Solid ball examination method and system based on ai

Abstract

The present application relates to the technical field of body examination score, specifically to a solid ball examination method and system based on AI, comprising the following steps: based on the video of the athlete throwing the solid ball, using the inter-frame difference method to calculate the pixel changes between consecutive video frames, identifying the motion area, separating the foreground and background, and determining the position and movement of the athlete and the solid ball, and generating motion area data.In the present application, by applying the inter-frame difference method and the optical flow method, the dynamic position of the athlete and the solid ball can be accurately identified and tracked, thereby providing more accurate and continuous motion trajectory information.The use of time convolution network and time series clustering makes the identification and classification of key actions more accurate, improves the accuracy and efficiency of action analysis, and the application of support vector machine and decision tree realizes more objective and detailed evaluation in technical execution quality scoring and comprehensive technical evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of physical examination scoring, and in particular to an AI-based shot put assessment method and system. Background Art

[0002] The field of physical assessment and scoring focuses on applying information technology and artificial intelligence to improve the accuracy and efficiency of assessments in sports training and competition. Within this technology, AI is used to analyze athlete performance, automate the scoring process, and provide training feedback. This includes quantitative assessment of athletic skills, real-time performance monitoring, and data-driven training improvements, with the goal of improving training effectiveness through accurate and objective data support.

[0003] The AI-based shot put assessment method utilizes machine learning and data analysis to evaluate and improve the performance and skills of shot put athletes. This method aims to provide an evaluation system that is more objective, accurate, and efficient than traditional scoring. This method enables a detailed analysis of an athlete's shot put skills, including throwing technique, power, speed, and accuracy.

[0004] Traditional shot put assessment methods rely on manual observation and recording, which is not only time-consuming and labor-intensive, but also susceptible to subjective judgment and operational errors, leading to inconsistent assessment results. Traditional methods lack sophisticated quantitative methods for motion capture and analysis, making it difficult to capture subtle movement changes and rapid movement sequences, limiting the depth of movement diagnosis and improvement. Traditional assessments also lack targeted training programs, making it impossible to provide improvement strategies for each athlete, resulting in inefficient training and limited technical improvement. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose an AI-based shot put assessment method and system.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: an AI-based shot put assessment method, comprising the following steps:

[0007] S1: Based on an athlete's shot-throwing video, the inter-frame difference method is used to calculate the pixel changes between consecutive video frames, identify the motion area, separate the foreground and background, and determine the position and movement of the athlete and the shot-throwing ball to generate motion area data;

[0008] S2: Based on the motion area data, using the optical flow method, analyze the movement pattern of each pixel, track the speed and direction changes of the motion sequence, and reconstruct the motion path based on the information to generate motion trajectory information;

[0009] S3: Based on the motion trajectory information, a temporal convolutional network is used to analyze the time series data, identify key action nodes in the motion sequence, and compare the nodes with standard action patterns to generate key recognition actions;

[0010] S4: Based on the key recognition actions, time series clustering is used to classify the time series of differentiated actions, and related action features and individual abnormal points are extracted to generate action pattern analysis results;

[0011] S5: Based on the movement pattern analysis results, a support vector machine is used in combination with the extracted movement features to rate the athlete's technical execution, and the movements are divided into multiple quality categories according to the ratings to generate a technical execution quality score;

[0012] S6: Based on the technical execution quality score, a decision tree is used to analyze and process the athlete's historical performance and current technical rating, and the athlete's technical evaluation is performed based on the analysis results to generate a comprehensive technical evaluation record;

[0013] S7: Based on the comprehensive technical evaluation records, a genetic algorithm is used to search and optimize the training plan according to the individual technical needs and abilities of the athletes, and the optimal training method is captured to generate a matching training plan.

[0014] As a further solution of the present invention, the motion area data includes the outline of the identified athlete's body parts, the outline of the solid ball, the position of the athlete and the solid ball, and the range of change of the motion area; the motion trajectory information includes the motion path coordinates, path shape, speed change curve and direction change index; the key identification actions include the preparation posture, leg movements, arm swings and release points of the ball during throwing; the action pattern analysis results include action pattern classification, athlete action deviation identification, and standard deviation of key action execution; the technical execution quality score includes technical action rating, scoring rules, and action score details; the comprehensive technical evaluation record includes the athlete's total technical level score, technical dimension score and technical improvement plan.

[0015] As a further embodiment of the present invention, based on a video of an athlete throwing a shot put, an inter-frame difference method is used to calculate pixel changes between consecutive video frames, identify motion regions, separate foreground and background, and determine the position and movement of the athlete and the shot put. The steps for generating motion region data are as follows:

[0016] S101: Based on a shot-throwing video of an athlete, a differential threshold method is used to compare pixels between two consecutive frames to distinguish between moving objects and static backgrounds, determine moving areas, and remove noise through threshold processing to generate foreground motion detection data.

[0017] S102: Based on the foreground motion detection data, a Gaussian mixture model method is used to perform hierarchical analysis on the motion region, and the athlete, the solid ball, and the background in the foreground are separated by statistical methods to refine the identification of the motion region and generate dynamic foreground separation data;

[0018] S103: Based on the dynamic foreground separation data, a template matching algorithm is used to search for areas similar to the defined template to determine the positions of the player and the shot, provide a tracking starting point, establish a motion path, and generate motion area data.

[0019] As a further solution of the present invention, based on the motion region data, an optical flow method is used to analyze the movement pattern of each pixel, track the speed and direction changes of the motion sequence, and reconstruct the motion path based on the information. The steps of generating motion trajectory information are specifically as follows:

[0020] S201: Based on the motion region data, using a dense optical flow method, analyzing each pixel in the scene, calculating the motion of the pixel between consecutive frames, identifying the motion direction and speed, and generating pixel-level motion information;

[0021] S202: Based on the pixel-level motion information, a vector field analysis method is used to aggregate the motion information of individual pixels into motion vectors. By establishing a vector field, support is provided for the continuity of the motion trajectory, and motion trend analysis is performed to generate motion vector analysis data.

[0022] S203: Based on the motion vector analysis data, a path reconstruction algorithm is used in combination with the motion vector to reconstruct the motion trajectories of the athlete and the shot put, and the integrity and matching of the trajectories are verified to generate motion trajectory information.

[0023] As a further solution of the present invention, based on the motion trajectory information, a temporal convolutional network is used to analyze the time series data, identify key action nodes in the motion sequence, and compare the nodes with the standard action pattern. The steps of generating key identified actions are as follows:

[0024] S301: Based on the motion trajectory information, a temporal convolutional network is used to analyze key actions and phase changes in the motion sequence, identify and extract key action features through network layers, and generate key action feature data;

[0025] S302: Based on the key action feature data, dynamic time warping is used to perform time series analysis between key action nodes and known standard action patterns, and similarities in the time series are calculated and aligned to identify deviations and consistencies in action execution, thereby generating action alignment and difference results.

