Capture recognition method for competitive posture analysis

Abstract

This invention discloses a capture and recognition method for analyzing competitive postures, relating to the field of video analysis technology. It acquires a competitive video stream and identifies the spatial coordinates of key human body points and their corresponding recognition confidence levels, forming key point coordinate data and key point confidence data respectively. An initial key point data set is then constructed from the key point coordinate data and key point confidence data. By performing temporal slicing on the co-occurring text of target entity pairs and constructing a global observation matrix, entity relationships can be dynamically represented in a continuous time dimension, thus solving the problem of existing relationship extraction methods lacking temporal modeling capabilities.

Description

Technical Field

[0001] This invention relates to the field of video analysis technology, specifically to a capture and recognition method for analyzing competitive postures. Background Technology

[0002] With the development of digital training and event analysis technology in competitive sports, the automatic analysis of athletes' postures using visual perception technology has gradually become an important means of competitive sports analysis. In sports such as gymnastics, martial arts, figure skating, and diving, athletes produce continuous and rapidly changing human posture structures when performing complex movements. In order to objectively analyze these posture changes, it is usually necessary to capture and structurally identify the positional relationships of key parts of the human body through video data. This typically relies on identifying and tracking key points of the human body in the competitive video stream, and then analyzing the human posture structure based on this to meet the application needs of competitive movement quality assessment, movement training feedback, and event auxiliary judgment.

[0003] In existing technologies, most competitive posture analysis relies on key point detection models to directly output human joint coordinates, and then uses simple time series smoothing or filtering to stabilize the key point trajectory.

[0004] However, in real-world competitive scenarios, athletes are often in high-speed motion, with limbs crossed and obstructed, or in extreme joint angles, such as somersaults, rotations, and rapid landings. In these situations, some key points are prone to decreased recognition confidence or shifted detection positions. Existing methods typically lack mechanisms for dynamically compensating for the reliability of key points, and also lack the ability to incorporate spatial constraints based on human biomechanical structures.

[0005] The lack of effective means to predict and correct the structure of low-confidence key points makes it easy for the posture recognition results in the existing technology to produce structural distortion, thus making it difficult to accurately reflect the real human posture state. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a capture and recognition method for analyzing competitive postures, thus solving the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a capture and recognition method for analyzing competitive postures, comprising the following steps:

[0008] S1. Acquire the competitive video stream and identify the spatial coordinates of key human body points and their corresponding recognition confidence levels, forming key point coordinate data Kcd and key point confidence data Kcf respectively, and constructing the initial key point data set Kps from the key point coordinate data Kcd and key point confidence data Kcf.

[0009] S2. Based on the connection relationship of the human skeleton, establish a human biomechanical constraint model, perform structural constraint processing on the key point coordinate data Kcd in the initial key point data set Kps, and generate a constraint posture data set Bps.

[0010] S3. Based on the trend of human body key point motion change between adjacent video frames, predict and compensate the key point coordinate data Kcd corresponding to the key point confidence data Kcf in the initial key point data set Kps being lower than the preset threshold, and generate the predicted posture data set Pps.

[0011] S4. Map the constrained attitude data set Bps and the predicted attitude data set Pps to the dynamic optimization space Dyn, and perform collaborative optimization processing on attitude deviation and human body structure constraints in the dynamic optimization space Dyn to generate the optimized attitude result set Ops.

[0012] S5. Reconstruct the human posture structure based on the optimized posture result set Ops, and output the competitive human posture structured result set Res.

[0013] Preferably, S1 includes S11;

[0014] S11. Deploy video acquisition equipment in the competitive sports scene to continuously capture video of the athletes' competitive actions, thereby obtaining the original competitive video stream. The original competitive video stream is processed by frame sequence parsing in chronological order, and the original competitive video stream is split into a continuous video frame image sequence, so that each video frame image corresponds to the moving human image information of a time sampling point.

[0015] Human target detection processing is performed on the video frame image sequence to identify the moving human body region in the video frame image. Specifically, the target detection calculation is performed on the video frame image based on the YouOnlyLookOnce target detection algorithm. The human body region in the video frame image is determined by the human body detection bounding box output by the algorithm, and the corresponding human body region image data is extracted according to the human body detection bounding box.

[0016] After obtaining human body region image data, human body posture key point recognition processing is performed on the human body region image data. The human body posture key point recognition processing is based on the OpenPose human body posture recognition algorithm to extract features and calculate key points from the human body region image data, and generate key point heatmaps corresponding to human body key points.

[0017] After obtaining the human body key point heatmap, peak position detection is performed on each key point heatmap to extract the two-dimensional position coordinates of each human body key point in the image coordinate system. Then, the position coordinates of each human body key point are organized according to the human skeletal structure order, and all human body key point coordinates are uniformly summarized to form key point coordinate data Kcd.

[0018] Preferably, S1 further includes S12;

[0019] S12. After obtaining the key point coordinate data Kcd, the recognition reliability assessment of the human key point recognition results is performed: Specifically, based on the key point heat map response results output by the OpenPose human pose recognition algorithm, the response intensity value of the key point heat map corresponding to each human key point is read, and the response intensity value corresponding to the peak position in the key point heat map is obtained, and the response intensity value is used as the recognition confidence of the corresponding human key point.

[0020] Then, according to the order of human body key points, the recognition confidence scores corresponding to all human body key points are sorted out and key point confidence data Kcf is formed;

[0021] After obtaining the key point confidence data Kcf, the key point coordinate data Kcd and the key point confidence data Kcf are associated one by one according to the index order of human key points, so that each human key point has corresponding spatial coordinate information and recognition confidence information at the same time.

[0022] After completing the corresponding association, the key point coordinate data Kcd and the key point confidence data Kcf are combined according to a unified data structure to construct the initial key point data set Kps.

[0023] Preferably, S2 includes S21;

[0024] S21. After obtaining the initial set of key point data Kps, read the key point coordinate data Kcd and analyze the connection relationship between human key points according to the human skeleton topology: construct the skeleton connection relationship of each human key point in the key point coordinate data Kcd according to the connection order between each human key point in the human skeleton structure, so as to obtain the skeleton connection link in the human skeleton structure.

[0025] The skeletal connection links are used to represent the skeletal connection relationships between key points in the human body;

[0026] After completing the construction of the skeletal connection relationship, based on the connection relationship between each adjacent human key point in the skeletal connection relationship, the bone length between adjacent human key points in the key point coordinate data Kcd is calculated, and the stability of human bone length is used as the constraint basis to establish a human biomechanical constraint model for the human skeletal structure.

[0027] The human biomechanical constraint model is used to represent the range of length variation of each bone segment in the human skeletal structure and the kinematic structural relationship between adjacent key points in the human body.

[0028] Preferably, S2 further includes S22;

[0029] S22. After establishing the human biomechanical constraint model, input the key point coordinate data Kcd from the initial key point data set Kps into the human biomechanical constraint model to perform structural rationality detection processing on the human skeletal structure.

