Basketball game analysis platform and method

Abstract

The application discloses a basketball movement analysis platform and method, relates to the technical field of intelligent basketball movement analysis, and is used for solving the problems of single data dimension, difficult accurate identification of action causes and inaccurate positioning of technical short boards in basketball training analysis; through collection of video data, movement sensing data and auxiliary marking data, basketball movement time sequence data is constructed; combined with accurate marking of human body key points and basketball key points, action posture, basketball control, action rhythm and physical state are coupled and analyzed; dominant factors and associated factors are further identified, action standardization, technical ability, physical state and technical short board positioning results are output, and comprehensive scores, ability levels and progress trends are formed. The scheme can improve the accuracy, comprehensiveness and pertinence of basketball training analysis.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for basketball, and more specifically, to a basketball analysis platform and method. Background Technology

[0002] With the development of digitalization and intelligentization in sports training, analyzing basketball processes using video recognition, sensor acquisition, and data modeling technologies has become an important means to improve training quality and competitive level. Current basketball training analysis methods largely rely on coach's experience and observation, single video replays, or judgments based on partial sensor data. While these methods can reflect the athlete's performance to some extent, they still suffer from problems such as limited analytical dimensions, insufficient data correlation, and biased evaluation results. Especially in continuous action scenarios such as shooting, dribbling, passing, and defense, changes in body posture, basketball trajectory, movement rhythm, and physical load are often intertwined. Judging based solely on a single image feature, trajectory parameter, or physiological indicator can easily lead to misjudgments and make it difficult to accurately pinpoint technical weaknesses and the causes of anomalies. At the same time, traditional solutions typically lack analysis of state transition relationships between continuous action segments and lack hierarchical identification mechanisms for dominant and related factors, making it difficult for the analysis results to support refined training guidance.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a basketball sports analysis platform and method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a preferred embodiment, it includes: The raw video data, raw motion sensing data, and raw auxiliary label data are collected synchronously, and time alignment is completed using the effective video frame time as the reference time axis. The unified mapping of human body position and basketball position is completed in the field coordinate system. Then, the image posture information, motion physics information, and event information are cleaned, standardized, and weighted and fused to generate basketball motion time series data. The current action segment in the basketball motion time series data is analyzed frame by frame, and a core key point annotation system is constructed based on the target localization of the human body area and the basketball area. Then, combined with annotation confidence, adjacent frame position constraints and motion trajectory constraints, the key point annotation results are corrected for errors and temporally correlated to generate accurate annotation data. Using action segments as analysis units, a correspondence is established between the precisely labeled data, the basketball motion time series data, and the basic information corresponding to the basketball motion object to construct motion analysis input data. Then, human posture representation data, basketball movement representation data, motion time rhythm representation data, and physiological load representation data are mapped to a unified parameter structure to establish intra-segment correspondence and inter-segment change relationships. Based on the state transition differences between adjacent action segments, composite fluctuating action segments are identified. Subsequently, the judgment thresholds corresponding to each state quantity in the composite fluctuating action segments are adaptively adjusted to distinguish the dominant factors that have the greatest impact on the result fluctuation and the related factors that exceed the threshold at the same time. A hierarchical attribution structure of dominant factors and related factors is established to generate basketball motion analysis results. The basketball sports analysis results are grouped and corresponding scoring indicators are extracted. The technical weakness identification results are converted into technical weakness impact values ​​according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific action. Then, the action standardization index score, technical ability index score, physical condition index score and technical weakness impact value are incorporated into a unified scoring structure to construct a comprehensive score value. The comprehensive score value is mapped to the corresponding ability level result according to the preset scoring interval. The scoring change relationship is established based on the difference between the comprehensive score value and the previous analysis period and the difference between the scores of each sub-item, and the basketball sports scoring result is generated.

[0006] In a preferred embodiment, raw video data, raw motion sensing data, and raw auxiliary marker data are collected simultaneously. The raw video data is used to extract human target position, basketball target position, and action area features to construct a video posture feature sequence. The raw motion sensing data is used to collect basketball trajectory, dribbling number, passing force, shooting speed, heart rate, and exercise duration, and sampling time markers are added to construct a motion physical feature sequence. The raw auxiliary marker data is used to perform event-based encoding to construct an event marker feature sequence.

[0007] In a preferred embodiment, the effective video frame time is used as the reference time axis. The video pose feature sequence, motion physics feature sequence and event marker feature sequence are time-aligned, and a field coordinate system is established to uniformly map the human body position and basketball position to the same spatial reference. Then, data cleaning, standardization and weighted fusion processing are performed on the three types of feature sequences to construct basketball motion time-series data that associates image pose information, motion physics information and event information in time order.

[0008] In a preferred embodiment, the current action segment in the basketball motion time series data is analyzed frame by frame to extract the video pose feature sequence, motion physical feature sequence, and event marker feature sequence corresponding to the current action segment. Target localization is performed on the human body region and basketball region in the video pose feature sequence to construct a core key point annotation system containing fifteen human body key points and three basketball key points. Then, based on the human body region features and basketball region features, initial annotations are performed on the head, neck, left and right shoulders, left and right elbows, left and right hands, waist, left and right knees, left and right feet, torso center, as well as the center of the ball, the upper vertex of the ball, and the lower vertex of the ball, obtaining the corresponding initial annotation coordinates and annotation confidence.

[0009] In a preferred embodiment, for annotation points to be corrected that are below a preset confidence threshold, adjacent frame position constraints and motion trajectory constraints are introduced for error correction. The adjacent frame position constraints are used to extract the corresponding positional relationship of the same key point in consecutive image frames. The motion trajectory constraints are used to call basketball trajectory data, dribbling rhythm data, passing force data, and shooting speed data to constrain the consistency of the positions of hand key points, basketball key points, and lower limb support key points. The corrected annotation coordinates are determined by weighted calculation of the initial annotation coordinates, the effective annotation coordinates of the previous image frame, and the trajectory reference coordinates. Based on this, the eighteen core key points in all image frames are associated in chronological order to establish a key point temporal annotation sequence. The connection relationship of human joints and the connection relationship of basketball trajectory are further calculated to generate accurate annotation data.

[0010] In a preferred embodiment, precise labeled data, basketball movement time-series data, and the age, height, weight, and skill level of the basketball player are summarized, and corresponding relationships are established according to movement segments to construct motion analysis input data. Specifically, for situations where changes in human key points, basketball key points, basketball trajectory, movement rhythm, and heart rate exist asynchronously, with different amplitudes, directions, and time sequences between adjacent movement segments during continuous or competitive training, the movement segment is used as the analysis unit. The positions of human key points, basketball key points, human joint connections, and basketball trajectory connections are integrated with basketball trajectory, dribbling frequency, passing force, shooting speed, heart rate, exercise duration, and movement event information. Combined with age, height, weight, and skill level, a motion state parameter group is constructed, including posture offset parameters, basketball control parameters, movement rhythm parameters, load fluctuation parameters, and result performance parameters. By mapping human posture representation data, basketball movement representation data, movement time rhythm representation data, and physiological load representation data to a unified parameter structure under the same movement segment, intra-segment correspondences and inter-segment change relationships between different types of state variables are established.

[0011] In a preferred embodiment, the following parameters are extracted: posture deviation (considered by joint angle deviation, center of gravity offset, release point position offset, and force direction offset); basketball control fluctuation (considered by basketball trajectory offset, basketball speed fluctuation, and basketball position dispersion); movement rhythm mismatch (considered by movement initiation duration deviation, movement transition duration deviation, movement completion duration deviation, and adjacent movement interval fluctuation); load coupling fluctuation (considered by instantaneous heart rate change, heart rate recovery rate deviation, movement duration deviation, and high-intensity movement density); and movement completion rate deviation, movement stability deviation, and movement result fluctuation. The resulting fluctuations are calculated, and a sequence of motion fluctuation states is constructed according to the order of motion segments. Then, the various state quantities between the current motion segment and the previous motion segment are compared, the state transition difference between adjacent motion segments is calculated, and the motion conflict judgment quantity is constructed by using the transition difference relationship between attitude offset, basketball control fluctuation, motion rhythm mismatch, load coupling fluctuation and result fluctuation. Motion segments with inconsistent relationships between changes in state quantities and changes in results are identified as composite fluctuation motion segments, thereby identifying multi-factor coupling situations where the same abnormal result corresponds to different causes and different abnormal results originate from the same potential factor from the continuous motion chain.

[0012] In a preferred embodiment, after identifying a composite undulating motion segment, adaptive adjustments are made to the judgment thresholds corresponding to the attitude offset and basketball control fluctuation based on the load coupling fluctuation and motion rhythm mismatch in the current motion segment. Corrected attitude thresholds and corrected basketball thresholds corresponding to the state changes of the current motion segment are constructed. Based on this, the dominant influence value corresponding to each state quantity is calculated according to the correspondence between attitude offset, basketball control fluctuation, motion rhythm mismatch, load coupling fluctuation, and their segment transfer differences and result fluctuation transfer differences. The state quantity with the greatest impact on the result fluctuation is identified as the dominant factor from multiple state quantities. Other state quantities in the same motion segment that simultaneously exceed the corresponding correction threshold are identified as related factors, establishing a hierarchical relationship between the dominant factor and related factors. Then, a correspondence is established between the dominant factor, related factors, and motion completion rate, motion stability, and motion result fluctuation, and the motion segment execution result is reconstructed by attribution. When the dominant factor is the attitude offset, it is combined with joint angle deviation and center of gravity deviation. The displacement, release point position offset, and force direction offset are used to identify movement posture imbalance anomalies. When the dominant factor is basketball control fluctuation, basketball trajectory offset, basketball speed fluctuation, and basketball position dispersion are combined to identify basketball control instability anomalies. When the dominant factor is movement rhythm mismatch, movement rhythm misalignment anomalies are combined with movement initiation duration deviation, movement transition duration deviation, and movement completion duration deviation. When the dominant factor is load coupling fluctuation, load coupling fluctuation anomalies are combined with instantaneous heart rate change, heart rate recovery rate deviation, movement duration deviation, and high-intensity movement density. Related factors are sequentially written into the attribution results to form a hierarchical attribution structure of dominant and related factors. Finally, the attribution results of all movement segments within the same analysis period are aggregated to generate movement standardization analysis results, technical ability analysis results, physical fitness analysis results, and technical weakness location results. The movement type, dominant factor name, related factor name, and corresponding movement segment location of the weakness are all merged into the basketball movement analysis results.

