Multi-dimensional tennis action analysis system

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-dimensional tennis action analysis system, which comprises a data acquisition module, a data preprocessing module, an evaluation analysis module and a feedback interaction module, and is characterized in that the data acquisition module comprises a posture data acquisition unit, a ball path and walking tracking unit and a motion parameter acquisition unit; the data preprocessing module carries out space-time calibration on posture, track and motion parameters based on a timestamp synchronization mechanism, a fusion data set is formed through data cleaning and standardization processing, and the evaluation analysis module is used for distinguishing action types, integrating scores of posture dimensions, power generation dimensions and drop point dimensions and generating a comprehensive evaluation report; and the feedback interaction module displays an evaluation result in real time on an interface based on characters and visual icons and outputs historical analysis. According to the invention, the data validity is guaranteed, the error interference is eliminated, the analysis efficiency and the professionality of the evaluation result are improved, the use threshold is reduced, and the universality and practicability of the system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a multi-dimensional tennis action analysis system. BACKGROUND

[0002] At present, many enthusiasts choose to learn tennis through network videos or imitation practice, but the self-study group generally has problems such as non-standard technical action, bad hitting habit, lack of scientific training concept, and weak injury prevention awareness, which seriously affects the movement effect and health safety.

[0003] To solve the above problems, some training institutions have tried to introduce intelligent technology to assist training, for example, the Smartshot system, which can automatically generate student hitting highlights based on computer vision technology, provide basic data analysis such as ball speed and hitting point, and preliminarily realize intelligent recording and feedback of the training process.

[0004] However, the deployment threshold of such products is high, and such systems mainly analyze macro statistical data, and have not covered micro biomechanical indicators such as swing trajectory, shoulder and elbow rotation amplitude, and trunk rotation angle, which are the key to evaluating action standardization and preventing sports injuries. Therefore, how to provide high-precision micro-motion analysis and feedback support at a low cost and build a scientific and personalized tennis training system is an important direction for current technology application and education practice to break through. In view of this, in order to solve the above problems, the present application provides a multi-dimensional tennis action analysis system. SUMMARY

[0005] In view of the above situation, in order to overcome the defects of the prior art, the present application provides a multi-dimensional tennis action analysis system, which comprises a data acquisition module, a data preprocessing module, an evaluation analysis module and a feedback interaction module:

[0006] The data acquisition module comprises a posture data acquisition unit, a ball route and walking position tracking unit, and a motion parameter acquisition unit. The posture data acquisition unit locks the student's movement trajectory by using a pan-tilt posture capture camera combined with a real-time tracking algorithm, extracts the dynamic characteristics of the key joints of the body (such as shoulders, elbows, wrists, hips, etc.) in the hitting process by using the MediaPipe machine learning framework, and generates a standardized posture vector. The ball route and walking position tracking unit completes the movement trajectory of the ball by relying on a full-field bird's eye camera combined with YOLO11 algorithm, synchronously records the walking trajectory of the student in the field, and realizes synchronous acquisition of the "student-ball" linkage information. The motion parameter acquisition unit acquires the racket motion signal through an acceleration sensor, and converts the acceleration data into speed parameters and three-dimensional motion trajectories by using integral operation;

[0007] The data preprocessing module synchronizes the posture, force, and landing point data based on a timestamp synchronization mechanism, forms a fusion data set through data cleaning and standardization processing, and uses a fixed-size sliding window (the window length can be set according to the motion characteristics). The sliding window continuously receives new data and eliminates old data, while maintaining a constant length. The key features of the acceleration data in the window are calculated to determine whether the middle position of the sliding window reaches the maximum acceleration value (the hitting threshold). If it does, the data in the sliding window at that time is the hitting frame.

[0008] The evaluation analysis module is used to distinguish the action type, integrate the scores of the posture dimension, force dimension, and landing point dimension, and generate a final total score.

[0009] The feedback interaction module displays the evaluation results in real time on the interface based on text and visual icons and outputs historical analysis.

[0010] Further, the pan-tilt posture capture camera uses Raspberry Pi as an edge computing node, carries a CSI HD camera and a double-rudder mechanism to form the hardware foundation, and realizes horizontal or vertical rotation through rudder driving. Relying on the YOLO target detection model, the coordinates of the target are outputted.

[0011] Further, the full-field bird's eye view camera uses a high-resolution wide-angle camera to generate a full-field bird's eye view without dead angles through hardware distortion correction and image stitching technology.

[0012] Further, the acceleration sensor is a WT931 high-precision three-axis acceleration sensor that collects acceleration data of the tennis racket in three orthogonal axes in real time. The image frame frequency of the pan-tilt posture capture camera is consistent with the sampling frequency of the WT931 three-axis acceleration sensor.

[0013] Further, the posture vector extracted by the MediaPipe machine learning framework includes the coordinates of the left and right shoulders, elbows, wrists, hips, knees, and ankles. Based on multiple coordinates, eight key angle features are generated, including the left arm angle, right arm angle, left arm distance from the body angle, right arm distance from the body angle, left shoulder-hip-knee angle, right shoulder-hip-knee angle, left leg angle, and right leg angle.