[0030] Specifically, based on the skeletal connection links and skeletal length constraint rules defined in the human biomechanical constraint model, the spatial distance between adjacent human key points in the key point coordinate data Kcd is calculated, and it is determined whether the spatial distance conforms to the human skeletal length constraint range.

[0031] When a bone segment is detected to exceed the human bone length constraint range, structural correction processing is performed on the position of the corresponding human key point in the key point coordinate data Kcd, so that the position of the human key point is adjusted to a reasonable bone length range while maintaining the bone connection link.

[0032] After completing the structural constraint detection and human key point position correction processing for all skeletal segments, the corrected key point coordinate data Kcd is uniformly organized, and the human posture structure data is reorganized according to the skeletal connection links to generate the constrained posture data set Bps.

[0033] Preferably, S3 includes S31;

[0034] S31. After obtaining the initial set of key point data Kps, read the key point coordinate data Kcd in the current video frame and the previous video frame, and calculate the motion change trend of the human body key points based on the changes in the position of human body key points between adjacent video frames.

[0035] Specifically, the key point coordinate data Kcd in the current video frame is matched one by one with the key point coordinate data Kcd in the previous video frame, and the position change of the same human body key point between consecutive video frames is calculated, so as to obtain the motion displacement information of the human body key point.

[0036] Then, based on the positional changes of human key points between consecutive video frames, the direction and amplitude of motion of human key points in image space are determined, thereby obtaining the motion change trend of human key points between adjacent video frames.

[0037] After calculating the movement trends of all key points on the human body, the movement trends of each key point are organized and compiled to form the movement trend data of key points on the human body.

[0038] Preferably, S3 further includes S32;

[0039] S32. After obtaining the motion trend data of human key points, the key point confidence data Kcf in the initial key point data set Kps is detected and processed to identify human key points with confidence levels lower than a preset threshold.

[0040] Specifically, based on the recognition confidence values ​​of each human key point in the key point confidence data Kcf, it is determined whether the recognition confidence of each human key point is lower than the preset confidence threshold.

[0041] When the confidence level of a certain human body key point recognition is lower than the preset confidence threshold, it is considered that the recognition of that human body key point is unstable.

[0042] Subsequently, based on the direction and amplitude of movement of the corresponding human key points in the human key point movement trend data, the position of the corresponding human key point in the key point coordinate data Kcd is predicted and calculated, thereby obtaining the predicted position coordinates of the human key point.

[0043] After completing the calculation of all low-confidence human keypoint prediction positions, the predicted human keypoint position data are organized according to the order of human skeletal structure and formed into a predicted pose data set Pps.

[0044] Preferably, S4 includes S41;

[0045] S41. Construct a dynamic optimization space Dyn, which is a multi-dimensional continuous vector space. Each dimension of the dynamic optimization space Dyn corresponds to the X / Y coordinates of human key points and the deviation of bone length constraints, and is used to uniformly represent the information of constrained posture and predicted compensated posture. The dynamic optimization space Dyn combines the key point coordinate data Kcd in the constrained posture data set Bps after bone structure constraint correction and the predicted compensated key point coordinate data Kcd in the predicted posture data set Pps according to the topological order of human bones to form a vector, realize vectorized representation and use it as optimization input.

[0046] Subsequently, the key point coordinate data Kcd of each key point in the constrained attitude data set Bps and the key point coordinate data Kcd of the corresponding predicted compensation key point in the predicted attitude data set Pps are integrated according to the topological order of the human skeleton and mapped to the dynamic optimization space Dyn to form an attitude state vector.

[0047] Preferably, S4 further includes S42;

[0048] S42. Perform co-optimization processing on the attitude state vector mapped to the dynamic optimization space Dyn. During the co-optimization process, calculate the attitude deviation loss (squared Euclidean distance between the predicted coordinates and the constraint coordinates) for each key point. At the same time, add the human skeleton length constraint and skeleton connection link as penalty terms to the optimization objective function. Implementers can use numerical optimization algorithms (such as gradient descent or quasi-Newton method) to iteratively adjust the attitude state vector to minimize the attitude deviation and satisfy the structural constraint conditions at the same time.

[0049] After optimization, the optimized posture state vector is converted back into key point coordinate data Kcd according to the human skeletal structure order, forming the optimized posture result set Ops.

[0050] Preferably, S5 includes S51;

[0051] S51. After obtaining the optimized posture result set Ops, the key point coordinate data Kcd of each key point in the set is reorganized according to the human skeleton topology order to reconstruct the complete human skeleton structure.

[0052] Specifically, the optimized key point coordinates Kcd are arranged according to the skeletal connection links, and key points are connected one by one to form skeletal segments. The skeletal structure of key parts of the human body such as the head, torso, and limbs is constructed to achieve complete spatial reconstruction of human posture.

[0053] After completing the reconstruction of the human skeletal structure, the reconstructed complete human skeletal structure and the spatial coordinates Kcd of each key point are organized into a set of structured results of competitive human posture, Res.

[0054] This invention provides a capture and recognition method for analyzing athletic postures, which has the following beneficial effects:

[0055] By performing temporal slicing on the co-occurring text of target entity pairs and constructing a global observation matrix, entity relationships can be dynamically represented in the continuous time dimension, thereby solving the problem that existing relationship extraction methods lack temporal modeling capabilities.

[0056] By performing singular value decomposition on the global observation matrix and extracting the dominant relation semantic trajectory, the main semantic information in the relation evolution process can be separated from noise disturbances, thus solving the problems of high noise and poor relation recognition stability in open domain web page data.

[0057] By combining discrete Friesian distance and Shannon entropy for dual verification of macroscopic geometric morphology and microscopic semantic determinism, the weighting results of relational predicates are made more reliable, thus solving the problem of insufficient accuracy of relational triples under traditional static prediction methods. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the capture and recognition method for analyzing competitive postures according to the present invention;

[0059] Figure 2 This is a flowchart of the skeleton constraint and key point prediction compensation process of the present invention. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0061] Example 1

[0062] This invention provides a capture and recognition method for analyzing athletic postures; please refer to [link / reference]. Figure 1 This includes the following steps:

[0063] S1. Acquire the competitive video stream and identify the spatial coordinates of key human body points and their corresponding recognition confidence levels, forming key point coordinate data Kcd and key point confidence data Kcf respectively, and constructing the initial key point data set Kps from the key point coordinate data Kcd and key point confidence data Kcf.

[0064] S2. Based on the connection relationship of the human skeleton, establish a human biomechanical constraint model, perform structural constraint processing on the key point coordinate data Kcd in the initial key point data set Kps, and generate a constraint posture data set Bps.

[0065] S3. Based on the trend of human body key point motion change between adjacent video frames, predict and compensate the key point coordinate data Kcd corresponding to the key point confidence data Kcf in the initial key point data set Kps being lower than the preset threshold, and generate the predicted posture data set Pps.

[0066] S4. Map the constrained attitude data set Bps and the predicted attitude data set Pps to the dynamic optimization space Dyn, and perform collaborative optimization processing on attitude deviation and human body structure constraints in the dynamic optimization space Dyn to generate the optimized attitude result set Ops.

[0067] S5. Reconstruct the human posture structure based on the optimized posture result set Ops, and output the competitive human posture structured result set Res.

