series fusion prediction model to model the dynamic changes in the physical state of the target football player: ,in, A function for modeling temporal states, which is implemented by one or more of the following: recurrent neural networks, gated recurrent unit networks, long short-term memory networks, temporal convolutional networks, attention networks, or Transformer networks; S44: Predict the target football player's real-time stamina: ,in, The real-time stamina value of the target football player at time t; The bias parameter of the physical fitness prediction layer is used to perform baseline translation correction on the mapping results between the time series state modeling results and the real-time physical fitness values; S45: Predict the trend of a target football player's physical strength over a future preset time period. , For a future time period.

[0013] As a preferred embodiment of this application, in S4, to reflect the reliability of the prediction results, the prediction confidence is calculated based on the quality confidence and the stability of the time-series fusion prediction model output: ,in, The prediction confidence level is used to indicate the reliability of the current prediction result. λ is the stability index output by the model, and λ is the trade-off parameter.

[0014] As a preferred embodiment of this application, in S4, the temporal fusion prediction model includes an input layer, a multi-source feature encoding layer, a quality confidence fusion layer, a temporal state modeling layer, and a physical strength prediction layer. The input layer receives a fused feature sequence over T consecutive time steps. The multi-source feature coding layer maps exercise load features, physiological response features, motion decay features, and competition state correction features into a coding vector of the same dimension; The quality confidence fusion layer assigns fusion weights based on the quality confidence of each data source; The temporal state modeling layer extracts the temporal dependencies of physical strength changes through one or more of GRU, LSTM, TCN, or Transformer. The physical strength prediction layer outputs real-time physical strength values ​​within a preset range, as well as a predicted curve of physical strength values, rate of decline of physical strength, and prediction confidence level within a preset future time range.

[0015] This application also provides a real-time physical fitness prediction system for football players based on multi-source sensing and individualized correction using any of the methods described above, comprising: The data acquisition module is used to acquire multi-source real-time data of the target football player during the game or training process; The data processing module is used to perform time synchronization, player identity matching, and abnormal data detection on the multi-source real-time data, and generate continuous time-series status data that matches the target football player. The feature construction module is used to extract physical state-related features from the continuous time-series state data based on a preset sliding time window, and to construct an individualized physical benchmark based on the target football player's historical physical data, on-field position attributes, and physical change curves during historical matches. The individualized correction module is used to normalize and correct the physical condition-related features based on the individualized physical fitness benchmark, and generate individualized physical fitness features that characterize the degree of deviation of the target football player's current physical fitness status relative to his / her normal range of physical fitness variation. The time-series fusion prediction module is used to dynamically weight and fuse the individualized physical condition features based on the quality confidence of data from different sources, and input the fused individualized physical condition features into the time-series fusion prediction model to predict the real-time physical condition value of the target football player and the trend of physical condition value change within a preset time period in the future. The results output module is used to output the real-time physical strength value of the target football player and the trend of physical strength value changes over a preset time period.

[0016] This application also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, it implements the real-time physical fitness prediction method for football players described in any of the above claims.

[0017] This application also provides a computer-readable storage medium storing a computer program; the computer program is used to implement any of the above-mentioned methods for predicting the real-time physical condition of a football player.

[0018] The technical solution described in this application has the following advantages over the prior art: 1. This application integrates real-time data from multiple sources, including motion trajectory data, physiological state data, action posture data, and match status data. It can comprehensively predict the physical strength of football players from multiple perspectives, such as external exercise load, internal physiological response, changes in action quality, and match context, avoiding judgment bias caused by relying on a single indicator. By constructing an individualized physical strength benchmark, the real-time status of the target football player is compared with their personal historical physical fitness level and on-field position attributes, thereby achieving individualized physical strength prediction and reducing misjudgments caused by differences in the physical fitness base of different athletes.

[0019] 2. This application uses a multi-timescale sliding window to extract physical state-related features, which can simultaneously characterize short-term explosive consumption, medium-term continuous exercise load, and long-term fatigue accumulation process, thereby improving the modeling ability of physical strength change patterns. It also introduces a data quality confidence assessment and dynamic weighted fusion mechanism, which can adjust the fusion weights of different data sources under conditions such as video occlusion, sensor anomalies, positioning drift, or data loss, thereby improving the stability and robustness of real-time physical strength value prediction.

[0020] 3. This application can not only output the current real-time physical strength value of the target football player, but also output the trend of physical strength value change, rate of physical strength decline and prediction confidence level within a future preset time period, which can provide quantitative basis for coaches to control training load, arrange substitutions and adjust game tactics.

[0021] 4. This application is applicable to real-time physical strength prediction during formal competitions, as well as training load control and physical recovery assessment during specialized training. In different scenarios, at least two types of data can be selected from video, positioning, inertial sensing, physiological monitoring, and competition status data based on the available data sources, thereby improving the applicability and ease of implementation of the solution.

