Method and system for analyzing fighting technique and tactics

Abstract

The invention belongs to the technical field of sports science and technology, and particularly relates to a fighting technique and tactical analysis system which comprises a target detection and skeletal muscle algorithm module, a space-time diagram neural network action recognition module, a striking effect evaluation module, an anti-interference mechanism module and a video processing and data storage module. The target detection and skeletal muscle algorithm module comprises data preparation and model training, the data preparation is to intercept a first preset number of pictures from common fight competition videos of Olympic Festival and Whole Festival, mark red and blue corners, a head, a trunk, arms and skeletal muscle key points in sequence by using a marking tool, and to simulate competition light changes, and to obtain a target detection and skeletal muscle algorithm; and randomly changing the image brightness. According to the method, full-automatic video analysis is realized through a target detection algorithm, a skeletal muscle recognition algorithm and a space-time diagram neural network, the analysis time is shortened from an hour level to a minute level, and subjective deviation can be eliminated; a space-time diagram neural network is adopted for modeling joint movement relevance, and fighting action features are optimized in a targeted mode.

Description

Technical Field

[0001] The present invention belongs to the field of sports science and technology, and in particular relates to a method and system for analyzing fighting techniques and tactics. Background Art

[0002] In recent years, computer vision and deep learning have been increasingly used in sports, primarily for athlete motion capture, technical and tactical analysis, and training optimization. Common technologies include target detection for athlete positioning, posture estimation for skeletal muscle keypoint detection, and action recognition for behavioral motion recognition. However, existing technologies primarily focus on single-action recognition (e.g., running, shooting), lacking comprehensive, dynamic, and competitive technical and tactical analysis solutions for combat sports like boxing and taekwondo.

[0003] Traditional combat tactical analysis relies on coaches to repeatedly watch videos and manually mark key movements. The analysis process is time-consuming and easily influenced by the coach's personal experience, making it difficult to quantify the athlete's performance. Currently, common action recognition systems are mostly based on general algorithms, and have low accuracy in recognizing complex actions such as combination punches, defense, counterattacks, and footwork. Current technologies are mostly focused on judging single actions, resulting in fragmented tactical analysis and difficulty in accurately analyzing action details. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for analyzing fighting techniques and tactics, which can realize fully automatic video analysis through target detection algorithm + skeletal muscle recognition algorithm + spatiotemporal graph neural network, shorten the analysis time from hours to minutes, and eliminate subjective bias; use spatiotemporal graph neural network to model the correlation of joint movements, and optimize the characteristics of fighting movements in a targeted manner; divide the fighting units based on the dynamic time planning algorithm, and automatically mark the attack and defense rounds; provide an interactive analysis module with frame-level playback + multiple speeds + data overlay.

[0005] The present invention provides a combat technique and tactics analysis system, which includes a target detection and skeletal muscle algorithm module, a spatiotemporal graph neural network action recognition module, a striking effect evaluation module, an anti-interference mechanism module, and a video processing and data storage module, wherein:

[0006] The target detection and skeletal muscle algorithm module collects fighting game videos, captures pictures from the videos, annotates the pictures using annotation tools, and trains according to the set model training parameters;

[0007] The action recognition module of the spatiotemporal graph neural network performs model training and deployment according to the input data, and recognizes the action based on the recognized skeletal muscle data;

[0008] The anti-interference mechanism module uses the Kalman filter algorithm and sliding window smoothing method to process interference images;

[0009] The hitting effect evaluation module evaluates the hitting effect of the identified action according to the part collision detection algorithm;

[0010] The video processing and data storage module processes the original video stream and stores the identified athlete data and image video data into a database.

[0011] Optionally, the motion recognition module of the spatiotemporal graph neural network inputs the processed data into the system in a specific [N, C, T, V, M] dimensional format, where N represents the number of samples participating in the training in each batch, C is the number of key features, T is the number of frames required for the analysis of the continuity and feature integrity of the fighting movements, V represents a second preset number of skeletal muscle joints, and M is the number of combatants; for the data of the above input dimensions, data labels for a third preset number of basic movements are constructed to perform model training and deployment, and movements are recognized based on the identified skeletal muscle data.

[0012] Optionally, the hitting effect evaluation module is used for part collision detection and biomechanical calculation. The part collision detection assumes a hit on the head and first calculates the intersection ratio of athlete A's wrist frame and athlete B's head frame to determine whether it is a valid hit:

[0013]

[0014] IoU is the intersection over union (IoU), which measures the degree of overlap between two frames. In this scenario, it is used to determine the overlap between athlete A's wrist frame and athlete B's head frame to determine whether the hit is effective.

[0015] represents the intersection area of ​​athlete A's wrist frame and athlete B's head frame, Represents athlete A's wrist frame, Represents the head frame of athlete B;

[0016]

[0017] Judgment result, when satisfied If the result is True, it means that the hit is valid; otherwise, it is False, which means that the hit is invalid. It represents the sum of the number of times the intersection-union ratio is greater than 0.3 in T calculations. When this sum is greater than or equal to 2, it is considered a valid hit;

[0018] Biomechanical calculations are used to calculate speed and impulse. These are combined with an impulse assessment algorithm model and a grading system for striking effects, including general hits, minor hits, important hits, the number "8," and knockout, to achieve a quantitative assessment of striking effects.