[0068] In this embodiment, through the implementation of this method, firstly, the initial keypoint data set Kps can accurately record the spatial coordinate data Kcd and the corresponding recognition confidence data Kcf of the human keypoints in each frame of video. This allows keypoints with low confidence to be clearly identified and marked during continuous competitive movements, solving the problem of unstable keypoint recognition leading to posture loss or large deviations in traditional methods. For example, when an athlete completes a jump, the hand or foot keypoints may have low confidence due to high-speed movement. By predicting and compensating the keypoint coordinate data Kcd that is lower than a preset threshold, the predicted posture data set Pps can restore the reasonable positions of the hand and foot keypoints, thereby maintaining the continuity of the movement. Secondly, the keypoint coordinate data Kcd is structurally constrained by a human biomechanical constraint model based on the human skeletal connection relationship, generating a constrained posture data set Bps. This effectively ensures the rationality of the length of the human skeleton and joint connections, avoiding abnormal situations such as bone segment stretching or joint misalignment under high-speed or complex movements. For example, it maintains the stable structure of the shoulder and hip keypoints during rapid rotation or bending movements. Subsequently, the constrained posture data set Bps and the predicted posture data set Pps are mapped to the dynamic optimization space Dyn, where posture deviations and human structural constraints are co-optimized to generate an optimized posture result set Ops. This optimizes both the predicted compensation position and structural constraints simultaneously, ensuring that the position of each keypoint conforms to both continuous motion trends and skeletal biomechanical constraints. Finally, the human skeletal structure is reconstructed based on the optimized posture result set Ops, outputting a complete structured result set Res for competitive human postures. This achieves high-precision, structurally sound, and continuously stable competitive human posture analysis. Overall, this method addresses issues such as keypoint loss, abnormal skeletal structure, and insufficient motion continuity in traditional video posture analysis. It provides a reliable data foundation for accurate posture analysis of high-speed movements in real-world competitive scenarios such as high jump, sprinting, and throwing, ensuring the accuracy of motion evaluation and training feedback.

[0069] Example 2

[0070] Specifically: S1 includes S11;

[0071] S11. Deploy video acquisition equipment in the competitive sports scene to continuously capture video of the athletes' competitive actions, thereby obtaining the original competitive video stream. The original competitive video stream is processed by frame sequence parsing in chronological order, and the original competitive video stream is split into a continuous video frame image sequence, so that each video frame image corresponds to the moving human image information of a time sampling point.

[0072] Human target detection processing is performed on the video frame image sequence to identify the moving human body region in the video frame image. Specifically, the target detection calculation is performed on the video frame image based on the YouOnlyLookOnce target detection algorithm. The human body region in the video frame image is determined by the human body detection bounding box output by the algorithm, and the corresponding human body region image data is extracted according to the human body detection bounding box.

[0073] After obtaining human body region image data, human body posture key point recognition processing is performed on the human body region image data. The human body posture key point recognition processing is based on the OpenPose human body posture recognition algorithm to extract features and calculate key points from the human body region image data, and generate key point heatmaps corresponding to human body key points.

[0074] After obtaining the human body key point heatmap, the peak position of each key point heatmap is detected, the two-dimensional position coordinates of each human body key point in the image coordinate system are extracted, and then the position coordinates of each human body key point are organized according to the human skeletal structure. Finally, all human body key point coordinates are summarized to form key point coordinate data Kcd.

[0075] It should be noted that:

[0076] The YouOnlyLookOnce target detection algorithm is an existing real-world target detection algorithm. It extracts features from the input image through a convolutional neural network and directly predicts the target category and target bounding box position on the feature map. In actual implementation, video frame images can be input into a pre-trained human detection network model. The human region position is determined by the human detection bounding box output by the network, and human region image data is cropped according to the bounding box position to obtain human region image data as subsequent human pose recognition input.

[0077] The OpenPose human pose recognition algorithm is an existing real human pose estimation algorithm. It extracts features from human images through a convolutional neural network and outputs a human key point heatmap. Each key point heatmap represents the probability distribution of a certain joint of the human body in the image space. By finding the position of the maximum response in the key point heatmap, the spatial position of the key point in the image coordinate system can be determined.

[0078] In practice, key points of the human body are usually defined as a fixed set of key joints according to the human skeletal structure. For example, the human skeletal structure may include the following key joint locations:

[0079] Key points for the head, neck, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle.

[0080] Implementers can read the key point heatmap output by the OpenPose human pose recognition algorithm and extract the peak position of the heatmap of the key joints to obtain the position coordinates of the human key points in the image space. All key point coordinates are then arranged in the order of human skeletal structure to form key point coordinate data Kcd.

[0081] The key point coordinate data Kcd represents the set of position coordinates of key human joints in the image space. This data reflects the spatial distribution of the human posture structure in the current video frame and serves as the basic input data for subsequent key point confidence data Kcf calculation and the construction of the initial key point data set Kps.

[0082] S1 further includes S12;

[0083] S12. After obtaining the key point coordinate data Kcd, the recognition reliability assessment of the human key point recognition results is performed: Specifically, based on the key point heat map response results output by the OpenPose human pose recognition algorithm, the response intensity value of the key point heat map corresponding to each human key point is read, and the response intensity value corresponding to the peak position in the key point heat map is obtained, and the response intensity value is used as the recognition confidence of the corresponding human key point.

[0084] Then, according to the order of human body key points, the recognition confidence scores corresponding to all human body key points are sorted out and key point confidence data Kcf is formed;

[0085] After obtaining the key point confidence data Kcf, the key point coordinate data Kcd and the key point confidence data Kcf are associated one by one according to the index order of human key points, so that each human key point has corresponding spatial coordinate information and recognition confidence information at the same time.

[0086] After completing the corresponding association, the key point coordinate data Kcd and the key point confidence data Kcf are combined according to a unified data structure to construct the initial key point data set Kps;

[0087] It should be noted that:

[0088] In the OpenPose human pose recognition algorithm, a key point heatmap is generated for each human body key point. The key point heatmap is a two-dimensional probability distribution matrix, and the value of each pixel position in the matrix represents the probability response intensity of that position belonging to the corresponding human body key point.

[0089] In the specific implementation process, the peak position of the key point heatmap and its response intensity can be determined according to the following processing method:

[0090] First, the key point heatmap is scanned pixel by pixel to read the response value of each pixel position in the key point heatmap;

[0091] Subsequently, by identifying the pixel location with the highest response value in the keypoint heatmap, this pixel location is determined as a candidate location for the corresponding human keypoint. This maximum response value is the peak response intensity of the keypoint heatmap.

[0092] After determining the peak position of the key point heatmap, the response value corresponding to the peak position is read, and the response value is used as the recognition confidence of the corresponding human key point, thereby obtaining the confidence data of the corresponding human key point;

[0093] Implementers can determine peak locations by performing a maximum value search operation on the keypoint heatmap and read the corresponding response values, thereby obtaining the keypoint recognition confidence level. This processing method converts the probability response information in the keypoint heatmap into recognition confidence data for the corresponding human keypoints in the keypoint confidence data (Kcf).