[0013] In a preferred embodiment, the results of the motion standardization analysis, technical ability analysis, physical fitness analysis, and technical weakness identification are grouped, and scoring indicators corresponding to each analysis result are extracted. Specifically, the technical weakness identification results obtained based on dominant factors, related factors, and a hierarchical attribution structure are used to construct a technical weakness impact value according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific action, transforming the multi-factor attribution results into quantifiable deduction items. Then, the motion standardization indicator scores, technical ability indicator scores, physical fitness indicator scores, and technical weakness impact values ​​are incorporated into a unified scoring structure to construct a comprehensive score value. This comprehensive score value is mapped to the corresponding ability level result according to a preset scoring interval. Furthermore, based on the difference in comprehensive score value between the current analysis period and the previous analysis period, as well as the difference in score values ​​for each sub-item, a scoring change relationship corresponding to changes in dominant factors and related factors is established, generating a basketball scoring result.

[0014] In a preferred embodiment, it includes: a basketball time-series data generation module, a precise labeled data generation module, a basketball analysis module, a basketball scoring module, and signal connections between the modules; The basketball motion time series data generation module is used to synchronously collect raw video data, raw motion sensor data, and raw auxiliary label data, and complete time alignment with the effective video frame time as the reference time axis, and complete the unified mapping of human body position and basketball position in the court coordinate system; then, the image posture information, motion physics information, and event information are cleaned, standardized, and weighted and fused to generate basketball motion time series data. The precise annotation data generation module is used to analyze the current action segment in the basketball motion time series data frame by frame, and to build a core key point annotation system based on the target localization of the human body area and the basketball area; then, combined with annotation confidence, adjacent frame position constraints and motion trajectory constraints, the key point annotation results are corrected for errors and temporally correlated to generate precise annotation data. The basketball motion analysis module uses motion segments as analysis units to establish a correspondence between the precisely labeled data, the basketball motion time series data, and the basic information corresponding to the basketball motion object, constructing motion analysis input data. Then, it maps human posture representation data, basketball movement representation data, motion time rhythm representation data, and physiological load representation data into a unified parameter structure, establishing intra-segment correspondences and inter-segment change relationships. Based on the state transition differences between adjacent motion segments, it identifies composite fluctuating motion segments. Subsequently, it adaptively adjusts the judgment thresholds corresponding to each state quantity in the composite fluctuating motion segments, distinguishing between the dominant factors that have the greatest impact on the result fluctuation and the associated factors that simultaneously exceed the threshold, establishing a hierarchical attribution structure for the dominant and associated factors, and generating basketball motion analysis results. The basketball scoring module is used to group the basketball analysis results and extract corresponding scoring indicators. The technical weakness identification results are converted into technical weakness impact values ​​according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific action. Then, the action standardization index score, technical ability index score, physical condition index score, and technical weakness impact value are incorporated into a unified scoring structure to construct a comprehensive score value. The comprehensive score value is mapped to the corresponding ability level result according to the preset scoring interval. The scoring change relationship is established based on the difference between the comprehensive score value and the previous analysis cycle and the difference between the scores of each sub-item, and the basketball scoring result is generated.

[0015] The technical effects and advantages of the basketball sports analysis platform and method of this invention are as follows: This invention achieves comprehensive and three-dimensional analysis of basketball movements by integrating multi-source data such as video, sensors, and auxiliary markers, thereby improving data utilization and analytical completeness. Secondly, it enhances the accuracy of movement recognition and trajectory analysis through precise keypoint annotation, error correction, and temporal correlation processing. Thirdly, by coupling and modeling posture deviation, basketball control, movement rhythm, and load fluctuations, it can more accurately identify the true causes of abnormal movements, reducing biases caused by single-indicator judgments. Fourthly, it can output results such as movement standardization, technical ability, physical condition, technical weakness identification, comprehensive score, ability level, and progress trend, providing more detailed and targeted decision-making basis for basketball training, thus possessing high practical and promotional value. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the comprehensive scoring and progress trend of the basketball sports analysis platform and method of the present invention.

[0017] Figure 2 This is a schematic diagram illustrating the precise annotation and error correction of key points in the basketball sports analysis platform and method of this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In this embodiment, the present invention discloses a basketball motion analysis method, such as... Figure 1 As shown, it includes: Step 1: Synchronously collect raw video data, raw motion sensor data, and raw auxiliary label data, and complete time alignment using the effective video frame time as the reference time axis to complete the unified mapping of human body position and basketball position in the field coordinate system; then perform data cleaning, standardization, and weighted fusion on image posture information, motion physics information, and event information to generate basketball motion time series data; Step 2: Analyze the current action segment in the basketball motion time series data frame by frame, and construct a core key point annotation system based on the target localization of the human body area and basketball area; then, combine the annotation confidence, adjacent frame position constraints and motion trajectory constraints to perform error correction and temporal correlation on the key point annotation results to generate accurate annotation data; Step 3: Using action segments as analysis units, establish a correspondence between the precisely labeled data, the basketball motion time series data, and the basic information corresponding to the basketball motion object to construct motion analysis input data; then map human posture representation data, basketball movement representation data, motion time rhythm representation data, and physiological load representation data into a unified parameter structure to establish intra-segment correspondence and inter-segment change relationships, and identify composite fluctuating action segments based on the state transition differences between adjacent action segments; subsequently, adaptively adjust the judgment thresholds corresponding to each state quantity in the composite fluctuating action segments, distinguish the dominant factors that have the greatest impact on the result fluctuation and the related factors that exceed the threshold at the same time, establish a hierarchical attribution structure for dominant factors and related factors, and generate basketball motion analysis results; Step 4: Group the basketball sports analysis results and extract corresponding scoring indicators. Convert the technical weakness location results into technical weakness impact values ​​according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific action. Then, incorporate the action standardization index score, technical ability index score, physical condition index score, and technical weakness impact value into a unified scoring structure to construct a comprehensive score value. Map the comprehensive score value to the corresponding ability level result according to the preset scoring interval. Establish the scoring change relationship based on the difference between the comprehensive score value and the previous analysis cycle and the difference between the scores of each sub-item, and generate the basketball sports scoring result.

[0020] In step one, raw video data, raw motion sensor data, and raw auxiliary label data are collected simultaneously for the basketball motion process to be analyzed. Then, the three types of raw data are processed sequentially, including effective frame extraction, time alignment, coordinate unification, data cleaning, feature standardization, and weighted fusion, to generate basketball motion time series data.

[0021] The raw video data is collected by mobile or fixed camera equipment, capturing continuous images of basketball players performing actions such as shooting, dribbling, passing, defense, jumping, landing, and turning. Raw motion sensor data is collected by smart basketballs, fitness trackers, and heart rate monitoring devices, including basketball trajectory, number of dribbles, passing force, shooting speed, heart rate, and duration of movement. Raw auxiliary labeling data is entered by coaches or referees, including evaluations of action standardization, fouls, and tactical execution. The raw video data forms the raw input for video posture features, the raw motion sensor data forms the raw input for motion physical features, and the raw auxiliary labeling data forms the raw input for event labeling features.

[0022] The original video data undergoes effective frame extraction. Effective frame extraction includes blurred frame filtering and redundant frame filtering. Blurred frame filtering removes image frames where the human and basketball outlines are indistinguishable due to defocusing, jitter, or motion blur. Redundant frame filtering removes image frames with repetitive content that does not demonstrate changes in motion. Redundant frame filtering is implemented using the inter-frame difference method, which calculates the degree of difference between adjacent image frames and filters out image frames with a difference level below a preset threshold. After effective frame extraction, the retained image frames are scaled and normalized to ensure a uniform input size. Then, human target position, basketball target position, and motion region features are extracted from each image frame to form a video pose feature sequence arranged chronologically. .

[0023] The raw motion sensing data is collected according to the acquisition type to form a sequence of motion physical characteristics. Among them, basketball trajectory data represents the spatial movement path of the basketball during the movement; the number of dribbles represents the rhythm of the dribbling action; passing force and shooting speed represent the force characteristics of the action; and heart rate and exercise duration represent the execution load of the action. Sampling time markers are added to each raw motion sensor data, establishing a correspondence between each sampled value and its corresponding sampling time, and these are arranged in chronological order to form a sequence of motion physical characteristics. .

[0024] The original auxiliary labeled data is encoded into events to form event-labeled feature sequences. Specifically, the evaluation of action standardization is converted into evaluation tags for corresponding action segments, foul situations are converted into foul event tags for corresponding time periods, and tactical execution is converted into tactical event tags for corresponding coordinated action segments. Each tag is associated with the recording time or the action segment it belongs to, and they are arranged in chronological order to form an event marker feature sequence. .

[0025] The action segment refers to a continuous action interval divided according to action event information, video posture feature sequence, motion physical feature sequence and event marker feature sequence, used to characterize the continuous time period corresponding to a basketball action object completing a single shot, dribble, pass or defensive action.