[0014] Further, the eight key angle features of the MediaPipe machine learning framework are extracted, and the StandardScaler standardization is used to eliminate the dimension difference. The standardization formula is:

[0015]

[0016] where, ​the original feature value, the feature mean value, the feature standard deviation, taking the data at the moment of hitting the ball, calculating the mean value, maximum value, minimum value and standard deviation of each action sequence, constructing a 32-dimensional feature vector, and then performing support vector machine model construction and training, the SVM model finds the optimal hyperplane to distinguish the forehand (label 0) and backhand (label 1) actions by maximizing the classification interval, in processing nonlinear classification problems, the SVM maps the input data to a higher dimensional feature space through the RBF kernel function, the definition of the RBF kernel function is:

[0017]

[0018] wherein, control the width of the RBF kernel function, that is , the model training target is to minimize the following regularization objective function:

[0019]

[0020] wherein, is the hyperplane normal vector, is the bias, is the slack variable, is the regularization parameter, balancing the classification interval and the penalty for misclassification, input the standardized vector into the trained support vector machine (SVM) model, and finally output the classification label: 0 represents "forehand", and 1 represents "backhand".

[0021] Further, in the evaluation and analysis module, the posture dimension score is based on the action similarity score algorithm of dynamic time warping (DTW), which integrates body angle and swing acceleration sequence, and realizes objective scoring and visual feedback of tennis hitting action specification through time series similarity measurement, outputting a score of 0-100, and based on the technology anomaly recognition algorithm of one-class support vector machine (One-Class SVM), a posture specification detection model is constructed, the boundary distribution of the action feature space is learned using standard forehand / backhand data, and the new input action is judged for anomaly, and then the key angle deviation is located and personalized technical feedback is generated.

[0022] Further, in the evaluation and analysis module, the power dimension score introduces a dynamics evaluation algorithm, which automatically identifies the data segment most likely to be the moment of hitting the ball from multiple sliding time windows, based on the identification result, combines acceleration and speed data, and evaluates the athlete's power timing and power efficiency.

[0023] Further, in the evaluation analysis module, the score of the landing point dimension is based on the preset ideal landing point area of the ball hitting type, the Euclidean distance between the actual landing point coordinates and the center of the ideal landing area is calculated, and an exponential decay function is used to convert the Euclidean distance into a quantitative score of 0 to 100 points.

[0024] Further, the evaluation analysis module introduces an analytic hierarchy process (AHP) to construct a comprehensive quality score model for the ball hitting, by decomposing the ball hitting process into three evaluation dimensions of posture, force, and landing point, and combining expert experience and technical specifications to assign appropriate weights to each dimension, thereby realizing multi-dimensional comprehensive quantitative evaluation of the ball hitting effect. The weights of the posture dimension, force dimension, and landing point dimension in the analytic hierarchy process are 0.45, 0.30, and 0.25, respectively. The final comprehensive quality score is obtained by linear combination of the scores of the three dimensions according to the weights.

[0025] Further, the feedback interaction module uses the graffiti cloud platform to store the posture data, acceleration sensor data, and comprehensive quality score of the student, and uses the edge computing capability of the graffiti T5 AI_BOARD development board to deploy a lightweight AI agent model. The AI agent model is used for personalized action recognition, dynamic adjustment of multi-dimensional weights, and intelligent push of improvement suggestions. The AI agent supports users to retrieve and analyze personal historical data in the form of text or dialogue.

[0026] Further, the feedback interaction module includes visual icons, including posture structure diagrams, acceleration curve diagrams, ball hitting trajectory diagrams, heat maps, evaluation statistical pie charts, and bird's eye view landing point diagrams.

[0027] The above technical solutions have the following beneficial effects:

[0028] (1) The data acquisition module covers the full dimension of training through multi-source data collaboration, ensures data effectiveness through dynamic tracking and precise positioning, and reduces scene dependence through lightweight deployment.

[0029] (2) The data preprocessing module eliminates error interference through multi-source data synchronization calibration, improves analysis efficiency through precise extraction of key frames, and provides targeted input for the evaluation module by converting raw data into structured features that can be directly used for evaluation.

[0030] (3) The evaluation analysis module covers the core indicators of training through multi-dimensional evaluation, improves the professionalism of the evaluation results based on scientific and reliable algorithms, and improves the universality and practicality of the system by dynamically adapting to the needs of different people.

[0031] (4) The feedback interaction module solves the problem of traditional training feedback lag through real-time feedback, adapts to individualized improvement needs through the establishment of a personalized think tank, realizes long-term improvement needs, reduces the use threshold through multi-modal interaction, and expands the use scenarios through cloud collaboration. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A flowchart of a multi-dimensional tennis action analysis system of the present application;

[0033] Figure 2 A data acquisition and preprocessing flowchart of a multi-dimensional tennis action analysis system of the present application;

[0034] Figure 3 A forehand and backhand determination flowchart of a multi-dimensional tennis action analysis system of the present application;

[0035] Figure 4 A comprehensive scoring flowchart of a multi-dimensional tennis action analysis system of the present application

[0036] Figure 5 A feedback interaction module workflow diagram of a multi-dimensional tennis action analysis system of the present application

[0037] Figure 6 An acceleration curve diagram of a multi-dimensional tennis action analysis system of the present application;

[0038] Figure 7 A ball hitting trajectory diagram of a multi-dimensional tennis action analysis system of the present application;

[0039] Figure 8 A heat map of a multi-dimensional tennis action analysis system of the present application;

[0040] Figure 9 An evaluation statistical pie chart of a multi-dimensional tennis action analysis system of the present application;

[0041] Figure 10 A bird's eye view of a multi-dimensional tennis action analysis system of the present application. DETAILED DESCRIPTION

[0042] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are only used to explain the present application and cannot be understood as a limitation of the present application.

