Taekwondo technical and tactical feature extraction method, electronic equipment and medium

Abstract

The invention belongs to the technical field of video analysis, and provides a taekwondo technical and tactical feature extraction method, electronic equipment and a medium, and the method comprises the steps: firstly obtaining image data of taekwondo competition of a target person; feature extraction is carried out, and skeleton features of a plurality of time sequence stages are determined; determining a three-dimensional human body posture estimation result of each time sequence stage according to the skeleton features of each time sequence stage; and finally, performing feature extraction on the image data according to the three-dimensional human body posture estimation result of each time sequence stage and a feature extraction model to obtain competitive features of the target personnel in each time sequence stage. According to the method, feature extraction is carried out for each stage, and action features in different time periods can be comprehensively captured. The method is not limited to simple analysis of actions any more, unique technical and tactical expressions of each stage are deeply mined, meanwhile, time sequence characteristics of the taekwondo competition process are fully considered, the actions are closely associated with the time dimension, and the accuracy of technical and tactical feature extraction is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of video analysis, and in particular relates to a method for extracting taekwondo technical and tactical features, an electronic device and a medium. Background Art

[0002] In Taekwondo competitions, time significantly influences the outcome, and top athletes adjust their techniques and tactics based on time and score. The application of techniques and tactics at different times, as well as the score gains and losses, are crucial to the overall outcome of a match.

[0003] The process of Taekwondo competition has obvious temporal characteristics. The application of techniques and tactics in different time periods and the score situation are crucial to the outcome of the entire game. At present, Taekwondo technique and tactics analysis generally only simply analyzes the action characteristics, the extraction of action features is insufficient and the temporal characteristics are ignored, resulting in poor extraction of technique and tactics features. Summary of the Invention

[0004] In view of this, the present invention provides a Taekwondo technical and tactical feature extraction method, electronic equipment and medium, aiming to solve the problem of poor technical and tactical feature extraction effect.

[0005] A first aspect of the present invention provides a method for extracting Taekwondo technical and tactical features, comprising:

[0006] Acquire image data of the target person's Taekwondo competition;

[0007] Perform feature extraction on image data to determine skeleton features at multiple time stages;

[0008] Determine the 3D human posture estimation result of each time sequence stage according to the skeleton features of each time sequence stage;

[0009] According to the 3D human posture estimation results and feature extraction model at each time stage, feature extraction is performed on the image data to obtain the competitive characteristics of the target person at each time stage.

[0010] In one possible implementation, feature extraction is performed on the image data to determine skeleton features at multiple time stages, including:

[0011] The image data is divided into time series to obtain the image data of the sudden attack segment, sudden counterattack segment, stalemate segment and counterpulsation segment.

[0012] The image data of each temporal stage are input into the Transformer model respectively to obtain the skeleton features of each temporal stage.

[0013] In one possible implementation, feature extraction is performed on the image data to determine skeleton features at multiple time stages, including:

[0014] Performing time series division on the image data to obtain the image data of the sudden attack segment, sudden counterattack segment, stalemate segment and counterpulsation segment;

[0015] The image data of each temporal stage is input into the first graph neural network model corresponding to each temporal stage to obtain the local features of human body movement in each temporal stage;

[0016] The image data of each temporal stage are input into the Transformer model to obtain the global features of human motion at each temporal stage;

[0017] The skeleton features of each temporal stage are determined according to the global features and local features of each temporal stage.

[0018] In a possible implementation, determining the skeleton features of each time series stage based on the global features and local features of each time series stage includes:

[0019] The global features and local features of each temporal stage are input into a feature fusion model based on the attention mechanism to obtain the skeleton features of each temporal stage; among them, the feature fusion model based on the attention mechanism pays different attention to global features and local features in each temporal stage.

[0020] In one possible implementation, determining a 3D human pose estimation result at each time sequence stage based on the skeleton features at each time sequence stage includes:

[0021] The skeleton features of each time sequence stage are input into the multi-scale pose estimation model to obtain the 3D human pose estimation results of each time sequence stage.

[0022] In one possible implementation, determining a 3D human pose estimation result at each time sequence stage based on the skeleton features at each time sequence stage includes:

[0023] Get the scale setting information for each timing stage;

[0024] The multi-scale attitude estimation model is adjusted according to the scale setting information to obtain the multi-scale attitude estimation model corresponding to each time series stage;

[0025] The skeleton features of each temporal stage are input into the multi-scale pose estimation model corresponding to each temporal stage to obtain the three-dimensional human pose estimation result of each temporal stage.