[0019] (Δd = joint position, Δt = frame interval)

[0020]

[0021] Calculate velocity and impulse separately, v represents velocity, F avg represents the average force, m represents the mass, and v impact is the input speed, v retract is the rotation speed of the fighter's arm, and Δt is the time interval.

[0022] Optionally, the anti-interference mechanism module is used for short-term occlusion prediction, false touch determination, and boundary jitter smoothing.

[0023] During the short-term occlusion prediction process, based on video analysis, when the confidence level of the detected target is less than a fourth preset value, the system activates the Kalman filter mechanism and uses the prediction model of the Kalman filter algorithm to predict the position of the currently occluded joint point;

[0024] During the false touch identification process, the system uses a motion direction consistency detection method. At the moment of the strike, it calculates the angle between the wrist velocity vector and the line connecting the center of the target part. If the angle is less than the fifth preset value, the contact is determined to be a false touch; if the angle is greater than or equal to the sixth preset value, it is considered a valid strike action.

[0025] During the boundary jitter smoothing process, the system adopts a sliding window smoothing method, selects the detection results of the latest seventh preset number of frames, and performs weighted averaging according to preset weights.

[0026] Optionally, the present invention further provides a fighting technique and tactics analysis method, which is applied to the fighting technique and tactics analysis system as described above, and the fighting technique and tactics analysis method comprises the following steps:

[0027] Step 1. Data preparation and model training;

[0028] Step 2. Action recognition;

[0029] Step 3. Evaluation of striking effect;

[0030] Step 4. Anti-interference processing;

[0031] Step 5. Video processing and data storage.

[0032] Optionally, the data preparation and model training include image acquisition and screening, fine annotation operations, and model training optimization, wherein:

[0033] During the image collection and screening process, fighting match videos are collected, and a first preset number of images are intercepted from the videos, covering different stages of the game, athlete movements and confrontation situations;

[0034] During the detailed annotation process, the professional annotation tool Labelme was used to meticulously annotate each image, marking the positions of red and blue corners to distinguish different athletes; annotating the outlines of the head, torso, and arms to provide a basis for subsequent analysis of striking locations; accurately locating a second preset number of skeletal muscle key points to accurately reflect changes in the athlete's body posture; and simulating the complex lighting environment of a fighting match, randomly changing the image brightness within a range of ±30% to enhance the model's adaptability to different lighting conditions.

[0035] During the model training and optimization process, the labeled images are used as training data for the target detection algorithm model and the skeletal muscle recognition algorithm model, respectively. During the training process, training parameters, including the learning rate and the number of iterations, are set according to the characteristics and performance requirements of different models. By continuously adjusting the parameters and optimizing the training, the target detection algorithm model can accurately identify the red and blue corners of the athletes, and the skeletal muscle recognition algorithm model can determine the positions of the key points of each skeletal muscle.

[0036] Optionally, the action recognition includes data standardization input, data label construction, model training and application, wherein:

[0037] During the data standardization input process, the processed data is input into the system in a specific [N, C, T, V, M] dimensional format, where N represents the number of samples participating in training in each batch, C is the number of key features, T is the number of frames required for analysis based on the coherence and feature integrity of combat movements, V represents a second preset number of skeletal muscle joints, and M is the number of combatants;

[0038] During the data label construction process, data labels for a third preset number of basic actions are constructed for the data of the above input dimensions. By accurately labeling the action labels corresponding to each data sample, an accurate learning target is provided for the spatiotemporal graph neural network, enabling it to learn the characteristic patterns of different actions.

[0039] During the training and application of the model, the constructed data labels and input data are used to train the spatiotemporal graph neural network, learn the mapping relationship between skeletal muscle data and action labels, and optimize network parameters. After the training is completed, it is deployed to the analysis system. When new fighting game video data is input, the system analyzes the athlete's skeletal muscle data based on the trained spatiotemporal graph neural network and accurately identifies the type of action the athlete is currently performing.

[0040] Optionally, the striking effect evaluation includes striking position determination and quantitative evaluation calculation, wherein:

[0041] During the process of determining the hitting position, the system uses a position collision detection algorithm based on precise geometric collision detection principles, combined with the target detection frame obtained by the target detection algorithm and the skeletal muscle data provided by the skeletal muscle recognition algorithm, to calculate the intersection ratio of the hitting position frame of athlete A and the hit position frame of athlete B. By setting a reasonable intersection ratio threshold, it is determined whether the hit is valid. If the intersection ratio reaches or exceeds the threshold, it is considered a valid hit; otherwise, it is considered an invalid hit.

[0042] During the quantitative evaluation calculation process, speed and impulse are calculated separately to quantify the striking effect. When calculating the impulse, if the user's arm weight is unknown, the arm weight is estimated based on the preset gender weight ratio according to gender differences and combined with the body weight. The calculated impulse data is combined with the preset striking effect grading standard to accurately grade the striking effect and achieve a quantitative evaluation of the striking effect.