[0022] 5. In an exemplary verification, data from 4 teaching matches and 8 specialized training sessions of 18 football players were selected as samples. The system updates physical fitness values ​​every 5 seconds and generates reference physical fitness labels by comprehensively considering physical fitness coach stage scores, repeated sprint ability decline rate, heart rate recovery speed, and post-training RPE scale. Compared with physical fitness estimation methods based solely on running distance and a uniform threshold, the mean absolute error of the method in this application is reduced from 9.8 to 6.1, the root mean square error is reduced from 12.6 to 8.2, and the correlation coefficient between the prediction results and the reference physical fitness labels is increased from 0.72 to 0.86. In scenarios simulating video occlusion, positioning drift, or short-term absence of physiological signals, without dynamic fusion of quality confidence, the mean absolute error of physical fitness value prediction is 11.4. After using the quality confidence assessment and dynamic weighted fusion of this application, the mean absolute error is reduced to 7.3, and the continuous output fluctuation amplitude is reduced by about 35%, indicating that this application can maintain good prediction stability in a competition environment with unstable multi-source data quality. The above data are exemplary test results used to illustrate the technical effects of this application. Actual values ​​may vary depending on the data scale, sampling equipment, and model training method. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the real-time physical fitness prediction method for football players provided in this application embodiment.

[0025] Figure 2 This is a flowchart illustrating the continuous time-series state data generation method provided in the embodiments of this application.

[0026] Figure 3 This is a schematic diagram of the multi-source data processing and time synchronization process provided in the embodiments of this application.

[0027] Figure 4 This is a flowchart illustrating the method for generating physical state-related features provided in this application embodiment.

[0028] Figure 5 This is a flowchart illustrating the individualized physical fitness benchmark generation method provided in this application embodiment.

[0029] Figure 6 This is a schematic diagram of the individualized physical fitness benchmark correction process provided in the embodiments of this application.

[0030] Figure 7 This is a schematic diagram of the temporal fusion prediction model structure provided in the embodiments of this application.

[0031] Figure 8 This is a schematic diagram of the module connections of the real-time physical fitness prediction system for football players provided in this application embodiment.

[0032] Figure 9 This is a schematic diagram of the coach's physical strength value visualization display interface provided in the embodiments of this application. Detailed Implementation

[0033] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0034] Before conducting real-time physical fitness prediction, at least the following data acquisition and computing equipment should be prepared: video acquisition equipment, positioning acquisition equipment, inertial sensing equipment, physiological monitoring equipment, match status recording equipment, data processing server, and coach terminal equipment. Specifically, video acquisition equipment is used to capture match or training footage; positioning acquisition equipment is used to obtain player position coordinates, speed, and acceleration; inertial sensing equipment is used to obtain motion status information such as sudden stops, changes of direction, and changes in body posture; physiological monitoring equipment is used to obtain physiological status information such as heart rate, heart rate variability, body temperature, and respiratory rate; the data processing server is used to perform time synchronization, feature extraction, individualized correction, dynamic fusion, and physical fitness prediction; and the coach terminal equipment is used to display real-time physical fitness values ​​and future trends.

[0035] refer to Figure 1 As shown in some embodiments of this application, the method for predicting the real-time physical condition of a football player based on multi-source perception and individualized correction includes the following steps: S1: Acquire multi-source real-time data of the target football player during the game or training process, and perform time synchronization, player identity matching and abnormal data detection on the multi-source real-time data to generate continuous time-series state data matching the target football player.

[0036] The multi-source real-time data includes at least two of the following: motion trajectory data, physiological state data, action posture data, and match status data. The motion trajectory data includes player two-dimensional or three-dimensional coordinates, instantaneous speed, average speed, acceleration, running distance, high-speed running distance, and number of sprints, with a sampling frequency preferably between 10Hz and 25Hz. The physiological state data includes heart rate, heart rate variability (HRV), respiratory rate, body temperature, and heart rate recovery rate, with a sampling frequency preferably between 1Hz and 5Hz. The action posture data includes joint angles, stride frequency, stride length, trunk tilt angle, lower limb swing amplitude, movement stability, and movement decay coefficient, with a sampling frequency preferably between 15Hz and 30Hz. The match status data includes match time, score, on-field position, offensive and defensive status, match rhythm, set-piece events, foul events, and substitution events, and can be updated based on event triggers.

[0037] refer to Figure 2 and Figure 3 As shown. Specifically, in S1, the method includes the following steps: S11: Collect multi-source real-time data of the target football player using one or more of the following: video acquisition equipment, positioning equipment, inertial sensing equipment, physiological monitoring equipment, and match status recording equipment. ,in For video data, For motion trajectory data, For action posture data, For physiological state data, This refers to match status data; S12: Based on the multi-source real-time data, detect speed anomalies, heart rate anomalies, trajectory anomalies, and posture anomalies, and process the anomalies using methods such as linear interpolation, previous frame filling, sliding window smoothing, or weight reduction.

[0038] Specifically, abnormal data detection rules may include: speed abnormality: instantaneous speed greater than 12m / s is considered abnormal; heart rate abnormality: heart rate greater than 220 minus age, or sudden changes greater than 30 beats / minute for 3 consecutive seconds are considered abnormal; trajectory abnormality: single frame displacement greater than 5 meters, positioning jumps, or prolonged stillness for more than 10 seconds are considered abnormal; posture abnormality: joint angles exceeding the physiological range of the human body or confidence of posture key points less than 0.3 are considered abnormal; abnormal data can be processed by linear interpolation, previous frame filling, sliding window smoothing, or weight reduction.

[0039] S13: Since the sampling frequency and timestamp of different data sources may be different, this embodiment performs time synchronization processing on multi-source real-time data, unifying the data from different sources to the same or similar time nodes. That is, based on the video frame timestamp or unified competition clock, all data in the multi-source real-time data are unified to the same time axis through linear interpolation, nearest neighbor interpolation or time window averaging.

[0040] Among these, the preferred synchronization error is no more than 100ms; for physiological state data with a low sampling frequency, alignment can be achieved between two adjacent sampling points using methods such as hold, interpolation, or window statistics.