[0026] Video pose feature sequence Motion physical characteristic sequence and event marker feature sequence Time alignment is performed. Time alignment uses the frame time of the effective video frame as the reference timeline, and the sequence of motion physical features is aligned. Each sampled value and event marker feature sequence Each event moment is mapped to a corresponding video frame moment. When a video frame moment does not have a directly corresponding sensor sample value, linear interpolation is used to generate an alignment value for that moment. Let the same motion sensing feature be at time... and time The effective sample values ​​are respectively and Then the alignment value at the intermediate time t for: ; in, The alignment value at time t. and These are the sampled values ​​from two adjacent valid sampling times. and This corresponds to the sampling time.

[0027] Coordinate unification is performed on data involving spatial location. First, a court coordinate system is established, with the midpoint of the baseline as the origin, the baseline direction as the x-axis, and the sideline direction as the y-axis. Then, fixed reference points such as the court sidelines, free throw line, or backboard edge are used to unify the video pose feature sequence. The positions of the human body and the basketball are mapped to the court coordinate system; simultaneously, based on the initial calibration parameters of the smart basketball or other sensing devices, the sequence of motion physical characteristics is... The basketball trajectory data was converted to the same court coordinate system. After coordinate unification, positional features were represented using the same spatial reference.

[0028] Data cleaning was performed on three types of feature sequences. Data cleaning included outlier removal and missing value completion. Outliers were abnormal sampling values ​​caused by accidental device touch, short-term disconnection, or image distortion; missing values ​​were local blank data caused by acquisition fluctuations. Outliers were directly removed, and missing values ​​were completed using linear interpolation with the same time alignment method to obtain continuous feature sequences.

[0029] Standardization is performed on quantifiable features in the three types of feature sequences. For continuous features, let the original feature value be x, and the minimum value of this feature within the statistical range of the training samples be x. The maximum value is Then the standardized value for: ; Where x is the original feature value, This is the minimum reference value for this feature within the statistical range of the training samples. This is the maximum reference value for this feature within the statistical range of the training samples. To standardize the results, event tagging features are converted into preset event encoding values ​​using a label mapping method.

[0030] Standardized video pose feature sequence Motion physical characteristic sequence and event marker feature sequence Weighted fusion is performed to generate time-series basketball data. Let the fusion result be F, then: ; in, For video pose feature weights, Weights for the physical characteristics of motion. Assign feature weights to events, satisfying: + + =1; The weights are determined based on the completeness, continuity, and reliability of each data source within the current action segment. When all three types of data are complete, the weight of the original motion sensor data is higher than that of the original video data, and the weight of the original video data is higher than that of the original auxiliary marker data. As a preferred implementation, the weight of the smart device data can be 60%, the weight of the video data can be 30%, and the weight of the auxiliary marker data can be 10%. The resulting F after fusion is the basketball motion time-series data, which is associated with image posture information, motion physics information, and event information in chronological order.

[0031] In step two, as Figure 2As shown, the basketball motion time-series data generated in step one is analyzed frame by frame. First, the video posture feature sequence, motion physical feature sequence, and event marker feature sequence corresponding to the current action segment are extracted from the basketball motion time-series data. Then, the human body region and basketball region in the video posture feature sequence are located, and the initial annotation of human body keypoints and basketball keypoints is performed based on the localization results. Human body keypoints refer to the position points used to characterize the body posture and action structure of the basketball motion object, while basketball keypoints refer to the position points used to characterize the spatial position and ball posture of the basketball. To ensure that the motion analysis results cover core actions such as shooting, dribbling, passing, and defense, this step defines eighteen core keypoints. Among them, there are fifteen human body keypoints, namely the head, neck, left shoulder, right shoulder, left elbow, right elbow, left hand, right hand, waist, left knee, right knee, left foot, right foot, and torso center; and three basketball keypoints, namely the center of the ball, the upper vertex of the ball, and the lower vertex of the ball. The above eighteen core keypoints are used to jointly describe the human body's action posture, upper limb force state, lower limb support state, and basketball movement state.

[0032] Each frame of the video pose feature sequence is first subjected to uniform size transformation and numerical normalization to ensure that the input image meets the preset input specifications. The processed image is then input into an improved single-stage object detection network to extract human body and basketball region features. In a preferred embodiment, the uniform size transformation scales the input image to 640×640 pixels, and the numerical normalization maps pixel values ​​to the 0-1 range.

[0033] The improved single-stage target detection network is formed by sequentially connecting a cross-stage local backbone feature extraction network, a feature fusion network, and a detection output layer. The cross-stage local backbone feature extraction network is used to extract multi-layer features from the input image, outputting human contour features, limb edge features, basketball edge features, and local motion texture features at different scales; the feature fusion network is used to upsample, downsample, and laterally connect features at different scales to form multi-scale fused features that take into account both global semantic information and local positional details; the detection output layer is used to output the target bounding box position, target category, and target confidence of the human and basketball regions based on the multi-scale fused features.

[0034] In the cross-stage local backbone feature extraction network, a convolutional block attention mechanism is introduced. This mechanism enhances the response intensity of the player action region and the basketball region in both channel and spatial dimensions, while suppressing interference from the court background and spectator background on the detection results. As a preferred implementation, the convolutional block attention mechanism is applied after the output of each level of backbone features to enhance both human body and basketball region features at multiple scales.

[0035] During the training phase, the improved single-stage object detection network was trained using pre-labeled human and basketball region samples. The training samples consisted of basketball motion images under different training environments, lighting conditions, occlusion conditions, and action types, with human and basketball region bounding boxes labeled in the samples. As a preferred implementation, the training objective function was a weighted sum of bounding box regression loss, object confidence loss, and object category loss. Data augmentation was performed using random flipping, random scaling, brightness perturbation, and slight rotation to improve the network's adaptability to occlusion, fast movement, and complex background scenes.

[0036] During the inference phase, the current image frame is input into the trained improved single-stage object detection network. The network first outputs candidate human regions and candidate basketball regions, then selects valid candidate regions based on the target confidence, and performs non-maximum suppression processing on overlapping candidate regions. Finally, the human region feature map and basketball region feature map of the current image frame are obtained.

[0037] After obtaining the human body region feature map and the basketball region feature map, the initial labeled coordinates of eighteen core keypoints in the current image frame are calculated using keypoint regression. Keypoint regression directly outputs the two-dimensional position coordinates of each keypoint in the current image frame, along with its corresponding label confidence score. The initial labeled coordinates are the keypoint positions directly calculated from the region features of the current image frame, and the label confidence score is a rating used to characterize the reliability of the keypoint position. For human body keypoints, the position coordinates of the head, neck, left shoulder, right shoulder, left elbow, right elbow, left hand, right hand, waist, left knee, right knee, left foot, right foot, and torso center are calculated; for basketball keypoints, the position coordinates of the center of the ball, the upper vertex of the ball, and the lower vertex of the ball are calculated. After this processing, each image frame generates corresponding initial labeling results and label confidence scores.

[0038] After obtaining the initial annotation results, the annotation confidence level of each keypoint is determined. Keypoints with an annotation confidence level not lower than the preset confidence threshold are identified as valid annotation points, while keypoints with an annotation confidence level lower than the preset confidence threshold are identified as annotation points to be corrected.

[0039] In a preferred embodiment, the preset confidence threshold is set according to the keypoint type. For human keypoints, the preset confidence threshold can be 0.70 to 0.90; for basketball keypoints, the preset confidence threshold can be 0.75 to 0.92. Further, when there is no occlusion and the image is clear, the preset confidence threshold for human keypoints is preferably 0.80, and the preset confidence threshold for basketball keypoints is preferably 0.85. In cases of partial occlusion, rapid arm swings, high-speed ball movement, or overlapping of multiple people, the preset confidence threshold for human keypoints can be lowered to 0.75, and the preset confidence threshold for basketball keypoints can be lowered to 0.80, to avoid directly misjudging correctable keypoints as invalid points.

[0040] The label confidence level is determined by the reliability of the key point position output by the key point regression. The larger the value, the closer the current position is to the true key point position. The smaller the value, the higher the possibility that the current position is offset, missing, or mislabeled.

[0041] The preset confidence threshold is used to distinguish whether the current position can be directly used as the final annotation result of the key point. The annotation point to be corrected is not directly discarded, but enters the error correction process. The reason for using confidence judgment is that in the case of occlusion, rapid arm swing, high-speed ball movement, or multiple people overlapping, the key point position directly output by a single frame image is prone to offset, omission, or mislabeling. Relying solely on the detection result of a single frame is insufficient to stably reflect the real action structure.

[0042] Annotation error correction is performed on the annotation points to be corrected. Annotation error correction includes adjacent frame position constraint correction and motion trajectory constraint correction. Adjacent frame position constraint correction refers to extracting the corresponding positions of the same key point in the previous and next image frames, and judging whether the initial annotation coordinates in the current image frame deviate from the continuous motion range of adjacent frames based on the continuous change relationship of the action; when the initial annotation coordinates in the current image frame exceed the continuous change range of adjacent frames, it is determined that the key point has a temporal offset. Motion trajectory constraint correction refers to calling the motion physical feature sequence formed in step one, extracting basketball trajectory data, dribbling rhythm data, passing force data, or shooting speed data corresponding to the current moment, and judging whether the current key point position is consistent with the actual motion trajectory; when the human hand key point, basketball key point, or lower limb support key point is inconsistent with the corresponding action trajectory, it is determined that the key point has a trajectory offset. The above two types of correction constrain the initial annotation results from the perspectives of temporal continuity and physical trajectory consistency, respectively.

