An intelligent evaluation method and system for boxing technical and tactical styles

Abstract

The present invention belongs to the field of intelligent evaluation technology, and discloses an intelligent evaluation method and system for technical and tactical styles of boxing. The method collects action video data of red and blue boxers, and obtains time series data of the red and blue boxers through the OpenPose deep learning algorithm and the pinhole imaging model principle in combination with the OpenPose confidence. The time series data is anomaly detected, eliminated, filled, and smoothed. Then, an action label is assigned to each timestamp of the pre-processed time series data. The boxer's technical and tactical style and characteristic analysis and evaluation module is used to implement technical and tactical analysis and evaluation. This solves the problem that the existing technology cannot provide real-time feedback and quantify technical and tactical means, does not fully consider defensive, counterattack and other technical and tactical techniques, and is not conducive to comprehensive technical and tactical evaluation.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent evaluation technology, and in particular relates to a method and system for intelligent evaluation of technical and tactical styles of boxing. Background Art

[0002] Boxers need to develop technical and tactical abilities through intensive training and extensive practical experience. Traditional evaluation methods for boxers rely on manual guidance and on-site sparring, which is inefficient and subjective, and cannot provide real-time feedback and personalized training plans.

[0003] Existing technologies rely on video action recognition but are unable to provide real-time feedback or quantify technical and tactical skills. While they improve a boxer's offensive skills and tactics by calculating metrics like punch angle and speed, they fail to fully consider defensive and counterattack techniques, hindering comprehensive evaluation. Existing technology is unable to provide real-time feedback to boxers, making it difficult for them to adjust their skills and tactics in a timely manner. Furthermore, traditional evaluation methods rely on human guidance and on-site sparring, which are susceptible to personal bias and lead to subjective evaluation results. Summary of the Invention

[0004] The present invention aims to solve the problems existing in the prior art and provides a method and system for intelligent evaluation of boxing technique and tactics.

[0005] The object of the present invention is achieved through the following technical solutions:

[0006] In a first aspect, the present invention provides a method for intelligently evaluating boxing skills and tactics, the method comprising the following steps:

[0007] Step S1: Deploy four cameras around the boxing ring to collect action video data of the red and blue boxers; calibrate the cameras to obtain the intrinsic and extrinsic parameter matrices of each camera;

[0008] Step S2: First, the OpenPose deep learning algorithm is used to detect the human joints in the acquired action video data, that is, to obtain the two-dimensional plane coordinates of the joints. Then, based on the principle of the pinhole imaging model and combined with the OpenPose confidence, the three-dimensional coordinates of the joints are obtained, that is, the time series data of the red and blue sides are obtained respectively;

[0009] Step S3, preprocessing the time series data: analyzing the time series data, identifying and marking the timestamps of the outliers in the time series data, then setting the data values ​​of the timestamps where the outliers are located to blank and filling the blank values, and then performing smoothing filtering on the filled time series data to obtain the preprocessed time series data;

[0010] Step S4, using the time series classification task to realize action recognition, and then assigning an action label to each timestamp of the preprocessed time series data;

[0011] Step S5: annotating the time series data with action labels, including sprint steps, circle steps, slide steps, and passive steps, punching hands and punching locations, and effective and ineffective punches in offensive and counterattack tactics;

[0012] Step S6, counting the boxer's skills and tactics: Based on the data obtained in step S5, the number and proportion of offensive, defensive, counterattack, and feint skills and tactics are calculated, as well as the statistics of the dominant hand used in the skills and tactics, the statistics of the hitting parts, the hit rate and the effective punch ratio, and the important parameters for evaluating the skills and tactics;

[0013] In step S7, the time series pre-processed in step S3 is labeled with styles and characteristics, and the labels are integrated with the important technical and tactical parameters obtained in step S6 to construct a classification model based on machine learning. The important technical and tactical parameters are input into the trained classification model based on machine learning to evaluate the boxer's technical and tactical style and characteristics in a round.

[0014] Preferably, the specific steps of step S2 are as follows:

[0015] Step S2-1: Use the OpenPose deep learning algorithm to detect the human joints in the acquired action video data and obtain the joint points P i The two-dimensional plane coordinates (u i , v i ), i=1, 2, 3 or 4;

[0016] Step S2-2, using the pinhole imaging model principle to calculate the two-dimensional coordinates of the joint point P in the ideal state, the three-dimensional coordinates P (X w ,Y w ,Z w ):

[0017]

[0018] Among them, f x and f y are the horizontal focal length and vertical focal length of the camera respectively; u0 and v0 are the pixel coordinates of the intersection of the camera optical axis and the imaging plane, M i is the external parameter matrix from the world coordinate system to the camera i coordinate system, x w 、y w 、z w is the three-dimensional coordinate of the joint point P in the ideal state, z c is the proportionality coefficient;

[0019] In step S2-3, to avoid noise, it is necessary to estimate the spatial position of the real joint point R. Considering the distance between the rays emitted by the camera, a spherical area is set. The radius of the sphere is 10 times the shortest distance between the rays. The three-dimensional coordinate formula of the real joint point R is as follows:

[0020]

[0021] Among them, l i is the ray emitted by the i-th camera, |P,l i | represents the ideal state of the joint point P to the ray l i The shortest distance, ρ i is the Openpose confidence parameter, n=4;

[0022] Step S2-4: The action video data covers multiple joints of the human body, and time series data is obtained according to the real three-dimensional coordinates of the multiple joints.