[0041] S14: Identify player numbers, team uniform colors, or initial positioning information, bind video trajectory IDs, and establish a correspondence between video trajectory IDs and various device numbers to build a unique player ID mapping table to ensure that data from different sources belong to the same target football player. When a player is obscured or the trajectory is temporarily lost, rematch is performed using the previous position, speed direction, and wearable device number.

[0042] S15: Based on the multi-source real-time data and the unique player ID mapping table, generate continuous time-series state data matching the target football player.

[0043] S2: Based on a preset sliding time window, extract physical condition-related features from the continuous time-series state data and construct an individualized physical condition benchmark based on the target football player's historical physical condition data, on-field position attributes, and physical condition change curves during historical matches.

[0044] The physical condition-related characteristics include one or more of the following: exercise load characteristics, physiological response characteristics, movement decay characteristics, and competition state correction characteristics.

[0045] The exercise load characteristics include instantaneous speed, average speed, maximum speed, percentage of high-speed running, running distance, mean acceleration, mean deceleration, number of sprints, number of changes of direction, and number of sudden stops, which are used to characterize the physical exertion of the target football player at the level of external load.

[0046] The physiological response characteristics include real-time heart rate, heart rate as a percentage of maximum heart rate, heart rate variability (HRV), heart rate recovery rate, respiratory rate, body temperature changes, and physiological load index, which are used to characterize the physical exertion of the target football player at the internal physiological response level.

[0047] Motion decay features characterize the decline in movement quality caused by fatigue. These include cadence decay rate, stride decay rate, postural joint angle change rate, trunk tilt fluctuation, lower limb joint angle change amplitude, and movement stability coefficient. Cadence decay rate is calculated by the difference between the current window cadence and the individual's historical average cadence for the same period. Stride decay rate is calculated by the difference between the current window stride and the individual's historical average stride. Postural joint angle change rate is calculated by the angle change amplitude of lower limb joints such as the hip, knee, and ankle in consecutive frames. The movement stability coefficient is determined by considering the fluctuation amplitude of postural key points in consecutive frames, the consistency of left and right lower limb swing, and joint angular velocity fluctuation. Generally, a higher motion decay feature value indicates a more significant decline in the target soccer player's movement quality, corresponding to a lower predicted physical strength value.

[0048] The game status correction features include game time, on-field position, offensive and defensive status, score difference, game intensity level, team ball control status, game rhythm, weather conditions, and field conditions, which are used to distinguish between normal tactical consumption and abnormal physical exhaustion.

[0049] The physical condition-related features are extracted using a multi-timescale sliding window. The short time window, set to the most recent 10 to 30 seconds, is used to extract explosive exercise consumption features such as acceleration, deceleration, number of sprints, number of sudden stops, and number of changes of direction. The medium time window, set to the most recent 1 to 5 minutes, is used to extract continuous exercise load features such as running distance, movement speed, high-speed running distance, and duration of sustained high-intensity exercise. The long time window, set to the most recent 5 to 15 minutes or from the start of the competition to the current moment, is used to extract overall fatigue accumulation features such as cumulative exercise load, heart rate recovery trend, trend of declining movement stability, and historical trend of physical decline. Different timescale windows can overlap and slide, with an update step length preferably between 1 and 5 seconds.

[0050] Therefore, through the above-mentioned multi-timescale characteristics, this application can simultaneously reflect instantaneous burst consumption, continuous exercise load, and long-term fatigue accumulation.

[0051] refer to Figure 4 As shown. Specifically, in S2, the method includes the following steps: S21: Set the sliding time window W with the current time t as the endpoint; S22: Calculate the number of sprints, sudden stops, changes of direction, acceleration, and deceleration within a short time window to generate explosive exercise consumption characteristics; S23: Calculate running distance, high-speed running distance, average speed, average heart rate, and heart rate recovery rate within a medium time window to generate continuous exercise load characteristics; S24: Calculate the cumulative load, duration of continuous high-intensity exercise, trend of decline in movement stability, and historical trend of decline in physical strength within a long time window to generate overall fatigue accumulation characteristics; S25: Based on the explosive exercise consumption characteristics, continuous exercise load characteristics, and overall fatigue accumulation characteristics, generate physical state-related features of multi-source feature vectors: ,in This is characterized by explosive energy expenditure within a short time window. Characteristics of continuous motion load within a medium time window. This represents the overall fatigue accumulation characteristics within a long time window. Adjust features for game status.

[0052] The individualized physical fitness benchmark is used to characterize the range of normal physical fitness changes of the target football player in different stages of the game, different positions on the field, different exercise intensities, and different game conditions. The individualized physical fitness benchmark is established based on the target football player's historical maximum speed, historical average running intensity, historical high-speed running ability, historical sprint recovery time, historical heart rate recovery ability, historical movement stability change pattern, physical fitness change curve in historical game stages, on-field position attributes, and one or more of the following: score, game time, offensive and defensive status, game rhythm, weather, or field conditions.

[0053] refer to Figure 5 As shown. Specifically, in S2, the method further includes the following steps: S26: Extract historical physical fitness data from the target football player's preset number of matches, preset number of special training sessions, and pre-match physical fitness tests. The historical physical fitness data includes one or more of the following: maximum speed, average running intensity, high-speed running distance, sprint recovery time, heart rate recovery ability, and movement stability variation patterns.