[0026] In one possible implementation, based on the 3D human pose estimation results and feature extraction model at each time stage, feature extraction is performed on the image data to obtain the competitive features of the target person at each time stage, including:

[0027] The three-dimensional human body posture estimation result of each time sequence stage and the image data are input into a second graph neural network model to obtain the competitive characteristics of the target person in each time sequence stage.

[0028] The second aspect of the present application provides a taekwondo skill and tactics feature extraction device, comprising:

[0029] The acquisition module is configured to acquire image data of taekwondo competition of the target person.

[0030] The extraction module is configured to perform feature extraction on the image data to determine skeleton features of multiple time sequence stages.

[0031] The estimation module is configured to determine a three-dimensional human body posture estimation result of each time sequence stage according to the skeleton features of each time sequence stage.

[0032] The output module is configured to perform feature extraction on the image data according to the three-dimensional human body posture estimation result of each time sequence stage and the feature extraction model to obtain competitive characteristics of the target person in each time sequence stage.

[0033] The third aspect of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the taekwondo skill and tactics feature extraction method of the first aspect.

[0034] The fourth aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the taekwondo skill and tactics feature extraction method of the first aspect.

[0035] The taekwondo skill and tactics feature extraction method, the electronic device and the medium provided by the embodiments of the present application first acquire image data of taekwondo competition of the target person, then perform feature extraction on the image data to determine skeleton features of multiple time sequence stages, then determine a three-dimensional human body posture estimation result of each time sequence stage according to the skeleton features of each time sequence stage, and finally perform feature extraction on the image data according to the three-dimensional human body posture estimation result of each time sequence stage and the feature extraction model to obtain competitive characteristics of the target person in each time sequence stage. The present application performs feature extraction for each stage respectively, can fully capture the action features of different time periods, is no longer limited to simple analysis of actions, but deeply excavates the unique skill and tactics performance of each stage, fully considers the time sequence characteristics of taekwondo competition process, closely associates the action with the time dimension, and guarantees the accuracy of skill and tactics feature extraction. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 This is a flowchart of the method for extracting Taekwondo technical and tactical features provided by an embodiment of the present invention;

[0038] Figure 2 It is a structural diagram of a Taekwondo technical and tactical feature extraction device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In the following description, specific details such as particular system structures and techniques are provided for purposes of illustration, not limitation, to facilitate a thorough understanding of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present invention with unnecessary detail.

[0040] Figure 1 This is a flow chart of the implementation of the Taekwondo technique and tactics feature extraction method provided by the embodiment of the present invention. Figure 1 As shown, in some embodiments, the method for extracting Taekwondo technical and tactical features includes:

[0041] S110, acquiring image data of a target person performing Taekwondo competition;

[0042] S120, extracting features from the image data to determine skeleton features at multiple time series stages;

[0043] S130, determining a 3D human body posture estimation result for each time sequence stage based on the skeleton features of each time sequence stage;

[0044] S140 , performing feature extraction on the image data based on the 3D human posture estimation result and feature extraction model at each time sequence stage, and obtaining the competitive features of the target person at each time sequence stage.

[0045] The taekwondo competition video data of the target person can be obtained from a taekwondo skill and tactics video resource intelligent service platform. The taekwondo skill and tactics video resource intelligent service platform is designed based on the skill and tactics index system of the taekwondo skill and tactics video data, constructs a taekwondo competition skill and tactics knowledge representation and knowledge description model, and realizes the digitization of the skill and tactics information in the taekwondo competition. The taekwondo high-level competition video resources are gathered, the resource islands between the resources and the coaches, the resources and the athletes, the resources and the management and service departments are broken, and a wide and open taekwondo skill and tactics resource service center is constructed. The implementation of the project will complete the cleaning and arrangement of the taekwondo high-level competition video data, form a taekwondo skill and tactics statistical analysis basic database, and provide a solid data basis for realizing artificial intelligence analysis.

[0046] In the embodiment of the application, the action feature data analysis, the opponent skill and tactics feature analysis, the referee rule analysis and other means are used to improve the skill and tactics analysis reasoning mechanism. The taekwondo skill and tactics data intelligent analysis method is established, the influence of different skill and tactics combinations on the taekwondo score is analyzed, the advantage skill and tactics combination and the disadvantage skill and tactics combination of the competitor are mined, the valuable skill and tactics feature clues such as the dynamic technical route of the competition process are found out, the competition diagnosis is helped for the coaches and athletes, the skill and tactics feature data of the main opponent is recognized and understood, and data support is provided for formulating scientific and reasonable competition strategy and tactics.