[0023] Preferably, the specific steps of step S3 are as follows:

[0024] Step S3-1, obtaining the boxer's motion data in an unobstructed environment, constructing a training data set that can reflect the real human motion patterns, and using the athlete training data collected in the boxing ring environment as the test data set;

[0025] Step S3-2: using a normalization method, the value range of the time series data in the training dataset and the test dataset is mapped to a real number space from 0 to 1;

[0026] Step S3-3, inputting the normalized training data set into the Transformer model, so that the Transformer model learns the normal pattern of boxing movement;

[0027] In step S3-4, the normalized test data set is input into the Transformer model obtained in step S3-3, and the anomaly score is constructed using the reconstruction error MSE. A threshold is set, and the top 5% of data with anomaly scores greater than the threshold are selected and marked as abnormal data segments, with timestamps as the unit. Otherwise, they are normal data segments.

[0028] Step S3-5: Empty the abnormal data segments and fill them with normal values ​​through interpolation or Transformer prediction;

[0029] Step S3-6, using adaptive filtering technology to perform smoothing filtering on the filled time series data.

[0030] Preferably, the abnormal data segments in step S3-5 are left blank and filled with normal values, specifically: when there are few abnormal data segments, an interpolation method is adopted, and adjacent normal data points are used to estimate the values ​​of the abnormal data segments through linear interpolation and spline interpolation; when there are many abnormal data segments, a trained Transformer model or regression model is used to predict the abnormal data segments and fill them with predicted values.

[0031] Preferably, the specific steps of step S4 are as follows:

[0032] Step S4-1, using the time series classification task to realize action recognition, classifying actions into offensive actions, defensive actions, counter-attack actions, and feints, annotating the pre-processed time series data, and dividing it into a training set and a test set;

[0033] Step S4-2, using the training set to train a multivariate time series classification model based on the LSTM architecture;

[0034] In step S4-3, the preprocessed time series data is input into the trained multivariate time series classification model based on the LSTM architecture, and a corresponding action label is assigned to each timestamp of the preprocessed time series data.

[0035] Preferably, the specific steps of step S5 are as follows:

[0036] Step S5-1: Filter out the timestamps corresponding to the movement categories in the action tags, classify the steps into sprint, circle, slide, and passive based on the displacement, speed, rotation angle, and acceleration of the joint movement, and perform fine-grained footwork annotation on the time series data of the movement categories in the action tags;

[0037] Step S5-2, filtering out the timestamps corresponding to the attack and counterattack in the action tags, labeling the puncher with the attack and counterattack time series data in the action tags according to the coordinate values, movement speed, and acceleration of the hands, and labeling the hitting parts with the attack and counterattack time series data according to the distance from the opponent's joint points;

[0038] Step S5-3, taking each interaction between the two parties as a unit, combined with the puncher and the hitting part in each interaction, identifies whether there is a hit or not. At the same time, the hitting effect of the time series with action labels of attack and defense is marked, that is, the effective punch and invalid punch are marked.

[0039] Preferably, the important parameters for evaluating techniques and tactics in step S6 include: pressure coefficient, distance coefficient, balance coefficient and movement coefficient.

[0040] Preferably, the specific steps of step S7 are as follows:

[0041] Step S7-1: Label the style and characteristics of the time series data obtained after preprocessing in step S3, and divide it into training and test sets. The style labels include active attack, defensive counterattack, offensive back-and-forth, and roaming. The characteristic labels include feint, direct attack, pressing, and attack first and then defense.

[0042] Step S7-2, integrating the statistics and percentage calculation of offensive, defensive, and counterattack techniques and tactics described in step S6, statistics of dominant hand used in techniques and tactics, statistics of striking locations, calculation of hit rate and effective punch ratio, and evaluation of important parameters of techniques and tactics, as well as style and characteristic labels, to construct and train a classification model based on machine learning;

[0043] Step S7-3: Input the important technical and tactical parameters into the trained machine learning-based classification model to evaluate the boxer's technical and tactical style and characteristics in a round.

[0044] In a second aspect, the present invention provides an intelligent evaluation system for boxing techniques and tactics, which is used to implement the above method and includes: an acquisition module, a motion capture module, a time series preprocessing module, a motion recognition module, a motion analysis module, a technique and tactics statistics module, and a technique and tactics style and characteristics analysis and evaluation module;

[0045] The acquisition module is connected to the motion capture model and obtains the action video data of the red and blue players around the boxing ring;

[0046] The motion capture module uses the OpenPose deep learning algorithm and the pinhole imaging model principle to obtain the red and blue action video data obtained through the input channel, combined with the OpenPose confidence parameter, to obtain the red and blue time series data.

[0047] The time series preprocessing module includes a time series anomaly detection submodule, an anomaly removal and filling submodule, and a smoothing filter submodule. The time series anomaly detection submodule analyzes the time series data, identifies and marks the timestamps where the outliers are located in the time series data, and the anomaly removal and filling submodule sets the data values ​​of the timestamps where the outliers are located to zero, and fills the empty values ​​by interpolation or Transformer prediction. The smoothing filter submodule performs smoothing filtering on the filled time series data to obtain the smoothed filtered time series data as the output of the time series preprocessing module;

[0048] The action recognition module uses the time series classification task to realize action recognition, and then assigns an action label to each timestamp of the preprocessed time series data;

[0049] The motion analysis labels the time series data with motion tags, including sprint steps, circle steps, slide steps and passive steps, punching hands and punching parts, and effective and ineffective punches in offensive and counterattack tactics.

[0050] The technical and tactical statistics module counts the boxer's technical and tactical usage, including the number and percentage calculation of offensive, defensive, and counterattack technical and tactical tactics; statistics on the boxer's dominant hand usage pattern; statistics on the hitting parts; calculation of hit rate and effective punch ratio; and important parameters for evaluating technical and tactical tactics.

[0051] The technical and tactical style and characteristic analysis and evaluation module constructs a classification model based on machine learning according to the output features of the technical and tactical statistics module combined with style and characteristic labels, and uses the classification model based on machine learning to realize the technical and tactical style evaluation of boxing.