[0094] Through the above processing, it can be ensured that the key point confidence data Kcf corresponds one-to-one with the human body key point positions in the key point coordinate data Kcd, thus providing a reliable data foundation for the subsequent construction of the initial key point dataset Kps.

[0095] In this embodiment, the method first deploys video capture equipment in the competitive sports scene to continuously capture and parse the athletes' competitive movements into a sequence of video frame images. Each video frame corresponds to the moving human image information at a specific time sampling point, thus ensuring the temporal continuity and integrity of motion capture. Subsequently, human target detection based on the YouOnlyLookOnce target detection algorithm is performed on the video frame image sequence. This accurately locates the moving human body region, and by extracting the human body region image data as input, subsequent pose analysis is ensured to process only the valid human body image region, effectively avoiding background interference and confusion between multiple human bodies. For example, in sprinting or gymnastics competitions, it can stably capture the athlete's full-body image at high speeds, ensuring that key parts are not obscured or lost. Next, the OpenPose human pose recognition algorithm is used to extract features and detect key points from the human body region image data, generating a key point heatmap. Through pixel-by-pixel scanning and peak response intensity calculation, the maximum response position in the key point heatmap is converted into key point coordinate data Kcd, while the peak response intensity is used as key point confidence data Kcf. The unique advantage of this processing method lies in its ability to not only accurately convert the fuzzy spatial information of the probability heatmap into specific two-dimensional coordinates, but also to quantify the recognition reliability of each keypoint. This allows for a clear distinction between high-confidence and low-confidence keypoints during subsequent processing. For example, in gymnastic flips, the wrist or ankle may produce blurred images during high-speed rotation. Low-confidence keypoints can be identified using the keypoint confidence data (Kcf), facilitating subsequent compensation or prediction and ensuring the integrity and traceability of the keypoint data. Finally, the keypoint coordinate data (Kcd) and keypoint confidence data (Kcf) are mapped one-to-one and organized according to the human skeletal structure to form the initial keypoint dataset (Kps), providing a reliable and structured data foundation for subsequent posture constraint processing and prediction compensation. This processing method significantly improves the accuracy and reliability of keypoint acquisition under high-speed athletic movements, ensuring that each frame of posture data reflects both spatial location and quantifies confidence, providing precise and quantifiable data support for athletic posture analysis.

[0096] Example 3

[0097] Please see Figure 2 Specifically: S2 includes S21;

[0098] S21. After obtaining the initial set of key point data Kps, read the key point coordinate data Kcd and analyze the connection relationship between human key points according to the human skeleton topology: construct the skeleton connection relationship of each human key point in the key point coordinate data Kcd according to the connection order between each human key point in the human skeleton structure, so as to obtain the skeleton connection link in the human skeleton structure.

[0099] The skeletal connection links are used to represent the skeletal connection relationships between key points in the human body;

[0100] After completing the construction of the skeletal connection relationship, based on the connection relationship between each adjacent human key point in the skeletal connection relationship, the bone length between adjacent human key points in the key point coordinate data Kcd is calculated, and the stability of human bone length is used as the constraint basis to establish a human biomechanical constraint model for the human skeletal structure.

[0101] The human biomechanical constraint model is used to represent the range of length variation of each bone segment in the human skeletal structure and the kinematic structural relationship between adjacent key points of the human body.

[0102] It should be noted that:

[0103] The connection relationships of the human skeleton can be defined according to the skeletal topology commonly used in the field of human pose estimation; for example, in the OpenPose human pose recognition algorithm, the human skeleton structure is defined as a fixed key joint connection relationship, in which adjacent joints are connected by bone segments.

[0104] For example, typical connections in the human skeletal structure include: the connection between the neck key point and the left shoulder key point, the connection between the neck key point and the right shoulder key point, the connection between the left shoulder key point and the left elbow key point, the connection between the left elbow key point and the left wrist key point, the connection between the right shoulder key point and the right elbow key point, the connection between the right elbow key point and the right wrist key point, the connection between the left hip key point and the left knee key point, the connection between the left knee key point and the left ankle key point, the connection between the right hip key point and the right knee key point, and the connection between the right knee key point and the right ankle key point.

[0105] According to the above skeletal topology, implementers can connect the spatial positions of adjacent key joints in the key point coordinate data Kcd to construct the human skeleton connection relationship.

[0106] Subsequently, the length of the bone segment was determined by calculating the Euclidean distance between adjacent key joints, and human bone length constraint rules were established based on the principle of human bone length stability, thus forming a human biomechanical constraint model.

[0107] The human biomechanical constraint model is based on the bone length stability of the human skeletal structure. During human movement, the bone segment lengths between adjacent key joints usually remain basically stable in a short period of time. Therefore, this characteristic can be used to establish bone length constraint relationships.

[0108] In the specific implementation process, the human biomechanical constraint model can be established in the following manner:

[0109] First, based on the skeletal connections, the spatial coordinates of adjacent key joints are read from the key point coordinate data (Kcd). For example, the spatial coordinates of the left shoulder key point and the left elbow key point, and the spatial coordinates of the left elbow key point and the left wrist key point are read.

[0110] Subsequently, the Euclidean distance between the spatial coordinates of each pair of adjacent key joints is calculated to obtain the corresponding bone segment length. By statistically analyzing the lengths of all bone segments, the set of bone segment lengths in the current human skeletal structure can be obtained.

[0111] Then, based on the principle of proportional stability of the human skeleton, an allowable range of variation is set for each bone segment. This allowable range of variation can be determined based on the length of the bone segment in the initial posture; for example, the initial bone segment length can be used as a reference length, and this length can be allowed to vary within a certain proportional range.

[0112] In actual implementation, the length of each bone segment can be calculated one by one according to the connection relationship of the human skeleton, and the calculated bone segment length can be compared with the reference bone segment length; when the length of a certain bone segment exceeds the allowable range of variation, it is considered that there is a structural deviation at the corresponding key joint position.

[0113] Subsequently, by adjusting the position of this key joint, the length of the bone segments between adjacent key joints is restored to the allowable range of variation, thereby enabling the human skeletal structure to meet the biomechanical constraints of the human body.

[0114] Through the above processing method, the bone length constraint relationship in the human skeletal structure can be established, and the constraint relationship can be used to constrain the structural rationality of the human posture structure, thereby forming the human biomechanical constraint model.

[0115] S2 further includes S22;

[0116] S22. After establishing the human biomechanical constraint model, input the key point coordinate data Kcd from the initial key point data set Kps into the human biomechanical constraint model to perform structural rationality detection processing on the human skeletal structure.

[0117] Specifically, based on the skeletal connection links and skeletal length constraint rules defined in the human biomechanical constraint model, the spatial distance between adjacent human key points in the key point coordinate data Kcd is calculated, and it is determined whether the spatial distance conforms to the human skeletal length constraint range.

[0118] When a bone segment is detected to exceed the human bone length constraint range, structural correction processing is performed on the position of the corresponding human key point in the key point coordinate data Kcd, so that the position of the human key point is adjusted to a reasonable bone length range while maintaining the bone connection link.