[0047] In the embodiment of the application, the taekwondo competition is usually 3 rounds, and each round is 2 minutes. In view of the importance of time in the competition, the higher the level of the taekwondo athlete, the more the skill and tactics will be adjusted in time according to the time and score in the competition. Therefore, the skill and tactics analysis of the athletes in different time periods is very important for the competition. The earlier the score is obtained, the earlier the competition initiative is occupied, and the final competition victory or defeat is greatly influenced. In the high-level competition, the score in the last decisive stage is very important. That is to say, the time period from the start of the competition to the first score not only creates the score, but also lays a foundation for the subsequent skill and tactics arrangement, which is very similar to the "three-stage evaluation theory" of table tennis. Specifically, the first half of the first round of the competition can be defined as the "surprise and rush stage". In the last period of each competition, the losing side wants to achieve a victory from a defeat, and will use a lot of all-or-nothing skill and tactics. The latter half of the third round of the competition can be defined as the "counterattack stage". After the "surprise and rush stage", the player will reflect during the two inter-round breaks, change the skill and tactics arrangement under the guidance of the coach or by himself / herself, so as to obtain an advantage. The latter half of the first round and the first half of the second round are defined as the "surprise and counterattack stage", and the latter half of the second round and the first half of the third round are defined as the "holding stage".

[0048] Specifically, cluster analysis can be used to identify the various temporal phases described above. However, given the differences in gender, level, and competition importance (event level) among taekwondo athletes, the specific temporal phase divisions will inevitably vary due to these factors. Therefore, clustering should be performed using video data that shares these characteristics. For example, a study examined the temporal segment division of the women's 49kg final at a particular Olympic Games. By observing and analyzing data from 19 matches, 285 scores and their corresponding times were obtained. Scoring times were categorized based on their intensity, resulting in four general categories. The average time per point was 0.94 seconds for the sprint phase, the highest of the four phases, while the stalemate phase had the lowest time, 0.57 seconds. This suggests that players initially test each other during the sprint phase, leading to longer scoring times, while the stalemate phase is the most fluid phase of the match. The sprint phase lasted from 0 to 30 seconds, the counterattack phase from 31 to 180 seconds, the stalemate phase from 181 to 300 seconds, and the counterattack phase from 301 to 360 seconds.

[0049] In some embodiments, feature extraction is performed on the image data to determine the skeleton features of multiple time series stages, including: temporal division of the image data to obtain image data of the attack segment, the counterattack segment, the stalemate segment, and the counterpulsation segment; and the image data of each time series stage is input into the Transformer model respectively to obtain the skeleton features of each time series stage.

[0050] In an embodiment of the present invention, the Transformer model is based on the self-attention mechanism and can effectively process long-distance dependencies in sequence data. The image data of each divided time series stage are input into the Transformer model in sequence. The model first encodes the input image data and converts it into a feature vector form suitable for model processing. During the encoding process, the self-attention mechanism calculates the degree of correlation between different areas and different time points in the image, highlighting the key information related to the skeleton features. For example, in the image data processing of the sudden grab segment, the model gradually extracts the skeleton features that can represent the human body movement in each time series stage. These features include information such as the position of the human body joints, the relative position relationship between the joints, and the dynamic changes during the movement process, providing a key basis for subsequent three-dimensional human posture estimation and technical and tactical analysis.

[0051] In some embodiments, feature extraction is performed on the video data to determine skeleton features of multiple time stages, including: performing time division on the video data to obtain video data of a sudden attack segment, a sudden defense segment, a stalemate segment and a counterattack segment; inputting the video data of each time stage into a first graph neural network model corresponding to each time stage to obtain local features of human motion in each time stage; inputting the video data of each time stage into a Transformer model respectively to obtain global features of human motion in each time stage; and determining skeleton features of each time stage according to the global features of each time stage and the local features of each time stage.

[0052] When processing video data, the Transformer model can capture the data dependency relationship of the global field by encoding the connection relationship between all joints, but it has some limitations. On the one hand, the Transformer model cannot model the local dependency relationship of the joint, and in Taekwondo, the local motion and mutual relationship of the joint are very important for accurately describing the action features of the athlete. On the other hand, the computational complexity of the Transformer model grows with the square of the sequence length, which will cause waste of computing resources. Therefore, the first graph neural network model is introduced to make up for the shortcomings of the Transformer model and ensure the ability to capture the local dependency relationship of the joint, while still using the Transformer model to ensure the ability to capture the global dependency relationship.

[0053] In the embodiments of the present application, the graph neural network (GNN) is suitable for processing data with a graph structure. In human motion analysis, the human body can be regarded as a graph structure composed of joint nodes (nodes) and joint connections (edges). The first graph neural network model is designed for the data characteristics of each time stage and can learn the local relationship and features between human joint nodes. The model captures the motion features of the local area of the human body in each time stage through the information transmission and aggregation mechanism between nodes, such as the motion range of a certain joint, the cooperative motion mode between adjacent joints, etc.