[0052] The present invention designs an intelligent evaluation method and system for boxing techniques and tactics. In the present invention, four cameras are used to collect video data of training or competition of boxers from both teams. In the motion capture module, the time series data of both teams are obtained by combining the principles of OpenPose and pinhole imaging models based on four perspective video streams and OpenPose confidence. The time series preprocessing module detects, removes, and fills the anomalies of the time series, and smoothes the time series. The LSTM-based action recognition module uses the time series classification task to obtain coarse classification labels such as actions. In addition, the present invention designs an action analysis module to fine-grainedly identify the footwork, effective punches, and single punch classification and striking locations of both teams, and constructs a standard action knowledge base. The technique and tactics statistics module collects statistics on the offensive, defensive, and counterattack techniques and tactics within a round of the game, and calculates the hit rate, effective punch ratio, pressure system, balance coefficient, distance coefficient, and movement coefficient. The boxing technique and tactics style and characteristics analysis and evaluation module integrates the features to achieve technical and tactical analysis and evaluation of the boxers.

[0053] The present invention has the following beneficial effects: (1) The present invention provides a method and system for intelligent evaluation of technical and tactical styles for boxing, which solves the problem that the existing technology cannot provide real-time feedback and means of quantifying technical and tactical styles, does not fully consider defensive, counterattack and other technical and tactical styles, and is not conducive to comprehensive evaluation of technical and tactical styles.

[0054] (2) The present invention collects video data through four cameras and uses OpenPose and the pinhole imaging model principle to capture motion, detecting the boxer's movements in all directions, which facilitates the subsequent comprehensive analysis of technical and tactical styles and characteristics.

[0055] (3) Based on the time series data output by the motion capture module, the present invention uses the LSTM-based time series classification task to complete the motion recognition task, avoiding the use of complex computer vision models. Through the inherent memory mechanism of LSTM, key motion features are captured, greatly improving the efficiency of motion recognition.

[0056] (4) The present invention adopts fine-grained recognition of the footwork of both red and blue teams, effective punches, and the sub-classification of single punches in combination punches, and the hitting parts, and constructs a standard action knowledge base. It counts the offensive, defensive, and counterattack tactics in a round of the game, calculates the hit rate, effective punch ratio, pressure system, balance coefficient, distance coefficient, and movement coefficient, and integrates the features through the boxing technique and tactical style and characteristic analysis and evaluation module, which can comprehensively analyze the boxer's techniques and tactics. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of an intelligent evaluation method for technical and tactical styles in boxing;

[0058] Figure 2 This is a framework diagram of an intelligent evaluation system for boxing techniques and tactics. DETAILED DESCRIPTION

[0059] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0060] A method for intelligently evaluating boxing technique and tactics, comprising the following steps:

[0061] Step S1, obtain the boxer's action video data; set up four cameras around the boxing ring to achieve multi-angle coverage. Use Zhang Zhengyou's binocular camera calibration method to determine the relative position relationship between two adjacent cameras. By calibrating the cameras, the intrinsic and extrinsic parameter matrices of each camera are obtained. The intrinsic parameter matrix contains information such as the focal length of the camera, and the extrinsic parameter matrix contains the conversion information of two different coordinate systems, that is, the transformation matrix. The four cameras are connected to the input channel of the motion capture module through a synchronization interface. The motion capture module supports one-click synchronous shooting. To ensure that the generated motion capture data is stable and the data obtained by the motion capture module does not have timestamp confusion, a master camera is selected from the four cameras and connected to the motion capture module. Multiple slaves and the host are synchronized through a dedicated synchronization line. The trigger signal is sent to the master camera by pressing a button, and the signal is transmitted to the slave through the synchronization line. The start time and end time of the shooting of the four cameras are completely synchronized, and the synchronization time error must be less than one frame.

[0062] Furthermore, by calibrating the camera, we can obtain the transformation matrix from the world coordinate system to each camera and the transformation matrix between any two cameras. The transformation matrix can reflect the conversion method from the coordinate system of one perspective to the coordinate system of another perspective. Through the transitivity of the transformation matrix, the transformation matrix between any camera and the other three cameras can be obtained, and the transformation matrix between multiple cameras can be obtained.

[0063] Step S2: The motion capture module captures four real-time video data streams through the input channel. Using the OpenPose deep learning algorithm and the pinhole imaging model principle, combined with the OpenPose confidence level, the time series of the red and blue sides are obtained. The specific implementation steps include:

[0064] By using the OpenPose deep learning algorithm, the key joints of the human body in four real-time video data streams are detected. The algorithm identifies and locates the key points in the video that reflect the athlete's body posture and movement trajectory, such as the head, shoulders, elbows, wrists, hips, knees and ankles, and obtains the joint point P i The two-dimensional plane coordinates (u i , v i ). The algorithm supports real-time video stream input and accurately locates the key joints of each frame in the video.

[0065] The imaging of a three-dimensional space point on a camera can be represented by a pinhole imaging model, which is expressed as follows:

[0066]

[0067] where f x and f yare the horizontal and vertical focal lengths of the camera respectively; u0, v0 are the pixel coordinates of the intersection of the camera optical axis and the imaging plane, f x , f y , u0, v0 are the internal parameters of the camera, M i is the external parameter matrix from the world coordinate system to the camera i coordinate system, x w ,y w , z w is the three-dimensional coordinate of the joint point P in the ideal state in space, z c is the proportional coefficient, which can be 1 by default.

[0068] For the i-th camera, the three-dimensional coordinate point P(X w ,Y w ,Z w ) and the two-dimensional plane coordinates of the point imaged by a certain camera.