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] This invention proposes a multi-dimensional tennis motion analysis system, including a data acquisition module, a data preprocessing module, an evaluation and analysis module, and a feedback interaction module:

[0045] Example 1, see Figure 1 and Figure 2 The data acquisition module includes a posture data acquisition unit, a ball trajectory and movement tracking unit, and a motion parameter acquisition unit. The posture data acquisition unit uses a gimbal posture capture camera combined with a real-time tracking algorithm to lock the trainee's movement trajectory. The gimbal posture capture camera uses a Raspberry Pi as an edge computing node, equipped with a CSI HD camera and a dual servo mechanism to form the hardware foundation, and achieves horizontal or vertical movement through servo drive. The system rotates and uses the YOLO target detection model to identify and locate the student, outputting the target's coordinate information. It also uses a PID algorithm to ensure the accuracy of the student's trajectory locking (deviation ≤ 5% of the screen area). Furthermore, it uses the MediaPipe machine learning framework to extract the coordinates of key joints (left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles) during the hitting process.

[0046] The ball trajectory and movement tracking unit relies on the full-field bird's-eye view camera combined with the YOLO11 algorithm to complete the ball's movement trajectory and simultaneously record the student's movement trajectory in the field, realizing the synchronous collection of "student-ball" linkage information. The full-field bird's-eye view camera adopts a high-resolution wide-angle camera and generates a full-field bird's-eye view without blind spots through hardware distortion correction and image stitching technology.

[0047] The motion parameter acquisition unit uses the WT931 high-precision triaxial accelerometer to collect acceleration data of the tennis racket in three orthogonal axes in real time.

[0048] Illustratively, a pan-tilt pose capture camera is erected at the midpoint of the court sideline, fixed on a tripod (height 1.8 meters), with the lens facing the center line of the court. The device uses Raspberry Pi 4B as the edge computing node, equipped with CSI HD camera and dual-rudder pan-tilt, with hardware support for horizontal and vertical ±90° rotation. A DJI 1080P wide-angle camera is installed 8 meters above the court, with the lens vertically downward, and the edge distortion correction technology is used to eliminate the deformation of the picture edge, generating an overhead view that covers the four corners of the court. A WT931 three-axis acceleration sensor module is embedded in the grip of the tennis racket, and data wireless transmission is realized through the ESP8266 chip. When the student starts to swing the racket, the camera captures the motion picture in real time, and the YOLO target detection model quickly identifies the human body outline in the picture, outputting its center coordinates in the picture (such as (320, 240) pixel points). The system calculates the deviation of the target from the center of the picture based on the coordinates, and drives the dual-rudder to adjust the angle of the lens through the PID control algorithm, ensuring that the student is always in the core area of the picture. In the tracking process, the MediaPipe machine learning framework synchronously analyzes the video stream, accurately locates the student's left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles, generating real-time coordinate sequences. For example, during the forehand lead-in stage, the left shoulder coordinates are (280, 180), and the right elbow coordinates are (350, 220), which are dynamically updated with the swing action. At the moment of the student's hitting the ball, the YOLO11 algorithm locks the ball in real time and traces its motion trajectory, recording the three-dimensional coordinate changes of the ball (such as from (5.2m, 3.1m, 0.8m) to (18.5m, 4.2m, 0.2m)), while marking the landing point as the deep area of the opposite court (1.2 meters from the baseline), and recording the student's running trajectory simultaneously, showing the student's walking path from the near center line of the court (4.1m, 2.5m) to the forehand position (3.8m, 5.3m), generating a "student-ball" linkage timeline. During the swing, the sensor collects X, Y, and Z axis acceleration data at a frequency of 50Hz, for example, the X-axis acceleration peak value reaches 12.8m / s², the Y-axis is 8.5m / s², and the Z-axis is -5.2m / s² at the moment of hitting the ball. After integration operation, it is converted into swing speed (maximum 32km / h) and three-dimensional motion trajectory (characteristic of "arc acceleration-peak-deceleration"), and the data is uploaded to the system processing end in real time. Through the coordinated operation of the three units, the system has synchronously obtained the dynamic coordinate sequence of the 12 key joints, the motion trajectory of the ball, the landing point, the student's walking heat map, the racket acceleration, the speed, and the swing trajectory data after the student completes a set of forehand hitting.