[0043] For offset annotation points, a weighted correction method is used to calculate the corrected annotation coordinates. Let the initial annotation coordinates of a certain annotation point to be corrected in the current image frame be... The effective labeled coordinates of the same key point in the previous image frame are: The trajectory reference coordinates obtained from the motion trajectory constraints are: Then the corrected coordinates of the point to be corrected are... Determine by the following formula: ; in, Provide the corrected coordinates for this keypoint in the current image frame. The initial coordinates of this keypoint in the current image frame. The valid labeled coordinates of this keypoint in the previous image frame. These are the trajectory reference coordinates calculated based on motion trajectory data or the continuity of actions; , and These are the initial annotation coordinate weights, adjacent frame position weights, and trajectory reference coordinate weights, respectively, and they satisfy the following: + + =1; The initial labeled coordinate weights are used to retain the direct detection results of the current image frame, the adjacent frame position weights are used to maintain the positional continuity of the same key point in consecutive image frames, and the trajectory reference coordinate weights are used to ensure that the corrected position is consistent with the trajectory of the basketball movement or the force trajectory of the action. For basketball key points, the trajectory reference coordinates are determined by the basketball trajectory data; for human hand key points, the trajectory reference coordinates are determined by the joint change relationship between the hand movement direction and the basketball movement direction; for lower limb key points, the trajectory reference coordinates are determined by the change direction of the body's center of gravity in consecutive image frames. After calculation using this formula, corrected labeled coordinates that satisfy the image detection results, temporal continuity, and motion trajectory relationships are obtained.

[0044] In a preferred embodiment, the initial annotation coordinate weight, adjacent frame position weight, and trajectory reference coordinate weight are determined based on the current keypoint type and the annotation confidence level of the current keypoint. When the current keypoint is a basketball keypoint, to enhance the continuity of basketball movement, the trajectory reference coordinate weight is preferably higher than the adjacent frame position weight; when the current keypoint is a human hand keypoint, the trajectory reference coordinate weight is preferably close to the adjacent frame position weight; when the current keypoint is a lower limb support keypoint, the adjacent frame position weight is preferably higher than the trajectory reference coordinate weight.

[0045] As a feasible set of parameter settings, for basketball key points, the initial labeled coordinate weight can be set to 0.20–0.35, the adjacent frame position weight to 0.20–0.30, and the trajectory reference coordinate weight to 0.40–0.55; for human hand key points, the initial labeled coordinate weight can be set to 0.25–0.40, the adjacent frame position weight to 0.25–0.40, and the trajectory reference coordinate weight to 0.25–0.40; for lower limb support key points, the initial labeled coordinate weight can be set to 0.25–0.35, the adjacent frame position weight to 0.40–0.55, and the trajectory reference coordinate weight to 0.15–0.30.

[0046] Furthermore, when the confidence level of the annotation point to be corrected is close to the preset confidence threshold, the weight of the initial annotation coordinates is increased; when the confidence level of the annotation point to be corrected is significantly lower than the preset confidence threshold, the weight of the initial annotation coordinates is decreased, and the weights of the adjacent frame positions and the trajectory reference coordinates are increased. As a preferred embodiment, when the difference between the annotation confidence level and the preset confidence threshold does not exceed 0.05, the weight of the initial annotation coordinates can be increased by 0.05 to 0.10; when the difference between the annotation confidence level and the preset confidence threshold exceeds 0.15, the weight of the initial annotation coordinates can be decreased by 0.05 to 0.15.

[0047] After completing the corrected annotation coordinate calculation, the annotation confidence of the corrected keypoints is recalculated, and the recalculated confidence is judged again. Keypoints whose recalculated confidence reaches the preset confidence threshold are determined as final valid annotation points; keypoints whose recalculated confidence is still below the preset confidence threshold undergo a re-examination process. The re-examination process involves re-extracting region features and performing keypoint regression calculations on the image region containing the keypoint, and then re-correcting the coordinates based on the corresponding keypoint positions in adjacent image frames. Keypoints whose positions cannot be stably determined after the re-examination process undergo manual review. Manual review involves annotators with experience in basketball motion recognition checking the extracted keypoint results frame by frame and manually correcting any missing or incorrectly labeled positions. The manually corrected keypoint positions are directly written into the keypoint results of the current image frame as the final annotation result for the corresponding keypoint in that image frame.

[0048] After keypoint analysis and annotation error correction are completed for all image frames, the eighteen core keypoints in each image frame are associated in chronological order to form a keypoint temporal annotation sequence. The keypoint temporal annotation sequence refers to the result of uniformly storing and numbering the positional changes of the head, neck, left shoulder, right shoulder, left elbow, right elbow, left hand, right hand, waist, left knee, right knee, left foot, right foot, torso center, ball center, upper vertices of the ball, and lower vertices of the ball within consecutive image frames in the same action segment. The keypoint temporal annotation sequence is further processed to calculate the human joint connection relationships and the basketball trajectory connection relationships, forming a structured annotation result for the action. The human joint connection relationships refer to the skeletal structure of the human action formed by the connections between human keypoints, while the basketball trajectory connection relationships refer to the basketball motion trajectory structure formed by the positional changes of basketball keypoints in consecutive image frames.

[0049] After the above processing is completed, precise annotation data is output. This precise annotation data includes the positions of eighteen core key points corresponding to each image frame within the action segment, the final annotation confidence of each key point, the temporal annotation sequence of the key points, the connection relationships between human joints, and the connection relationships between the basketball trajectory. This precise annotation data is used to comprehensively characterize the changes in body posture, basketball movement, and the collaborative relationship between human and basketball movements within the action segment.

[0050] In step three, the precisely labeled data generated in step two, the basketball motion time-series data generated in step one, and the basic information corresponding to the basketball motion object are summarized, and a correspondence is established according to motion segments to form motion analysis input data. The basic information refers to descriptive information used to characterize the physical condition and training stage of the basketball motion object, including age, height, weight, and skill level. The motion analysis input data consists of three parts: the first part is precisely labeled data, used to characterize human key points, basketball key points, human joint connections, and basketball trajectory connections; the second part is basketball motion time-series data, used to characterize image posture information, motion physics information, and event information; the third part is basic information, used to limit the age and skill levels corresponding to the motion analysis.

[0051] The standard movement database is used to provide reference movement parameters for different age groups, skill levels, and movement types. The reference data in the standard movement database can be composed of data obtained from multiple basketball players repeatedly performing shooting, dribbling, passing, and defensive movements in standard training scenarios. The data is then statistically processed to include the positions of key human body points, the relationships between human body joints, the relationships between basketball trajectory lines, the basketball trajectory, the movement rhythm, and the movement results for each movement.

[0052] The reference physical fitness database provides reference values ​​for heart rate, recovery rate, and movement duration for different age groups, skill levels, and training stages. The reference data in the database can be collected from multiple basketball players under the same training intensity and movement type conditions, and stored hierarchically according to age, height, weight, and skill level.

[0053] When calling the standard movement database and the reference physical fitness database, the reference data that is the same as or closest to the current basketball player's age, height, weight and skill level is selected as the corresponding reference value; when there is no completely consistent reference data, the reference data of the adjacent level is selected for weighted interpolation to obtain the reference value corresponding to the current basketball player.

[0054] Based on this, the key points of the human body, key points of the basketball, the connection relationship of human joints, the connection relationship of basketball trajectory, basketball trajectory, number of dribbles, passing force, shooting speed, heart rate, exercise duration and action event information in the motion analysis input data are extracted in a parameterized manner to form a motion state parameter group.

[0055] The parameter extraction includes calculating corresponding parameter values ​​for human posture representation data, basketball movement representation data, action time rhythm representation data, and physiological load representation data, and mapping parameters of different dimensions to a unified numerical range. As a preferred embodiment, the unified numerical range is the 0-1 interval.

[0056] Among them, the joint angle deviation value is determined based on the difference between the human joint angle in the current action segment and the corresponding joint angle reference value in the standard action database; the center of gravity offset is determined based on the offset distance between the projection position of the torso center in the field coordinate system and the corresponding center of gravity reference trajectory in the standard action database; the release point position offset is determined based on the offset distance between the key point position of the left or right hand at the moment of release and the corresponding release point reference position in the standard action database; and the force direction offset is determined based on the angle deviation between the force direction vector formed by the shoulder, elbow, and hand and the corresponding force direction reference vector in the standard action database.

[0057] The basketball trajectory offset is determined based on the average distance between the basketball trajectory in the current action segment and the basketball reference trajectory of the corresponding action in the standard action database; the basketball speed fluctuation is determined based on the dispersion of the basketball speed change amplitude at adjacent sampling times; and the basketball position dispersion is determined based on the dispersion of the basketball key point positions around the reference trajectory in the current action segment.

[0058] The deviation value of the action start time is determined based on the difference between the duration from the start of the current action segment to the start of the main force exertion action and the corresponding reference duration in the standard action database; the deviation value of the action transition time is determined based on the difference between the transition time between adjacent sub-actions within the current action segment and the corresponding reference transition time; the deviation value of the action completion time is determined based on the difference between the total duration of the current action segment and the corresponding reference total duration; the fluctuation of the interval between adjacent actions is determined based on the degree of dispersion of the time interval between consecutive action segments.

[0059] The instantaneous heart rate change is determined based on the difference in heart rate between the start and end of the current action segment; the heart rate recovery rate deviation is determined based on the difference between the heart rate drop rate during the preset recovery period after the end of the current action segment and the corresponding recovery rate reference value in the reference fitness database; the action duration deviation is determined based on the difference between the duration of the current action segment and the reference duration of similar actions; and the high-intensity action density is determined based on the proportion of high-intensity action segments to the total number of action segments within a preset time window.