[0069]

[0070] in

[0071] Ideally, the above mapping relationship is applicable to the four cameras, and the rays emitted by the four cameras can intersect at one point, namely P(x w ,y w ,z w However, due to the inevitable noise in the data, the rays li deviate, resulting in multiple rays not intersecting at a single point. Therefore, it is necessary to estimate the spatial position of the joint point P in space. This paper considers the distance between rays and sets a spherical region with a radius 10 times the shortest distance between rays. The moving point P' moves within the spherical region. In combination with the OpenPose confidence parameter, the spatial point with the minimum distance to multiple rays is used as the intersection point of the multiple rays.

[0072] If the rays are l1, l2, l3, l4, where |P,l i | represents the ideal state from the joint point P to the ray l i The shortest distance, the three-dimensional coordinates of the real joint point R are expressed as follows:

[0073]

[0074] Among them, the above formula indicates that the final joint point P in space is the point closest to the four rays. Since the rays have confidence parameters, when solving the position of the true joint point R, the spatial point should be close to the ray with higher confidence. In order to reduce the measurement error, the Openpose confidence parameter ρ is introduced i ; Then the three-dimensional coordinate formula of the real joint point R is as follows:

[0075]

[0076] Furthermore, the weighted distance and expression of the joint point to multiple rays in space and the three-dimensional coordinate expression of the space point are:

[0077]

[0078] where d i For ray l i The direction vector, c i The coordinates of the origin of the camera coordinate system in the world coordinate system.

[0079] As real-time video data streams are input, the OpenPose deep learning algorithm outputs a two-dimensional coordinate sequence. Through the above process, a three-dimensional multivariate time series covering the motion information of multiple joints of the human body can be further obtained, that is, time series data. The variable of the multivariate time series is a certain dimension of a certain joint point.

[0080] Step S3, the time series preprocessing module detects abnormal timestamps in the time series data through the time series anomaly detection submodule, removes them and fills them with null values ​​through the anomaly removal and filling submodule, and the smoothing filter submodule uses Kalman filtering to adaptively adjust the filter parameters under the MSE error constraint to smooth the filled time series data.

[0081] This invention recognizes boxing moves based on time series, rather than video, because deep learning models based on video data cannot provide real-time functionality for the scenarios it addresses, and these models are generally large. By using time series classification to identify boxing moves, the invention not only identifies the type of punch but also the opponent's defensive movements, providing a basis for subsequent analysis of boxing techniques, tactics, styles, and characteristics.

[0082] Due to the occlusion of the boxing ring during the video shooting process, the motion capture module cannot capture the athlete's movement trajectory within a specific time period, resulting in anomalies in some timestamps in the time series data, which affects the accuracy of data processing and motion analysis. For this reason, the present invention designs a time series preprocessing module before motion recognition, including a time series anomaly detection submodule, an anomaly removal and filling submodule, and a smoothing filter submodule. The time series anomaly detection submodule analyzes the time series data, identifies and marks the timestamps where the outliers are located; the anomaly removal submodule sets the data value of the abnormal timestamp to zero, and fills the empty value through interpolation or Transformer prediction; the smoothing filter submodule smoothes the filled time series data through Kalman filtering, adaptively adjusts the filter parameters under the MSE error constraint, and outputs the time series data after smoothing filtering. The specific implementation steps include:

[0083] It is known that the training dataset should not contain any anomalies. To this end, we first collect boxing athlete training data in an unobstructed environment to construct a training dataset that can reflect the real laws of human movement and ensure that the training data contains no anomalies. Subsequently, we collect athlete training data in an actual boxing ring environment as the test dataset to simulate data anomalies in real scenarios.

[0084] Normalization is used to map the value range of the time series of the training and test data to the real number space from 0 to 1, thereby eliminating the impact of different feature dimensions. Specifically, the min-max method is used to normalize the time series data of the training and test sets. Considering that time series are multivariate, this method needs to be applied to each time series sample and each variable. Therefore, x is the time series data under a certain variable, and the dimension is the time series length. The specific implementation is shown in the following formula.

[0085]

[0086] The time series anomaly detection submodule incorporates a Transformer reconstruction model. While the model accurately reconstructs normal movements, it can significantly deviate from the original input for anomalous data segments. This principle allows the location of anomalous timestamps to be identified. Using the Transformer deep learning model architecture, the model learns the normal patterns of the training data, including but not limited to: 1. The normal motion patterns of each joint during punching and retracting; 2. The acceptable range of velocity and acceleration within human motion; and 3. The acceptable range of joint angle variation.

[0087] Furthermore, the training dataset is fed into the Transformer model. The encoder in the Transformer model uses the self-attention mechanism to learn the relationship between different time steps in the time series and encode the entire sequence into a feature vector containing global information. The decoder reconstructs the original time series data from this feature vector and minimizes the reconstruction error, allowing the Transformer model to learn the normal pattern of boxing movement. The reconstruction error MSE during the training phase is shown in the following formula, where the dimension is the time series dimension, where x∈R L×M is the original data, To reconstruct the data, L is the length of the time series and m is the number of joint points.

[0088]

[0089] An anomaly score is constructed using reconstruction error to measure data anomalies. A larger anomaly score indicates that the model has difficulty accurately reconstructing the data segment, indicating that anomalies may exist within this time period due to occlusion. By setting an appropriate threshold, anomalous segments in the boxing ring data are identified. Specifically, model training is performed to minimize reconstruction error. However, during testing, the presence of anomalies prevents the normal reconstruction of the anomaly location, resulting in a larger reconstruction error at the anomaly location. Based on this principle, a larger reconstruction error equates to a larger anomaly score, which can be used to measure data anomalies and thus output the anomaly location. Regarding threshold setting, a higher threshold may mark normal data as anomalies, hindering data mining; a lower threshold may mark data at anomaly locations as normal. This embodiment of the present invention does not compromise the normal movement patterns of the data. To filter out anomalous locations, the anomaly scores are ranked, and the timestamp locations corresponding to the top 5% (i.e., the anomalous data segments) are output by the time series anomaly module.