[0119] After completing the structural constraint detection and human key point position correction processing of all bone segments, the corrected key point coordinate data Kcd is uniformly organized, and the human posture structure data is reorganized according to the bone connection link to generate the constraint posture data set Bps.

[0120] It should be noted that:

[0121] Skeletal length constraints in human biomechanical constraint models can be established based on the proportional relationships of human skeletons. For example, in the field of human posture analysis, the length of human skeletal segments usually remains stable over a short period of time, so the range of variation in the length of skeletal segments between adjacent key joints is usually small.

[0122] In the specific implementation process, the Euclidean distance between adjacent key joint positions in the key point coordinate data Kcd can be calculated to obtain the corresponding bone segment length. When the bone segment length exceeds the preset bone length range, the bone segment length can be restored to a reasonable range by adjusting the key joint position.

[0123] Personnel can inspect each bone segment according to the connection relationship of the human skeleton, and adjust the position of key joints when necessary, so that the human skeleton structure can meet the biomechanical constraints of the human body.

[0124] Specifically, adjusting to a reasonable bone length range involves: when a bone segment is detected to exceed the human bone length constraint range, structural correction can be performed on the corresponding key joint position using a key joint position projection adjustment method.

[0125] It should be noted that:

[0126] As a preferred method for structural correction, this invention employs a global constraint-based optimization adjustment approach, rather than independently correcting individual bone segments, to avoid cumulative errors and structural conflicts during the correction process. Specifically, when the human biomechanical constraint model detects that the lengths of multiple bone segments exceed a preset range, the following optimization objective function is constructed:

[0127] ;

[0128] in, This is a structural constraint term that represents a measure of bone length deviation, used to measure the sum of squared deviations between the current length and the reference length of all bone segments. This represents the set of all bone segments. and This represents the spatial coordinate vector of two adjacent key points on the human body. This is the reference length for the bone segment, which can be preset based on the athlete's height ratio or determined from the initial frame.

[0129] The optimization objective is to minimize the sum of squared deviations in length of all bone segments while preserving as much confidence as possible in the original identification location of keypoints. Therefore, a fidelity term is introduced to construct the total loss function:

[0130] ;

[0131] in, The total loss function is composed of structural constraint terms and... For the set of all key points, The initial identification coordinates of the key points, i.e., the key point coordinate data Kcd. This refers to the confidence level of the corresponding keypoint, i.e., the keypoint confidence data Kcf. Regularization coefficients to balance fidelity and structural constraints;

[0132] By solving the above least squares problem, for example using the Levenberg-Marquardt algorithm, the positions of all relevant key points can be adjusted at once, so that the entire human skeletal structure can satisfy the bone length constraint while relying as much as possible on the original recognition results, and finally obtain the constrained pose data set Bps.

[0133] In the specific implementation process, the following steps can be followed:

[0134] First, determine the positions of the two key joints of the current bone segment based on the connection relationship of the human skeleton. For example, read the corresponding upstream and downstream key joint coordinates in the key point coordinate data Kcd.

[0135] Subsequently, the current bone segment length between the two key joints is calculated, and the reference bone length corresponding to the bone segment is read. The reference bone length can be the bone segment length in the initial human posture as the reference length.

[0136] When the length of the current bone segment is detected to exceed the allowable range of change, the reference bone length is used as the target bone length, and the position of the upstream key joint remains unchanged. The position of the downstream key joint is adjusted along the direction of the bone segment so that the distance between the adjusted downstream key joint position and the upstream key joint position is equal to the target bone length.

[0137] In the specific implementation, the adjustment can be achieved in the following way: First, calculate the direction vector of the current bone segment, then scale the direction vector according to the target bone length, and recalculate the position of the downstream key joint while keeping the position of the upstream key joint unchanged, so as to obtain the new key joint position coordinates.

[0138] Using the above methods, structural modifications can be made to key joint positions that exceed the constraints of bone length while keeping the connection relationship of the human skeleton unchanged.

[0139] The implementers can sequentially perform the above-mentioned detection and adjustment processes on all bone segments in the human skeletal structure, so that the length of all bone segments in the human skeletal structure meets the human biomechanical constraints.

[0140] After the above-mentioned key joint structure correction process, the position data of human key joints that meet the bone length constraints can be obtained, and the corrected key joint positions are organized into a constrained posture data set Bps.

[0141] In this embodiment, after obtaining the initial keypoint data set Kps, the method can analyze the skeletal connection relationships between the keypoint coordinate data Kcd based on the human skeletal topology and establish a complete skeletal connection link, thereby constructing a human biomechanical constraint model. This constraint model sets the allowable range of variation for each skeletal segment by statistically analyzing the lengths of bone segments between adjacent keypoints and combining this with the principle of human skeletal proportional stability. This ensures that minor errors in keypoint positions during high-speed or complex movements do not lead to abnormal skeletal structures. For example, in the high jump, when the athlete's upper limbs swing rapidly, the original coordinates of the shoulder, elbow, and wrist keypoints may deviate due to camera angles or occlusion. However, through the human biomechanical constraint model, the downstream keypoint positions can be automatically adjusted to restore the bone segment lengths to a reasonable range while maintaining the skeletal connection link, ensuring the spatial structural rationality of the shoulder, elbow, and wrist. Furthermore, after completing the detection of all bone segment lengths and the correction of key joint structures, the resulting constrained posture data set Bps can provide the structurally stable position of each keypoint in space, providing a reliable foundation for subsequent posture prediction and optimization. The special advantage of this processing method is that it can not only correct the key point deviations caused by rapid movement or detection errors in a single frame, but also ensure that the entire human skeletal structure remains stable in a short period of time. This significantly improves the structural rationality and analysis accuracy of posture resolution. For example, in competitive gymnastics or sprint starting movements, it can ensure that the key points of the shoulders, hips and knees maintain the correct spatial relationship in continuous movements, avoid bone stretching or misalignment, and improve the reliability of movement evaluation and training feedback.

[0142] Example 4

[0143] Please see Figure 2 Specifically: S3 includes S31;

[0144] S31. After obtaining the initial set of key point data Kps, read the key point coordinate data Kcd in the current video frame and the previous video frame, and calculate the motion change trend of the human body key points based on the changes in the position of human body key points between adjacent video frames.

[0145] Specifically, the key point coordinate data Kcd in the current video frame is matched one by one with the key point coordinate data Kcd in the previous video frame, and the position change of the same human body key point between consecutive video frames is calculated, so as to obtain the motion displacement information of the human body key point.

[0146] Then, based on the positional changes of human key points between consecutive video frames, the direction and amplitude of motion of human key points in image space are determined, thereby obtaining the motion change trend of human key points between adjacent video frames.

[0147] After completing the calculation of the movement change trends of all human body key points, the movement change trends of each human body key point are uniformly organized to form human body key point movement trend data.

[0148] It should be noted that:

[0149] The motion change trend of the human body key points can be calculated based on the optical flow motion estimation method. By analyzing the changes in pixel positions between consecutive image frames, the motion direction and speed of the target in the image can be estimated.