[0054] The image data of each time sequence stage is input into a corresponding first graph neural network model. The model first processes the image data and converts it into a representation suitable for graph structure analysis to determine the positions of human body joints and the connection relationships between joints. Then, through the message passing algorithm of the graph neural network, each node (joint) updates its own features according to the node information connected thereto. For example, in the sudden attack stage, the model focuses on the local motion characteristics of the arm or leg joints when the athlete quickly punches or kicks, such as the angle change of the joint, the motion speed, etc.; in the stalemate stage, the relative position relationship and small adjustment characteristics between the joints of different parts of the body in the defensive posture are focused on. After multiple rounds of message passing and feature updating, the model outputs the local features of human motion in each time sequence stage, which describe the motion details of the human body in each local area in detail.

[0055] In the embodiment of the present application, the image data of each time sequence stage is input into a Transformer model respectively. The model encodes the input image data and converts it into a feature vector form. In the encoding process, the self-attention mechanism calculates the correlation degree between different regions and different time points in the image, highlighting the global features related to the overall motion of the human body. For example, in the sudden attack stage, the model focuses on the overall attack posture of the athlete, including the transfer of the body center of gravity, the coordination of the overall action, etc.; in the counterattack stage, the global features presented by the coordinated motion of the athlete's body parts under the overall tactical layout taken to achieve the reversal are focused on. Through multiple layers of attention calculation and feedforward neural network processing, the model extracts the global features of human motion in each time sequence stage, which describe the motion patterns and states of the human body in different time sequence stages from the overall level.

[0056] In the embodiment of the present application, an adjustment layer is arranged in the first graph neural network, which is used to adjust the node definition and the connection of the edge, and the specific adjustment manner needs to be determined according to the characteristics of different time sequence stages, so that the first graph neural network corresponding to different time sequence stages can be obtained.

[0057] For example, in the "sudden attack stage", since the players are probing each other, the actions are relatively simple and direct, and the focus is on the preliminary attack and defense exploration. The nodes of the graph can be set as the main joints of the human body, such as shoulder joints, hip joints, knee joints, etc. These joints play a leading role in the actions in the sudden attack stage, such as the movement of the shoulder joint when punching, the cooperation of the hip joint and the knee joint when kicking. By focusing on these key nodes, the main action features of this stage can be efficiently captured.

[0058] During the "attack and counterattack" phase, the action is more fluid, with complex and varied movements. During this phase, competitors adjust their techniques and tactics based on the previous phase, leading to increased movement variation. In addition to the major joints, wrists, ankles, and other joints can serve as nodes. These smaller joints help fine-tune the direction and force of movements. For example, wrist and ankle movements can have a crucial impact on adjusting the direction of a kick or the angle of a punch.

[0059] The stalemate phase of a match is characterized by a fast pace and frequent, complex movements. During this phase, human parts like the head and spine can also be set as nodes. Head rotation and spinal flexion and extension play important roles in maintaining balance, observing the opponent, and adjusting posture during the stalemate phase. Including these parts in nodes provides a more comprehensive description of the player's movements.

[0060] The lagging opponent will take more aggressive and risky actions during the counterattack phase, so the node settings can be further refined to include the finger and toe joints. In some special attacking or defensive movements, the grip of the fingers and the force of the toes may affect the stability and effectiveness of the movement, such as the grip strength of the fingers during a takedown.

[0061] For example, in the "snatch" section, when recording the arm punching action, the focus is on connecting the shoulder and elbow joints, because the coordinated movement of these two joints determines the direction and force of the punch; for the leg kicking action, the focus is on connecting the hip and knee joints. This connection method can highlight the close cooperation between adjacent joints in simple movements.

[0062] In the "reversal segment," in addition to preserving connections between adjacent joints, we also add cross-joint connections. For example, we connect the shoulder and wrist to capture the overall coherence of the arm during movement adjustment, and we connect the hip and ankle to record the coordinated changes in leg movement adjustment. This cross-joint connection helps capture more complex relationships during movement adjustment.

[0063] During the "stalemate phase," a denser connection network is constructed. This includes not only connections between adjacent joints and across joints, but also connections between non-directly related joints. For example, the head and hip joints are connected. When a player turns to dodge an attack or adjusts their center of gravity, the movements of the head and hip joints are correlated. This connection captures this potential synergy.

[0064] During the "counterattack phase," the connections between nodes related to key offensive and defensive movements are strengthened. For example, when performing a powerful kick, the connections between the hip, knee, ankle, and toe joints are strengthened to highlight the coordinated force generation relationship between these joints in key movements; when performing hand defense movements, the connections between the shoulder, elbow, wrist, and finger joints are strengthened.