[0090] The abnormal data segments are set to blank, and a variety of methods are used to fill the blank data segments with normal values, including but not limited to: 1. When there is less missing data, the interpolation method is used to use adjacent normal data points to estimate the values ​​of the abnormal data points through linear interpolation and spline interpolation; 2. When there is more missing data, the trained Transformer model is used to predict the abnormal data segments and fill them with predicted values.

[0091] Due to the high intensity and rapid changes in boxing, the data stream contains highly dynamic and rapidly changing characteristics, and the spatial information reflected in the time series cannot be destroyed. Therefore, the present invention uses adaptive filtering technology to smooth the padded time series data, using the most suitable filter for each time series sample to ensure that the coordinate values ​​do not change while the data is smoothed. This process is measured using the MSE error and the filtering parameters are adaptively adjusted.

[0092] Step S4: The action recognition module uses the time series classification task to assign an action label to each timestamp in the time series data. The specific implementation steps include:

[0093] The present invention realizes action recognition through time series classification tasks. This scenario involves: offensive actions, defensive actions, counterattack actions and feints, including but not limited to the following specific actions:

[0094] Offensive moves: straight punch, hook punch, uppercut, two-punch combination, three-punch combination, multi-punch combination;

[0095] Defensive actions: holding the head, blocking, sticking to each other, dodging, slapping, moving, and swinging the arms;

[0096] Counterattack: counter punch, counter hook, counter hook, counter combination punch;

[0097] Feint: A move by a boxer to mislead his opponent into reacting.

[0098] Professional sports science practitioners and computer-related practitioners annotated the time series data output by the time preprocessing module and split it into training and test sets in a ratio of 8:2.

[0099] The action recognition module uses a multivariate time series classification model based on the LSTM architecture. To facilitate model training, data augmentation is used to segment the time series using overlapping time windows and feed them into the model for training. This multivariate time series classification model based on the LSTM architecture has multiple LSTM layers that capture temporal and variable dependencies in the preprocessed time series data. This is followed by a fully connected layer that maps the LSTM layer outputs to the probability distribution of the target class. The output layer uses a softmax activation function to output the probability of each class.

[0100] The cross entropy loss function is used to measure the difference between the model prediction results and the true label. By calculating the difference between the predicted probability distribution and the true label distribution, the prediction error of the model is quantified, thereby guiding the model to continuously optimize the parameters during the training process.

[0101] The preprocessed time series data is input into the trained multivariate time series classification model based on the LSTM architecture, and a corresponding action label is assigned to each timestamp of the preprocessed time series data.

[0102] Step S5, the action analysis module further analyzes the time series with action tags: analyzes the footwork of the movement category; identifies each single punch in the combination punch; identifies the hitting part of the punch; and identifies effective and ineffective punches.

[0103] The movement category analyzes footwork and classifies boxers' movement tags into fine-grained categories, including sprint, circle, slide, and passive. When a boxer uses high speed and large angles to quickly close the distance with the opponent and initiate an attack, it is classified as a sprint. At moderate speeds and angles, moving around the opponent to control the field and seek offensive opportunities, it is classified as a circle. When a boxer uses smooth speed and small angles, which is more often used for defense, it is classified as a slide. Passive movement occurs when a boxer is forced to move under attack.

[0104] The combination punch recognizes each single punch and classifies it into combination punches and counter-combination punches in a fine-grained manner to ensure that the combination punch label remains unchanged. It has a secondary label to distinguish between straight punches, swinging punches, and hooks.

[0105] Single-punch puncher identification determines the puncher based on the coordinate values, speed, and acceleration fluctuations of both hands during attack and counterattack techniques and tactics.

[0106] Identify the impact part of a punch, and based on the distance from the punch to the opponent's joint point during offensive and counterattack tactics, determine whether it hits the opponent and output the impact part.

[0107] The identification of valid and invalid punches, in a competition scenario, takes each interaction between the two sides as a unit, combines the punching parameters and the hitting part of the attacker during each interaction, and identifies whether the attacker's punch is a valid punch or an invalid punch.

[0108] The specific implementation steps are as follows:

[0109] The time series data with action labels contains information on 21 joints, including important joints for measuring boxing action information, such as hand joints, wrist joints, elbow joints, knee joints, ankle joints and foot joints. The displacement, velocity, rotation angle and acceleration of these important joints are calculated and summarized into a data set. Professional sports science practitioners and computer-related practitioners perform fine-grained annotation on the time series with action labels, including sprint steps, circle steps, sliding steps and passive annotations in footwork; single punch annotations in combination punches and counterattack combination punches; and annotations of effective punches and ineffective punches in offensive and counterattack techniques and tactics.

[0110] Specifically, using the red team's perspective as an example, we filter out the timestamps corresponding to the red team's major movement categories and retrieve the blue team's data at the same timestamps. We construct a new multivariate time series dataset containing the xyz coordinates of the knee, ankle, and foot joints, as well as the velocities, rotation angles, and accelerations of these joints. We further derive the distance information between the red and blue team's knee, ankle, and foot joints. This new multivariate time series dataset with derived features is used for fine-grained classification of gait patterns within major movement categories.

[0111] The dataset was further split into training and test sets in an 8:2 ratio. Based on this data, an LSTM-based fine-grained gait classification model was trained to provide corresponding fine-grained gait labels for each timestamp of the movement categories, including sprints, circles, slides, and passives.

[0112] In particular, the timestamps of combination punches and counter-combination punches are screened out, and the xyz coordinate values ​​of the hand joints, wrist joints, and elbow joints, as well as the speed, rotation angle, and acceleration of these joints are combined to construct a new multivariate time series dataset for the classification of single punches in combination punches.