[0026] S303: Based on the action alignment and difference results, a K-means clustering algorithm is used to analyze and classify the action execution patterns, aggregate similar action executions, distinguish normal from abnormal action patterns, and generate key recognition actions.

[0027] As a further solution of the present invention, based on the key recognition actions, time series clustering is used to classify the time series of differentiated actions, and the associated action features and individual abnormal points are extracted to generate the action pattern analysis results. Specifically, the steps are as follows:

[0028] S401: Based on the key recognition actions, principal component analysis is used to perform dimensionality reduction processing on the clustered action features, optimize the data dimension and retain the key action information, and generate feature data after dimensionality reduction;

[0029] S402: Based on the feature data after dimensionality reduction, a local anomaly factor algorithm is used to analyze the local density deviation of each action feature, identify abnormal data points compared with surrounding neighboring points during action execution, and generate an action abnormality index;

[0030] S403: Based on the action abnormality index, a support vector machine is used to classify and evaluate the execution quality of multiple actions, and the actions are divided into different quality levels according to the abnormality degree of the actions to generate action pattern analysis results.

[0031] As a further solution of the present invention, based on the results of the movement pattern analysis, a support vector machine is used in combination with the extracted movement features to rate the athlete's technical execution, and the movements are divided into multiple quality categories based on the ratings. The steps of generating the technical execution quality score are specifically as follows:

[0032] S501: Based on the motion pattern analysis results, a support vector machine is used to perform feature analysis on each motion, and a classification model is constructed and trained to distinguish various quality levels of the motion. The motion is classified according to the feature vectors and classification boundaries of the motion to generate motion quality classification data;

[0033] S502: Based on the movement quality classification data, a weighted average scoring method is used to evaluate the movement execution. By considering the frequency of each type of movement and its criticality in technical execution, corresponding weights are assigned, and a weighted score for each category is calculated to generate a weighted movement scoring result.

[0034] S503: Based on the weighted action scoring results, a comprehensive scoring method is used to combine the athlete's weighted score and personal historical performance to evaluate the technical ability, quantify the athlete's technical execution level, and generate a technical execution quality score.

[0035] As a further scheme of the present application, based on the technical execution quality score, a decision tree is used to analyze and process the historical performance and current technical rating of the athlete, and a technical evaluation of the athlete is made according to the analysis result, and a comprehensive technical evaluation record is generated, and the steps are specifically as follows:

[0036] S601: Based on the technical execution quality score, a decision tree algorithm is applied to analyze the technical performance and historical data of the athlete, a decision model is constructed to identify key factors affecting the technical performance, and key factor analysis data is generated;

[0037] S602: Based on the key factor analysis data, a correlation rule mining method is used to extract rules and patterns affecting technical improvement from the analysis results, including action improvement points or training methods, and improvement measures and strategy plans are generated;

[0038] S603: Based on the improvement measures and strategy plans, a comprehensive evaluation model is used to evaluate the technical level of the athlete by combining the existing technical score, historical performance and improvement measures of the athlete, and a comprehensive technical evaluation record is generated.

[0039] As a further scheme of the present application, based on the comprehensive technical evaluation record, a genetic algorithm is used to search and optimize the training scheme according to the individual technical needs and abilities of the athlete, and the optimal training method is captured, and a matching training scheme is generated, and the steps are specifically as follows:

[0040] S701: Based on the comprehensive technical evaluation record, a genetic algorithm is used to initialize the training scheme as an initial population, and the fitness of each scheme is calculated according to the technical needs and individual abilities of the athlete, and an initial training scheme set is generated;

[0041] S702: Based on the initial training scheme set, a genetic algorithm is used to capture the optimal scheme by using selection, crossover and mutation mechanisms, and the training scheme is optimized and adjusted, and an adjusted training scheme set is generated;

[0042] S703: Based on the adjusted training scheme set, the termination condition of the genetic algorithm is applied to perform screening processing, including reaching the target iteration number or the amplitude of fitness being less than the threshold value, selecting the key scheme, and generating a matching training scheme.

[0043] The AI-based solid ball examination system is used to perform the above-mentioned AI-based solid ball examination method, and the system includes a motion detection module, a trajectory analysis module, an action recognition module, an action analysis module, a technical rating module, and a scheme generation module.

[0044] The motion detection module uses an inter-frame difference method based on an athlete's shot-throwing video to calculate pixel changes between consecutive frames, uses binarization to distinguish moving areas from static backgrounds in the image, and performs morphological operations to optimize the representation of moving areas. It then uses a Gaussian mixture model to refine the separation of foreground and background to generate moving area data.

[0045] The trajectory analysis module uses dense optical flow method based on motion area data to analyze the movement of each pixel in the scene, uses vector field method to convert the movement information into motion vectors, and integrates the vectors through trajectory reconstruction algorithm to generate motion trajectory information;

[0046] The action recognition module uses a temporal convolutional network based on motion trajectory information to extract key action features from sequence data, adopts a dynamic time warping method to align and analyze the action features in time series, and applies a K-means clustering algorithm to classify and label the actions to generate key recognition actions.

[0047] The action analysis module is based on key recognition actions, performs dimensionality reduction processing on action feature data through principal component analysis, uses local anomaly factor algorithm to perform local anomaly detection on the reduced data, adopts support vector machine to classify the detection results into action quality, and generates action pattern analysis results;

[0048] The technical rating module uses a support vector machine to analyze and classify multiple action features based on the results of the action pattern analysis, quantifies the execution quality of each action through a weighted average scoring method, and uses a comprehensive scoring model based on the quantified results to evaluate the athlete's technical execution and generate a technical execution quality score;

[0049] The program generation module analyzes the athletes' technical performance and historical data based on the technical execution quality score and applies a decision tree. It uses association rule mining to extract optimization measures from the analysis results, and uses a genetic algorithm to search and optimize the training program according to the measures to generate a matching training program.

[0050] Compared with the prior art, the advantages and positive effects of the present invention are:

[0051] In this invention, by applying inter-frame differencing and optical flow methods, the dynamic positions of athletes and shot puts can be accurately identified and tracked, providing more accurate and continuous motion trajectory information. The use of temporal convolutional networks and time series clustering enables more accurate identification and classification of key movements, improving the accuracy and efficiency of motion analysis. The application of support vector machines and decision trees enables more objective and detailed evaluation of technical execution quality and comprehensive technical assessment. Genetic algorithms provide athletes with targeted, optimized training programs, improving training effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0053] Figure 2 This is a detailed flow chart of S1 of the present invention;

[0054] Figure 3 This is a detailed flow chart of S2 of the present invention;

[0055] Figure 4 This is a detailed flow chart of S3 of the present invention;

[0056] Figure 5 This is a detailed flow chart of S4 of the present invention;

[0057] Figure 6 This is a detailed flow chart of S5 of the present invention;

[0058] Figure 7 This is a detailed flow chart of S6 of the present invention;

[0059] Figure 8 This is a detailed flow chart of S7 of the present invention;

[0060] Figure 9 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0062] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0063] Example 1