[0065] In some embodiments, the skeleton features of each time stage are determined based on the global features of each time stage and the local features of each time stage, including: inputting the global features of each time stage and the local features of each time stage into a feature fusion model based on the attention mechanism to obtain the skeleton features of each time stage; wherein the feature fusion model based on the attention mechanism pays different attention to the global features and the local features in each time stage.

[0066] In the embodiment of the present invention, the key to accurately determining the skeleton features at each temporal stage during the extraction of Taekwondo technical and tactical features lies in the effective fusion of global and local features. A feature fusion model based on an attention mechanism can dynamically allocate attention to global and local features based on the characteristics of different temporal stages, thereby obtaining accurate skeleton features. The feature fusion model can be a convolutional neural network model.

[0067] During the scramble phase, global features reflect the athlete's overall attacking momentum, such as the shift of the body's center of gravity and the coordination of overall movements. Local features focus on the rapid movements of key joints, such as the movement of the shoulder and elbow joints during a punch, or the movement of the hip and knee joints during a kick. These global and local features are input into a feature fusion model based on an attention mechanism. During this stage, the model prioritizes local features. During the scramble phase, the athlete's primary goal is to quickly launch an attack and score, and the local movements of key joints play a decisive role in the effectiveness of the attack. The attention mechanism assigns higher weights to these local features to more accurately capture the details of the attack. For example, when determining the power of an athlete's punch or the accuracy of a kick, the model focuses on analyzing the movement characteristics of the shoulder, elbow, hip, and knee joints. While global features are also considered, they are given a relatively lower weight because the overall attacking momentum is composed of a combination of local movements, and the accuracy and power of these local movements are more critical.

[0068] Global features during the attack phase reflect the athlete's overall technical and tactical layout, adjusted based on the previous phase, such as adjustments to body posture and changes in movement rhythm. Local features capture subtle changes in small joints like the wrist and ankle during movement adjustments, as well as the coordination of major joints in the new movement. These features are input into the feature fusion model. During the attack phase, the model must balance attention to global and local features. This is because athletes must adjust their overall technical and tactical layout while also fine-tuning small joints to achieve precise movement changes. The attention mechanism dynamically assigns weights based on the specific situation. For example, when an athlete makes significant adjustments to their body posture, global features receive increased attention; while when fine-tuning movement details, local features receive increased weight. This ensures that the model fully captures the athlete's movement adjustments during the attack phase.

[0069] Global features during the stalemate phase describe the athlete's overall movement rhythm, body balance, and defensive layout during sustained confrontation. Local features focus on the specific movements of various body joints during rapid, frequent movements, such as head rotation and spinal flexion and extension. These global and local features are input into the model. During this stage, the model prioritizes global features. Maintaining overall movement rhythm and body balance is crucial for athletes during the stalemate phase, characterized by a fast tempo and frequent movements. The attention mechanism assigns greater weight to global features to capture overall movement patterns and defensive layout. At the same time, local features are not neglected, as subtle movements of individual joints can affect overall balance and performance. The model dynamically adjusts its focus on local features based on the specific movement.

[0070] The global features of the counterattack phase reflect the overall aggressive tactical layout adopted by the lagging athlete to achieve a comeback, such as rapid body movement and the focus of the attack. Local features focus on the force applied to the finger and toe joints during risky moves, as well as the coordinated force generated by the major joints during powerful attacks. These features are input into the feature fusion model. During the counterattack phase, the model places greater emphasis on global features. This is because the lagging athlete needs to utilize a comprehensive tactical layout to launch an effective counterattack. The overall attack direction and movement coordination are crucial for a reversal of momentum. The attention mechanism assigns a higher weight to global features to capture the overall counterattack strategy. At the same time, attention is also given to the components of the local features related to key attacking movements, such as the force applied to the finger and toe joints, to ensure detailed analysis of the attacking movements.

[0071] In some embodiments, a three-dimensional human body posture estimation result of each timing stage is determined based on the skeleton features of each timing stage, including: inputting the skeleton features of each timing stage into a multi-scale posture estimation model to obtain a three-dimensional human body posture estimation result of each timing stage.