[0113] Furthermore, the dataset was split into training and test sets in an 8:2 ratio. Based on this data, an LSTM-based fine-grained combination punch classification model was trained to provide a corresponding single punch label for each timestamp of the combination punch.

[0114] The timestamps of offensive and counterattack tactics are filtered out, and the opponent's data at the same timestamps is retrieved. A new multivariate time series dataset is constructed, containing the xyz coordinates, velocity, rotation angle, and acceleration of the attacker's hand, wrist, and elbow joints. The distances between the attacker's hands and the defender's head, hand, wrist, elbow, and hip joints are derived, as well as the displacement, velocity, and acceleration of the head, hand, elbow, shoulder, knee, ankle, and foot joints. This new multivariate time series dataset is used to identify the puncher of a single punch, the impact site of a punch, and the identification of effective and ineffective punches.

[0115] Furthermore, the coordinate values, speed, rotation angle, and acceleration of a pair of hand joints, wrist joints, and elbow joints are used to determine that the corresponding hand with the larger fluctuation is the punching hand, and a punching hand label is given to each time series sample.

[0116] Furthermore, thresholds are set based on the distances between the attacker's punching hand and the defender's head, hand, elbow, and hip joints to identify the impact location. If multiple distances reach the threshold, the smallest distance is selected as the impact location. Impact locations include: head, hand, elbow, and abdomen. The corresponding timestamps are labeled with the impact location and the no-impact label.

[0117] Furthermore, the average speed and acceleration of the hand joints, wrist joints, and elbow joints of the offensive puncher for 10 consecutive time stamps are calculated, and the average speed and acceleration of the 10 consecutive time stamps are guaranteed to be the largest, as well as the average displacement, speed, and acceleration of the head joints, hand joints, elbow joints, shoulder joints, knee joints, ankle joints, and foot joints of the defensive side at the corresponding time stamps. Combined with the offensive punch label, hit label, and valid / invalid punch label, a machine learning-based binary classification model is trained. The alternative models include K-nearest neighbor, decision tree, random forest, XGBoost and other algorithms, and the model with the best classification accuracy is selected as the valid punch and invalid punch recognition model.

[0118] Step S6, the technical and tactical statistics module, counts the boxer's technical and tactical usage: statistics on the number and percentage of offensive, defensive, and counterattack technical and tactical tactics; statistics on the boxer's dominant hand usage pattern; statistics on the hitting parts; calculation of hit rate and effective punch ratio; and calculation of important parameters for evaluating technical and tactical skills, including pressure coefficient, distance coefficient, balance coefficient, and movement coefficient. The specific implementation steps are as follows:

[0119] Statistics and percentage calculation of offensive, defensive, and counter-attack tactics. Taking the game round as the unit, the number of offensive, defensive, counter-attack, and feint tactics in each round is counted.

[0120] Furthermore, we calculated the proportion of the number of occurrences of straight punch, swing punch, hook punch, and combination punch labels in offensive techniques and tactics; the proportion of the number of occurrences of various defensive action labels such as head holding and arm swinging in defensive techniques and tactics; the proportion of the number of occurrences of refined footwork action labels in movement; the proportion of the number of occurrences of counterattack single punch and counterattack combination punch labels in counterattack techniques and tactics; the proportion of offensive technique and tactics timestamps in the entire game; the proportion of defensive technique and tactics timestamps in the entire game; the proportion of counterattack technique and tactics timestamps in the entire game; and the proportion of fake action timestamps in the entire game.

[0121] Statistics on the dominant hand of boxing techniques and tactics are used to analyze boxers' dominant hand usage patterns. The offensive and counterattack technique and tactics puncher tags are counted per round, and the proportion of left-handed tags and right-handed tags among all puncher tags is calculated.

[0122] Hit location statistics. Count the number of hit location labels in a round of the game scene, taking the game round as the unit.

[0123] Calculation of hit rate and effective punch ratio. Taking the round as the unit, the hit rate is calculated as the ratio of the number of offensive and counter-attack tactics that hit the target to the total number of offensive and counter-attack tactics. The effective punch ratio is calculated as the ratio of effective punches to the number of offensive and counter-attack tactics that hit the target.

[0124] Calculation of important parameters for evaluating technical and tactical skills, including:

[0125] Calculate the pressure system of the red and blue teams for a round. Taking the round as the unit, we combine the proportion of offensive and counterattack technique and tactical timestamps in the entire match, hit rate, effective punch ratio, and the number of sprint step label appearances in the proportion of movement actions. We calculate the weighted average with weights of 0.5, 0.2, 0.2, and 0.1 to obtain the pressure system of the red and blue teams.

[0126] Distance coefficient calculation: Calculate the average distance between the red and blue foot joints in a round as the distance coefficient.

[0127] Balance coefficient calculation: Calculate the average area of ​​the triangle formed by the two hand joints and the head joint within a round as the balance coefficient.

[0128] Movement coefficient calculation: calculate the ratio of the timestamp of the movement tag in a round to the entire game as the movement coefficient.

[0129] Step S7, the technical and tactical style and characteristics analysis and evaluation module, uses the machine learning classification module and integrates the features output by the technical and tactical statistics module in S6 to evaluate the athlete's technical and tactical style and characteristics in a round. The specific implementation steps are as follows:

[0130] Based on the parameters output by the technical and tactical statistics module, the features used as the analysis and evaluation of technical and tactical styles and characteristics include: the number of offensive, defensive, and counterattack techniques and tactics in a round; the proportion of the number of straight punches, swinging punches, hooks, and combination punches in offensive techniques and tactics; the proportion of the number of various defensive action labels such as head holding and arm swinging in defensive techniques and tactics; the proportion of the number of refined footwork action labels in movement; the proportion of the number of counterattack single punches and counterattack combination punches in counterattack techniques and tactics, and combined with the hit rate, effective punch ratio, pressure system, distance coefficient, balance coefficient, and movement coefficient.