[0060] The deviation value of the action completion rate is determined by the difference between the proportion of successful actions to the total number of actions in the current analysis period and the corresponding reference completion rate; the deviation value of the action stability is determined by the difference between the dispersion of the results of similar actions in the current analysis period and the corresponding reference dispersion; the fluctuation of the action result is determined by the mean and variance of the change amplitude of the action result in continuous action segments.

[0061] For each of the above parameters, the original deviation is first calculated based on the corresponding reference value, and then mapped to the 0-1 range through standardization processing, ultimately forming attitude offset parameters, basketball control parameters, motion rhythm parameters, load fluctuation parameters, and result performance parameters.

[0062] It should be noted that during continuous or competitive training of basketball players, the changes in key human body points, key basketball points, basketball trajectory, movement rhythm, and heart rate in adjacent action segments are not synchronized. Specifically, in some action segments, the deviation in the posture formed by key human body points is small, but the trajectory of the basketball corresponding to the key points fluctuates significantly; in some action segments, the trajectory of the basketball remains continuous, but the supporting posture formed by the torso center, left knee, right knee, left foot, and right foot experiences phased imbalances; in some action segments, the shooting speed or passing force remains within the normal range, but the heart rate changes and the duration of the movement show abnormal coupling fluctuations; and in some action segments, the same basketball player exhibits a stable dribbling rhythm in the previous action segment, but in the subsequent action segment, the number of dribbles does not decrease, but the accuracy of basketball position control declines.

[0063] Meanwhile, abnormal performance in basketball is not caused by a single factor, but by the combined effects of changes in body posture, basketball trajectory, movement rhythm, and load status. Because the amplitude, direction, and timing of changes in different types of data vary across different movement segments, the following situations may arise: judging solely based on the position of key points in a single movement segment might misjudge basketball trajectory deviations as posture problems; judging solely based on changes in basketball trajectory might misjudge decreased movement control caused by energy fluctuations as purely technical issues; and judging solely based on heart rate or exercise duration might fail to distinguish whether the abnormality is caused by an imbalance in force exertion or by rhythm mismatch during continuous movement transitions.

[0064] Furthermore, the same abnormal result may correspond to different causes in different segments of a movement, while different abnormal results may originate from the same underlying factor. For example, the same manifestation of unstable shooting results may be caused by fluctuations in the release point position in one segment of a movement, while in another segment it may be caused by a combination of trunk center shift and lower limb support imbalance; similarly, the same manifestation of decreased passing control may be caused by changes in the direction of force in one segment of a movement, while in another segment it may be caused by instability in the movement rhythm after a rapid increase in heart rate.

[0065] Therefore, in this embodiment, motion segment coupling analysis processing is performed on the motion analysis input data. Motion segment coupling analysis processing refers to using motion segments as analysis units, integrating the positions of key human body points, key basketball points, human joint connections, and basketball trajectory connections from the precisely labeled data with basketball trajectory, dribbling frequency, passing force, shooting speed, heart rate, exercise duration, and motion event information from the basketball motion time-series data. This is then combined with basic information such as age, height, weight, and skill level to construct the motion state parameter set corresponding to the current motion segment. The motion state parameter set includes posture deviation parameters, basketball control parameters, motion rhythm parameters, load fluctuation parameters, and result performance parameters. Specifically, the posture deviation parameter characterizes the degree of deviation of the human posture from the standard motion; the basketball control parameter characterizes the degree of dispersion of the basketball's running state from the stable state; the motion rhythm parameter characterizes the degree of time matching between motion initiation, motion transition, and motion completion; the load fluctuation parameter characterizes the degree of change in physiological load during the motion; and the result performance parameters characterize the completion status of the motion corresponding to the current motion segment.

[0066] The posture offset parameters are extracted from the motion state parameter group. These parameters consist of joint angle deviations, center of gravity offsets, release point position offsets, and force direction offsets. Joint angle deviations are obtained by comparing the angles formed by adjacent key points in the precisely annotated data with the corresponding standard joint angles in the standard motion database. Center of gravity offsets are obtained by the positional offset of the torso center relative to the center of the foot support area. Release point position offsets are obtained by the spatial offset of the ball's center relative to the head, shoulder, and torso centers when the basketball leaves the hand. Force direction offsets are obtained by the angle between the line connecting the upper limb key points and the direction of the basketball's movement. After mapping all these offsets to the same numerical range, the posture offset of the current motion segment is calculated. : ; in, G represents the posture offset, M represents the joint angle deviation, R represents the center of gravity offset, R represents the release point position offset, and F represents the force direction offset. , , and These are the parameter weights for the corresponding offsets. The parameter weights are determined based on the current action type; when the current action is a shooting action, the parameter weights corresponding to the release point position offset and the force direction offset are increased; when the current action is a defensive action, the parameter weights corresponding to the center of gravity offset are increased.

[0067] Basketball control parameters are extracted from the motion state parameter group. These parameters consist of basketball trajectory offset, basketball speed fluctuation, and basketball position dispersion. The basketball trajectory offset is obtained by comparing the actual trajectory formed by the basketball key points with the reference trajectory of the corresponding action in the standard motion database; the basketball speed fluctuation is obtained by the amplitude of basketball speed changes between adjacent sampling times; and the basketball position dispersion is obtained by the degree of dispersion of the ball's center position in consecutive image frames. After mapping these parameters to the same numerical range, the basketball control fluctuation of the current motion segment is calculated. : ; in, Let T be the basketball control fluctuation, V be the basketball trajectory deviation, V be the basketball velocity fluctuation, and C be the basketball position dispersion. , and These are the parameter weights for the corresponding parameters. The parameter weights are determined based on the current action type; when the current action is a dribbling action, the parameter weight corresponding to the basketball position dispersion is increased; when the current action is a passing action or a shooting action, the parameter weight corresponding to the basketball trajectory offset is increased.

[0068] The motion rhythm parameters are extracted from the motion state parameter group. These parameters consist of motion initiation duration, motion transition duration, motion completion duration, and fluctuations in the interval between adjacent motions. Motion initiation duration refers to the time from the motion start marker to the start of the main force exertion motion; motion transition duration refers to the time from the start of the main force exertion motion to the release of the ball or change of direction; motion completion duration refers to the time from the current motion segment entering the completion phase to the appearance of the release, catch, or end marker; and fluctuations in the interval between adjacent motions refer to the degree of change in the time interval between similar motion segments. After mapping these parameters to the same numerical range, the motion rhythm mismatch of the current motion segment is calculated. : ; in, Let A be the movement rhythm mismatch, B be the movement initiation time deviation, E be the movement transition time deviation, and I be the movement completion time deviation. , , and These are the parameter weights for the corresponding parameters. The deviation values ​​for action start time, action transition time, and action completion time are all obtained from the difference between the corresponding duration of the current action segment and the reference duration in the standard action database.

[0069] Load fluctuation parameters are extracted from the action state parameter group. These parameters consist of instantaneous heart rate change, heart rate recovery rate deviation, action duration deviation, and high-intensity action density. Instantaneous heart rate change is obtained from the difference in heart rate between the start and end times of the current action segment; heart rate recovery rate deviation is obtained from the difference between the heart rate recovery rate and the reference recovery rate within a preset recovery observation period after the current action segment ends; action duration deviation is obtained from the difference between the actual duration and the reference duration of the current action segment; and high-intensity action density is obtained from the number of high-intensity action segments per unit time. After mapping these parameters to the same numerical range, the load coupling fluctuation of the current action segment is calculated. : ; in, Here, H represents the load-coupled fluctuation, J represents the instantaneous heart rate change, L represents the heart rate recovery rate deviation, and K represents the action duration deviation. , , and These are the parameter weights for the corresponding parameters. The reference recovery rate in the heart rate recovery rate deviation value is provided by a reference fitness database for the same age group and skill level.

[0070] The reference physical fitness database is used to store reference physical fitness values ​​and reference recovery rates corresponding to different age groups and skill levels.

[0071] The performance parameters of the action state parameter group are extracted. These performance parameters consist of action completion rate, action stability, and action result fluctuation. Action completion rate is obtained by comparing the number of successful completions of the current action segment with the total number of attempts; action stability is obtained by the consistency of results for similar action segments within the same analysis period; and action result fluctuation is obtained by the variation in completion results between consecutive action segments. After mapping these parameters to the same numerical range, the result fluctuation of the current action segment is calculated. : ; in, For the fluctuation of the result, This represents the deviation value of the action completion rate. This represents the deviation value of motion stability. The fluctuation of the action result. , and These are the parameter weights for the corresponding parameters. The action completion rate deviation is obtained by the difference between the current action completion rate and the reference completion rate, and the action stability deviation is obtained by the difference between the current action stability and the reference stability.

[0072] After obtaining the attitude offset Basketball control fluctuation Movement rhythm mismatch Load coupling fluctuation and result volatility Then, a motion fluctuation state sequence is constructed according to the order of the motion segments. This sequence characterizes the state changes of a basketball motion object within consecutive motion segments. The state variables between the current and previous motion segments are compared to obtain the state transition differences between them. Let a certain state variable corresponding to the current motion segment be... The state variable corresponding to the previous action segment is Then the fragment transition difference of this state variable for: ; in, This represents the state of the current action segment. This represents the state of the previous action segment. The state variable represents the state transition difference between adjacent action segments. It can be any of the following: attitude offset, basketball control fluctuation, motion rhythm mismatch, load coupling fluctuation, or result fluctuation. The direction and magnitude of change of each state variable between consecutive motion segments can be obtained through the state transition difference.