[0043] Optionally, the anti-interference processing includes short-term occlusion prediction, false touch determination and recognition, and boundary jitter smoothing processing, wherein:

[0044] During the short-term occlusion prediction process, based on video analysis, when the confidence level of the detected target is less than a fourth preset value, the system activates the Kalman filter mechanism and uses the prediction model of the Kalman filter algorithm to predict the position of the currently occluded joint point;

[0045] During the false touch identification process, the system uses a motion direction consistency detection method to calculate the angle between the wrist velocity vector and the line connecting the center of the target part at the moment of hitting. If the angle is less than a fifth preset value, the contact is determined to be a false touch; if the angle is greater than or equal to a sixth preset value, it is determined to be a valid hitting action;

[0046] During the boundary jitter smoothing process, the system adopts a sliding window smoothing method, selects the detection results of the latest seventh preset number of frames, and performs weighted averaging according to preset weights.

[0047] Optionally, the video processing and data storage includes flexible video processing and data storage management, wherein:

[0048] During the flexible video processing, the system uses multiple speed setting architectures for the original fighting game video stream;

[0049] During the data storage management process, the system organizes and stores the athlete data identified during the analysis process, including position, movement, hitting effect information, and image and video data, and establishes a data storage structure and database management system.

[0050] The technical effects achieved by the present invention are:

[0051] The model system's combat technique and tactical analysis architecture: a cascaded model architecture of target detection algorithm + skeletal muscle recognition algorithm + spatiotemporal graph neural network, enabling end-to-end analysis from target detection -> skeletal muscle recognition -> motion recognition, optimizing joint connections and dynamic weight calculation for combat movements. High-precision impact location detection: a geometric collision detection-based impact determination algorithm combines the target detection algorithm's target detection frame with the skeletal muscle data from the skeletal muscle recognition algorithm, along with Kalman filter prediction and smoothing algorithms, for precise impact location detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a fighting technique and tactics analysis system of the present invention.

[0053] Figure 2 It is an overall flow chart of the execution of the fighting technique and tactics analysis system of the present invention.

[0054] Figure 3 This is the overall framework diagram of target detection and skeletal muscle recognition of the present invention.

[0055] Figure 4 It is a schematic diagram of the spatiotemporal graph neural network action recognition of the present invention. DETAILED DESCRIPTION

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0057] See also Figure 1-4 As shown, in one embodiment, the present invention provides a combat technique and tactics analysis system, including a target detection and skeletal muscle algorithm module, a spatiotemporal graph neural network action recognition module, a striking effect evaluation module, an anti-interference mechanism module, and a video processing and data storage module, wherein:

[0058] The target detection and skeletal muscle algorithm module collects fighting game videos, captures pictures from the videos, annotates the pictures using annotation tools, and trains according to the set model training parameters;

[0059] The action recognition module of the spatiotemporal graph neural network performs model training and deployment according to the input data, and recognizes the action based on the recognized skeletal muscle data;

[0060] The anti-interference mechanism module uses the Kalman filter algorithm and sliding window smoothing method to process interference images;

[0061] The hitting effect evaluation module evaluates the hitting effect of the identified action according to the part collision detection algorithm;

[0062] The video processing and data storage module processes the original video stream and stores the identified athlete data and image video data into a database.

[0063] Exemplarily, the target detection and skeletal muscle algorithm modules include data preparation and model training. The data preparation is to capture 2,000 pictures from common fighting competition videos of the Olympics and National Games, and use the labelme annotation tool to annotate the red and blue corners, head, torso, arms and skeletal muscle key points in sequence. In order to simulate the changes in competition lighting, the image brightness is randomly changed. The model training is to train according to the set model training parameters, use the target detection algorithm model to identify the red and blue corners of the athletes, and use the skeletal muscle recognition algorithm to determine the position of each skeletal muscle key point. The video processing and data storage module processes the original video stream, supports speed setting and frame-level control, and saves the identified athlete data and image video data into the database.

[0064] Optionally, the motion recognition module of the spatiotemporal graph neural network inputs the processed data into the system in a specific [N, C, T, V, M] dimensional format, where N represents the number of samples participating in the training in each batch, C is the number of key features, T is the number of frames required for the analysis of the continuity and feature integrity of the fighting movements, V represents a second preset number of skeletal muscle joints, and M is the number of combatants; for the data of the above input dimensions, data labels for a third preset number of basic movements are constructed to perform model training and deployment, and movements are recognized based on the identified skeletal muscle data.

[0065] Exemplarily, the action recognition module of the spatiotemporal graph neural network includes input data processing, data labeling, and model training deployment. The input data dimension is set to [N, C, T, V, M], where N is the number of training batches, batch is set to 256, C represents joint features, and there are 3 features of x, y, and acc of 2D skeletal muscle joints. T is the number of frames, and one action takes 20 frames. V is 18 skeletal muscle joints, and M is the number of people in a frame. Two people are taken for a two-player fight. The data labeling and model training deployment are to perform data labeling on 15 basic actions, straight punches, and swing punches for the above input dimensions, and then perform model training and deployment to recognize actions based on the identified skeletal muscle data.