[0049] Example two, refer to Figure 1 and Figure 2In the data preprocessing module, based on the multiple coordinates extracted by the MediaPipe machine learning framework, eight key angle features are generated, including left arm angle, right arm angle, left arm-body angle, right arm-body angle, left shoulder-hip-knee angle, right shoulder-hip-knee angle, left leg angle, and right leg angle, which are the core data for posture evaluation;

[0050] For the picture transmitted by the full-view bird's-eye camera, the YOLO11 model is used to track the player and the tennis data: the player is accurately identified and tracked, the motion coordinates of the player in the whole process are obtained, the Roboflow open source vision dataset tennis dataset and the self-owned tennis match pictures are selected to construct the dataset, the pre-trained model of YOLO11 is used for training, the training condition is RTX 4090 24GB GPU, the tracking of the tennis in the movement is realized, and then the motion trajectory and landing position of the tennis are obtained;

[0051] The acceleration data collected by the acceleration sensor is subjected to one-time integration to obtain the ball hitting speed, and two-time integration is performed to realize the reconstruction of the racquet hitting trajectory;

[0052] The image frame frequency of the pan-tilt posture capture camera and the sampling frequency of the WT931 three-axis acceleration sensor are consistent, both are set to 50HZ, and then the data is synchronously received and sent through MQTT, further ensuring the synchronization of the data, based on the system clock (accuracy 1ms) for marking the posture, force, and landing point three data, when the deviation is more than 5ms, it is calibrated through NTP, through data cleaning and standardization processing to form a fusion dataset, a fixed size sliding window (window length can be set according to the motion characteristics) is used, the sliding window continuously receives new data and eliminates old data, the length remains unchanged, the key features of the acceleration data in the window are calculated, and it is judged whether the middle position of the sliding window reaches the maximum value of acceleration (ball hitting threshold), if it reaches, the data of the sliding window at this time is the ball hitting frame.

[0053] For example, the gimbal posture capture camera outputs the coordinates of 12 key joints of the student during the backhand shot through the MediaPipe framework in real time. The coordinates are in units of screen pixels (e.g., left shoulder (290, 175), left elbow (310, 210), left wrist (330, 240), etc.). The left arm angle is calculated based on the coordinates of the left shoulder, left elbow, and left wrist using the vector angle formula. The angle during the lead-in stage is 145°, and it increases to 160° at the moment of hitting the ball. The angle between the left arm and the body is calculated based on the coordinates of the left shoulder, left hip, and left elbow, reflecting the opening and closing degree of the arm and the torso. The angle is 75° (standard range: 60°-80°) at the moment of hitting the ball. The right shoulder-hip-knee angle is calculated based on the coordinates of the right shoulder, right hip, and right knee, representing the coordinated rotation of the lower limbs and the torso. This student's angle is 150°, slightly lower than the standard value of 160°. The remaining five angles are calculated synchronously, forming an 8-dimensional angle feature sequence, updated every 0.02 seconds (50Hz frame rate). In the overhead view captured by the overhead camera, the YOLO11 model identifies the student's position in real time, generating motion coordinates in the court coordinate system (unit: meters). For example, the student moves from the initial position (4.2, 2.5) to the backhand position (3.8, 6.1) and then back to (4.0, 2.7). The trajectory data is recorded at a frequency of 50Hz, forming a continuous walking curve of the student. The system uses a YOLO11 model trained on a Roboflow open-source dataset and 1000 self-owned competition images (training environment: RTX4090 24GB GPU, 300 iterations), achieving a tennis recognition accuracy of 98%. After the student hits the ball, the model tracks the ball's trajectory from the hitting point (5.1, 3.2) to the landing point (19.3, 5.8), records key coordinate points (e.g., rising period (7.2, 3.5), highest point (12.0, 4.1)), calculates the distance from the landing point to the baseline as 0.8 meters (ideal range ≤1 meter), and obtains the trajectory of the ball. The WT931 sensor collects the racket's X / Y / Z axis acceleration data (unit: m / s²), such as the peak values of X: -11.2, Y: 9.5, and Z: -3.8 at the moment of hitting the ball. The acceleration data is subjected to a single integration to obtain the hitting speed, with a maximum speed of 28km / h. Based on the speed data, a three-dimensional swing trajectory is reconstructed through a double integration, showing an arc-shaped trajectory of "backswing-forward swing-following swing". The curvature radius of the trajectory at the hitting point is 0.8 meters; the system clock (with an accuracy of 1 ms) is marked with three types of data, if the timestamp of a frame of posture data deviates from the acceleration data by 8 ms, the NTP calibration is automatically triggered, the deviation after correction is less than or equal to 3 ms, the posture data, walking data and motion parameters are mapped to a unified three-dimensional coordinate system, the spatial deviation is eliminated, the abnormal values caused by sensor jitter (such as a frame of acceleration Z-axis data suddenly changes to-20 m / s², which exceeds the reasonable range (±15 m / s²), and is judged as abnormal and replaced with the mean value of the adjacent frame) are removed, the window length is 0.2 seconds (including 10 data points, 50 Hz sampling rate), the step length is 0.02 seconds, the peak value of the acceleration module length in the window is calculated, when the peak value of the middle frame reaches 12 m / s² and the window symmetry meets the characteristics of the hitting action, the window is determined as the hitting frame, and the posture, ball path and speed data in the corresponding time period are extracted as the key analysis objects.

[0054] Example three, refer to Figure 1 and Figure 3 , 8 key angle features of the MediaPipe machine learning framework are extracted, and the StandardScaler standardization is used to eliminate the dimension difference, and the standardization formula is:

[0055]

[0056] wherein, is the original feature value, is the feature mean value, is the feature standard deviation.