[0064] See also Figure 1 The present invention provides a technical solution: an AI-based shot put assessment method, comprising the following steps:

[0065] S1: Based on an athlete's shot-throwing video, the inter-frame difference method is used to calculate the pixel changes between consecutive video frames, identify the motion area, separate the foreground and background, and determine the position and movement of the athlete and the shot-throwing ball to generate motion area data;

[0066] S2: Based on the motion area data, the optical flow method is used to analyze the movement pattern of each pixel, track the speed and direction changes of the motion sequence, and reconstruct the motion path based on the information to generate motion trajectory information;

[0067] S3: Based on motion trajectory information, a temporal convolutional network is used to analyze time series data, identify key action nodes in the motion sequence, and compare the nodes with standard action patterns to generate key recognition actions;

[0068] S4: Based on the key recognition actions, time series clustering is used to classify the time series of differentiated actions, and the associated action features and individual abnormal points are extracted to generate action pattern analysis results;

[0069] S5: Based on the results of the movement pattern analysis, support vector machines are used in combination with the extracted movement features to rate the athletes' technical execution. The movements are then divided into multiple quality categories based on the ratings to generate a technical execution quality score.

[0070] S6: Based on the technical execution quality score, a decision tree is used to analyze and process the athlete's historical performance and current technical rating. The athlete's technical evaluation is then evaluated based on the analysis results to generate a comprehensive technical evaluation record.

[0071] S7: Based on comprehensive technical evaluation records, a genetic algorithm is used to search and optimize training plans according to the individual technical needs and abilities of athletes, and to capture the optimal training methods to generate matching training plans.

[0072] The motion area data includes the outline of the identified athlete's body parts, the outline of the solid ball, the position of the athlete and the solid ball, and the range of change of the motion area. The motion trajectory information includes the motion path coordinates, path shape, speed change curve and direction change indicators. The key identification actions include the preparation posture, leg movements, arm swings and release points of the ball during throwing. The action pattern analysis results include action pattern classification, athlete action deviation identification, and standard deviation of key action execution. The technical execution quality score includes technical action rating, scoring rules, and action scoring details. The comprehensive technical evaluation record includes the athlete's total technical level score, technical dimension score and technical improvement plan.

[0073] In step S1, the shot put throwing video is processed using the inter-frame difference method, which involves pixel-by-pixel comparison between the two frames of video and calculation of the change in pixel brightness to identify the motion area. When the algorithm is executed, a threshold is first set to determine whether the pixel has changed; if the change in the corresponding pixel between the two frames exceeds the threshold, it is marked as a moving pixel. Subsequently, morphological operations such as dilation and erosion are used to remove noise, and connected region analysis is used to distinguish the foreground (athlete and shot put) from the background. Through this series of operations, the algorithm accurately locates the position of the athlete and shot put and their movement trajectory, thereby generating motion area data containing the outlines of the athlete's body parts, the outline of the shot put and their relative positions.

[0074] In step S2, the pixel movement pattern based on the motion area data is analyzed by the optical flow method. The optical flow method tracks the movement of each pixel by calculating the movement vector between consecutive frames. In the specific implementation, this method evaluates the position change of the pixel point in the time series and calculates the velocity vector of each pixel by minimizing the equation of the pixel brightness change. Then, using the velocity vector, the algorithm reconstructs the movement trajectory of the athlete and the solid ball, including the trajectory coordinates, shape, and changes in speed and direction. The generated motion trajectory information not only reflects the speed and direction of the action in detail, but also provides key data for evaluating technical execution.

[0075] In step S3, the motion trajectory information is processed by a temporal convolutional network. The temporal convolutional network is specially designed to analyze time series data and can identify and extract key features and action nodes in motion sequences. In this step, the network processes the motion trajectory through multiple layers of convolution and pooling layers to identify key action nodes, such as throwing preparation, leg movements, and arm swings. The nodes are then compared with predefined standard action patterns to determine the technical accuracy of the athlete's execution. The generated key identified action information provides coaches and athletes with clear feedback, indicating which actions need improvement.

[0076] In step S4, the motion pattern analysis is refined through the time series clustering method. This method divides the athlete's motion time series into different categories by calculating and comparing the similarities between different motion sequences. In the specific implementation, the algorithm first calculates the characteristics of each time series, such as average speed, acceleration, and motion amplitude, and then uses the characteristics to classify the movements into different categories. This process helps to identify individual anomalies in the athlete's movements and their deviations from the standard motion pattern. The generated motion pattern analysis results not only reveal the athlete's technical characteristics, but also provide a basis for improving the athlete's movements.

[0077] In step S5, the athletes' technical execution is rated using a support vector machine. Support vector machine is a supervised learning algorithm used for classification and regression analysis. In this step, the algorithm trains a model based on the movement features extracted from the previous steps, such as movement speed, accuracy, and coherence. The model is able to classify technical execution into different quality categories. By defining boundaries and calculating the maximum interval between different categories, the algorithm ensures a high degree of classification accuracy. The generated technical execution quality score provides a clear indicator of the technical level of each athlete.

[0078] In step S6, technical evaluation is performed through decision tree analysis. The decision tree is a tree-structured algorithm used for decision analysis and prediction. In this step, the algorithm combines the athlete's historical performance data and current technical execution score, and analyzes through a series of conditional branches to conduct a comprehensive technical evaluation of the athlete. The decision tree provides guidance and suggestions for the athlete's technical improvement and training by analyzing the relationship between different technical indicators and performance results. The generated comprehensive technical evaluation record records the athlete's technical level and room for improvement in detail, providing coaches and athletes with a specific improvement plan.

[0079] In step S7, the training plan is optimized through a genetic algorithm. The genetic algorithm is a search algorithm used to solve optimization and search problems. In this step, the algorithm simulates natural selection and genetic mechanisms, such as selection, crossover, and mutation, to search for and optimize training plans tailored to the athlete's individual technical needs. In specific implementation, the algorithm is based on the athlete's technical evaluation records, iteratively generates multiple training plans, evaluates their effectiveness, and selects the optimal solution. The generated training plan is designed to provide training suggestions that best suit the athlete's current technical status and improvement needs.

[0080] See also Figure 2 Based on the athlete's shot-throwing video, the inter-frame difference method is used to calculate the pixel changes between consecutive video frames, identify the motion area, separate the foreground and background, and determine the position and movement of the athlete and the shot-throwing ball. The specific steps for generating motion area data are as follows:

[0081] S101: Based on a shot-throwing video of an athlete, a differential threshold method is used to compare pixels between two consecutive frames to distinguish between moving objects and static backgrounds, determine moving areas, and remove noise through threshold processing to generate foreground motion detection data.

[0082] S102: Based on the foreground motion detection data, a Gaussian mixture model method is used to perform hierarchical analysis on the motion region. Through statistical methods, the athlete, the solid ball, and the background in the foreground are separated, the identification of the motion region is refined, and dynamic foreground separation data is generated.