[0054] For example, historical physical fitness data of the target football player can be extracted from the last 5 matches, recent specialized training sessions, and pre-match physical fitness tests. The specific settings are determined by the designers based on actual needs.

[0055] S27: Segment the statistics according to the competition stage, preferably into 0-15min, 15-30min, 30-45min, 45-60min, 60-75min and 75-90min, or other segments can be set according to the training or competition type; S28: Calculate the typical physical fatigue decay curve for each position according to the on-field position attributes, wherein the on-field position attributes include at least forwards, midfielders, defenders and goalkeepers, and different positions correspond to different running patterns and physical fatigue benchmarks; S29: Combining correction factors including score difference, offensive and defensive status, game rhythm, weather conditions and field conditions, generate individual normal physical strength ranges, mean and standard deviations for different times, positions, intensities and game conditions to form individualized physical strength benchmarks.

[0056] Therefore, it can avoid misjudgment caused by using a uniform threshold to evaluate the physical condition of different athletes, and improve the stability of physical condition prediction under different positions and different stages of the competition.

[0057] S3: Based on the individualized physical fitness benchmark, normalize and correct the physical fitness status-related features to generate individualized physical fitness status features that characterize the degree of deviation of the target football player's current physical fitness status relative to his / her normal range of physical fitness changes.

[0058] refer to Figure 6 As shown. Specifically, in S3, the method further includes the following steps: Based on the individualized physical fitness benchmark, the physical fitness-related features are normalized and corrected using the Z-score method to generate individualized physical fitness features: ,in, For the target football player's normal physical fitness benchmark under the current stage of the game and position attributes, preferably, This data can be obtained by statistically analyzing the athlete's recent match data, training test data, and pre-match physical fitness test data, segmented by competition stage. To correspond to the historical fluctuation scale, the standard deviation of historical characteristics under the same competition stage and the same position task is taken; ε is a constant to prevent the denominator from being zero, and its preferred value is [value missing]. ; Used to characterize the degree of deviation of a target football player's current physical condition from their historical normal level; where, when When the absolute value of increases, it indicates that the current state deviates more from the individual's normal range of physical strength variation.

[0059] The normalization correction includes comparing the physical condition-related features of the target football player at the current moment with the individualized physical condition benchmark to obtain standardized offset features; the standardized offset features are used to reflect the degree of deviation of the target football player's current state from its own historical normal state.

[0060] Specifically, the system compares real-time physical condition features with individualized physical fitness benchmarks to obtain the physical condition deviation of the target soccer player relative to their historical physical fitness level. For features such as speed, heart rate, number of sprints, or movement decay that exceed the historical normal exertion range, the system maps them to a higher exertion deviation; for features within the individual's historical normal range, the system maps them to a lower exertion deviation. The resulting individualized physical condition features include both real-time athletic performance and the degree of deviation from the individual's normal state.

[0061] S4: Based on the quality confidence of data from different sources, the individualized physical condition features are dynamically weighted and fused, and the fused individualized physical condition features are input into the time series fusion prediction model to predict and output the real-time physical condition value of the target football player and the trend of physical condition value change within a preset time period in the future.

[0062] The quality confidence level is determined based on one or more of the following: data integrity, temporal continuity, numerical fluctuation amplitude, sensor signal stability, video occlusion degree, player trajectory continuity, and anomaly detection results; and the motion trajectory features, physiological response features, action decay features, and match state correction features are dynamically weighted and fused based on the quality confidence level.

[0063] The process involves evaluating the quality and confidence of data from different sources. If the target soccer player in the video is occluded, causing instability in posture recognition, the quality and confidence of the corresponding motion posture data are reduced. Similarly, if the positioning trajectory shows significant jumps or discontinuities, the quality and confidence of the corresponding motion trajectory data are reduced. Furthermore, if heart rate data shows abnormal abrupt changes or prolonged absence, the quality and confidence of the corresponding physiological state data are reduced. The quality and confidence can be determined jointly by data reliability, completeness, temporal continuity, degree of occlusion, and anomaly detection results.

[0064] Specifically, when outputting the real-time physical strength value of the target football player and the trend of physical strength value change within a preset time period, the data is output to a terminal device, which can be a coach's tablet computer, a match analysis terminal, a training monitoring platform, or a back-end server; the specific output content includes the current real-time physical strength value, the predicted curve of future physical strength value, the rate of decline in physical strength, the prediction confidence level, and optional abnormal data prompts.

[0065] refer to Figure 7 As shown. Specifically, in S4, the method includes the following steps: S41: Set quality confidence level: ,in, For the reliability of the k-th type of data source, For the integrity of the k-th type of data source, For the time continuity of the k-th type of data source, The degree of occlusion or anomaly of the k-th type of data source. These are weight parameters; In a preferred embodiment of this application, it is possible to And all weights are non-negative and satisfy the following conditions: Alternatively, the weights can be adaptively adjusted using validation sets or historical match data. The fusion weights are calculated based on the quality confidence level.

[0066] S42: Calculate the fusion weights based on the quality confidence level: Generate the fused individual physical condition characteristics: , For the feature encoding function corresponding to the k-th type of data source, For the k-th type of data source, the features are after individualized correction. The characteristics of the individualized physical condition after integration; Let be the quality confidence score of the j-th type of data source at time t.

[0067] By normalizing the fusion weights ,make Furthermore, the sum of the fusion weights of all data sources is 1, thereby automatically reducing the impact of a data source on the final physical strength prediction result when a data source is obscured, drifted, missing, or experiences abnormal fluctuations.