[0057] Take the hitting instant data, calculate the mean, maximum, minimum and standard deviation of each action sequence, construct a 32-dimensional feature vector, and then perform support vector machine model construction and training. The SVM model maximizes the classification interval to find the optimal hyperplane to distinguish the forehand (label 0) and backhand (label 1) actions. In dealing with nonlinear classification problems, SVM maps the input data to a higher dimensional feature space through the RBF kernel function. The definition of the RBF kernel function is:

[0058]

[0059] wherein, controls the width of the RBF kernel function, i.e. .

[0060] The model training target is to minimize the following regularization objective function:

[0061]

[0062] wherein, is a hyperplane normal vector, is a bias, is a slack variable, is a regularization parameter.

[0063] Balancing the classification margin and the misclassification penalty, the standardized vector is input into the trained Support Vector Machine (SVM) model, and finally the classification labels "0" and "1" are output: "0" represents "forehand", and "1" represents "backhand".

[0064] For example, 10 students are selected, each of whom completes 20 groups of forehand shots and 20 groups of backhand shots, a total of 400 groups of valid action data. The system captures the XX student's shot picture in real time through the pan-tilt gesture capture camera, locates the coordinates of 12 key joints such as left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles relying on the MediaPipe machine learning framework, calculates 8 core angle features, and the specific data is as follows:

[0065] Angle characteristics Forehand shot instant value (°) Forehand shot instant value (°) Left arm angle 155 140 Right arm angle 160 135 Left arm to body angle 70 85 Right arm to body angle 65 90 Left shoulder-hip-knee angle 150 145 Right shoulder-hip-knee angle 160 155 Left leg angle 170 165 Right leg angle 165 160

[0066] The system extracts key angle features through MediaPipe and uses StandardScaler standardization to eliminate dimension differences. Taking "right arm angle" as an example:

[0067] The mean of "right arm angle" in all actions , and the standard deviation ;

[0068] The original value of "right arm angle" of a certain forehand shot , and after standardization ;

[0069] The original value of "right arm angle" of a certain backhand shot , and after standardization .

[0070] The remaining 7 angle features are standardized according to the same logic to obtain an 8-dimensional standardized feature vector for each group of actions. For each group of action sequences (including 0.5 seconds before hitting the ball, the moment of hitting the ball, and 0.5 seconds after hitting the ball, a total of 50 frames of data), the 8 standardized angle features at the moment of hitting the ball are extracted, and the mean, maximum, minimum, and standard deviation of each feature in the sequence are calculated to form a 32-dimensional feature vector. Taking a group of forehand hits as an example: the left arm angle: mean 152°, maximum 158°, minimum 148°, and standard deviation 3.2°, and the remaining 7 angle features generate 4-dimensional data, and finally combine into a 32-dimensional feature vector, which is used for SVM model training and reasoning. Input a test sample, the model maps the data to a high-dimensional feature space through the RBF kernel function, calculates the distance between the sample and the hyperplane, and if the distance is greater than or equal to 0, outputs the classification label "1" (determined as backhand); if the distance is less than 0, outputs the label "0" (determined as forehand).

[0071] Example Four, see Figure 1 and Figure 4 The posture dimension score is based on the action similarity score algorithm of Dynamic Time Warping (DTW), which combines body angle and swing acceleration sequence, and realizes objective scoring and visual feedback of tennis hitting action standardization through time series similarity measurement, outputting a score of 0-100. First, input the 8 key angle feature sequences and 3 acceleration sequences. To reduce the influence of periodic angle jumps, all angle data is normalized to [0, 180) by taking the modulus, and DTW is used for distance calculation and multi-dimensional aggregation. The DTW algorithm calculates the minimum distance between two sequences through dynamic programming, considering the nonlinear alignment on the time axis, and the distance is defined as:

[0072]

[0073] wherein, is the key angle feature sequence, is the acceleration sequence, is the Euclidean distance, for the 8 key angle features and 3 acceleration features, the DTW distance between the student sequence and the standard sequence is calculated respectively, and the total distance is obtained by accumulation, which is mapped to the standardization score The exponential decay function is used to control the sensitivity of the score:

[0074]

[0075] wherein, is the attenuation coefficient, which adjusts the sensitivity of the score, and the score is limited to 0-100 with two decimal places.

[0076] A technical abnormality recognition algorithm based on One-Class SVM is used to construct a posture normality detection model. The boundary distribution of the action feature space is learned using standard forehand / backhand data, and the new input action is judged for abnormality. Then, the key angle deviation is located and individualized technical feedback is generated. The input data comes from multiple frames of hitting sequences, containing time series features of 8 key angles. Each hitting action is composed of multiple frames. To improve computational efficiency and robustness, statistical quantities of sequences in the time dimension are extracted as feature vectors :

[0077]

[0078] This way compresses dynamic sequences into static features, which is suitable for boundary learning and fast discrimination. The following transformation is used for standardization to eliminate the scale effect of different angle dimensions:

[0079]

[0080] Two independent One-Class SVM models are constructed for forehand and backhand actions, respectively, and only standard action data is used for training. The core goal is to learn the closed boundary of the action space. To improve the model's ability to effectively deal with complex human action patterns in high-dimensional feature space, the optimization form is as follows:

[0081]

[0082] Constraints:

[0083]

[0084] where, is the hyperplane normal vector, is the boundary offset, is the slack variable, controls the proportion of support vectors, is the sample size. If the prediction result is abnormal, the algorithm further compares the action's urgent time point features with the standard mean value to determine the deviation of each angle through a threshold. The angle whose deviation exceeds the threshold will trigger professional evaluation and generate specific suggestions, such as "left shoulder not opened" or "hitting point biased forward". If the action is judged to be normal, it returns "action normal, technique in place".