[0083] S103: Based on the dynamic foreground separation data, a template matching algorithm is used to search for areas similar to the defined template to determine the positions of the athlete and the shot, provide a tracking starting point, establish a motion path, and generate motion area data.

[0084] In sub-step S101, the athlete's shot put video is processed by the differential threshold method. First, two consecutive frames of video images are collected and the pixels are compared one by one. In specific implementation, for each pair of adjacent frames, the difference between the corresponding pixels is calculated. If the difference exceeds the preset threshold, it is identified as a moving pixel; otherwise, it is regarded as the background. The setting of this threshold is based on experimental data and aims to balance the sensitivity and false alarm rate of motion detection. It is usually determined through iterative testing. The differential method can effectively distinguish between static background and moving objects, but is easily interfered by noise. To solve this problem, threshold processing is introduced. By setting an appropriate threshold, areas with small pixel differences, i.e., noise, are filtered out, thereby improving the accuracy of motion detection. The foreground motion detection data generated by this process is a binary image, in which the white part represents the moving area and the black part represents the static background. This data format is convenient for subsequent processing and provides a basis for motion analysis in dynamic scenes.

[0085] In the S102 sub-step, the motion region recognition is refined by the Gaussian Mixture Model method, which is based on statistical principles to analyze the motion region in multiple levels, so as to distinguish the athletes, solid balls and background. Specifically, the Gaussian Mixture Model (GMM) simulates the color distribution in the image by assigning different Gaussian distributions to each pixel point, and then identifies and separates different moving foreground objects. In the algorithm implementation, the observation value of each pixel point is updated using the Maximum Likelihood Estimation (MLE) to update the parameters of the Gaussian distribution, including the mean, variance and mixing coefficient, to best match the data distribution. By comparing the matching degree of each pixel value with each Gaussian distribution, it is determined which category it belongs to. Through iteration of this process, the algorithm can gradually refine the separation of athletes and solid balls in the foreground, and the generated dynamic foreground separation data is displayed in the form of layered images.

[0086] In the S103 sub-step, the template matching algorithm is used to further accurately determine the positions of the athletes and the solid balls. Based on the dynamic foreground separation data, the algorithm is implemented by searching for the region matching the predefined template. In the template matching process, the feature templates of the athletes and the solid balls are defined, and each region in the dynamic foreground separation data is scanned to calculate the similarity between it and the template. The similarity calculation usually uses the correlation coefficient, Euclidean distance or other statistical measurement methods to quantify the matching degree. The region with the highest matching degree is identified as the position of the target object. This process not only provides the starting point for motion tracking, but also establishes the motion path through the position changes between consecutive frames. The generated motion region data is stored in the form of path vectors, which records the motion trajectories of the athletes and the solid balls in the video sequence in detail. This data plays an important role in analyzing the motion technology and evaluating the motion performance, and provides a quantitative basis for subsequent motion evaluation and improvement.

[0087] Please refer to Figure 3 Based on the motion region data, the optical flow method is used to analyze the movement pattern of each pixel, track the speed and direction changes of the motion sequence, and reconstruct the motion path according to the information to generate the motion trajectory information. The steps are as follows:

[0088] S201: Based on the motion region data, the dense optical flow method is used to analyze each pixel point in the scene, identify the motion direction and speed by calculating the motion between consecutive frames, and generate pixel-level motion information;

[0089] S202: Based on the pixel-level motion information, the vector field analysis method is used to summarize the motion information of individual pixel points into motion vectors, establish a vector field to support the continuity of the motion trajectory, and perform motion trend analysis to generate motion vector analysis data;

[0090] S203: Based on the motion vector analysis data, a path reconstruction algorithm is used in combination with the motion vector to reconstruct the motion trajectories of the athlete and the shot put, and the integrity and matching of the trajectories are verified to generate motion trajectory information.

[0091] In sub-step S201, the motion of each pixel in the scene between consecutive frames is analyzed by the dense optical flow method. This method can identify the motion direction and speed of the pixel and generate detailed motion information for each pixel. The specific operation process involves calculating the pixel displacement between two consecutive frames and estimating the velocity vector of each pixel by evaluating the time change of the pixel intensity. The dense optical flow method is usually implemented by using the optical flow equation, combined with gradient descent or other optimization algorithms, to solve the motion vector of each pixel. In this process, the algorithm takes into account the intensity changes of the pixel in space and time to ensure the accuracy of the motion information. The generated pixel-level motion information is represented in the form of a vector, which includes the motion direction and speed of all pixels in the scene. This data format facilitates subsequent analysis and processing.

[0092] In sub-step S202, the pixel-level motion information is further processed using the vector field analysis method, the motion information of a single pixel is aggregated into a motion vector, and a vector field is established to support the continuity analysis and motion trend analysis of the motion trajectory. In this process, the algorithm constructs a vector field for the entire scene by analyzing the motion vector of each pixel. The vector field reflects the motion trend and pattern of the moving objects in the entire scene. The establishment of the vector field takes into account the direction and size of the pixel motion vector, and describes the overall distribution and change trend of the motion through a mathematical model. This method can not only reveal the global characteristics of the motion, but also identify local variations in the motion pattern, providing an effective tool for in-depth analysis of the motion trajectory and prediction of the motion trend. The generated motion vector analysis data is recorded in the form of graphs or data tables, showing the motion trend and the distribution of motion vectors, which is of great value for understanding the motion behavior in dynamic scenes and guiding subsequent motion analysis.

[0093] In sub-step S203, the motion trajectories of the athlete and the shot put are reconstructed by combining the motion vector analysis data with a path reconstruction algorithm, and the integrity and matching of the trajectories are verified. The path reconstruction algorithm simulates and reconstructs the motion trajectory using a mathematical model by comprehensively considering the direction and speed of the motion vector. During this process, the algorithm integrates and analyzes the collected motion vectors, identifies key motion nodes, and reconstructs a coherent motion trajectory through the nodes. The reconstructed trajectory is optimized to ensure that it matches the actual observed motion trajectory. At the same time, the integrity and accuracy of the data are taken into account to ensure that the trajectory reflects the actual motion pattern of the moving object. The generated motion trajectory information is displayed in digital or graphical form, which records in detail the movement path of the moving object in the entire scene, providing accurate basic data for further motion analysis, performance evaluation and technical improvement.

[0094] See also Figure 4 Based on the motion trajectory information, a temporal convolutional network is used to analyze the time series data, identify the key action nodes in the motion sequence, and compare the nodes with the standard action patterns. The specific steps for generating key recognition actions are as follows:

[0095] S301: Based on the motion trajectory information, a temporal convolutional network is used to analyze the key actions and phase changes in the motion sequence, identify and extract key action features through the network layer, and generate key action feature data;

[0096] S302: Based on the key action feature data, dynamic time warping is used to perform time series analysis between key action nodes and known standard action patterns. By calculating and aligning the similarities in the time series, deviations and consistencies in action execution are identified, and action alignment and difference results are generated.