[0150] In the specific implementation process, the Lucas-Kanade optical flow algorithm can be used to perform motion estimation calculations on continuous video frames. The implementers can use the previous video frame and the current video frame as input, and obtain the motion vector of each pixel position in the image through optical flow calculation. Then, based on the pixel position of the human body key point, the corresponding optical flow vector is read to obtain the motion displacement information of the human body key point.

[0151] By reading the optical flow vectors of the corresponding pixel positions of human key points, the direction and amplitude of motion of human key points in image space can be determined, and this motion information can be used as the trend of motion change of human key points.

[0152] Using the above processing method, implementers can calculate the motion change trend of human key points based on the position changes of human key points in continuous video frames, thereby providing motion reference for subsequent key point prediction and compensation processing.

[0153] S3 further includes S32;

[0154] S32. After obtaining the motion trend data of human key points, the key point confidence data Kcf in the initial key point data set Kps is detected and processed to identify human key points with confidence levels lower than a preset threshold.

[0155] Specifically, based on the recognition confidence values ​​of each human key point in the key point confidence data Kcf, it is determined whether the recognition confidence of each human key point is lower than the preset confidence threshold.

[0156] When the confidence level of a certain human body key point recognition is lower than the preset confidence threshold, it is considered that the recognition of that human body key point is unstable.

[0157] Subsequently, based on the direction and amplitude of movement of the corresponding human key points in the human key point movement trend data, the position of the corresponding human key point in the key point coordinate data Kcd is predicted and calculated, thereby obtaining the predicted position coordinates of the human key point.

[0158] After completing the calculation of all low-confidence human keypoint prediction positions, the predicted human keypoint position data are organized according to the order of human skeletal structure and formed into a predicted pose data set Pps.

[0159] To achieve more robust key point prediction in high-speed competitive sports, this invention not only employs simple displacement superposition but also introduces a prediction method based on motion models.

[0160] As a preferred implementation, Kalman filtering is used to estimate and predict the state of confidence key points;

[0161] The specific steps include:

[0162] The state vector of each key point is defined as follows: It includes position, velocity, and acceleration information; it uses a constant acceleration model (CA model) as the state transition equation to predict the state of key points in the current frame; for key frames with high confidence, the identified coordinates are used as observations to update the filter and correct the prediction error of the motion model; when the confidence of a key point is detected to be lower than a preset threshold, it no longer relies on unreliable observations, but directly uses the position coordinates predicted by the Kalman filter based on historical motion trends as the predicted position of the key point.

[0163] In addition, prediction methods based on historical trajectories, such as multi-top fitting or long short-term memory networks, can be used to fit the motion curve of the key point in multiple consecutive frames and extrapolate its position in the current frame to cope with complex and ever-changing competitive actions.

[0164] It should be noted that:

[0165] The location of key points can be predicted and calculated by following these steps;

[0166] After obtaining the motion trend data of human key points, the implementers first traverse and detect the key point confidence data Kcf in the initial key point dataset Kps. Key points with recognition confidence lower than a preset threshold (e.g., 0.5, which can be set as an empirical trade-off between key point recognition success rate and accuracy: 0.5 to 0.7 based on experimental data or statistical analysis of human key point recognition performance) are marked as low-confidence key points, thereby clarifying which human key points need to be predicted and compensated.

[0167] Subsequently, the key point coordinate data Kcd of the corresponding low-confidence human key point in the previous video frame is read, and combined with the corresponding horizontal and vertical displacement in the human key point motion trend data, the prediction calculation is performed by taking advantage of the continuity of position changes in consecutive frames. That is, the corresponding motion displacement is superimposed on the coordinates of the previous frame to obtain the predicted position coordinates of the low-confidence key point in the current video frame.

[0168] After calculating the predicted locations of all low-confidence keypoints, the predicted low-confidence keypoint locations are combined with the keypoint coordinate data Kcd, which identifies keypoint locations with confidence levels higher than the threshold, according to the order of the human skeletal structure. The coordinates of each keypoint are then stored in a unified data structure unit (such as a list or array), thus forming a complete predicted pose data set Pps. This set contains both the original high-confidence keypoints and the predicted and compensated low-confidence keypoints, which can stably reflect human pose and provide reliable input for subsequent steps.

[0169] In this embodiment, after obtaining the initial keypoint data set Kps, the method calculates the motion trend of human keypoints based on the changes in the coordinate data Kcd of human keypoints in consecutive video frames, forming human keypoint motion trend data. This trend data is then used to perform predictive compensation processing on low-confidence keypoints to generate a predicted pose data set Pps. Specifically, by matching the coordinates Kcd of keypoints one by one in consecutive frames and calculating the positional changes, the motion direction and amplitude of each keypoint in the image space can be obtained, thereby quantifying the displacement pattern of keypoints in continuous motion. Subsequently, based on the keypoint confidence data Kcf, keypoints with confidence levels below a preset confidence threshold (e.g., 0.5–0.7) are identified. Combined with the motion trend, the low-confidence keypoints are predicted to obtain predicted position coordinates, which are then integrated with high-confidence keypoints according to the skeletal topology to form the predicted pose data set Pps. The special advantage of this processing method is that it can maintain the integrity and continuity of keypoint data through continuous frame motion trends and predictive compensation even in high-speed motion or when keypoints are temporarily occluded. For example, in gymnastic flips or long jumps, when the recognition confidence of key points such as the wrist or ankle decreases due to high-speed rotation or occlusion, prediction compensation using key point motion trend data can effectively restore the key point positions, making the entire posture smooth and continuous in the video frame sequence, avoiding abrupt changes or loss of key points, thus providing reliable input for subsequent posture optimization and skeletal structure analysis, and improving the accuracy and stability of posture analysis under high-speed competitive movements.

[0170] Example 5

[0171] Specifically: S4 includes S41;

[0172] S41. Construct a dynamic optimization space Dyn, which is a multi-dimensional continuous vector space. Each dimension of the dynamic optimization space Dyn corresponds to the X / Y coordinates of human key points and the deviation of bone length constraints, and is used to uniformly represent the information of constrained posture and predicted compensated posture. The dynamic optimization space Dyn combines the key point coordinate data Kcd in the constrained posture data set Bps after bone structure constraint correction and the predicted compensated key point coordinate data Kcd in the predicted posture data set Pps according to the topological order of human bones to form a vector, realize vectorized representation and use it as optimization input.

[0173] Subsequently, the key point coordinate data Kcd of each key point in the constrained attitude data set Bps and the key point coordinate data Kcd of the corresponding predicted compensation key point in the predicted attitude data set Pps are integrated according to the topological order of the human skeleton and mapped to the dynamic optimization space Dyn to form an attitude state vector.

[0174] By mapping the attitude state vector, the constrained attitude information and the predicted compensation information are digitally represented in the same optimization space, providing a unified input for subsequent optimization calculations;

[0175] It should be noted that:

[0176] The dynamic optimization space Dyn is a practically implementable multi-dimensional vector space, where each dimension represents the spatial coordinates of key human body points and the deviation from bone length constraints. Implementers can construct it as follows:

[0177] Determine the dimensions: Each human body key point occupies two dimensions (X and Y coordinates), and the length deviation of each bone segment is used as an additional dimension. All key points are arranged in the topological order of the human skeleton to form a posture state vector.