[0073] After obtaining the segment transition differences of each state variable, a motion conflict determination factor is constructed. The motion conflict determination factor is used to characterize whether there is an inconsistency between the change direction of multiple state variables and the resulting fluctuation. Let the segment transition differences of attitude offset, basketball control fluctuation, motion rhythm mismatch, and load coupling fluctuation in the current motion segment be respectively... , , and The result is that the segment transfer difference of the fluctuation is Then the action conflict determination quantity of the current action segment for: ; in, This is the action conflict determination value. A larger value indicates a more inconsistent relationship between the state variables and the result fluctuations in the current action segment; a smaller value indicates a more consistent relationship between the state variables and the result fluctuations in the current action segment. When the action conflict determination value is higher than the preset conflict threshold, the current action segment is determined to be a composite fluctuation action segment.

[0074] The preset conflict threshold is determined based on the statistical results of the training samples. In a preferred embodiment, the action conflict determination quantity is first calculated for action segments known to have multi-factor coupling anomalies and action segments known not to have multi-factor coupling anomalies in the training samples. Then, the preset conflict threshold is determined based on the distribution interval of the action conflict determination quantities for the two types of action segments. Further, the mean of the action conflict determination quantities for action segments known not to have multi-factor coupling anomalies plus one standard deviation can be taken as the initial preset conflict threshold, and then adjusted to a position with higher recognition accuracy based on validation samples.

[0075] As a set of feasible parameter settings, the preset conflict threshold can be set to 0.15 to 0.35; for training scenarios with short continuous action chains and smooth action transitions, the preset conflict threshold is preferably set to 0.18 to 0.25; for adversarial training scenarios with long continuous action chains and frequent action transitions, the preset conflict threshold is preferably set to 0.25 to 0.32.

[0076] The composite fluctuation action segment refers to an action segment in which at least two of the following state variables—attitude offset, basketball control fluctuation, action rhythm mismatch, and load coupling fluctuation—exhibit significant fluctuations, and these fluctuations jointly affect the action result fluctuation.

[0077] Adaptive adjustment of state thresholds is performed on composite undulating motion segments. Adaptive state threshold adjustment refers to dynamically correcting the judgment thresholds for attitude offset and basketball control fluctuations based on the load coupling fluctuation and motion rhythm mismatch of the current motion segment. Let the basic judgment threshold for attitude offset be... The basic threshold for determining the fluctuation of basketball control is Then the corrected pose threshold corresponding to the current action segment and corrected basketball threshold They are respectively: ; ; in, To correct the attitude threshold, To correct the basketball threshold, This serves as the basic threshold for determining attitude offset. This serves as the basic threshold for determining the fluctuation amount in basketball control. This refers to load coupling fluctuations. This refers to the amount of mismatch in movement rhythm. , , and This parameter adjusts the threshold. Through this process, when heart rate fluctuations are large or movement rhythm changes significantly, the fixed judgment criteria are no longer directly applied; instead, the corresponding judgment threshold is automatically adjusted based on the changes in the current movement segment's state.

[0078] After completing the adaptive adjustment of the state threshold, the dominant factor identification and associated factor identification are performed on the composite oscillating motion segment. Dominant factor identification refers to identifying the state variables that have the greatest impact on the resulting oscillation from among attitude offset, basketball control oscillation, motion rhythm mismatch, and load coupling oscillation. Associated factor identification refers to identifying other state variables that interact with the dominant factor and amplify the resulting oscillation. Let the four state variables in the current motion segment be... , , and The corresponding fragment transfer differences are respectively , , and Then the corresponding dominant value , , and They are respectively: ; ; ; ; in, , , and These represent the dominant values ​​corresponding to attitude deviation, basketball control fluctuation, motion rhythm mismatch, and load coupling fluctuation, respectively. To prevent extremely small positive numbers with a denominator of zero, the state variable with the largest dominant effect value is taken as the dominant factor; other state variables that simultaneously exceed the corresponding correction threshold in the same action segment as the dominant factor are identified as related factors.

[0079] After identifying the dominant and related factors, the execution results of the action segment are reconstructed through attribution. Attribution reconstruction involves establishing a correspondence between the dominant and related factors and the action completion rate, action stability, and action result fluctuation in the result performance parameters to form the attribution result for the current action segment. When the dominant factor is posture offset, combining joint angle deviation, center of gravity offset, release point position offset, and force direction offset, the current abnormal performance is determined to be a posture imbalance type anomaly. When the dominant factor is basketball control fluctuation, combining basketball trajectory offset, basketball speed fluctuation, and basketball position dispersion, the current abnormal performance is determined to be a basketball control instability type anomaly. When the dominant factor is action rhythm mismatch, combining action start duration deviation, action transition duration deviation, and action completion duration deviation, the current abnormal performance is determined to be an action rhythm misalignment type anomaly. When the dominant factor is load coupling fluctuation, combining instantaneous heart rate change, heart rate recovery rate deviation, action duration deviation, and high-intensity action density, the current abnormal performance is determined to be a load coupling fluctuation type anomaly. For action segments with related factors, the related factors are written into the attribution results in sequence, forming a hierarchical attribution structure of dominant factor-related factor.

[0080] After completing the result attribution reconstruction, the attribution results of all motion segments within the same analysis period are aggregated to generate motion standardization analysis results, technical ability analysis results, physical fitness analysis results, and technical weakness location results.

[0081] When aggregating the attribution results of all motion segments within the same analysis period, the frequency of occurrence of dominant factors, the frequency of occurrence of related factors, the location distribution of corresponding motion segments, and the correspondence between them and the fluctuation of motion results are statistically analyzed according to motion type. When the number of consecutive occurrences of the same dominant factor under the same motion type reaches a preset number, the abnormal performance corresponding to the dominant factor is identified as a stability technical weakness; when the same dominant factor only occurs concentratedly under high load coupled fluctuation conditions, the abnormal performance corresponding to the dominant factor is identified as a physical fitness coupled technical weakness.

[0082] As a preferred embodiment, the preset number of times can be 3 times, 5 times or 8 times; 3 times is preferred when the analysis period is short, and 5 times or 8 times is preferred when the analysis period is long.

[0083] The results of the motion standardization analysis are generated based on posture deviation and its attribution, including shooting posture standard, dribbling stability, passing motion standard, and defensive posture standard. The results of the technical ability analysis are generated based on basketball control fluctuation, result performance parameters, and their attribution, including shooting percentage, dribbling success rate, passing accuracy, and defensive success rate. The results of the physical condition analysis are generated based on load coupling fluctuation and its attribution, including sports endurance evaluation value, explosive power evaluation value, and physical recovery speed evaluation value. The results of the technical weakness location are generated based on the hierarchical attribution structure of dominant factors and related factors, including the motion type of the weakness, the name of the dominant factor, the name of the related factor, and the position of the corresponding motion segment.

[0084] Finally, the basketball motion analysis results are output. These results include analysis of movement standardization, technical ability, physical condition, and identification of technical weaknesses. The identification of technical weaknesses includes not only deviations but also the dominant and related factors corresponding to those deviations, as well as the distribution of the corresponding movement segments within the analysis period.

[0085] After identifying technical weaknesses, the analysis results are categorized into four types: movement standardization, technical ability, physical condition, and technical weaknesses, generating basketball sports analysis results. These results include at least: shooting form standardization, dribbling stability, passing form standardization, defensive form standardization; shooting percentage, dribbling success rate, passing accuracy, defensive success rate; sports endurance evaluation value, explosiveness evaluation value, physical recovery speed evaluation value; and technical weaknesses corresponding to the above analysis indicators.

[0086] In step four, the basketball motion analysis results generated in step three are categorized and organized. These results are grouped according to the analysis results of movement standardization, technical ability, physical condition, and technical weakness identification. Scoring indicators corresponding to the quantitative scores are then extracted for each group. These scoring indicators include shooting posture standardization, dribbling stability, passing motion standardization, defensive posture standardization, shooting percentage, dribbling success rate, passing accuracy, defensive success rate, sports endurance evaluation value, explosiveness evaluation value, physical recovery speed evaluation value, and the impact value of technical weaknesses.

[0087] All scoring indicators undergo a standardized scoring process. Standardized scoring means converting different forms of analysis results into indicator scores within the same scoring range. For analysis results already expressed as percentages, including shooting percentage, dribbling success rate, passing accuracy, and defensive success rate, the corresponding percentage is directly used as the indicator score. For results from movement standardization analysis and physical condition analysis, the degree of deviation is first determined based on the magnitude of the deviation between the corresponding analysis value and the reference standard value, and then the degree of deviation is mapped to the indicator score. For the impact value of technical shortcomings, the deduction value is determined based on the degree of impact of the technical shortcomings on the corresponding specific movements. After standardized scoring, all scoring indicators are converted into scoring components that can participate in the comprehensive calculation.

[0088] The current specific action refers to the type of shooting, dribbling, passing, or defensive action corresponding to the current analysis object.

[0089] Scores are calculated for the results of the motion standardization analysis. The results include shooting posture accuracy, dribbling stability, passing motion accuracy, and defensive posture accuracy. For any motion standardization indicator, the current motion standardization analysis value is set to B, and the corresponding standard reference value for the age group and skill level is [value missing]. Then the deviation value of the action standardization index for: ; Where B is the current action standardization analysis value, For standard reference value, This represents the deviation value for proper movement. Then, based on the deviation value and a preset deviation mapping rule, a score for the proper movement index is determined. The smaller the deviation value, the higher the score; conversely, the larger the deviation value, the lower the score. The standard reference values ​​are derived from a database of standard movements that match the age and skill levels of basketball participants.

[0090] The standard movement database is used to store reference values, reference durations, and reference rhythm parameters for standard movements corresponding to different age and technical levels.