[0072] In the embodiments of the present application, the model first performs multi-scale analysis on the input skeleton features. It will construct feature representations of different scales, similar to constructing a feature pyramid. At a finer scale, the model can capture detailed information of human poses, such as the precise positions of the joints and the slight changes in the angles, which is very important for accurately describing the fine actions of the human body at each time stage, such as the subtle movements of the finger joints when the hand punches in the sudden attack segment. At a coarser scale, the model focuses on the overall structure and general posture of the human body, such as the overall orientation of the body and the relative position relationship of the main joints, which helps to grasp the overall action trend at different time stages, such as the overall defensive posture of the athlete in the holding segment. Through the fusion between features of different scales, the model can comprehensively utilize detailed information and overall information for pose estimation. The features of finer scales provide the basis for accurate estimation of poses, while the features of coarser scales help the model understand the pose of the human body in a global range, avoiding errors caused by interference of local details. The model uses these fused features for reasoning, learns the mapping relationship between the skeleton features and the three-dimensional human poses through the internal neural network structure. This learning of mapping relationship is based on a large amount of training data, so that the model can accurately predict the corresponding three-dimensional human pose according to the input skeleton features. The multi-scale pose estimation model can be Hourglass Network, HRNet, etc.

[0073] In some embodiments, according to the skeleton features of each time stage, the three-dimensional human pose estimation result of each time stage is determined, including: obtaining scale setting information of each time stage; adjusting the multi-scale pose estimation model according to the scale setting information to obtain a multi-scale pose estimation model corresponding to each time stage; inputting the skeleton features of each time stage into the multi-scale pose estimation model corresponding to each time stage to obtain the three-dimensional human pose estimation result of each time stage.

[0074] In this embodiment of the present invention, different temporal phases, such as the attack phase, counterattack phase, stalemate phase, and counterattack phase, each possess unique motion characteristics. The attack phase is swift and critical, often involving rapid attack initiation. Therefore, the model must focus on capturing subtle local movements and the overall attack dynamics. In the attack phase, players adjust their techniques and tactics, employing diverse movements that involve changes in body posture and fine-tuning of movement details. Therefore, the model must balance overall strategic adjustments with subtle changes in local movements. The stalemate phase features a fast-paced, high-action phase, requiring a comprehensive grasp of overall movement rhythm and body balance, as well as the specific movements of each joint during high-frequency movements. In the counterattack phase, players employ aggressive and risky moves. Therefore, the model must focus on the overall counterattack layout and the detailed force generation of key joints during powerful movements. Based on the characteristics of the movements in each phase, the corresponding scale settings are determined. For the attack phase, due to the rapidity of the attack and the importance of local joint movements, a higher proportion of fine scales may be used to accurately capture the rapid joint movements and subtle changes. Simultaneously, a certain proportion of coarse scales is used to grasp the overall attack direction and body center of gravity shift. For example, a 70% fine scale and a 30% coarse scale may be used. During the counterattack phase, considering both overall posture adjustments and local movement details, the scale setting may be relatively balanced, such as 40% fine scale, 30% medium scale, and 30% coarse scale, to comprehensively capture movement information at different levels. During the stalemate phase, due to the high frequency of movement and the need to maintain body balance, a greater emphasis may be placed on medium and coarse scales to grasp overall rhythm and balance, with a setting of 30% fine scale, 40% medium scale, and 30% coarse scale. During the counterattack phase, given the overall aggressive tactical layout and powerful movements at key joints, a larger number of coarse scales and a moderate amount of fine scales may be set, such as 20% fine scale and 80% coarse scale, to highlight the overall counterattack strategy and the power of key joints.

[0075] Adjust the model's parameters based on the scale setting. In the sudden phase, due to the high proportion of fine-scale features, the parameter weights related to fine-scale feature processing can be appropriately increased. For example, in the model's loss function, the weight of the fine-scale feature prediction error is set to 0.7, and the coarse scale is 0.3. During training, adjust the learning rate strategy and set a slightly larger learning rate for the parameters of the fine-scale feature extraction part so that it can learn the key local motion features of the sudden phase more quickly. In other stages, such as the sudden-reversal phase, based on the balance of the scale setting, the weights of the fine, medium, and coarse-scale feature prediction errors in the loss function are set relatively balanced, such as 0.33 for each, and a uniform and moderate learning rate is used to ensure that the model comprehensively learns motion adjustment information at different scales.

[0076] Taking the tackling phase as an example, after receiving skeleton feature data, the adjusted model uses multi-layer convolution operations on fine-scale paths to extract features from subtle joint movements, capturing details such as the minute angle changes of the finger joints during fist clenching and the rapid rotation of the wrist during punching. On the coarse-scale path, convolution kernels with larger receptive fields are used to analyze the overall attacking dynamics, such as the rapid shift of the body's center of gravity and changes in overall body orientation. Then, through the model's internal feature fusion mechanism, features extracted at different scales are combined, comprehensively considering local details and overall dynamics to infer the 3D human pose during the tackling phase. Similar results apply to other temporal phases, such as the stalemate phase. Based on the adjusted structure and parameters, the model focuses on analyzing the overall rhythm and balance maintenance of the body during high-frequency movements on mesoscale and coarse-scale paths. Combined with the specific joint motion features on the fine-scale path, it outputs accurate 3D human pose estimation for the stalemate phase. In this way, the model corresponding to each temporal phase leverages the advantages of its respective scale setting to accurately estimate the 3D human pose for that phase, providing precise posture data for subsequent in-depth analysis of the technical and tactical performance of taekwondo athletes at different stages.