[0131] One round generates two samples, namely the red boxer sample and the blue boxer sample, and each sample contains the above features.

[0132] Professional sports science practitioners annotate each sample with its style and characteristics. The annotation labels include:

[0133] Technical and tactical style: active attack, defensive counterattack, offensive back and forth, roaming

[0134] Technical and tactical characteristics: feints, direct attacks, pressing, attack first and then defend

[0135] By integrating the above features and labels, a classification model based on machine learning is constructed. Alternative models include K-nearest neighbor, decision tree, random forest and other algorithms. The model with the best classification performance is selected as the technical and tactical style and characteristics evaluation model.

[0136] Important technical and tactical parameters are input into the trained machine learning-based classification model to evaluate the boxer's technical and tactical style and characteristics within a round.

[0137] Furthermore, a standard action knowledge base is established, a standard execution method of boxing actions is set, the action data of professional boxers are preprocessed and action recognition is performed, and important parameters such as the speed, angle, acceleration, etc. of the athlete when performing boxing actions are calculated.

[0138] Using the standard movement knowledge base, we can compare a boxer's actual punching data with the standard movements for similarity. By analyzing the deviations between the angles, speeds, and other parameters of the boxing moves and the corresponding standard movements, we can analyze and evaluate the quality of the athletes' movements.

[0139] Furthermore, the dynamic time warping (DTW) algorithm is used to process the similarity comparison of time series data and calculate the deviations of various parameters between the actual punching data of ordinary boxers and the standard action data.

[0140] The above description is only a preferred embodiment of the present invention, but is not intended to limit the present invention. Those skilled in the art can use the technical content disclosed above to make possible changes and modifications to the present invention without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment with equivalent changes. Without departing from the content of the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. An intelligent evaluation method for boxing technical and tactical styles, characterized by: The method comprises the following steps: Step S1: Deploy four cameras around the boxing ring to collect action video data of the red and blue boxers; calibrate the cameras to obtain the intrinsic and extrinsic parameter matrices of each camera; Step S2: First, the OpenPose deep learning algorithm is used to detect the human joints in the acquired action video data, that is, to obtain the two-dimensional plane coordinates of the joints. Then, based on the principle of the pinhole imaging model and combined with the OpenPose confidence, the three-dimensional coordinates of the joints are obtained, that is, the time series data of the red and blue sides are obtained respectively; Step S3, preprocessing the time series data: analyzing the time series data, identifying and marking the timestamps of the outliers in the time series data, then setting the data values ​​of the timestamps where the outliers are located to blank and filling the blank values, and then performing smoothing filtering on the filled time series data to obtain the preprocessed time series data; The specific steps of step S3 are as follows: Step S3-1, obtaining the boxer's motion data in an unobstructed environment, constructing a training data set that can reflect the real human motion patterns, and using the athlete training data collected in the boxing ring environment as the test data set; Step S3-2: using a normalization method, the value range of the time series data in the training dataset and the test dataset is mapped to a real number space from 0 to 1; Step S3-3, inputting the normalized training data set into the Transformer model, so that the Transformer model learns the normal pattern of boxing movement; In step S3-4, the normalized test data set is input into the Transformer model obtained in step S3-3, and the anomaly score is constructed using the reconstruction error MSE. A threshold is set, and the top 5% of data with anomaly scores greater than the threshold are selected and marked as abnormal data segments, with timestamps as the unit. Otherwise, they are normal data segments. Step S3-5: Empty the abnormal data segments and fill them with normal values ​​through interpolation or Transformer prediction; Step S3-6, using adaptive filtering technology to smooth the filled time series data; Step S4, using the time series classification task to realize action recognition, and then assigning an action label to each timestamp of the preprocessed time series data; Step S5: annotating the time series data with action labels, including sprint steps, circle steps, slide steps, and passive steps, punching hands and punching locations, and effective and ineffective punches in offensive and counterattack tactics; Step S6, counting the boxer's skills and tactics: Based on the data obtained in step S5, the number and proportion of offensive, defensive, counterattack, and feint skills and tactics are calculated, as well as the statistics of the dominant hand used in the skills and tactics, the statistics of the hitting parts, the hit rate and the effective punch ratio, and the important parameters for evaluating the skills and tactics; In step S7, the time series pre-processed in step S3 is labeled with styles and characteristics, and the labels are integrated with the important technical and tactical parameters obtained in step S6 to construct a classification model based on machine learning. The important technical and tactical parameters are input into the trained classification model based on machine learning to evaluate the boxer's technical and tactical style and characteristics in a round.

2. The intelligent evaluation method for boxing technique and tactics according to claim 1 is characterized in that: The specific steps of step S2 are as follows: Step S2-1: Use the OpenPose deep learning algorithm to detect the human joints in the acquired action video data and obtain the joint points P i The two-dimensional plane coordinates (u i , v i ), i=1, 2, 3 or 4; Step S2-2, using the pinhole imaging model principle to calculate the two-dimensional coordinates of the joint point P in the ideal state, the three-dimensional coordinates P (X w ,Y w ,Z w ): Among them, f x and f y are the horizontal focal length and vertical focal length of the camera respectively; u0 and v0 are the pixel coordinates of the intersection of the camera optical axis and the imaging plane, M i is the external parameter matrix from the world coordinate system to the camera i coordinate system, x w 、y w 、z w is the three-dimensional coordinate of the joint point P in the ideal state, z c is the proportionality coefficient; In step S2-3, to avoid noise, it is necessary to estimate the spatial position of the real joint point R. Considering the distance between the rays emitted by the camera, a spherical area is set. The radius of the sphere is 10 times the shortest distance between the rays. The three-dimensional coordinate formula of the real joint point R is as follows: Among them, l i is the ray emitted by the i-th camera, |P,l i | represents the ideal state from the joint point P to the ray l i The shortest distance, ρ i is the Openpose confidence parameter, n=4; Step S2-4: The action video data covers multiple joints of the human body, and time series data is obtained according to the real three-dimensional coordinates of the multiple joints.