[0091] Scores are calculated for the technical ability analysis results. These results include shooting percentage, dribbling success rate, passing accuracy, and defensive success rate. For the completion rate-related analysis results, the percentage value is directly used as the base score for the corresponding technical ability indicator. When differences in the quality of movement completion need to be reflected, adjustments are made based on the characteristics of the release trajectory, force direction, center of gravity shift, and body balance as described in step three. Let the base score for a certain technical ability indicator be... The quality correction value is Then the final score C for this technical capability indicator is: in, The base score is directly obtained from the completion rate analysis results. C represents the final score for this technical ability indicator, calculated as an additional quality correction value based on the quality of the action completion. The quality correction value is used to distinguish between situations where "the result is the same but the quality of the action differs." For example, a shooting segment with the same shooting percentage but a more stable release trajectory will have a higher quality correction value; similarly, a passing segment with the same passing accuracy but a smaller deviation in the direction of force will have a higher quality correction value.

[0092] Scores are calculated for the physical fitness analysis results. These results include performance evaluation values ​​for exercise endurance, explosive power, and recovery speed. For any physical fitness index, let the current physical fitness analysis value be P, and the reference physical fitness value for the same age group and skill level be [value missing]. The difference in this physical fitness index for: ; Where P is the current physical fitness analysis value. For reference physical fitness values, This represents the difference in physical fitness indicators. Then, based on the difference in physical fitness indicators and the preset scoring mapping rules, the scores for exercise endurance, explosive power, and physical recovery speed are determined. When the difference in physical fitness indicators is positive, it indicates that the current physical fitness level is higher than the reference level, and the corresponding indicator score increases; when the difference in physical fitness indicators is negative, it indicates that the current physical fitness level is lower than the reference level, and the corresponding indicator score decreases.

[0093] Next, the impact value of the identified technical shortcomings is calculated. The impact value of a technical shortcoming refers to the deduction that a particular technical shortcoming causes to the overall score. Let the absolute value of the difference in the analytical indicators corresponding to a certain technical shortcoming be... The impact weight of this technological shortcoming in the current special operation is: Then the impact value Q of this technological shortcoming is: ; Where Q represents the impact value of technological shortcomings. Technical shortcomings affect weighting. This represents the absolute value of the difference in the analytical indicators corresponding to the technical weakness. The influence weight of the technical weakness is used to characterize the degree of impact of the technical weakness on overall performance. For example, a low shooting release point has a higher influence weight on shooting-specific scores than on defensive-specific scores, and insufficient physical endurance has a higher influence weight on physical fitness-specific scores than on single-pass scores. By calculating the influence value of the technical weakness, the technical weakness identification results in step three can be converted into quantifiable deduction items.

[0094] After calculating the scores for each scoring indicator, a comprehensive weighted average is applied to all indicators to generate a composite score. Let the score for the action standardization indicator be... The technical capability index score is The physical fitness index score is If the impact value of the technical shortcomings is Q, then the overall score value S is: ; in, , and The scoring weights are respectively for movement standardization, technical ability, and physical fitness, and satisfy the following conditions: ; The scoring weights for movement standardization and physical fitness represent the proportion of influence of the degree of movement standardization on the overall score. The scoring weights for technical ability and physical fitness represent the proportion of influence of the level of completion of a specific movement on the overall score. The scoring weights are determined based on the age and skill level of the basketball player. When the player is in the basic training stage, the scoring weight for movement standardization is increased; when the player is in the specialized training stage, the scoring weight for technical ability is increased; and when the player is in the high-intensity training stage, the scoring weight for physical fitness is increased.

[0095] After obtaining the overall score, the score is categorized into levels. Level categorization refers to mapping the overall score to a corresponding ability level based on a preset scoring range. The ability level characterizes the overall training level of the basketball player within the current analysis period. If the overall score is categorized into Excellent, Good, Pass, and Needs Improvement levels, the corresponding level result is output based on the scoring range the overall score falls into. The scoring range is preset according to age group, skill level, and training stage to ensure the comparability of scoring results across different age groups and skill levels.

[0096] After completing the grading, the improvement trend of the overall score and individual score is calculated. The improvement trend refers to the direction and magnitude of score changes for basketball athletes across different analysis periods. Let the overall score for the current analysis period be... The overall score for the previous analysis period was Then the change in the overall score ΔS is: ; Where ΔS is the change in the overall score. This is the overall score for the current analysis period. This represents the overall score for the previous analysis period. A positive change in the overall score indicates improved overall performance; a negative change indicates a decline; and a change close to zero indicates relatively stable overall performance. Using the same method, the corresponding changes in shooting, dribbling, passing, defense, and physical fitness scores are calculated to obtain the individual improvement trend results.

[0097] After calculating the overall score, classifying the levels, and calculating the progress trend, the results are merged to generate a basketball performance score. This score includes sub-scores, an overall score, ability levels, and a progress trend. Sub-scores include shooting, dribbling, passing, defense, and physical fitness scores; the overall score is the total score value; the ability level is the ability level corresponding to the overall score value; and the progress trend includes the change in the overall score and the changes in each sub-score.

[0098] This invention also proposes a basketball sports analysis platform, including: a basketball sports time-series data generation module, a precise labeled data generation module, a basketball sports analysis module, a basketball sports scoring module, and signal connections between the modules; The basketball motion time series data generation module is used to synchronously collect raw video data, raw motion sensor data, and raw auxiliary label data, and complete time alignment with the effective video frame time as the reference time axis, and complete the unified mapping of human body position and basketball position in the court coordinate system; then, the image posture information, motion physics information, and event information are cleaned, standardized, and weighted and fused to generate basketball motion time series data. The precise annotation data generation module is used to analyze the current action segment in the basketball motion time series data frame by frame, and to build a core key point annotation system based on the target localization of the human body area and the basketball area; then, combined with annotation confidence, adjacent frame position constraints and motion trajectory constraints, the key point annotation results are corrected for errors and temporally correlated to generate precise annotation data. The basketball motion analysis module uses motion segments as analysis units to establish a correspondence between the precisely labeled data, the basketball motion time series data, and the basic information corresponding to the basketball motion object, constructing motion analysis input data. Then, it maps human posture representation data, basketball movement representation data, motion time rhythm representation data, and physiological load representation data into a unified parameter structure, establishing intra-segment correspondences and inter-segment change relationships. Based on the state transition differences between adjacent motion segments, it identifies composite fluctuating motion segments. Subsequently, it adaptively adjusts the judgment thresholds corresponding to each state quantity in the composite fluctuating motion segments, distinguishing between the dominant factors that have the greatest impact on the result fluctuation and the associated factors that simultaneously exceed the threshold, establishing a hierarchical attribution structure for the dominant and associated factors, and generating basketball motion analysis results. The basketball scoring module is used to group the basketball analysis results and extract corresponding scoring indicators. The technical weakness identification results are converted into technical weakness impact values ​​according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific action. Then, the action standardization index score, technical ability index score, physical condition index score, and technical weakness impact value are incorporated into a unified scoring structure to construct a comprehensive score value. The comprehensive score value is mapped to the corresponding ability level result according to the preset scoring interval. The scoring change relationship is established based on the difference between the comprehensive score value and the previous analysis cycle and the difference between the scores of each sub-item, and the basketball scoring result is generated.

[0099] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0100] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0101] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0102] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0103] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A basketball motion analysis method, characterized in that, include: The raw video data, raw motion sensing data, and raw auxiliary label data are collected synchronously, and time alignment is completed using the effective video frame time as the reference time axis. The unified mapping of human body position and basketball position is completed in the field coordinate system. Then, the image posture information, motion physics information, and event information are cleaned, standardized, and weighted and fused to generate basketball motion time series data. The current action segment in the basketball motion time series data is analyzed frame by frame, and a core key point annotation system is constructed based on the target localization of the human body area and the basketball area; By combining annotation confidence, adjacent frame position constraints, and motion trajectory constraints, error correction and temporal correlation are performed on the key point annotation results to generate accurate annotation data; Using action segments as analysis units, a correspondence is established between the precisely labeled data, the basketball movement time sequence data, and the basic information corresponding to the basketball movement object to construct action analysis input data; Then, human posture representation data, basketball movement representation data, action time rhythm representation data, and physiological load representation data are mapped into a unified parameter structure to establish intra-segment correspondence and inter-segment change relationships. Based on the state transition differences between adjacent action segments, composite fluctuating action segments are identified. Subsequently, the judgment thresholds corresponding to each state quantity in the composite fluctuating action segments are adaptively adjusted to distinguish the dominant factors that have the greatest impact on the result fluctuation and the associated factors that exceed the threshold at the same time. A hierarchical attribution structure of dominant factors and associated factors is established to generate basketball motion analysis results. The basketball sports analysis results are grouped and corresponding scoring indicators are extracted. The technical weakness identification results are converted into technical weakness impact values ​​according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific action. Then, the action standardization index score, technical ability index score, physical condition index score and technical weakness impact value are incorporated into a unified scoring structure to construct a comprehensive score value. The comprehensive score value is mapped to the corresponding ability level result according to the preset scoring interval. The scoring change relationship is established based on the difference between the comprehensive score value and the previous analysis period and the difference between the scores of each sub-item, and the basketball sports scoring result is generated.

2. The basketball motion analysis method according to claim 1, characterized in that: By simultaneously collecting raw video data, raw motion sensing data, and raw auxiliary labeling data, human target position, basketball target position, and action area features are extracted from the raw video data to construct a video posture feature sequence. The original motion sensor data is collected to include basketball trajectory, number of dribbles, passing force, shooting speed, heart rate, and exercise duration, and sampling time markers are added to construct a sequence of motion physical characteristics. The original auxiliary labeled data is encoded into events to construct an event-labeled feature sequence.