[0068] In some examples of this application, in S4, the method further includes the following steps: S43: Input the fused features within the continuous time window into the time-series fusion prediction model to model the dynamic changes in the physical state of the target football player: ,in, A function for modeling temporal states, which is implemented by one or more of the following: recurrent neural networks, gated recurrent unit networks, long short-term memory networks, temporal convolutional networks, attention networks, or Transformer networks; S44: Predict the target football player's real-time stamina: ,in, The real-time stamina value of the target football player at time t is output in the range of [0,100], which can be understood as the remaining stamina value in percentage form. The higher the value, the stronger the current remaining athletic ability. The bias parameter for the physical fitness prediction layer is used to perform baseline translation correction on the mapping results between the time-series state modeling results and the real-time physical fitness values.

[0069] The real-time physical strength value is a quantitative value used to characterize the current remaining athletic ability of the target football player. The real-time physical strength value is expressed in a percentage system, a grade system, or a normalized numerical form.

[0070] S45: Predict the trend of a target football player's physical strength over a future preset time period. ,in For a predetermined time period in the future, the trend of physical strength changes within that predetermined time period includes one or more of the following: a future physical strength prediction curve, a rate of decline in physical strength, an inflection point in physical strength decline, and a prediction confidence level; when When the value is negative and the absolute value is large, it indicates that the target football player's stamina value will decrease rapidly within the preset time period in the future.

[0071] In some examples of this application, in S4, to reflect the reliability of the prediction results, the prediction confidence is calculated based on the quality confidence and the stability of the temporal fusion prediction model output: ,in, The prediction confidence level is used to indicate the reliability of the current prediction result. λ is the stability index output by the model, and λ is the trade-off parameter.

[0072] In a preferred embodiment of this application, λ can be taken as 0.7; when the fluctuation range of the model output within three consecutive update cycles is less than 5%, it can be considered that... It is in a highly stable state; the prediction confidence level can be output together with the real-time physical strength value and the future trend of physical strength change to indicate the reliability of the current prediction results to the coach or back-end system.

[0073] refer to Figure 7As shown. In some examples of this application, in S4, the temporal fusion prediction model includes an input layer, a multi-source feature encoding layer, a quality confidence fusion layer, a temporal state modeling layer, and a physical strength prediction layer; The input layer receives a fused feature sequence over T consecutive time steps. The multi-source feature coding layer maps exercise load features, physiological response features, motion decay features, and competition state correction features into a coding vector of the same dimension; The quality confidence fusion layer assigns fusion weights based on the quality confidence of each data source; The temporal state modeling layer extracts the temporal dependencies of physical strength changes through one or more of GRU, LSTM, TCN, or Transformer. The stamina prediction layer outputs real-time stamina values ​​within the range of [0,100], a stamina prediction curve for the next 3 to 10 minutes, the rate of stamina decline, and prediction confidence.

[0074] When training the time-series fusion prediction model, reference physical fitness labels can be constructed by comprehensively considering factors such as physical fitness coach stage scores, post-training RPE scale, heart rate recovery speed, repeated sprint ability decline rate, and staged running ability decline rate. Supervised training is then performed using mean squared error loss, mean absolute error loss, or a weighted combination of both. To avoid overfitting, the model can be trained using a training set, validation set, and test set. The validation set is used to determine the quality confidence weights, time window length, and time-series model parameters.

[0075] It should be noted that the core of this application lies in the real-time prediction of football players' physical fitness. The substitution reference, training load control, or tactical adjustment prompts in the output results are all application scenarios of the physical fitness prediction results and do not constitute a limitation on the core technical solution of this application.

[0076] refer to Figure 8 As shown. This application also relates to a real-time physical fitness prediction system for football players based on multi-source sensing and individualized correction, using the method described in any of the above embodiments, comprising: The data acquisition module is used to acquire multi-source real-time data of the target football player during the game or training process; The data processing module is used to perform time synchronization, player identity matching, and abnormal data detection on the multi-source real-time data, and generate continuous time-series status data that matches the target football player. The feature construction module is used to extract physical state-related features from the continuous time-series state data based on a preset sliding time window, and to construct an individualized physical benchmark based on the target football player's historical physical data, on-field position attributes, and physical change curves during historical matches. The individualized correction module is used to normalize and correct the physical condition-related features based on the individualized physical fitness benchmark, and generate individualized physical fitness features that characterize the degree of deviation of the target football player's current physical fitness status relative to his / her normal range of physical fitness variation. The time-series fusion prediction module is used to dynamically weight and fuse the individualized physical condition features based on the quality confidence of data from different sources, and input the fused individualized physical condition features into the time-series fusion prediction model to predict the real-time physical condition value of the target football player and the trend of physical condition value change within a preset time period in the future. The results output module is used to output the real-time physical strength value of the target football player and the trend of physical strength value changes over a preset time period.

[0077] Therefore, in some embodiments of this application, the system updates the physical condition of the target football player every 5 seconds, outputting the current real-time physical condition value, the predicted physical condition value for the next 5 minutes, the rate of physical condition decline, and the prediction confidence level; when the current real-time physical condition value of the 8th midfielder is 72, the predicted physical condition value for the next 5 minutes is 63, the rate of physical condition decline is -1.8 / minute, and the prediction confidence level is 0.86, the system can display the player's current physical condition, historical physical condition change curve, and future physical condition change trend on the coach's terminal device.