[0097] S303: Based on the action alignment and difference results, the K-means clustering algorithm is used to analyze and classify the action execution patterns, aggregate similar action executions, distinguish normal from abnormal action patterns, and generate key recognition actions.

[0098] In sub-step S301, the key actions and phase changes in the motion sequence are analyzed through the temporal convolutional network (TCN). The temporal convolutional network is a deep learning model designed specifically for processing time series data and has the ability to capture long-term dependencies. TCN processes the input motion trajectory information through stacked convolutional layers. Each convolution layer captures the features of different time scales in the time series through an expanded receptive field, effectively identifying and extracting key action features. In this process, the parameters of the convolutional layer, such as the convolution kernel size, step size, and padding, are carefully designed to ensure that the model can adapt to the dynamic changes of the motion sequence. Through forward propagation and backpropagation of the network layer, the key action feature data learned by the model is represented in vector form, which contains the key features and phase change information of the actions in the time series. The generated key action feature data provides an accurate basis for subsequent action analysis and can reveal the detailed features of the athletes when performing techniques.

[0099] In sub-step S302, dynamic time warping (DTW) is used to perform time series analysis between key action nodes and known standard action patterns. DTW is an algorithm that calculates and aligns the similarity between two time series. It finds the best match between the two sequences by warping the time series. During the analysis process, the algorithm calculates the distance between each pair of time points and constructs a cumulative distance matrix to find the path with the lowest cost, that is, the best alignment between the two sequences. In this way, DTW can identify deviations and consistencies in movement execution, especially in the case of time expansion and contraction changes. The generated movement alignment and difference results record in detail the time alignment and differences between the athlete's movement execution and the standard movement pattern.

[0100] In sub-step S303, the K-means clustering algorithm is applied to analyze and classify the action execution patterns. K-means clustering is a partitioning clustering algorithm that minimizes the distance between each point and the center of its cluster by assigning data points to K clusters. At the same time, the data points within the cluster are as close as possible, and the data points between clusters are as dispersed as possible. When processing action alignment and difference results, the algorithm first determines the number of clusters K, and then randomly selects K initial centers. The distance from each data point to each cluster center is calculated based on the characteristic data of the action execution, and the clusters are reallocated according to the principle of the nearest distance. The cluster centers are iteratively updated until the clustering results are stable. The generated key recognition action data is presented in the form of clustering results, which clearly distinguishes normal and abnormal action patterns.

[0101] See also Figure 5 ,Based on the key recognition actions, time series clustering is used to classify the time series of ,differentiated actions, and extract the associated action features and ,individual anomalies.,The steps to generate the action pattern analysis results are as follows:

[0102] S401: Based on the key recognition actions, principal component analysis is used to perform dimensionality reduction processing on the clustered action features, optimize the data dimension and retain the key action information, and generate feature data after dimensionality reduction;

[0103] S402: Based on the feature data after dimensionality reduction, a local anomaly factor algorithm is used to analyze the local density deviation of each action feature, identify abnormal data points compared with surrounding neighboring points during action execution, and generate an action abnormality index;

[0104] S403: Based on the action abnormality index, a support vector machine is used to classify and evaluate the execution quality of multiple actions, and the actions are divided into different quality levels according to the abnormality degree of the actions to generate action pattern analysis results.

[0105] In sub-step S401, principal component analysis (PCA) is used to reduce the dimensionality of the clustered action features, aiming to optimize the data dimension while retaining key action information. Principal component analysis is a statistical method used to convert high-dimensional data sets into lower-dimensional forms, and to achieve dimensionality reduction by extracting the most important features in the data. In the specific execution process, the algorithm first calculates the covariance matrix of the data set, and then finds the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvectors represent the main direction of change in the data set, while the eigenvalues ​​indicate the importance of each principal component. By selecting the eigenvector with the largest eigenvalue, PCA can retain the most important action features and effectively reduce the data dimension. The generated reduced-dimensional feature data is represented in matrix form, where each row corresponds to an action instance and each column represents a principal component. This data format simplifies the complexity of subsequent processing while ensuring the retention of key action information, providing efficient data input for action recognition and analysis.

[0106] In sub-step S402, the local outlier factor (LOF) algorithm is used to analyze the local density deviation of each action feature and identify abnormal data points in the action execution. The local outlier factor algorithm is a density-based anomaly detection method. By comparing the local density of a given point with its neighboring points, the degree of abnormality of the point is evaluated. During the execution process, the algorithm first calculates the distance between each point and its nearest neighbor, and then estimates the local density of each point. Then, by comparing the local density of a point with the local density of its neighbors, the LOF score is calculated. The higher the score, the greater the density difference between the point and its surrounding neighbors, that is, the more potential anomaly point. The generated action abnormality index is expressed in vector form, and each element corresponds to the abnormality score of an action instance. The analysis not only reveals the abnormal pattern in the action execution, but also provides a basis for the evaluation and optimization of action quality.

[0107] In the S403 sub-step, the execution quality of multiple actions is classified and evaluated by a support vector machine (SVM), and different quality levels are divided according to the abnormality of the action. The support vector machine is a supervised learning algorithm used to solve classification and regression problems. It separates different categories of data by finding an optimal hyperplane. In this process, the algorithm uses action abnormality indicators as input features to classify action instances into different quality levels. SVM determines the separation hyperplane by maximizing the separation between classes. The optimization process involves kernel tricks to handle non-linearly separable cases. The generated action pattern analysis results provide detailed quality classification for each action instance, providing quantitative action quality evaluation for coaches and athletes to guide the improvement of sports technology and the development of training plans.

[0108] Please refer to Figure 6 Based on the action pattern analysis results, the support vector machine is used to rate the technical execution of athletes combined with the extracted action features, and the action is divided into multiple quality categories according to the rating. The steps to generate the technical execution quality score are as follows:

[0109] S501: Based on the action pattern analysis results, the support vector machine is used to analyze the features of each action. By constructing and training a classification model, the multiple quality levels of the action are distinguished, and the action is classified according to the action feature vector and classification boundary to generate action quality classification data.

[0110] S502: Based on the action quality classification data, the weighted average scoring method is used to evaluate the action execution. By considering the frequency of each type of action and its key role in technical execution, the corresponding weight is assigned, and the weighted score of each category is calculated to generate the weighted action score result.

[0111] S503: Based on the weighted action score result, the comprehensive scoring method is used to evaluate the technical ability of athletes combined with their weighted score and historical performance, and to quantitatively evaluate the technical execution level of athletes to generate the technical execution quality score.

[0112] In sub-step S501, the support vector machine (SVM) algorithm is used to perform feature analysis on each action. The SVM algorithm constructs and trains a classification model, uses the action's feature vector and pre-labeled quality level data, and finds the optimal classification boundary between different action quality levels. Specifically, SVM searches for a hyperplane in high-dimensional space, uses this hyperplane as the decision boundary, and assigns data points to different categories according to their feature vectors. During the training process, the original feature space is mapped to a higher-dimensional space through the kernel technique to handle nonlinear separable situations. After the model training is completed, the new action data is classified, and the quality level of the action is determined based on its feature vector. The generated action quality classification data contains the identification label of each action and its quality level, providing a basis for subsequent evaluation and analysis.