[0066] Optionally, the striking effect evaluation module is used for part collision detection and biomechanical calculations.

[0067] When determining the strike location, the system uses a collision detection algorithm based on precise geometric collision detection principles. It combines the target detection frame obtained by the target detection algorithm with the skeletal muscle data provided by the skeletal muscle recognition algorithm to calculate the intersection ratio between athlete A's striking location frame and athlete B's struck location frame. By setting a reasonable intersection ratio threshold, the system determines whether the strike is valid. If the intersection ratio reaches or exceeds the threshold, the strike is considered valid; otherwise, it is considered invalid.

[0068] During the quantitative evaluation calculation process, speed and impulse are calculated separately to quantify the striking effect. When calculating the impulse, if the user's arm weight is unknown, the arm weight is estimated based on the preset gender weight ratio according to gender differences and combined with the body weight. The calculated impulse data is combined with the preset striking effect grading standard to accurately grade the striking effect and achieve a quantitative evaluation of the striking effect.

[0069] For example, in the part collision detection, assuming a hit on the head, the intersection ratio of athlete A's wrist frame and athlete B's head frame is first calculated to determine whether it is a valid hit:

[0070]

[0071] IoU is the intersection over union (IoU), which measures the degree of overlap between two frames. In this scenario, it is used to determine the overlap between athlete A's wrist frame and athlete B's head frame to determine whether the hit is effective.

[0072] represents the intersection area of ​​athlete A's wrist frame and athlete B's head frame, Represents athlete A's wrist frame, represents the head frame of athlete B, and ∩ represents the intersection operation;

[0073]

[0074] Judgment result, when satisfied If the result is True, it means that the hit is valid; otherwise, it is False, which means that the hit is invalid. It represents the sum of the number of times the intersection-union ratio is greater than 0.3 in T calculations. When this sum is greater than or equal to 2, it is considered a valid hit;

[0075] Biomechanical calculations are used to calculate speed and impulse. These are combined with an impulse assessment algorithm model and a grading system for striking effects, including general hits, minor hits, important hits, the number "8," and knockout, to achieve a quantitative assessment of striking effects.

[0076] (Δd = joint position, Δt = frame interval)

[0077]

[0078] Calculate velocity and impulse separately, v represents velocity, F avg represents the average force, m represents the mass, and v impact is the input speed, v retract is the rotation speed of the fighter's arm, and Δt is the time interval.

[0079] Optionally, the anti-interference mechanism module is used for short-term occlusion prediction, false touch determination, and boundary jitter smoothing.

[0080] During the short-term occlusion prediction process, based on video analysis, when the confidence level of the detected target is less than a fourth preset value, the system activates the Kalman filter mechanism and uses the prediction model of the Kalman filter algorithm to predict the position of the currently occluded joint point;

[0081] During the false touch identification process, the system uses a motion direction consistency detection method. At the moment of the strike, it calculates the angle between the wrist velocity vector and the line connecting the center of the target part. If the angle is less than the fifth preset value, the contact is determined to be a false touch; if the angle is greater than or equal to the sixth preset value, it is considered a valid strike action.

[0082] During the boundary jitter smoothing process, the system adopts a sliding window smoothing method, selects the detection results of the latest seventh preset number of frames, and performs weighted averaging according to preset weights.

[0083] For example, short-term occlusion prediction is to use Kalman filtering when the detection confidence conf<0.5, and predict the current joint point position based on the joint point vectors of the previous 5 frames. False touch is determined through motion direction consistency detection. If the angle between the wrist velocity vector and the center of the target part at the moment of hitting is <45°, a false touch is determined. Boundary jitter smoothing is to use sliding window smoothing, and the weighted average of the detection results of the last 3 frames is taken, with weights of 0.2, 0.6, and 0.2 respectively, to reduce boundary jitter interference.

[0084] Optionally, a fighting technique and tactics analysis method is applied to the fighting technique and tactics analysis system as described above, and the fighting technique and tactics analysis method includes the following steps:

[0085] Step 1. Data preparation and model training;

[0086] Step 2. Action recognition;

[0087] Step 3. Evaluation of striking effect;

[0088] Step 4. Anti-interference processing;

[0089] Step 5. Video processing and data storage.

[0090] Optionally, the data preparation and model training include image acquisition and screening, fine annotation operations, and model training optimization, wherein:

[0091] During the image collection and screening process, fighting match videos are collected, and a first preset number of images are intercepted from the videos, covering different stages of the game, athlete movements and confrontation situations;

[0092] During the detailed annotation process, the professional annotation tool Labelme was used to meticulously annotate each image, marking the positions of red and blue corners to distinguish different athletes; annotating the outlines of the head, torso, and arms to provide a basis for subsequent analysis of striking locations; accurately locating a second preset number of skeletal muscle key points to accurately reflect changes in athlete posture; and simulating the complex lighting environment of a fighting match, randomly changing the image brightness to enhance the model's adaptability to different lighting conditions.