[0077] In some embodiments, based on the three-dimensional human posture estimation results and feature extraction model of each time stage, feature extraction is performed on the image data to obtain the competitive characteristics of the target person at each time stage, including: inputting the three-dimensional human posture estimation results and image data of each time stage into the second graph neural network model to obtain the competitive characteristics of the target person at each time stage.

[0078] In an embodiment of the present invention, the sorted three-dimensional human posture estimation results are fused with the pre-processed image data and input into the second graph neural network model. During the fusion process, it is necessary to accurately map the joint positions in the posture estimation results to the corresponding human body parts in the image, so that the model can use the posture information and image information for feature extraction at the same time. The second graph neural network model is specially designed to process this graph structure data that integrates posture and image information. The nodes of the model can be defined as the joints of the human body, key areas in the image (such as different parts of the athlete's body, areas of light and shadow changes caused by movements, etc.), and the edges are used to connect nodes with associated relationships, such as connections between joints, connections between joints and corresponding image areas, etc. Through this graph structure, the model can effectively capture the spatial relationship and interaction between the various elements in the posture and image data.

[0079] First, initial feature extraction is performed on the input fused data. For the pose estimation results, the model analyzes the positions and angles of joints to extract features reflecting the athlete's movement pattern, amplitude, and speed. For example, the speed and force of the attacking movement are extracted from the posture information of the rapidly extending arm joints during the attack phase. For the image data, operations such as convolution are used to extract visual features from the image, such as color, texture, and shape. For example, the texture features of the athlete's clothing and the changes in light and shadow caused by the movement are extracted from the image. These features may be related to the athlete's movement style and technical characteristics. Through the graph propagation mechanism, nodes transfer information to each other, allowing each node to integrate features from neighboring nodes. For example, a joint node can receive visual features from the corresponding image region node and simultaneously transmit its own pose features to the image region node. In this process, the model continuously updates and integrates features from different nodes, gradually generating more representative competitive features. For example, during the attack and counterattack phase, the model uses graph propagation to fuse the changing features of the joints as the athlete adjusts their posture with the visual features of the body posture changes shown in the image, extracting comprehensive features reflecting technical and tactical adjustments.

[0080] After multiple rounds of graph propagation and feature fusion, the model globally integrates features across the entire graph structure. By comprehensively considering posture and image features at each temporal stage, it extracts features that fully reflect the target player's competitive state at that stage. For example, during the stalemate phase, the model integrates the body's posture stability during high-frequency movements, the coordination of joint movements, and the spatial relationships and interaction between athletes shown in the image. This yields global features that capture the competitive characteristics of the stalemate phase, such as defensive strategies and the utilization of offensive opportunities.

[0081] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0082] Figure 2 FIG. 1 is a schematic diagram of the structure of the Taekwondo technique and tactics feature extraction device provided by an embodiment of the present invention. Figure 2 As shown, the Taekwondo technique and tactics feature extraction device 2 includes:

[0083] An acquisition module 210 is used to acquire Taekwondo competitive image data of a target person;

[0084] Extraction module 220, for performing feature extraction on the image data and determining skeleton features at multiple time stages;

[0085] An estimation module 230 is configured to determine a 3D human body posture estimation result at each time sequence stage based on the skeleton features at each time sequence stage;

[0086] The output module 240 is used to extract features from the image data based on the 3D human posture estimation results and feature extraction model at each time sequence stage, and obtain the competitive features of the target person at each time sequence stage.

[0087] Optionally, the extraction module 220 is used to divide the image data into time series to obtain image data of the sudden attack segment, the sudden counterattack segment, the stalemate segment and the counterpulsation segment; and input the image data of each time series stage into the Transformer model respectively to obtain the skeleton features of each time series stage.

[0088] Optionally, the extraction module 220 is used to divide the image data into time series to obtain image data of the attack segment, the counterattack segment, the stalemate segment and the counterpulsation segment; the image data of each time series stage is input into the first graph neural network model corresponding to each time series stage to obtain the local features of the human body movement in each time series stage; the image data of each time series stage is input into the Transformer model respectively to obtain the global features of the human body movement in each time series stage; the skeleton features of each time series stage are determined according to the global features of each time series stage and the local features of each time series stage.

[0089] Optionally, the extraction module 220 is used to input the global features of each time series stage and the local features of each time series stage into a feature fusion model based on the attention mechanism to obtain the skeleton features of each time series stage; wherein, the feature fusion model based on the attention mechanism pays different attention to the global features and local features in each time series stage.