[0113] In sub-step S502, a weighted average scoring method is used to evaluate the execution of movements. Based on the movement quality classification data, this method considers the frequency and criticality of each type of movement in technical execution and assigns corresponding weights. The weight allocation is based on the importance of the movement in the overall technical system and its impact on the athlete's performance. The weighted score of each category is obtained by calculating the sum of the scores of all movements in the category multiplied by their weights. The weighted movement scoring results generated by this process record the scoring of different movement categories in detail, provide a quantitative assessment of the quality of athletes' movement execution, and facilitate coaches and athletes to understand the advantages and improvement points in technical execution.

[0114] In sub-step S503, a comprehensive scoring method is used to evaluate technical ability. Combining the athlete's weighted movement score and personal historical performance, the athlete's performance in different movement categories is comprehensively analyzed to calculate the athlete's technical execution quality score. This method not only considers the athlete's current movement execution quality, but also takes their historical performance as part of the evaluation to ensure the comprehensiveness and accuracy of the scoring results. The generated technical execution quality score is expressed in numerical form, quantifying the athlete's technical level. It plays an important role in guiding the formulation of training plans, monitoring athletes' technical progress, and adjusting training strategies. Through this evaluation process, coaches and athletes can more clearly understand the overall level of technical execution and the specific needs of each aspect, providing a scientific basis for improving technical performance and formulating training plans.

[0115] See also Figure 7 Based on the technical execution quality score, a decision tree is used to analyze and process the athlete's historical performance and current technical rating. The athlete's technical evaluation is then evaluated based on the analysis results. The steps for generating a comprehensive technical evaluation record are as follows:

[0116] S601: Based on the technical execution quality score, apply the decision tree algorithm to analyze the athlete's technical performance and historical data, build a decision model, identify the key factors affecting technical performance, and generate key factor analysis data;

[0117] S602: Analyze data based on key factors and apply association rule mining methods to extract rules and patterns that influence technology improvement from the analysis results, including action improvement points or training methods, and generate improvement measures and strategic plans;

[0118] S603: Based on the improvement measures and strategic plans, a comprehensive evaluation model is used to evaluate the athlete's technical level, combining the athlete's existing technical scores, historical performance and improvement measures, and generate a comprehensive technical evaluation record.

[0119] In sub-step S601, a decision tree algorithm is applied to analyze the athletes' technical performance and historical data. The key factors affecting technical performance are identified through the constructed decision model. The decision tree algorithm analyzes the relationship between the technical execution quality score and the athletes' historical performance data, gradually splits the data set to form a tree structure. Each node represents a decision rule, and the rules together reveal the factors affecting technical performance. In the process of constructing the decision tree, information gain or Gini impurity is selected as the splitting criterion to maximize the purity improvement of the subset after each split. The process is iterated until the preset depth is reached or the amount of data contained in the node is less than the threshold, and the model is constructed. The generated key factor analysis data is presented in the form of a tree diagram, which details the decisive factors behind technical performance at each level, providing coaches and athletes with a basis for targeted improvement of technical performance.

[0120] In sub-step S602, the association rule mining method is used to conduct an in-depth analysis of the key factor analysis data to extract the rules and patterns that affect technical improvement. Association rule mining analyzes the relationship between technical performance and various influencing factors, identifies frequently occurring factor combinations and the strong correlations between them, including calculating the support and confidence of item sets. The support indicates the frequency of item sets in the data set, while the confidence measures the reliability of a rule. By setting thresholds, rules with high support and high confidence are screened out to ensure that the mined rules have statistical significance and practical application value. The generated improvement measures and strategic plans record the action improvement points or training methods in detail, providing athletes with practical and feasible technical improvement strategies aimed at optimizing training effects and improving technical performance.

[0121] In sub-step S603, a comprehensive evaluation model is used to comprehensively evaluate the athlete's technical level by combining the athlete's existing technical score, historical performance and improvement measures. The comprehensive evaluation model analyzes various aspects of the athlete's technical execution and comprehensively considers multiple factors that affect the technical level, including technical scores, historical performance trends and the potential impact of improvement measures. When constructing the model, a weighted method is used to integrate different factors, and the weight allocation is based on the importance of each factor's impact on the technical level. Through this method, a quantitative and comprehensive score that reflects the athlete's technical level can be obtained. The score not only reflects the athlete's current technical status, but also takes into account the potential for future improvement. The generated comprehensive technical evaluation record is presented in the form of a report, which provides athletes and coaches with a comprehensive technical level analysis and evaluation, and guides the athlete's training plan formulation and technical improvement strategy.

[0122] See also Figure 8 Based on comprehensive technical evaluation records, a genetic algorithm is used to search and optimize training programs according to the individual technical needs and abilities of athletes, and to capture the optimal training methods. The specific steps for generating a matching training program are as follows:

[0123] S701: Based on the comprehensive technical evaluation records, a genetic algorithm is used to initialize the training program as the initial population, and the fitness of each program is calculated according to the athlete's technical requirements and personal ability to generate an initial training program set;

[0124] S702: Based on the initial training solution set, using a genetic algorithm, applying selection, crossover, and mutation mechanisms, capturing the optimal solution, and optimizing and adjusting the training solution to generate an adjusted training solution set;

[0125] S703: Based on the adjusted training scheme set, the termination condition of the genetic algorithm is applied to perform screening processing, including reaching the target number of iterations or the fitness amplitude being less than a threshold, selecting key schemes, and generating matching training schemes.

[0126] In sub-step S701, the training plan is initialized as the initial population through the genetic algorithm. The genetic algorithm is a heuristic search algorithm that simulates natural selection and genetic principles and is used to solve optimization problems. The initial training plan set consists of multiple training plans. Each plan is designed according to the athlete's technical needs and personal abilities and is encoded as a "chromosome". The chromosome represents different parameters of the training plan, such as training frequency, intensity, and type of technical exercises. The fitness of each plan is calculated, that is, the potential effect of the plan on the athlete's technical improvement is evaluated. The fitness evaluation is based on the comprehensiveness, adaptability, and satisfaction of the technical needs of the plan. Plans with high fitness are more likely to be selected for subsequent genetic operations. The generated initial training plan set provides a variety of training options and provides a basis for further optimization.

[0127] In sub-step S702, the selection, crossover, and mutation mechanisms of the genetic algorithm are used to optimize and adjust the training scheme. During the selection process, the better scheme is selected as the parent according to the fitness of the scheme. Strategies such as the roulette wheel method or tournament selection are used to ensure that the scheme with high fitness is selected first. The crossover operation simulates chromosome crossover in biological genetics. A new training scheme is generated by combining chromosome fragments of two parent schemes to increase the diversity of the population. The mutation operation introduces new features by randomly changing certain genes (training parameters) in the chromosomes to prevent the algorithm from converging to the local optimal solution too early. Through this series of operations, an adjusted training scheme set is generated. This set reflects the training scheme optimized by genetic operations and aims to provide a more personalized and effective training plan.