3. The basketball motion analysis method according to claim 2, characterized in that: Using the effective video frame time as the reference time axis, the video pose feature sequence, motion physics feature sequence and event marker feature sequence are time-aligned, and a field coordinate system is established to uniformly map the human body position and basketball position to the same spatial reference. Then, data cleaning, standardization and weighted fusion processing are performed on the three types of feature sequences to construct basketball motion time-series data that associates image pose information, motion physics information and event information in time order.

4. The basketball motion analysis method according to claim 3, characterized in that: The current action segment in basketball motion time series data is analyzed frame by frame to extract the video pose feature sequence, motion physics feature sequence, and event marker feature sequence corresponding to the current action segment. Target localization of the human body region and basketball region is performed in the video pose feature sequence, and a core key point annotation system containing fifteen human body key points and three basketball key points is constructed. Based on the human body region features and basketball region features, initial annotations are performed on the head, neck, left and right shoulders, left and right elbows, left and right hands, waist, left and right knees, left and right feet, torso center, as well as the center of the ball, the upper vertex of the ball, and the lower vertex of the ball, obtaining the corresponding initial annotation coordinates and annotation confidence.

5. The basketball motion analysis method according to claim 4, characterized in that: For annotation points below a preset confidence threshold, adjacent frame position constraints and motion trajectory constraints are introduced for error correction. Adjacent frame position constraints are used to extract the corresponding positional relationship of the same key point in consecutive image frames. Motion trajectory constraints are used to call basketball trajectory data, dribbling rhythm data, passing force data, and shooting speed data to constrain the consistency of the positions of hand key points, basketball key points, and lower limb support key points. The corrected annotation coordinates are determined by weighted calculation of the initial annotation coordinates, the effective annotation coordinates of the previous image frame, and the trajectory reference coordinates. Based on this, the eighteen core key points in all image frames are associated in chronological order to establish a key point temporal annotation sequence. The connection relationship of human joints and the connection relationship of basketball trajectory are further calculated to generate accurate annotation data.

6. The basketball motion analysis method according to claim 5, characterized in that: By summarizing precisely labeled data, basketball movement time-series data, and the age, height, weight, and skill level of basketball players, and establishing correspondences according to movement segments, motion analysis input data is constructed. Specifically, for situations where changes in human key points, basketball key points, basketball trajectory, movement rhythm, and heart rate exist asynchronously, with different amplitudes, directions, and time sequences between adjacent movement segments during continuous or competitive training, movement segments are used as analysis units. The positions of human key points, basketball key points, human joint connections, and basketball trajectory connections are integrated with basketball trajectory, dribbling frequency, passing force, shooting speed, heart rate, exercise duration, and movement event information. Combined with age, height, weight, and skill level, a motion state parameter group is constructed, including posture deviation parameters, basketball control parameters, movement rhythm parameters, load fluctuation parameters, and result performance parameters. By mapping human posture representation data, basketball movement representation data, movement time rhythm representation data, and physiological load representation data to a unified parameter structure under the same movement segment, intra-segment correspondences and inter-segment change relationships between different types of state variables are established.

7. The basketball motion analysis method according to claim 6, characterized in that: The following parameters are extracted: posture deviation (considered by joint angle deviation, center of gravity shift, release point position shift, and force direction shift); basketball control fluctuation (considered by basketball trajectory shift, basketball speed fluctuation, and basketball position dispersion); movement rhythm mismatch (considered by movement initiation duration deviation, movement transition duration deviation, movement completion duration deviation, and adjacent movement interval fluctuation); load coupling fluctuation (considered by instantaneous heart rate change, heart rate recovery rate deviation, movement duration deviation, and high-intensity movement density); and result fluctuation (considered by movement completion rate deviation, movement stability deviation, and movement result fluctuation). The system measures and constructs a sequence of motion fluctuation states according to the order of motion segments. Then, it compares the various state quantities between the current motion segment and the previous motion segment, calculates the state transition difference between adjacent motion segments, and constructs a motion conflict judgment quantity using the transition difference relationship between posture offset, basketball control fluctuation, motion rhythm mismatch, load coupling fluctuation, and result fluctuation. Motion segments with inconsistent relationships between changes in various state quantities and changes in results are identified as composite fluctuation motion segments. This allows the identification of multi-factor coupling situations where the same abnormal result corresponds to different causes and different abnormal results originate from the same potential factor from the continuous motion chain.

8. The basketball motion analysis method according to claim 7, characterized in that: After identifying the composite undulating motion segment, adaptive adjustments are made to the judgment thresholds corresponding to the posture offset and basketball control fluctuation based on the load coupling fluctuation and motion rhythm mismatch in the current motion segment. Corrected posture thresholds and corrected basketball thresholds corresponding to the state changes of the current motion segment are constructed. Based on this, the dominant influence value corresponding to each state quantity is calculated according to the correspondence between posture offset, basketball control fluctuation, motion rhythm mismatch, load coupling fluctuation, and their segment transfer differences and result fluctuation transfer differences. The state quantity with the greatest impact on the result fluctuation is identified as the dominant factor from multiple state quantities. Other state quantities in the same motion segment that simultaneously exceed the corresponding correction threshold are identified as related factors, establishing a hierarchical relationship between the dominant factor and related factors. Then, a correspondence is established between the dominant factor, related factors, and motion completion rate, motion stability, and motion result fluctuation, attributing and reconstructing the motion segment execution result. Specifically, when the dominant factor is the posture offset, it is combined with joint angle deviation, center of gravity offset, and release point. Positional offset and force direction offset are used to identify movement posture imbalance anomalies. When the dominant factor is basketball control fluctuation, basketball trajectory offset, basketball speed fluctuation, and basketball position dispersion are combined to identify basketball control instability anomalies. When the dominant factor is movement rhythm mismatch, movement rhythm misalignment anomalies are combined with movement initiation duration deviation, movement transition duration deviation, and movement completion duration deviation. When the dominant factor is load coupling fluctuation, load coupling fluctuation anomalies are combined with instantaneous heart rate change, heart rate recovery rate deviation, movement duration deviation, and high-intensity movement density. Related factors are sequentially written into the attribution results to form a hierarchical attribution structure of dominant and related factors. Finally, the attribution results of all movement segments within the same analysis period are aggregated to generate movement standardization analysis results, technical ability analysis results, physical fitness analysis results, and technical weakness location results. The movement type, dominant factor name, related factor name, and corresponding movement segment position of the weakness are all merged into the basketball movement analysis results.

9. The basketball motion analysis method according to claim 8, characterized in that: The results of the analysis of movement standardization, technical ability, physical fitness, and technical weakness are grouped, and scoring indicators corresponding to each analysis result are extracted. Specifically, the technical weakness identification results obtained based on dominant factors, related factors, and hierarchical attribution structures are used to construct a technical weakness impact value according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific movement, transforming multi-factor attribution results into quantifiable deduction items. Then, the movement standardization index score, technical ability index score, physical fitness index score, and technical weakness impact value are incorporated into a unified scoring structure to construct a comprehensive score. This comprehensive score is then mapped to the corresponding ability level result according to a preset scoring interval. Based on the difference in comprehensive score between the current analysis period and the previous analysis period, as well as the difference in score for each sub-item, a scoring change relationship corresponding to changes in dominant factors and related factors is established, generating a basketball scoring result.

10. A basketball motion analysis platform, used to implement the basketball motion analysis method according to any one of claims 1-9, characterized in that, include: The system includes a basketball time-series data generation module, a precise labeled data generation module, a basketball analysis module, a basketball scoring module, and signal connections between these modules. The basketball motion time series data generation module is used to synchronously collect raw video data, raw motion sensor data, and raw auxiliary label data, and complete time alignment with the effective video frame time as the reference time axis, and complete the unified mapping of human body position and basketball position in the court coordinate system; then, the image posture information, motion physics information, and event information are cleaned, standardized, and weighted and fused to generate basketball motion time series data. The precise annotation data generation module is used to analyze the current action segment in the basketball motion time series data frame by frame, and to build a core key point annotation system based on the target localization of the human body area and the basketball area. By combining annotation confidence, adjacent frame position constraints, and motion trajectory constraints, error correction and temporal correlation are performed on the key point annotation results to generate accurate annotation data; The basketball motion analysis module is used to construct motion analysis input data by using motion segments as analysis units and establishing a correspondence between the precise labeled data, the basketball motion time sequence data, and the basic information corresponding to the basketball motion object. Then, human posture representation data, basketball movement representation data, action time rhythm representation data, and physiological load representation data are mapped into a unified parameter structure to establish intra-segment correspondence and inter-segment change relationships. Based on the state transition differences between adjacent action segments, composite fluctuating action segments are identified. Subsequently, the judgment thresholds corresponding to each state quantity in the composite fluctuating action segments are adaptively adjusted to distinguish the dominant factors that have the greatest impact on the result fluctuation and the associated factors that exceed the threshold at the same time. A hierarchical attribution structure of dominant factors and associated factors is established to generate basketball motion analysis results. The basketball scoring module is used to group the basketball analysis results and extract corresponding scoring indicators. The technical weakness identification results are converted into technical weakness impact values ​​according to the absolute value of the difference between the corresponding analysis indicators and their influence weight in the current specific action. Then, the action standardization index score, technical ability index score, physical condition index score, and technical weakness impact value are incorporated into a unified scoring structure to construct a comprehensive score value. The comprehensive score value is mapped to the corresponding ability level result according to the preset scoring interval. The scoring change relationship is established based on the difference between the comprehensive score value and the previous analysis cycle and the difference between the scores of each sub-item, and the basketball scoring result is generated.

Images