[0090] Optionally, the estimation module 230 is used to input the skeleton features of each time sequence stage into the multi-scale posture estimation model to obtain the three-dimensional human posture estimation result of each time sequence stage.

[0091] Optionally, the estimation module 230 is used to obtain the scale setting information of each timing stage; adjust the multi-scale posture estimation model according to the scale setting information to obtain the multi-scale posture estimation model corresponding to each timing stage; input the skeleton features of each timing stage into the multi-scale posture estimation model corresponding to each timing stage to obtain the three-dimensional human body posture estimation result of each timing stage.

[0092] Optionally, the output module 240 is used to input the three-dimensional human posture estimation results and image data of each time sequence stage into the second graph neural network model to obtain the competitive characteristics of the target person in each time sequence stage.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for extracting Taekwondo technical and tactical features, characterized in that: include: Acquire image data of the target person's Taekwondo competition; Performing feature extraction on the image data to determine skeleton features at multiple time series stages; Determine the 3D human posture estimation result of each time sequence stage according to the skeleton features of each time sequence stage; According to the three-dimensional human posture estimation result and the feature extraction model at each time sequence stage, feature extraction is performed on the image data to obtain the competitive features of the target person at each time sequence stage.

2. The method for extracting Taekwondo technical and tactical features according to claim 1, wherein: Feature extraction is performed on the image data to determine skeleton features at multiple time stages, including: Performing time series division on the image data to obtain image data of a sudden attack segment, a sudden counterattack segment, a stalemate segment, and a counterpulsation segment; The image data of each temporal stage are input into the Transformer model respectively to obtain the skeleton features of each temporal stage.

3. The method for extracting Taekwondo technical and tactical features according to claim 2, wherein: Feature extraction is performed on the image data to determine skeleton features at multiple time stages, including: Performing time series division on the image data to obtain image data of a sudden attack segment, a sudden counterattack segment, a stalemate segment, and a counterpulsation segment; The image data of each temporal stage is input into the first graph neural network model corresponding to each temporal stage to obtain the local features of human body movement in each temporal stage; The image data of each temporal stage are input into the Transformer model to obtain the global features of human motion at each temporal stage; The skeleton features of each temporal stage are determined according to the global features and local features of each temporal stage.

4. The method for extracting Taekwondo technical and tactical features according to claim 3, wherein: Based on the global features and local features of each time series stage, the skeleton features of each time series stage are determined, including: The global features and local features of each temporal stage are input into a feature fusion model based on the attention mechanism to obtain the skeleton features of each temporal stage; among them, the feature fusion model based on the attention mechanism pays different attention to global features and local features in each temporal stage.

5. The method for extracting Taekwondo technical and tactical features according to claim 1, wherein: Based on the skeleton features of each time sequence stage, the 3D human pose estimation result of each time sequence stage is determined, including: The skeleton features of each time sequence stage are input into the multi-scale pose estimation model to obtain the 3D human pose estimation results of each time sequence stage.

6. The method for extracting Taekwondo technical and tactical features according to claim 1, wherein: Based on the skeleton features of each time sequence stage, the 3D human pose estimation result of each time sequence stage is determined, including: Get the scale setting information for each timing stage; Adjusting the multi-scale attitude estimation model according to the scale setting information to obtain a multi-scale attitude estimation model corresponding to each time sequence stage; The skeleton features of each temporal stage are input into the multi-scale pose estimation model corresponding to each temporal stage to obtain the three-dimensional human pose estimation result of each temporal stage.

7. The method for extracting Taekwondo technical and tactical features according to claim 1, wherein: Based on the 3D human posture estimation results and feature extraction model at each time sequence stage, feature extraction is performed on the image data to obtain the competitive features of the target person at each time sequence stage, including: The three-dimensional human posture estimation results and the image data of each time sequence stage are input into the second graph neural network model to obtain the competitive characteristics of the target person in each time sequence stage.

8. A Taekwondo technical and tactical feature extraction device, characterized in that: include: An acquisition module, used to acquire Taekwondo competitive image data of a target person; An extraction module, configured to extract features from the image data and determine skeleton features at multiple time series stages; An estimation module, configured to determine a 3D human pose estimation result at each time sequence stage based on the skeleton features at each time sequence stage; The output module is used to extract features from the image data based on the three-dimensional human posture estimation results and feature extraction model at each time sequence stage to obtain the competitive features of the target person at each time sequence stage.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the Taekwondo technical and tactical feature extraction method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the Taekwondo technical and tactical feature extraction method as described in any one of claims 1 to 7 are implemented.

Images