[0178] Mapping method: Combine the coordinates Kcd of each keypoint and the bone length constraint difference in the constrained pose data set Bps with the coordinates Kcd of each predicted keypoint in the predicted pose data set Pps into the same vector, and arrange them in order in the vector to form a unified pose state representation.

[0179] Optimize the input: Input the attitude state vector into a numerical optimization algorithm (such as gradient descent or quasi-Newton method), calculate the attitude deviation loss and structural constraint penalty term in each iteration, and adjust the coordinates of each key point.

[0180] Generate optimization results: After iterative convergence, the vectors are converted back to human keypoint coordinate data Kcd to form the optimized pose result set Ops.

[0181] S4 also includes S42;

[0182] S42. Perform co-optimization processing on the attitude state vector mapped to the dynamic optimization space Dyn. During the co-optimization process, calculate the attitude deviation loss (squared Euclidean distance between the predicted coordinates and the constraint coordinates) for each key point. At the same time, add the human skeleton length constraint and skeleton connection link as penalty terms to the optimization objective function. Implementers can use numerical optimization algorithms (such as gradient descent or quasi-Newton method) to iteratively adjust the attitude state vector to minimize the attitude deviation and satisfy the structural constraint conditions at the same time.

[0183] After optimization, the optimized posture state vector is converted back into key point coordinate data Kcd according to the human skeletal structure order, forming the optimized posture result set Ops.

[0184] S5 includes S51;

[0185] S51. After obtaining the optimized posture result set Ops, the key point coordinate data Kcd of each key point in the set is reorganized according to the human skeleton topology order to reconstruct the complete human skeleton structure.

[0186] Specifically, the optimized key point coordinates Kcd are arranged according to the skeletal connection links, and key points are connected one by one to form skeletal segments. The skeletal structure of key parts of the human body such as the head, torso, and limbs is constructed to achieve complete spatial reconstruction of human posture.

[0187] After the human skeleton structure is reconstructed, the reconstructed complete human skeleton structure and the spatial coordinates Kcd of each key point are organized into a competitive human posture structured result set Res. The competitive human posture structured result set Res contains the final position coordinates of each key point and its topological relationship information in the skeleton structure, which can be directly used for posture analysis, motion evaluation or other subsequent applications in competitive scenarios.

[0188] It should be noted that:

[0189] The keypoint coordinate data (Kcd) of each keypoint in the optimized pose result set (Ops) is arranged according to the topological order of the human skeleton and processed in conjunction with the skeleton connection links. Implementers can deploy and execute the process as follows:

[0190] Keypoint traversal: Sequentially read the coordinates (Kcd) of each keypoint in the optimized pose result set (Ops) and its index relationship in the skeleton connection link. Each keypoint element contains X, Y (or Z in a 3D scene) coordinate information and a skeleton connection index, which is used to determine the upstream and downstream nodes connected to the keypoint.

[0191] Skeletal Segment Generation: Based on the skeletal connection links, each keypoint is combined with its connected keypoints to form a skeletal segment. For two-dimensional or three-dimensional space, this can be accomplished by drawing line segments or generating vectors in coordinate space. Implementers can combine all skeletal segments according to the topological order of the human body to form a complete skeletal framework, including major skeletal parts such as the head, neck, torso, shoulders, arms, and legs.

[0192] Skeletal structure integration: All bone segments are combined according to the topological order of the human skeleton to ensure that the relative position of each key point in space is consistent with the original optimization result, while retaining the index information and topological relationship of each key point to form a complete skeletal structure model.

[0193] In this embodiment, after obtaining the constrained posture data set Bps and the predicted posture data set Pps, the two types of posture data can be mapped to the dynamic optimization space Dyn to form a posture state vector. Within this space, posture deviations and human structural constraints are collaboratively optimized to generate an optimized posture result set Ops. Specifically, in Dyn, the keypoint coordinate data Kcd from Bps, corrected for skeletal constraints, and the predicted and compensated keypoint coordinate data Kcd from Pps are combined according to the topological order of the human skeleton to form a multidimensional vector representation. This achieves a unified digital representation of posture information and serves as input to the optimization algorithm. Subsequently, the posture state vector is iteratively adjusted using a numerical optimization algorithm to minimize the deviation between the predicted and constrained keypoints while maintaining skeletal connection links and bone length constraints. After obtaining the optimized posture result set Ops, the optimized keypoint coordinates Kcd can be reconstructed according to the skeletal topological order to form a complete human skeletal structure. This generates a structured result set Res for competitive human posture, which includes the final position coordinates Kcd of each keypoint and the topological relationships within the skeletal structure, achieving complete human posture reconstruction. For example, in gymnastics movements, optimized keypoint coordinates (Kcd) can maintain the reasonable spatial positions of key points in the shoulders, hips, and knees / ankles during high-speed flips and rapid arm and foot swings, while ensuring stable bone segment lengths. This guarantees consistency in posture reconstruction between consecutive frames of each movement, thus providing a precise and quantifiable data foundation for competitive movement analysis, movement scoring, or training feedback. A unique advantage of this method is that it can not only integrate constrained posture information and predictive compensation information for unified optimization, but also directly generate structured human posture data that can be used for subsequent analysis and visualization, improving the accuracy and reliability of posture analysis under high-speed competitive movements.

[0194] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.

Claims

1. A method for capture recognition for athletic posture analysis, characterized by: Includes the following steps: S1. Acquire the competitive video stream and identify the spatial coordinates of key human body points and their corresponding recognition confidence levels, forming key point coordinate data Kcd and key point confidence data Kcf respectively, and constructing the initial key point data set Kps from the key point coordinate data Kcd and key point confidence data Kcf. S2. Based on the connection relationship of the human skeleton, establish a human biomechanical constraint model, perform structural constraint processing on the key point coordinate data Kcd in the initial key point data set Kps, and generate a constraint posture data set Bps. S3. Based on the trend of human body key point motion change between adjacent video frames, predict and compensate the key point coordinate data Kcd corresponding to the key point confidence data Kcf in the initial key point data set Kps being lower than the preset threshold, and generate the predicted posture data set Pps. S4. Map the constrained attitude data set Bps and the predicted attitude data set Pps to the dynamic optimization space Dyn, and perform collaborative optimization processing on attitude deviation and human body structure constraints in the dynamic optimization space Dyn to generate the optimized attitude result set Ops. S5. Reconstruct the human posture structure based on the optimized posture result set Ops, and output the competitive human posture structured result set Res.