[0085] For example, 8 key angle feature sequences (such as left arm angle sequence: 140°-155°-148°…) and 3 acceleration sequences (X-axis: 8.2-12.5-9.8…m / s²) are input, the angle data is normalized to [0, 180) by modulo specification, and the DTW distance between the 8 angle sequences and 3 acceleration sequences of the student and the standard sequence of the forehand is calculated by using the DTW algorithm, such as the left arm angle DTW distance 12.3 and the X-axis acceleration DTW distance 8.5, and the total distance is accumulated , assuming the attenuation coefficient , substituting the exponential decay formula, the score is calculated points (0-100 points), indicating that the action deviates greatly from the standard; the time statistics of the 8 angle sequences of the student are extracted to construct a feature vector, which is processed by a standardization formula to eliminate the scale effect, and the standardized vector is input into the pre-trained forehand One-Class SVM model. Compared with the angle at the moment of hitting the ball and the standard mean value, it is found that the angle of the left arm away from the body (95°) exceeds the threshold value (60°-80°), and the suggestion is generated: “the angle of the left arm away from the body is too large, the arm extension amplitude needs to be reduced, and the power chain needs to be optimized”.

[0086] In example five, refer to Figure 1 and Figure 4 , the score of the power dimension is processed in real time by using a sliding window mechanism on the original acceleration sequence, three physical index systems of acceleration module length peak value, jerk, and window symmetry and offset analysis are constructed for each window, a normalized weighted scoring function is used to sort each window, and the highest scorer is selected as the “hitting window”. The form of the scoring function is as follows:

[0087]

[0088] Among them, is the total score of the i-th window, is the acceleration module length peak value, is the jerk peak value, is the offset degree of the peak value relative to the center of the window, , , is an empirical weighting parameter, after identifying the hitting window, the system further extracts the acceleration and speed information within about 0.3 seconds before and after the moment of hitting the ball, evaluates the hitting force, the rationality of the power time, and the power efficiency, and provides technical feedback according to the analysis results.

[0089] For example, the sliding window length is 0.2 seconds, the step length is 0.02 seconds, and a total of 20 candidate windows are generated. For each window, the acceleration module length peak value (such as the acceleration module length peak value of window 5 ), the jerk peak value (peak of window 5 ), peak offset degree (peak of window 5 , near center), set , , , substitute into scoring function, window 5 score is 0.82 (highest), determined as "hitting window", extract hitting window and 0.3 seconds of data before and after, get hitting moment acceleration peak 13.2 m / s2, speed 29 km / h, hitting strength meets standard (standard > 10 m / s2), force timing deviation (jerk peak lags behind acceleration peak 0.05 seconds), force efficiency is medium (speed and acceleration correlation coefficient 0.65, standard > 0.7), force dimension score is 72 points, output feedback suggestion: "force timing is late, need to concentrate force 0.05 seconds earlier, improve speed and acceleration synergy".

[0090] Example six, refer to Figure 1 and Figure 4 , scoring of landing point dimension is according to hitting type (forehand or backhand) to preset ideal landing point area, calculate Euclidean distance between actual landing point coordinates and ideal area center, use exponential decay function to convert distance into quantitative score of 0-100 points, the closer the distance, the higher the score.

[0091] For example, ideal landing point for forehand attack is opponent's deep area (1.5 m from baseline, 2 m right of center line), area center coordinates are set as (18.2 m, 5.1 m), camera captures actual landing point coordinates of tennis ball as (17.8 m, 5.5 m), calculate Euclidean distance from ideal center: , set decay coefficient λ = 50, substitute into exponential decay formula to get score points, system feedback: "landing point is shallow and right, need to adjust racquet face angle, improve landing point accuracy to opponent's deep area".

[0092] Example seven, refer to Figure 1 and Figure 4 , comprehensive quality score introduces Analytic Hierarchy Process (AHP) to build comprehensive hitting quality scoring model, decompose hitting process into three levels of target layer (comprehensive hitting quality score), criterion layer (action specification, hitting speed, landing point accuracy), and index layer (action deviation angle and other specific indexes), based on professional experience, use pairwise comparison method to score importance of criterion layer indexes and build judgment matrix, determine weight proportion (posture dimension 0.45, force dimension 0.30, landing point dimension 0.25), original data under each dimension is converted into standard score of interval 0-100 through linear normalization or segmented scoring function, to eliminate dimension influence, finally total score is obtained by linear combination of dimension scores according to weight, formula is:

[0093]

[0094] wherein, is the weight of the ith dimension, is its normalized score.