[0078] refer to Figure 7 and Figure 9 As shown, in some embodiments of this application, a set of exemplary effect verifications are performed; data from 4 teaching matches and 8 specialized training sessions of 18 football players are collected. The data includes motion trajectory data at 10Hz-25Hz, physiological state data at 1Hz-5Hz, action posture data at 15Hz-30Hz, and match status data updated according to event triggers. The system outputs physical strength values ​​using the aforementioned time synchronization, identity matching, feature extraction, individualized correction, quality confidence fusion, and temporal prediction steps with a 5-second update cycle.

[0079] In this exemplary verification, the reference fitness label is derived from a combination of strength and conditioning coach stage ratings, repeated sprint performance decline rate, heart rate recovery rate, and the post-training RPE scale. A comparison of the method in this application with fitness estimation methods based solely on running distance and a uniform threshold shows that the uniform threshold method has a mean absolute error of 9.8, a root mean square error of 12.6, and a correlation coefficient of 0.72; while the method in this application has a mean absolute error of 6.1, a root mean square error of 8.2, and a correlation coefficient of 0.86.

[0080] Furthermore, under simulated video occlusion, positioning drift, and short-term absence of physiological signals, the mean absolute error of physical fitness prediction was 11.4 without dynamic fusion based on quality confidence. After applying the dynamic fusion based on quality confidence of this application, the mean absolute error of physical fitness prediction was reduced to 7.3, and the fluctuation amplitude of continuous output was reduced by approximately 35%. This example illustrates that this application can maintain good stability and robustness in physical fitness prediction in a competition environment with unstable multi-source data quality. The above verification data is used to illustrate the technical effects of this application and is not intended to limit the scope of protection of this application.

[0081] To further illustrate the specific implementation process of this application in different application scenarios, the following embodiments are all based on the processing flow of S1 to S4 described above, and respectively illustrate the specific data collection, feature correction, dynamic fusion and result output methods in formal competition scenarios and special training scenarios.

[0082] refer to Figure 1 , Figure 6 and Figure 9 As shown, in some embodiments of this application, the application is used to predict the real-time physical fitness of target players during official matches. Before the start of the match, the system imports the target player's data from the last 6 official matches, the last 3 specialized training sessions, and pre-match physical fitness tests, and establishes individualized physical fitness benchmarks according to match stages such as 0-15min, 15-30min, 30-45min, 45-60min, 60-75min, and 75-90min. Simultaneously, it records the player's historical maximum speed, high-speed running distance, sprint recovery time, heart rate recovery ability, and movement stability changes under different positions such as midfielder, winger, or full-back. Taking midfielder number 8 as an example, the system uses his historical high-speed running ability, heart rate recovery amplitude 30 seconds after sprint, lower limb movement stability, and physical fitness decline curve in the second half of the match as individual benchmark parameters, instead of directly using a uniform threshold for the entire team.

[0083] At the 62nd minute of the match, the data acquisition module receives the target player's location trajectory data, physiological state data, inertial sensor data, and match status data. The data processing module synchronizes the time using a unified match clock and binds the video trajectory ID, positioning device number, heart rate belt number, and inertial sensor number to the same player ID. The feature construction module extracts features such as the number of sprints, the number of sudden stops, the distance covered at high speed, average heart rate, heart rate recovery rate, stride frequency decline rate, trunk tilt fluctuation, and the trend of declining movement stability within sliding windows of the most recent 20 seconds, the most recent 3 minutes, and the most recent 10 minutes, respectively. These features are then combined with the current score, the frequency of offensive and defensive transitions, and the team's ball possession status to generate match status correction features.

[0084] At this moment, if the target player completes two consecutive high-speed defensive runs and experiences a brief video obstruction, the system can detect that their recent 3-minute high-speed running distance is approximately 35% higher than their personal baseline for the same period, their post-sprint heart rate recovery is lower than their historical normal range, and their movement stability decreases by approximately 18%. Simultaneously, due to the reduced confidence of posture key points caused by video obstruction, the quality confidence of the movement posture data is lowered to 0.55, while the quality confidence of the positioning trajectory data, physiological state data, and inertial sensor data remain at 0.91, 0.88, and 0.84, respectively. Based on this, the quality confidence fusion unit reduces the weight of posture data in the final fused features and increases the relative weight of positioning, physiological, and inertial sensor data, thereby preventing posture errors caused by obstruction from being directly amplified into physical fitness prediction errors.

[0085] After receiving the fusion features within a continuous time window, the temporal fusion prediction module outputs the player's current real-time stamina value as 58, the predicted stamina value range for the next 5 minutes as 50-52, the stamina decline rate as -1.4 / minute, and the prediction confidence level as 0.83. The results output module displays these results on the coach's terminal device as curves, numerical values, and confidence level indicators, and can simultaneously display explanatory prompts such as "High energy expenditure during the second half of the high-intensity defensive retreat," "Heart rate recovery below the individual's normal range," and "Posture data is obscured and weighted." These prompts are intended to assist coaches in understanding the reasons for changes in stamina values. Substitution arrangements or tactical adjustments are merely application scenarios for the stamina value prediction results and do not limit the scope of protection of this application.