2. The capture recognition method for competitive posture analysis according to claim 1, characterized in that: S1 includes S11; S11. Deploy video capture equipment in competitive sports scenarios to continuously capture video of athletes' competitive actions, thereby obtaining the original competitive video stream; Human target detection processing is performed on the video frame image sequence to identify the moving human body region in the video frame image. Specifically, the target detection calculation is performed on the video frame image based on the YouOnlyLookOnce target detection algorithm. The human body region in the video frame image is determined by the human body detection bounding box output by the algorithm, and the corresponding human body region image data is extracted according to the human body detection bounding box. After obtaining human body region image data, human body posture key point recognition processing is performed on the human body region image data, and key point heatmaps of corresponding human body key points are generated. After obtaining the human body key point heatmap, peak position detection is performed on each key point heatmap to extract the two-dimensional position coordinates of each human body key point in the image coordinate system. Then, the position coordinates of each human body key point are organized according to the human skeletal structure order, and all human body key point coordinates are uniformly summarized to form key point coordinate data Kcd.

3. The capture recognition method for competitive posture analysis according to claim 2, characterized in that: S1 further includes S12; S12. After obtaining the key point coordinate data Kcd, the recognition reliability assessment of the human key point recognition results is performed: Specifically, based on the key point heat map response results output by the OpenPose human pose recognition algorithm, the response intensity value of the key point heat map corresponding to each human key point is read, and the response intensity value corresponding to the peak position in the key point heat map is obtained, and the response intensity value is used as the recognition confidence of the corresponding human key point. Then, according to the order of human body key points, the recognition confidence scores corresponding to all human body key points are sorted out and key point confidence data Kcf is formed; After obtaining the key point confidence data Kcf, the key point coordinate data Kcd and the key point confidence data Kcf are associated one by one according to the index order of human key points, so that each human key point has corresponding spatial coordinate information and recognition confidence information at the same time. After completing the corresponding association, the key point coordinate data Kcd and the key point confidence data Kcf are combined according to a unified data structure to construct the initial key point data set Kps.

4. The capture recognition method for competitive posture analysis according to claim 3, characterized in that: S2 includes S21; S21. After obtaining the initial set of key point data Kps, read the key point coordinate data Kcd and analyze the connection relationship between human key points according to the human skeleton topology: construct the skeleton connection relationship of each human key point in the key point coordinate data Kcd according to the connection order between each human key point in the human skeleton structure, so as to obtain the skeleton connection link in the human skeleton structure. After completing the construction of the skeletal connection relationship, based on the connection relationship between each adjacent human key point in the skeletal connection relationship, the bone length between adjacent human key points in the key point coordinate data Kcd is calculated, and the stability of human bone length is used as the constraint basis to establish a human biomechanical constraint model for the human skeletal structure.

5. The capture and recognition method for analyzing competitive postures according to claim 4, characterized in that: S2 further includes S22; S22. After establishing the human biomechanical constraint model, input the key point coordinate data Kcd from the initial key point data set Kps into the human biomechanical constraint model to perform structural rationality detection processing on the human skeletal structure. When a bone segment is detected to exceed the human bone length constraint range, structural correction processing is performed on the position of the corresponding human key point in the key point coordinate data Kcd, so that the position of the human key point is adjusted to a reasonable bone length range while maintaining the bone connection link. After completing the structural constraint detection and human key point position correction processing for all skeletal segments, the corrected key point coordinate data Kcd is uniformly organized, and the human posture structure data is reorganized according to the skeletal connection links to generate the constrained posture data set Bps.

6. The capture and recognition method for analyzing competitive postures according to claim 5, characterized in that: S3 includes S31; S31. After obtaining the initial set of key point data Kps, read the key point coordinate data Kcd in the current video frame and the previous video frame, and calculate the motion change trend of the human body key points based on the changes in the position of human body key points between adjacent video frames. Specifically, the key point coordinate data Kcd in the current video frame is matched one by one with the key point coordinate data Kcd in the previous video frame, and the position change of the same human body key point between consecutive video frames is calculated, so as to obtain the motion displacement information of the human body key point. Then, based on the positional changes of human key points between consecutive video frames, the direction and amplitude of motion of human key points in image space are determined, thereby obtaining the motion change trend of human key points between adjacent video frames. After calculating the movement trends of all key points on the human body, the movement trends of each key point are organized and compiled to form the movement trend data of key points on the human body.

7. The capture and recognition method for analyzing competitive postures according to claim 6, characterized in that: S3 further includes S32; S32. After obtaining the motion trend data of human key points, the key point confidence data Kcf in the initial key point data set Kps is detected and processed to identify human key points with confidence levels lower than a preset threshold. Specifically, based on the recognition confidence values ​​of each human key point in the key point confidence data Kcf, it is determined whether the recognition confidence of each human key point is lower than the preset confidence threshold. When the confidence level of a certain human body key point recognition is lower than the preset confidence threshold, it is considered that the recognition of that human body key point is unstable. Subsequently, based on the direction and amplitude of movement of the corresponding human key points in the human key point movement trend data, the position of the corresponding human key point in the key point coordinate data Kcd is predicted and calculated, thereby obtaining the predicted position coordinates of the human key point. After completing the calculation of all low-confidence human keypoint prediction positions, the predicted human keypoint position data are organized according to the order of human skeletal structure and formed into a predicted pose data set Pps.

8. The capture and recognition method for analyzing competitive postures according to claim 7, characterized in that: S4 includes S41; S41. Construct the dynamic optimization space Dyn. The dynamic optimization space Dyn is formed by combining the key point coordinate data Kcd in the constrained posture data set Bps after skeletal structure constraint correction with the predicted compensated key point coordinate data Kcd in the predicted posture data set Pps according to the topological order of the human skeleton to form a vector, realize the vectorized representation and use it as optimization input. Subsequently, the key point coordinate data Kcd of each key point in the constrained attitude data set Bps and the key point coordinate data Kcd of the corresponding predicted compensation key point in the predicted attitude data set Pps are integrated according to the topological order of the human skeleton and mapped to the dynamic optimization space Dyn to form an attitude state vector.

9. The capture and recognition method for analyzing competitive postures according to claim 8, characterized in that: S4 further includes S42; S42. Perform collaborative optimization processing on the attitude state vector mapped to the dynamic optimization space Dyn. During the collaborative optimization process, calculate the attitude deviation loss for each key point. At the same time, add the human skeleton length constraint and skeleton connection link as penalty terms to the optimization objective function, and use numerical optimization algorithm to iteratively adjust the attitude state vector to minimize the attitude deviation and simultaneously satisfy the structural constraint conditions. After optimization, the optimized posture state vector is converted back into key point coordinate data Kcd according to the human skeletal structure order, forming the optimized posture result set Ops.

10. The capture and recognition method for analyzing competitive postures according to claim 9, characterized in that: S5 includes S51; S51. After obtaining the optimized posture result set Ops, the key point coordinate data Kcd of each key point in the set is reorganized according to the human skeleton topology order to reconstruct the complete human skeleton structure. Specifically, the optimized key point coordinates Kcd are arranged according to the skeletal connection links, and key points are connected one by one to form skeletal segments. The skeletal structure of key parts of the human body, such as the head, torso and limbs, is constructed to achieve complete spatial reconstruction of human posture. After completing the reconstruction of the human skeletal structure, the reconstructed complete human skeletal structure and the spatial coordinates Kcd of each key point are organized into a set of structured results of competitive human posture, Res.

Images