[0128] In sub-step S703, the termination conditions of the genetic algorithm are applied for screening and a matching training plan is selected. The termination conditions include reaching a predetermined target number of iterations or the fitness improvement of the best plan in the population being less than a specific threshold, which means that the algorithm has stabilized and further iterations are unlikely to produce a better solution. When the termination conditions are met, the training plan with the highest fitness is selected from the current population as the final plan. This process ensures that the selected training plan can not only meet the needs of the athletes but also has a high optimization effect. The generated matching training plan provides athletes with a scientific and systematic training plan that is optimized based on the athletes' existing technical level, technical needs and personal abilities, and aims to maximize training efficiency and the speed of technical improvement.

[0129] See also Figure 9 , AI-based shot put assessment system, the AI-based shot put assessment system is used to implement the above-mentioned AI-based shot put assessment method, the system includes a motion detection module, a trajectory analysis module, an action recognition module, an action analysis module, a technology rating module, and a solution generation module;

[0130] The motion detection module uses inter-frame difference analysis to calculate pixel changes between consecutive frames based on a shot-throwing video. It then uses binarization to distinguish moving areas from static backgrounds in the image. Morphological operations are then performed to optimize the representation of moving areas. A Gaussian mixture model is then used to refine the separation between foreground and background, generating moving area data.

[0131] The trajectory analysis module uses dense optical flow method based on motion area data to analyze the movement of each pixel in the scene. It uses the vector field method to convert the movement information into motion vectors, and integrates the vectors through the trajectory reconstruction algorithm to generate motion trajectory information.

[0132] The action recognition module uses a temporal convolutional network based on motion trajectory information to extract key action features from sequence data. It then uses a dynamic time warping method to align and analyze the action features in time series. It then applies a K-means clustering algorithm to classify and label the actions and generate key recognition actions.

[0133] The motion analysis module is based on key recognition actions. It uses principal component analysis to reduce the dimension of motion feature data. It uses the local anomaly factor algorithm to detect local anomalies on the reduced data. It uses support vector machines to classify the detection results into motion quality and generate motion pattern analysis results.

[0134] The technical rating module uses support vector machines to analyze and classify multiple movement features based on the results of movement pattern analysis. It then quantifies the execution quality of each movement using a weighted average scoring method. Combining the quantified results, it uses a comprehensive scoring model to evaluate the athlete's technical execution and generate a technical execution quality score.

[0135] The program generation module uses a decision tree to analyze the athletes' technical performance and historical data based on the technical execution quality score, uses association rule mining to extract optimization measures from the analysis results, and uses genetic algorithms to search and optimize training programs based on the measures to generate matching training programs.

[0136] The motion detection module identifies the athlete's motion area during the shot-throwing process, effectively distinguishing between motion and static backgrounds, and provides accurate input data for subsequent trajectory analysis and action recognition. This process not only improves the accuracy of motion analysis, but also provides athletes and coaches with intuitive feedback on motion execution.

[0137] The combined use of the trajectory analysis module and the action recognition module can deeply analyze the athlete's action execution trajectory and key action characteristics, and then accurately identify deviations and deficiencies in technical execution through action alignment and classification. This function plays an important role in guiding athletes to make targeted technical improvements and adjustments, making training more targeted and efficient.

[0138] The motion analysis module further uses advanced algorithms to conduct in-depth analysis of motion execution and identify abnormal patterns in the movements, thereby providing athletes with more detailed and specific technical adjustment suggestions. This process helps athletes optimize their techniques in detail and improve the accuracy and quality of technical execution.

[0139] The application of the technical rating module provides athletes with a clear technical level positioning through quantitative evaluation of athletes' technical execution, helps athletes and coaches to objectively understand the athletes' technical level, and provides a scientific basis for formulating training plans and technical improvement goals.

[0140] The program generation module uses algorithms to generate training plans by comprehensively considering the athletes' technical needs and historical performance. This function not only improves the scientific nature and targetedness of the training plan, but also greatly improves the training efficiency and effectiveness, helping athletes achieve significant improvements in their technical level in a relatively short period of time.

[0141] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The AI-based shot put assessment method is characterized by: The following steps are involved: Based on the athlete's shot-throwing video, the inter-frame difference method is used to calculate the pixel changes between consecutive video frames, identify the motion area, separate the foreground and background, and determine the position and movement of the athlete and the shot-throwing ball to generate motion area data; Based on the motion area data, an optical flow method is used to analyze the movement pattern of each pixel, track the speed and direction changes of the motion sequence, and reconstruct the motion path based on the information to generate motion trajectory information; Based on the motion trajectory information, a temporal convolutional network is used to analyze the time series data, identify key action nodes in the motion sequence, and compare the nodes with the standard action pattern to generate key recognition actions, specifically: Based on the motion trajectory information, a temporal convolutional network is used to analyze key actions and phase changes in the motion sequence, identify and extract key action features through network layers, and generate key action feature data; Based on the key action feature data, dynamic time warping is used to perform time series analysis between key action nodes and known standard action patterns, and by calculating and aligning the similarities in the time series, deviations and consistencies in action execution are identified, generating action alignment and difference results. Based on the action alignment and difference results, a K-means clustering algorithm is used to analyze and classify the action execution patterns, aggregate similar action executions, distinguish normal and abnormal action patterns, and generate key recognition actions; Based on the key recognition actions, time series clustering is used to classify the time series of differentiated actions, and related action features and individual abnormal points are extracted to generate action pattern analysis results; Based on the movement pattern analysis results, a support vector machine is used in combination with the extracted movement features to rate the athletes' technical executions, and the movements are divided into multiple quality categories according to the ratings to generate a technical execution quality score; Based on the technical execution quality score, a decision tree is used to analyze and process the athlete's historical performance and current technical rating. Based on the analysis results, the athlete's technical evaluation is performed to generate a comprehensive technical evaluation record, specifically: Based on the technical execution quality score, a decision tree algorithm is applied to analyze the athlete's technical performance and historical data, and by building a decision model, key factors affecting technical performance are identified and key factor analysis data is generated; Analyze data based on the key factors and use association rule mining methods to extract rules and patterns that affect technology improvement from the analysis results, including action improvement points or training methods, and generate improvement measures and strategic plans; Based on the improvement measures and strategic plans, a comprehensive evaluation model is used to evaluate the athlete's technical level, combining the athlete's existing technical scores, historical performance, and improvement measures, to generate a comprehensive technical evaluation record; Based on the comprehensive technical evaluation records, a genetic algorithm is used to search and optimize training plans according to the athletes' personal technical needs and abilities, and to capture the optimal training methods to generate matching training plans.