[0086] In some embodiments of this application, the application is used for training load control and physical recovery assessment during specialized training. Before training begins, the system reads each athlete's historical training test data, pre-competition physical fitness test data, and positional attributes, and establishes corresponding individualized physical fitness benchmarks based on the type of training program. The training program can be repetitive sprint training, high-intensity interval running training, offensive and defensive transition training, or small-field sparring training. Taking repetitive sprint training as an example, the training content includes 6 sets of 30-second high-intensity sprints and 90-second low-intensity recovery. The system collects at least positioning trajectory data and physiological state data, and can further collect inertial sensor data or video posture data.

[0087] During each training session, the feature construction module calculates the maximum speed, peak acceleration, number of sprints, and number of sudden stops within a short time window; the average speed, high-speed running distance, heart rate as a percentage of maximum heart rate, and heart rate recovery rate within a medium time window; and the cumulative load, trend of decreased movement stability, and inter-set recovery trend within a long time window. The individualized correction module compares these features with the athlete's historical normal range in similar training sessions to obtain standardized offset features relative to the individual baseline. For example, for a forward, if after the fourth training session their maximum speed is still close to their historical normal level, but their heart rate recovery rate is significantly lower than the historical normal range, and their stride frequency decline rate and movement stability decline rate continue to increase, the system can determine that their external running performance has not yet declined significantly, but their internal physiological load and movement quality have already shown a high fatigue offset.

[0088] If a heart rate monitor detaches briefly or physiological data is missing for 12 consecutive seconds during training, the data processing module marks this period as incomplete physiological data and temporarily compensates for it by interpolating before and after windows or maintaining the previous stable value. The quality confidence calculation unit reduces the quality confidence of the physiological state data based on completeness and temporal continuity, while increasing the weight of positioning trajectory data and inertial sensor data in the fusion result. Once the physiological data stabilizes, the system gradually increases their fusion weight. Through this process, even if a data source in the training environment experiences a short-term anomaly, the system can still continuously output relatively stable physical strength values ​​and fatigue trends.

[0089] In this training example, the system can output the athlete's real-time physical fitness value every 5 seconds and output interim results after each training session. If a player's current real-time physical fitness value is 47, and the predicted physical fitness value may drop below 40 in the next 3 minutes with a prediction confidence level of 0.81, the result output module can display the player's current physical fitness value, inter-set recovery trend, sources of abnormal data, and rate of physical fitness decline on the training monitoring platform. After training, the system can also compare the predicted physical fitness value curve with the coach's interim score, the post-training RPE scale, and heart rate recovery speed; when the prediction confidence level is higher than a preset threshold and the data quality meets the requirements, the training data is used as reference data for subsequent individualized physical fitness benchmark updates. Therefore, this embodiment demonstrates that this application is not only applicable to real-time monitoring in official competitions but also to load management, recovery assessment, and continuous correction of individual physical fitness benchmarks in daily training.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting the real-time physical fitness of football players based on multi-source sensing and individualized correction, characterized in that, Includes the following steps: S1: Acquire multi-source real-time data of the target football player during the game or training process, and perform time synchronization, player identity matching and abnormal data detection on the multi-source real-time data to generate continuous time-series state data matching the target football player. S2: Based on a preset sliding time window, extract physical state-related features from the continuous time-series state data and construct an individualized physical benchmark according to the target football player's historical physical data, on-field position attributes, and physical change curves during historical matches. S3: Based on the individualized physical fitness benchmark, the physical fitness-related features are normalized and corrected to generate individualized physical fitness features that characterize the degree of deviation of the target football player's current physical fitness status relative to his / her normal range of physical fitness changes. S4: Based on the quality confidence of data from different sources, the individualized physical condition features are dynamically weighted and fused, and the fused individualized physical condition features are input into the time series fusion prediction model to predict and output the real-time physical condition value of the target football player and the trend of physical condition value change within a preset time period in the future.

2. The method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 1, characterized in that, In S1, the method includes the following steps: S11: Collect multi-source real-time data of the target football player using one or more of the following: video acquisition equipment, positioning equipment, inertial sensing equipment, physiological monitoring equipment, and match status recording equipment. ,in For video data, For motion trajectory data, For action posture data, For physiological state data, This refers to match status data; S12: Based on the multi-source real-time data, detect speed anomalies, heart rate anomalies, trajectory anomalies, and posture anomalies, and process the anomalies using methods such as linear interpolation, previous frame filling, sliding window smoothing, or weight reduction. S13: Using video frame timestamps or a unified competition clock as a reference, all data in the multi-source real-time data are unified to the same time axis through linear interpolation, nearest neighbor interpolation, or time window averaging. S14: Identify player numbers, team uniform colors, or initial positioning information, bind video trajectory IDs, establish a correspondence between video trajectory IDs and various device numbers, and construct a unique player ID mapping table to ensure that data from different sources all belong to the same target football player; S15: Based on the multi-source real-time data and the unique player ID mapping table, generate continuous time-series state data matching the target football player.

3. A method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 1 or 2, characterized in that, In S2, the method includes the following steps: S21: Set the sliding time window W with the current time t as the endpoint; S22: Calculate the number of sprints, sudden stops, changes of direction, acceleration, and deceleration within a short time window to generate explosive exercise consumption characteristics; S23: Calculate running distance, high-speed running distance, average speed, average heart rate, and heart rate recovery rate within a medium time window to generate continuous exercise load characteristics; S24: Calculate the cumulative load, duration of continuous high-intensity exercise, trend of decline in movement stability, and historical trend of decline in physical strength within a long time window to generate overall fatigue accumulation characteristics; S25: Based on the explosive exercise consumption characteristics, continuous exercise load characteristics, and overall fatigue accumulation characteristics, generate physical state-related features of multi-source feature vectors: ,in This is characterized by explosive energy expenditure within a short time window. Characteristics of continuous motion load within a medium time window. This represents the overall fatigue accumulation characteristics within a long time window. Adjust features for game status.