[0093] During the model training and optimization process, the labeled images are used as training data for the target detection algorithm model and the skeletal muscle recognition algorithm model, respectively. During the training process, training parameters, including the learning rate and the number of iterations, are set according to the characteristics and performance requirements of different models. By continuously adjusting the parameters and optimizing the training, the target detection algorithm model can accurately identify the red and blue corners of the athletes, and the skeletal muscle recognition algorithm model can determine the positions of the key points of each skeletal muscle.

[0094] Exemplarily, data preparation and model training include image acquisition and screening, fine annotation operations, and model training optimization.

[0095] Image acquisition and screening is to widely collect fighting competition videos from the Olympic Games and the National Games. The video quality is clear and the movements are standardized, which fully reflects the various technical and tactical application scenarios of fighting sports. The first preset number of representative pictures are carefully intercepted from the massive videos, such as selecting 2,000 pictures covering different stages of the game, athlete movements and confrontation situations. The fine annotation operation is to use the professional annotation tool labelme to carefully annotate each picture, accurately mark the position of the red and blue corners, and facilitate the distinction between different athletes. The head, torso, and arm contours are marked in detail to provide a basis for subsequent analysis of the striking parts, and accurately locate 18 skeletal muscle key points, including the left eye, right eye, left ear, right ear, nose, neck, shoulders, elbows, wrists, and hips. , knees, and ankles to accurately reflect the changes in athlete's body posture. In order to simulate the complex lighting environment at the fighting competition, the image brightness is randomly changed, and the brightness change range is controlled at ±30% to enhance the model's adaptability to different lighting conditions. The model training is optimized to use the labeled pictures as training data, which are used for the training of the target detection algorithm model and the skeletal muscle recognition algorithm model respectively. During the training process, the training parameters are set according to the characteristics and performance requirements of different models, including the learning rate and the number of iterations. By continuously adjusting the parameters and optimizing the training process, the target detection algorithm model can accurately identify the red and blue corners of the athletes, and the skeletal muscle recognition algorithm model can determine the position of the key points of each skeletal muscle, laying a solid foundation for subsequent action recognition and technical and tactical analysis.

[0096] Optionally, the action recognition includes data standardization input, data label construction, model training and application, wherein:

[0097] During the data standardization input process, the processed data is input into the system in a specific [N, C, T, V, M] dimensional format, where N represents the number of samples participating in training in each batch, C is the number of key features, T is the number of frames required for analysis based on the coherence and feature integrity of combat movements, V represents a second preset number of skeletal muscle joints, and M is the number of combatants;

[0098] During the data label construction process, data labels for a third preset number of basic actions are constructed for the data of the above input dimensions. By accurately labeling the action labels corresponding to each data sample, an accurate learning target is provided for the spatiotemporal graph neural network, enabling it to learn the characteristic patterns of different actions.

[0099] During the training and application of the model, the constructed data labels and input data are used to train the spatiotemporal graph neural network, learn the mapping relationship between skeletal muscle data and action labels, and optimize network parameters. After the training is completed, it is deployed to the analysis system. When new fighting game video data is input, the system analyzes the athlete's skeletal muscle data based on the trained spatiotemporal graph neural network and accurately identifies the type of action the athlete is currently performing.

[0100] Exemplarily, action recognition includes data standardization input, data label construction, model training and application.

[0101] The data standardization input is to input the processed data into the system in a specific [N, C, T, V, M] dimensional format, where N is set to 256, representing the number of samples participating in the training in each batch. This value is determined after multiple experiments and performance evaluations. C is 3, corresponding to the three key features of the x-coordinate, y-coordinate and accuracy acc of the 2D skeletal muscle joints. T selects 20 frames, which is determined based on the research on the continuity and feature integrity of fighting movements. V represents 18 skeletal muscle joints, covering the main movement joints of the human body, and comprehensively reflecting the athlete's physical movement state. M is 2, which is in line with the actual scene of two-person fighting. Data label construction is to construct data labels for 15 basic actions based on the data of the above input dimensions. These 15 basic actions include straight punches, swing punches, hook punches, forward slides, and backward slides. The data is processed by the neural network, and the spatial-temporal graph neural network is trained to detect the movement of forward step, left slide step, right slide step, as well as the compound movement of sprint step, backward step, forward slide step, backward slide step, left circle step, right circle step, cross step, and the state of not moving. By accurately marking the action label corresponding to each data sample, the spatiotemporal graph neural network is provided with accurate learning targets, so that it can learn the characteristic patterns of different actions. Model training and application are to use the constructed data labels and input data to train the spatiotemporal graph neural network. During the training process, the network continuously learns the mapping relationship between skeletal muscle data and action labels, and optimizes the network parameters. After the training is completed, it is deployed to the analysis system. When new fighting game video data is input, the system analyzes the athlete's skeletal muscle data based on the trained spatiotemporal graph neural network to accurately identify the type of action currently performed by the athlete.