2. The AI-based shot put assessment method according to claim 1, characterized in that: The motion area data includes the outline of the identified athlete's body parts, the outline of the solid ball, the position of the athlete and the solid ball, and the range of change of the motion area. The motion trajectory information includes the motion path coordinates, path shape, speed change curve and direction change index. The key identification actions include the preparation posture, leg movements, arm swings and release points of the ball during throwing. The action pattern analysis results include action pattern classification, athlete action deviation identification, and standard deviation of key action execution. The technical execution quality score includes technical action rating, scoring rules, and action score details. The comprehensive technical evaluation record includes the athlete's total technical level score, technical dimension score and technical improvement plan.

3. The AI-based shot put assessment method according to claim 1, characterized in that: Based on an athlete's shot-throwing video, we use the inter-frame difference method to calculate the pixel changes between consecutive video frames, identify motion regions, separate foreground and background, and determine the position and movement of the athlete and shot. The steps for generating motion region data are as follows: Based on a shot-throwing video of an athlete, the differential threshold method is used to compare pixels between two consecutive frames to distinguish between moving objects and static backgrounds, determine the moving area, and remove noise through threshold processing to generate foreground motion detection data. Based on the foreground motion detection data, a Gaussian mixture model method is used to perform hierarchical analysis on the motion area, and the athlete, the solid ball and the background in the foreground are separated by statistical methods to refine the identification of the motion area and generate dynamic foreground separation data; Based on the dynamic foreground separation data, a template matching algorithm is used to search for areas similar to the defined template to determine the positions of the players and the shot, provide a tracking starting point, establish a motion path, and generate motion area data.

4. The AI-based shot put assessment method according to claim 1, characterized in that: Based on the motion area data, the optical flow method is used to analyze the movement pattern of each pixel, track the speed and direction changes of the motion sequence, and reconstruct the motion path based on the information. The steps of generating motion trajectory information are as follows: Based on the motion region data, a dense optical flow method is used to analyze each pixel in the scene, and the motion direction and speed are identified by calculating the motion of the pixel between consecutive frames to generate pixel-level motion information; Based on the pixel-level motion information, a vector field analysis method is used to aggregate the motion information of individual pixels into motion vectors. By establishing a vector field, support is provided for the continuity of the motion trajectory, and motion trend analysis is performed to generate motion vector analysis data. Based on the motion vector analysis data, a path reconstruction algorithm is used in combination with the motion vectors to reconstruct the motion trajectories of the athlete and the solid ball, and the integrity and matching of the trajectories are verified to generate motion trajectory information.

5. The AI-based shot put assessment method according to claim 1, characterized in that: Based on the key recognition actions, time series clustering is used to classify the time series of differentiated actions, and the associated action features and individual abnormal points are extracted to generate the action pattern analysis results. The specific steps are as follows: Based on the key recognition actions, principal component analysis is used to perform dimensionality reduction processing on the clustered action features, optimize the data dimension and retain the key action information, and generate feature data after dimensionality reduction; Based on the feature data after dimensionality reduction, a local anomaly factor algorithm is used to analyze the local density deviation of each action feature, identify abnormal data points compared with surrounding neighboring points during action execution, and generate an action abnormality index; Based on the action abnormality index, a support vector machine is used to classify and evaluate the execution quality of multiple actions, and the actions are divided into different quality levels according to the abnormality degree of the actions to generate action pattern analysis results.

6. The AI-based shot put assessment method according to claim 1, characterized in that: Based on the results of the movement pattern analysis, a support vector machine is used in combination with the extracted movement features to rate the athletes' technical execution. The movements are then divided into multiple quality categories based on the ratings. The specific steps for generating the technical execution quality score are as follows: Based on the motion pattern analysis results, a support vector machine is used to perform feature analysis on each motion, and a classification model is constructed and trained to distinguish various quality levels of the motion. The motion is classified according to the feature vectors and classification boundaries of the motion to generate motion quality classification data; Based on the movement quality classification data, a weighted average scoring method is used to evaluate the execution of the movement. By considering the frequency of each type of movement and its criticality in technical execution, corresponding weights are assigned, and a weighted score for each category is calculated to generate a weighted movement score result; Based on the weighted action scoring results, a comprehensive scoring method is used to combine the athlete's weighted score and personal historical performance to evaluate technical ability, quantify the athlete's technical execution level, and generate a technical execution quality score.

7. The AI-based shot put assessment method according to claim 1, characterized in that: Based on the comprehensive technical evaluation records, a genetic algorithm is used to search and optimize training programs according to the individual technical needs and abilities of athletes, and the optimal training method is captured. The specific steps for generating a matching training program are as follows: Based on the comprehensive technical evaluation records, a genetic algorithm is used to initialize the training plans as an initial population, and fitness calculation is performed on each plan according to the technical requirements and personal abilities of the athletes to generate an initial training plan set; Based on the initial training scheme set, using a genetic algorithm, applying selection, crossover and mutation mechanisms, capturing the optimal scheme, and optimizing and adjusting the training scheme to generate an adjusted training scheme set; Based on the adjusted training scheme set, the termination condition of the genetic algorithm is applied to perform screening processing, including reaching the target number of iterations or the fitness amplitude being less than a threshold, selecting key schemes, and generating matching training schemes.

8. The AI-based shot put assessment system is characterized by: According to the AI-based shot put assessment method according to any one of claims 1 to 7, the system includes a motion detection module, a trajectory analysis module, an action recognition module, an action analysis module, a technical rating module, and a solution generation module; The motion detection module uses an inter-frame difference method based on an athlete's shot-throwing video to calculate pixel changes between consecutive frames, uses binarization to distinguish moving areas from static backgrounds in the image, and performs morphological operations to optimize the representation of moving areas. It then uses a Gaussian mixture model to refine the separation of foreground and background to generate moving area data. The trajectory analysis module uses dense optical flow method based on motion area data to analyze the movement of each pixel in the scene, uses vector field method to convert the movement information into motion vectors, and integrates the vectors through trajectory reconstruction algorithm to generate motion trajectory information; The action recognition module uses a temporal convolutional network based on motion trajectory information to extract key action features from sequence data, adopts a dynamic time warping method to align and analyze the action features in time series, and applies a K-means clustering algorithm to classify and label the actions to generate key recognition actions. The action analysis module is based on key recognition actions, performs dimensionality reduction processing on action feature data through principal component analysis, uses local anomaly factor algorithm to perform local anomaly detection on the reduced data, adopts support vector machine to classify the detection results into action quality, and generates action pattern analysis results; The technical rating module uses a support vector machine to analyze and classify multiple action features based on the results of the action pattern analysis, quantifies the execution quality of each action through a weighted average scoring method, and uses a comprehensive scoring model based on the quantified results to evaluate the athlete's technical execution and generate a technical execution quality score; The program generation module analyzes the athletes' technical performance and historical data based on the technical execution quality score and applies a decision tree. It uses association rule mining to extract optimization measures from the analysis results, and uses a genetic algorithm to search and optimize the training program according to the measures to generate a matching training program.

Images