3. The intelligent evaluation method for boxing technique and tactics according to claim 1 is characterized in that: In step S3-5, the abnormal data segments are set to empty and filled with normal values. Specifically, when there are few abnormal data segments, the interpolation method is adopted, and the adjacent normal data points are used to estimate the values ​​of the abnormal data segments through linear interpolation and spline interpolation; when there are many abnormal data segments, the trained Transformer model or regression model is used to predict the abnormal data segments and fill them with predicted values.

4. The intelligent evaluation method for boxing technique and tactics according to claim 1, characterized in that: The specific steps of step S4 are as follows: Step S4-1, using the time series classification task to realize action recognition, classifying actions into offensive actions, defensive actions, counter-attack actions, and feints, annotating the pre-processed time series data, and dividing it into a training set and a test set; Step S4-2, using the training set to train a multivariate time series classification model based on the LSTM architecture; In step S4-3, the preprocessed time series data is input into the trained multivariate time series classification model based on the LSTM architecture, and a corresponding action label is assigned to each timestamp of the preprocessed time series data.

5. The intelligent evaluation method for boxing technique and tactics according to claim 1 is characterized in that: The specific steps of step S5 are as follows: Step S5-1: Filter out the timestamps corresponding to the movement categories in the action tags, classify the steps into sprint, circle, slide, and passive based on the displacement, speed, rotation angle, and acceleration of the joint movement, and perform fine-grained footwork annotation on the time series data of the movement categories in the action tags; Step S5-2, filtering out the timestamps corresponding to the attack and counterattack in the action tags, labeling the puncher with the attack and counterattack time series data in the action tags according to the coordinate values, movement speed, and acceleration of the hands, and labeling the hitting parts with the attack and counterattack time series data according to the distance from the opponent's joint points; Step S5-3, taking each interaction between the two parties as a unit, combined with the puncher and the hitting part in each interaction, identifies whether there is a hit or not. At the same time, the hitting effect of the time series with action labels of attack and defense is marked, that is, the effective punch and invalid punch are marked.

6. The intelligent evaluation method for boxing technique and tactics according to claim 1, characterized in that: The important parameters for evaluating techniques and tactics in step S6 include: pressure coefficient, distance coefficient, balance coefficient and movement coefficient.

7. The intelligent evaluation method for boxing technique and tactics according to claim 1, characterized in that: The specific steps of step S7 are as follows: Step S7-1: Label the style and characteristics of the time series data obtained after preprocessing in step S3, and divide it into training and test sets. The style labels include active attack, defensive counterattack, offensive back-and-forth, and roaming. The characteristic labels include feint, direct attack, pressing, and attack first and then defense. Step S7-2, integrating the statistics and percentage calculation of offensive, defensive, and counterattack techniques and tactics described in step S6, statistics of dominant hand used in techniques and tactics, statistics of striking locations, calculation of hit rate and effective punch ratio, and evaluation of important parameters of techniques and tactics, as well as style and characteristic labels, to construct and train a classification model based on machine learning; Step S7-3: Input the important technical and tactical parameters into the trained machine learning-based classification model to evaluate the boxer's technical and tactical style and characteristics in a round.

8. An intelligent evaluation system for boxing techniques and tactics, used to implement the method according to any one of claims 1 to 7, characterized in that: include: Acquisition module, motion capture module, time series preprocessing module, motion recognition module, motion analysis module, technical and tactical statistics module, technical and tactical style and characteristics analysis and evaluation module; The acquisition module is connected to the motion capture model and obtains the action video data of the red and blue players around the boxing ring; The motion capture module uses the OpenPose deep learning algorithm and the pinhole imaging model principle to obtain the red and blue action video data obtained through the input channel, combined with the OpenPose confidence parameter, to obtain the red and blue time series data. The time series preprocessing module includes a time series anomaly detection submodule, an anomaly removal and filling submodule, and a smoothing filter submodule. The time series anomaly detection submodule analyzes the time series data, identifies and marks the timestamps where the outliers are located in the time series data, and the anomaly removal and filling submodule sets the data values ​​of the timestamps where the outliers are located to zero, and fills the empty values ​​by interpolation or Transformer prediction. The smoothing filter submodule performs smoothing filtering on the filled time series data to obtain the smoothed filtered time series data as the output of the time series preprocessing module; The action recognition module uses the time series classification task to realize action recognition, and then assigns an action label to each timestamp of the preprocessed time series data; The motion analysis labels the time series data with motion tags, including sprint steps, circle steps, slide steps and passive steps, punching hands and punching parts, and effective and ineffective punches in offensive and counterattack tactics. The technical and tactical statistics module counts the boxer's technical and tactical usage, including the number and percentage calculation of offensive, defensive, and counterattack technical and tactical tactics; the boxer's dominant hand usage pattern is counted by the technical and tactical punching dominant hand; and the statistics of the striking parts; Calculation of hit rate and effective punch ratio; important parameters for evaluating techniques and tactics; The technical and tactical style and characteristic analysis and evaluation module constructs a classification model based on machine learning according to the output features of the technical and tactical statistics module combined with style and characteristic labels, and uses the classification model based on machine learning to realize the technical and tactical style evaluation of boxing.

Images