[0102] Optionally, the striking effect evaluation includes striking position determination and quantitative evaluation calculation, wherein:

[0103] During the process of determining the hitting position, the system uses a position collision detection algorithm based on precise geometric collision detection principles, combined with the target detection frame obtained by the target detection algorithm and the skeletal muscle data provided by the skeletal muscle recognition algorithm, to calculate the intersection ratio of the hitting position frame of athlete A and the hit position frame of athlete B. By setting a reasonable intersection ratio threshold, it is determined whether the hit is valid. If the intersection ratio reaches or exceeds the threshold, it is considered a valid hit; otherwise, it is considered an invalid hit.

[0104] During the quantitative evaluation calculation process, speed and impulse are calculated separately to quantify the striking effect. When calculating the impulse, if the user's arm weight is unknown, the arm weight is estimated based on the preset gender weight ratio according to gender differences and combined with the body weight. The calculated impulse data is combined with the preset striking effect grading standard to accurately grade the striking effect and achieve a quantitative evaluation of the striking effect.

[0105] Exemplarily, the striking effect evaluation includes striking location judgment and quantitative evaluation calculation.

[0106] For strike location determination, assuming a head hit is an example, the system uses a collision detection algorithm to calculate the intersection ratio between athlete A's wrist frame and athlete B's head frame. By setting a reasonable intersection ratio threshold, the system determines whether the strike is valid. If the intersection ratio reaches or exceeds the threshold, the strike is considered valid; otherwise, it is considered invalid. This determination method is based on precise geometric collision detection principles, combining the target detection frame obtained by the target detection algorithm and the skeletal muscle data provided by the skeletal muscle recognition algorithm to accurately determine the strike location. Quantitative evaluation involves calculating velocity and impulse separately to quantify the strike effect. When calculating impulse, if the user's arm weight is unknown, the arm weight is estimated based on gender differences, with the arm weight generally accounting for 0.057 of body weight for men and 0.0497 of body weight for women. The calculated impulse data is then combined with the preset strike effect grading criteria, including general hit, minor hit, significant hit, count "8", and knockout, to accurately grade the strike effect and achieve a quantitative assessment of the strike effect.

[0107] Optionally, the anti-interference processing includes short-term occlusion prediction, false touch determination and recognition, and boundary jitter smoothing processing, wherein:

[0108] During the short-term occlusion prediction process, based on video analysis, when the confidence level of the detected target is less than a fourth preset value, the system activates the Kalman filter mechanism and uses the prediction model of the Kalman filter algorithm to predict the position of the currently occluded joint point;

[0109] During the false touch identification process, the system uses a motion direction consistency detection method to calculate the angle between the wrist velocity vector and the line connecting the center of the target part at the moment of hitting. If the angle is less than a fifth preset value, the contact is determined to be a false touch; if the angle is greater than or equal to a sixth preset value, it is determined to be a valid hitting action;

[0110] During the boundary jitter smoothing process, the system adopts a sliding window smoothing method, selects the detection results of the latest seventh preset number of frames, and performs weighted averaging according to preset weights.

[0111] Exemplarily, the anti-interference processing includes short-term occlusion prediction, false touch determination and recognition, and boundary jitter smoothing processing.

[0112] For short-term occlusion prediction, when the confidence level conf of the detected target is less than 0.5 during video analysis, indicating that short-term occlusion may occur, the system activates the Kalman filter mechanism. This mechanism predicts the position of the currently occluded joint point based on the joint point vector information of the previous 5 frames and the prediction model of the Kalman filter algorithm, thereby reducing the impact of occlusion on the analysis results. For false touch identification, the system uses motion direction consistency detection. At the moment of striking, it calculates the angle between the wrist velocity vector and the center of the target part. If the angle is less than 45°, it is determined that the contact may be a false touch and needs to be further combined with other information for comprehensive judgment. If the angle is greater than or equal to 45°, it is more likely to be considered as part of a valid striking action. For boundary jitter smoothing, the system uses a sliding window smoothing method to address possible boundary jitter problems. The detection results of the last 3 frames are selected and weighted averaged with weights of 0.2, 0.6, and 0.2.

[0113] Optionally, the video processing and data storage includes flexible video processing and data storage management, wherein:

[0114] During the flexible video processing, the system uses multiple speed setting architectures for the original fighting game video stream;

[0115] During the data storage management process, the system organizes and stores the athlete data identified during the analysis process, including position, movement, hitting effect information, and image and video data, and establishes a data storage structure and database management system.

[0116] Exemplarily, video processing and data storage include flexible video processing and data storage management. Flexible video processing is for the original fighting game video stream. The system provides rich processing functions and supports multiple speed settings, including 0.25, 0.5, 0.75, 1.25, and 1.5. Data storage management is for the system to organize and store athlete data identified during the analysis process, including position, movement, hitting effect information, and image video data, and establish a data storage structure and database management system.