4. The method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 3, characterized in that, In S2, the method further includes the following steps: S26: Extract historical physical fitness data from the target football player's preset number of matches, preset number of special training sessions, and pre-match physical fitness tests. The historical physical fitness data includes one or more of the following: maximum speed, average running intensity, high-speed running distance, sprint recovery time, heart rate recovery ability, and movement stability variation patterns. S27: Statistics are segmented according to competition stage, training, or competition type; S28: Calculate the typical physical fatigue decay curve for each position according to the on-field position attributes, wherein the on-field position attributes include at least forwards, midfielders, defenders and goalkeepers, and different positions correspond to different running patterns and physical fatigue benchmarks; S29: Combining correction factors including score difference, offensive and defensive status, game rhythm, weather conditions and field conditions, generate individual normal physical strength ranges, mean and standard deviations for different times, positions, intensities and game conditions to form individualized physical strength benchmarks.

5. The method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 4, characterized in that, In S3, the method further includes the following steps: Based on the individualized physical fitness benchmark, the physical fitness-related features are normalized and corrected to generate individualized physical fitness features: ,in, This serves as a baseline for the normal physical condition of the target football player under the current stage of the match and their position attributes. To correspond to the historical fluctuation scale, the standard deviation of historical characteristics under the same competition stage and the same position task is taken; ε is a constant to prevent the denominator from being zero. Used to characterize the degree of deviation of a target football player's current physical condition from their historical normal level; where, when When the absolute value of increases, it indicates that the current state deviates more from the individual's normal range of physical strength variation.

6. A method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 1 or 5, characterized in that, In S4, the method includes the following steps: S41: Set quality confidence level: ,in, For the reliability of the k-th type of data source, For the integrity of the k-th type of data source, For the time continuity of the k-th type of data source, The degree of occlusion or anomaly of the k-th type of data source. These are weight parameters; S42: Calculate the fusion weights based on the quality confidence level: Generate the fused individual physical condition characteristics: , For the feature encoding function corresponding to the k-th type of data source, For the k-th type of data source, the features are after individualized correction. The characteristics of the individualized physical condition after integration; Let be the quality confidence score of the j-th type of data source at time t.

7. The method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 6, characterized in that, In S4, the method further includes the following steps: S43: Input the fused features within the continuous time window into the time-series fusion prediction model to model the dynamic changes in the physical state of the target football player: ,in, A function for modeling temporal states, which is implemented by one or more of the following: recurrent neural networks, gated recurrent unit networks, long short-term memory networks, temporal convolutional networks, attention networks, or Transformer networks; S44: Predict the target football player's real-time stamina: ,in, The real-time stamina value of the target football player at time t; The bias parameter of the physical fitness prediction layer is used to perform baseline translation correction on the mapping results between the time series state modeling results and the real-time physical fitness values; S45: Predict the trend of a target football player's physical strength over a future preset time period. , For a future time period.

8. The method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 7, characterized in that, In S4, to reflect the reliability of the prediction results, the prediction confidence is calculated based on the quality confidence and the stability of the time-series fusion prediction model output: ,in, The prediction confidence level is used to indicate the reliability of the current prediction result. λ is the stability index output by the model, and λ is the trade-off parameter.

9. A method for predicting the real-time physical condition of a football player based on multi-source sensing and individualized correction according to claim 7 or 8, characterized in that, In S4, the temporal fusion prediction model includes an input layer, a multi-source feature encoding layer, a quality confidence fusion layer, a temporal state modeling layer, and a physical strength prediction layer. The input layer receives a fused feature sequence over T consecutive time steps. The multi-source feature coding layer maps exercise load features, physiological response features, motion decay features, and competition state correction features into a coding vector of the same dimension; The quality confidence fusion layer assigns fusion weights based on the quality confidence of each data source; The temporal state modeling layer extracts the temporal dependencies of physical strength changes through one or more of GRU, LSTM, TCN, or Transformer. The physical strength prediction layer outputs real-time physical strength values ​​within a preset range, as well as a predicted curve of physical strength values, rate of decline of physical strength, and prediction confidence level within a preset future time range.

10. A real-time physical fitness prediction system for football players based on multi-source sensing and individualized correction, using the method described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-source real-time data of the target football player during the game or training process; The data processing module is used to perform time synchronization, player identity matching, and abnormal data detection on the multi-source real-time data, and generate continuous time-series status data that matches the target football player. The feature construction module is used to extract physical state-related features from the continuous time-series state data based on a preset sliding time window, and to construct an individualized physical benchmark based on the target football player's historical physical data, on-field position attributes, and physical change curves during historical matches. The individualized correction module is used to normalize and correct the physical condition-related features based on the individualized physical fitness benchmark, and generate individualized physical fitness features that characterize the degree of deviation of the target football player's current physical fitness status relative to his / her normal range of physical fitness variation. The time-series fusion prediction module is used to dynamically weight and fuse the individualized physical condition features based on the quality confidence of data from different sources, and input the fused individualized physical condition features into the time-series fusion prediction model to predict the real-time physical condition value of the target football player and the trend of physical condition value change within a preset time period in the future. The results output module is used to output the real-time physical strength value of the target football player and the trend of physical strength value changes over a preset time period.

Images