[0095] For example, after evaluation by the DTW algorithm and One-Class SVM, the posture dimension normalized score is 78 (weight 0.45), the force dimension normalized score is 82 (weight 0.30) through the sliding window mechanism evaluation, and the landing dimension normalized score is 65 (weight 0.25) through the calculation of the Euclidean distance mapping score. Substituting into the comprehensive score formula , the system feedback is: “overall performance is good, posture standardization and force efficiency meet the standards, but landing accuracy needs to be improved, it is suggested to strengthen the training of the racket surface control and optimize the hitting point to the ideal area”.

[0096] Embodiment eight, refer to Figure 1 and Figure 5 The feedback interaction module stores the posture data, acceleration sensor data and comprehensive quality score of the student by using the graffiti cloud platform, deploys a lightweight AI agent model by using the edge computing capability of the graffiti T5 AI_BOARD development board, the AI agent model is used for personalized action recognition, dynamically adjusts the weight of multiple dimensions, intelligently pushes improvement suggestions, and the AI agent supports users to retrieve and analyze personal historical data in the form of text or dialogue.

[0097] For example, XX student uses the system for a period of 2 weeks of forehand hitting training (1 group per day, a total of 14 groups of data), after each training, the 12 joint coordinate sequences, racket acceleration curve and comprehensive score of Li are uploaded to the graffiti cloud platform in real time, are classified and archived according to “date-action type”, personalized interactive services are provided by relying on the AI agent deployed by T5 AI_BOARD, based on the first 5 groups of data, the action characteristics of “small forehand rotation amplitude” are accurately identified, because the student has poor force stability, the model temporarily increases the force dimension weight from 0.30 to 0.35, and preferentially pushes the improvement scheme related to “improve the rotation amplitude” and “optimize the force timing”; Li asks through the front-end text question: “how is the comprehensive score of the forehand in the past 3 days?”, the AI agent searches the cloud data and feedbacks: “the score in the past 3 days is 78→82→80, the posture score has increased significantly, but the landing score fluctuates greatly, it is suggested to strengthen the training of the racket surface angle control”; also through voice dialogue: “how to improve the force?”, the agent outputs: “the jerk peak of hitting the ball lags behind 0.04 seconds, the force needs to be started 0.03 seconds earlier, which can be adjusted through slow motion swing practice”.

[0098] Embodiment nine, refer to Figure 1 andFigure 5 The feedback interaction module includes a visual icon, and the visual icon includes a posture structure diagram, an acceleration curve diagram, a hitting trajectory diagram, a heat map, an evaluation statistical pie chart, and a bird's eye view landing point diagram:

[0099] (1) The posture structure diagram obtains joint data through a Vue component integrated with a MediaPipe Pose API, and uses a Canvas 2D technology to draw a skeleton line on a camera screen in real time, to realize intuitive display of a posture structure;

[0100] (2) Referring to Figure 6 The acceleration curve diagram is based on a Vue responsive data and a Chart.js, receives real-time data through a WebSocket, dynamically draws a three-axis curve and automatically labels a hitting peak value and a power interval, displays a three-axis acceleration change trend (x / y / z) of each hitting, is used for assisting in analyzing a power rhythm and a timing of a user, and verifying whether a power output is completed at an optimal hitting point, is matched with a DTW algorithm to output a score, and the diagram provides intuitive evidence for a coach or a trainer to identify whether the power is effective;

[0101] (3) Referring to Figure 7 The hitting trajectory diagram adopts a Three.js to build a 3D scene, uses racket sensor and ball positioning data to build a three-dimensional trajectory diagram of a hitting path, intuitively presents a spatial trajectory in each swing process, and analyzes completeness and path rationality of a technical action;

[0102] (4) Referring to Figure 8 The heat map is a cumulative stay heat distribution diagram of a training site based on historical position data of a user's body, reflects a moving range and a training active area of an athlete, uses a Heatmap.js to calculate a density distribution through a Gaussian kernel function and visually display a color gradient, and is used for analyzing a site coverage capability and a tactical execution situation;

[0103] (5) Referring to Figure 9 The evaluation statistical pie chart combines a machine learning action phase identification, collects a statistical result of a score of all hits in this training, is divided into categories such as “excellent”, “good”, and “to be improved”, calculates a proportion of each category, uses Chart.js to realize real-time statistics of a time proportion of a paddle, acceleration and the like, and displays the time proportion in a sector partition, to help a user grasp an overall performance level of this training from a global perspective;

[0104] (6) Referring to Figure 10 The bird's eye view landing point diagram maps a court and a screen coordinate through a Canvas, simulates a top-down perspective, draws a paddle head horizontal projection trajectory and key node labels in real time, visually labels and clusters landing points of the ball on the court, analyzes a distribution rule and a deviation trend of the hitting landing points, and is convenient for finding problems of control force and precision.

[0105] (7) The text prompt function renders the AI-generated personalized improvement suggestions to the specified area of the interface in real time through Vue data binding, realizing the instant presentation of training feedback.

[0106] The above describes the basic principles and main features of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments. The above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application. These changes and improvements are all within the scope of the claimed application. The scope of the application is defined by the appended claims and their equivalents.