[0117] In the present invention, the video input is a fighting game video (official or recorded). To improve the accuracy of recognition, the video specification is at least 1080P, 30FPS;

[0118] Athlete detection: Using target detection algorithms, combined with labeled training sets, to detect the red and blue corners of athletes;

[0119] Skeletal muscle key point extraction: Using the skeletal muscle recognition algorithm, the data of 18 skeletal muscle joints of athletes (left eye, right eye, left ear, right ear, nose, neck, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, right ankle) are recognized;

[0120] Action classification algorithm: The joint point data of the past 0.5 seconds is input into the spatiotemporal graph neural network, and the action label (such as boxing method, step, etc.) is output. Based on the judgment of the punch, the target detection algorithm is used to identify key parts such as the head, torso, and arms. Combined with the key points of skeletal muscles, the punch trajectory is calculated, the hitting part and punching speed are predicted, and the hitting effect is calculated.

[0121] This application uses a cascaded model architecture of a target detection algorithm, a skeletal muscle recognition algorithm, and a spatiotemporal graph neural network to achieve end-to-end analysis from target detection to skeletal muscle recognition to action recognition, optimizing joint connection relationships and dynamic weight calculation for combat movements. High-precision detection of striking locations: A striking determination algorithm based on geometric collision detection combines the target detection algorithm's target detection frame with the skeletal muscle data from the skeletal muscle recognition algorithm, along with Kalman filter prediction and smoothing algorithms to achieve precise detection of striking locations.

[0122] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. A combat technique and tactics analysis system, characterized by: It includes target detection and skeletal muscle algorithm module, spatiotemporal graph neural network action recognition module, striking effect evaluation module, anti-interference mechanism module, video processing and data storage module, among which: The target detection and skeletal muscle algorithm module collects fighting game videos, captures pictures from the videos, annotates the pictures using annotation tools, and trains according to the set model training parameters; The action recognition module of the spatiotemporal graph neural network performs model training and deployment according to the input data, and recognizes the action based on the recognized skeletal muscle data; The anti-interference mechanism module uses the Kalman filter algorithm and sliding window smoothing method to process interference images; The hitting effect evaluation module evaluates the hitting effect of the identified action according to the part collision detection algorithm; The video processing and data storage module processes the original video stream and stores the identified athlete data and image video data into a database.

2. A combat technique and tactics analysis system according to claim 1, characterized in that: The action recognition module of the spatiotemporal graph neural network inputs the processed data into the system in a specific [N, C, T, V, M] dimensional format, where N represents the number of samples participating in the training in each batch, C is the number of key features, T is the number of frames required for the analysis of the continuity and feature integrity of the fighting movements, V represents a second preset number of skeletal muscle joints, and M is the number of combatants; for the data of the above input dimensions, data labels for a third preset number of basic movements are constructed to perform model training and deployment, and actions are recognized based on the identified skeletal muscle data.

3. A combat technique and tactics analysis system according to claim 1, characterized in that: The hitting effect evaluation module is used for part collision detection and biomechanical calculation. The part collision detection assumes a hit on the head. The intersection ratio of athlete A's wrist frame and athlete B's head frame is first calculated to determine whether it is a valid hit: IoU is the intersection over union (IoU), which measures the degree of overlap between two frames. In this scenario, it is used to determine the overlap between athlete A's wrist frame and athlete B's head frame to determine whether the hit is effective. represents the intersection area of ​​athlete A's wrist frame and athlete B's head frame, Represents athlete A's wrist frame, Represents the head frame of athlete B; Judgment result, when satisfied If the result is True, it means that the hit is valid; otherwise, it is False, which means that the hit is invalid. It represents the sum of the number of times the intersection-union ratio is greater than 0.3 in T calculations. When this sum is greater than or equal to 2, it is considered a valid hit; Biomechanical calculations are used to calculate speed and impulse. These are combined with an impulse assessment algorithm model and a striking effect classification system, including general hits, minor hits, important hits, "8" counts, and knockouts, to achieve a quantitative assessment of striking effects. Calculate velocity and impulse separately, v represents velocity, F avg represents the average force, m represents the mass, and v impact is the input speed, v retract is the rotation speed of the fighter's arm, and Δt is the time interval.

4. A combat technique and tactics analysis system according to claim 1, characterized in that: The anti-interference mechanism module is used for short-term occlusion prediction, false touch determination, and boundary jitter smoothing. During the short-term occlusion prediction process, based on video analysis, when the confidence level of the detected target is less than a fourth preset value, the system activates the Kalman filter mechanism and uses the prediction model of the Kalman filter algorithm to predict the position of the currently occluded joint point; During the false touch identification process, the system uses a motion direction consistency detection method. At the moment of impact, it calculates the angle between the wrist velocity vector and the line connecting the center of the target part. If the angle is less than the fifth preset value, the contact is determined to be a false touch. If the included angle is greater than or equal to a sixth preset value, it is considered a valid striking action; During the boundary jitter smoothing process, the system adopts a sliding window smoothing method, selects the detection results of the latest seventh preset number of frames, and performs weighted averaging according to preset weights.