[0107] In the description of the present application, the description of the terms "some embodiments", "optional embodiments" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0108] Although embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and purposes of the present application. The scope of the present application is defined by the claims and their equivalents.

Claims

1. A multi-dimensional tennis motion analysis system, characterized by, It comprises a data acquisition module, a data preprocessing module, an evaluation analysis module and a feedback interaction module. The data acquisition module comprises a posture data acquisition unit, a ball path and walking position tracking unit and a motion parameter acquisition unit, the posture data acquisition unit locks the player's motion trajectory by using a gimbal posture capture camera combined with a real-time tracking algorithm, extracts the dynamic characteristics of the body joints and generates a standardized posture vector by using a MediaPipe machine learning framework, the ball path and walking position tracking unit completes the motion trajectory of the ball by relying on a full-field bird's eye camera combined with a YOLO11 algorithm, and the motion parameter acquisition unit acquires the racket motion signal by using an acceleration sensor and converts the acceleration data into speed parameters and three-dimensional motion trajectories by using integral operation; The data preprocessing module performs space-time calibration on the posture, trajectory and motion parameters based on a timestamp synchronization mechanism, forms a fusion data set through data cleaning and standardization processing; The evaluation analysis module is used for distinguishing the action types, integrating the scores of the posture dimension, the force dimension and the landing point dimension and generating a comprehensive evaluation report; The feedback interaction module displays the evaluation results in real time on the interface based on text and visual icons and outputs historical analysis.

2. The multi-dimensional tennis motion analysis system according to claim 1, wherein, The gimbal posture capture camera takes Raspberry Pi as an edge computing node, carries a CSI HD camera and a double rudder mechanism to form a hardware foundation, realizes ±90° rotation in the horizontal or vertical direction through rudder driving, relies on a YOLO target detection model to complete the identification and positioning of the player and outputs the coordinate information of the target.

3. The multi-dimensional tennis motion analysis system according to claim 1, wherein, The full-field bird's eye camera adopts a wide-angle camera, generates a full-field bird's eye view without dead angle through hardware distortion correction and image stitching technology.

4. The multi-dimensional tennis motion analysis system according to claim 1, wherein, The acceleration sensor is a WT931 three-axis acceleration sensor, the image frame frequency of the gimbal posture capture camera is consistent with the sampling frequency of the WT931 three-axis acceleration sensor.

5. The multi-dimensional tennis motion analysis system according to claim 1, wherein, The posture vector extracted by the MediaPipe machine learning framework comprises the coordinates of the left and right shoulders, the left and right elbows, the left and right wrists, the left and right hips, the left and right knees and the left and right ankles, and eight key angle features, i.e., the left arm angle, the right arm angle, the left arm angle from the body, the right arm angle from the body, the left shoulder-hip-knee angle, the right shoulder-hip-knee angle, the left leg angle and the right leg angle, are generated based on a plurality of the coordinates.

6. A multi-dimensional tennis motion analysis system according to claim 5, wherein, The eight key angle features of the MediaPipe machine learning framework are extracted, the data at the moment of hitting the ball is taken, the mean value, the maximum value, the minimum value and the standard deviation of each action sequence are calculated, a 32-dimensional feature vector is constructed, the dimension difference is eliminated by using StandardScaler standardization, the standardized vector is input into the trained SVM model, and the classification label is output: 0 represents forehand, and 1 represents backhand.

7. A multi-dimensional tennis motion analysis system according to claim 6, wherein, In the evaluation analysis module, the score of the posture dimension is based on the action similarity score algorithm of DTW, and the non-standard action is evaluated in words based on the technical anomaly identification algorithm of One-Class SVM; the score of the force dimension introduces a kinetics evaluation algorithm to identify the hitting moment and evaluate the force quality; the score of the landing point dimension is based on the ideal landing point area preset according to the type of the shot, calculates the Euclidean distance between the actual landing point coordinates and the center of the ideal landing area, and converts the Euclidean distance into a quantitative score of 0 to 100 points by using an exponential decay function.

8. A multi-dimensional tennis motion analysis system according to claim 7, wherein, The evaluation analysis module introduces the analytic hierarchy process to construct a comprehensive score model of the quality of the shot, and the weights of the action standardization, the hitting speed and the landing point accuracy in the analytic hierarchy process are 0.45, 0.30 and 0.25 respectively, and the comprehensive quality score is generated by linear combination of the scores of each dimension according to the weights.

9. The multi-dimensional tennis motion analysis system according to claim 1, wherein, The feedback interaction module stores the posture data, acceleration sensor data and comprehensive quality score of the player by using the graffiti cloud platform, deploys a lightweight AI agent model by using the edge computing capability of the graffiti T5AI_BOARD development board, and the AI agent model is used for personalized action recognition, dynamic adjustment of the weights of multiple dimensions, intelligent push of improvement suggestions, and the AI agent supports the user to retrieve and analyze personal historical data in the form of text or dialogue.

10. The multi-dimensional tennis motion analysis system according to claim 1, wherein, The feedback interaction module includes visual icons, and the visual icons include a posture structure diagram, an acceleration curve diagram, a shot trajectory diagram, a heat map, an evaluation statistical pie chart and a bird's eye view landing point diagram.

Images