5. A combat technique and tactics analysis method, applied to a combat technique and tactics analysis system according to any one of claims 1 to 4, characterized in that: The fighting technique and tactics analysis method comprises the following steps: Step 1. Data preparation and model training; Step 2. Action recognition; Step 3. Evaluation of striking effect; Step 4. Anti-interference processing; Step 5. Video processing and data storage.

6. A combat technique and tactics analysis method according to claim 5, characterized in that: The data preparation and model training include image acquisition and screening, fine annotation operations, and model training optimization, among which: During the image collection and screening process, fighting match videos are collected, and a first preset number of images are intercepted from the videos, covering different stages of the game, athlete movements and confrontation situations; During the detailed annotation process, the professional annotation tool Labelme was used to meticulously annotate each image, marking the positions of red and blue corners to distinguish different athletes; annotating the outlines of the head, torso, and arms to provide a basis for subsequent analysis of striking locations; accurately locating a second preset number of skeletal muscle key points to accurately reflect changes in athlete posture; and simulating the complex lighting environment of a fighting match, randomly changing the image brightness to enhance the model's adaptability to different lighting conditions. During the model training and optimization process, the labeled images are used as training data for the target detection algorithm model and the skeletal muscle recognition algorithm model, respectively. During the training process, training parameters, including the learning rate and the number of iterations, are set according to the characteristics and performance requirements of different models. By continuously adjusting the parameters and optimizing the training, the target detection algorithm model can accurately identify the red and blue corners of the athletes, and the skeletal muscle recognition algorithm model can determine the positions of the key points of each skeletal muscle.

7. A combat technique and tactics analysis method according to claim 5, characterized in that: The action recognition process includes data standardization input, data label construction, model training and application, among which: During the data standardization input process, the processed data is input into the system in a specific [N, C, T, V, M] dimensional format, where N represents the number of samples participating in training in each batch, C is the number of key features, T is the number of frames required for analysis based on the coherence and feature integrity of combat movements, V represents a second preset number of skeletal muscle joints, and M is the number of combatants; During the data label construction process, data labels for a third preset number of basic actions are constructed for the data of the above input dimensions. By accurately labeling the action labels corresponding to each data sample, an accurate learning target is provided for the spatiotemporal graph neural network, enabling it to learn the characteristic patterns of different actions. During the training and application of the model, the constructed data labels and input data are used to train the spatiotemporal graph neural network, learn the mapping relationship between skeletal muscle data and action labels, and optimize network parameters. After the training is completed, it is deployed to the analysis system. When new fighting game video data is input, the system analyzes the athlete's skeletal muscle data based on the trained spatiotemporal graph neural network and accurately identifies the type of action the athlete is currently performing.

8. A combat technique and tactics analysis method according to claim 5, characterized in that: The striking effect evaluation includes striking position determination and quantitative evaluation calculation, wherein: During the process of determining the hitting position, the system uses a position collision detection algorithm based on precise geometric collision detection principles, combined with the target detection frame obtained by the target detection algorithm and the skeletal muscle data provided by the skeletal muscle recognition algorithm, to calculate the intersection ratio of the hitting position frame of athlete A and the hit position frame of athlete B. By setting a reasonable intersection ratio threshold, it is determined whether the hit is valid. If the intersection ratio reaches or exceeds the threshold, it is considered a valid hit; otherwise, it is considered an invalid hit. During the quantitative evaluation calculation process, speed and impulse are calculated separately to quantify the striking effect. When calculating the impulse, if the user's arm weight is unknown, the arm weight is estimated based on the preset gender weight ratio according to gender differences and combined with the body weight. The calculated impulse data is combined with the preset striking effect grading standard to accurately grade the striking effect and achieve a quantitative evaluation of the striking effect.

9. The method for analyzing fighting techniques and tactics according to claim 1, characterized in that: The anti-interference processing includes short-term occlusion prediction, false touch determination and recognition, and boundary jitter smoothing processing, wherein: During the short-term occlusion prediction process, based on video analysis, when the confidence level of the detected target is less than a fourth preset value, the system activates the Kalman filter mechanism and uses the prediction model of the Kalman filter algorithm to predict the position of the currently occluded joint point; During the false touch identification process, the system uses a motion direction consistency detection method to calculate the angle between the wrist velocity vector and the line connecting the center of the target part at the moment of hitting. If the angle is less than a fifth preset value, the contact is determined to be a false touch; if the angle is greater than or equal to a sixth preset value, it is determined to be a valid hitting action; During the boundary jitter smoothing process, the system adopts a sliding window smoothing method, selects the detection results of the latest seventh preset number of frames, and performs weighted averaging according to preset weights.

10. The method for analyzing fighting techniques and tactics according to claim 1, characterized in that: The video processing and data storage include flexible video processing and data storage management, wherein: During the flexible video processing, the system uses multiple speed setting architectures for the original fighting game video stream; During the data storage management process, the system organizes and stores the athlete data identified during the analysis process, including position, movement, hitting effect information, and image and video data, and establishes a data storage structure and database management system.

Images