An intelligent sports state analysis and posture estimation method, system, device and medium based on multi-module collaborative optimization

Abstract

The application discloses a kind of intelligent sports movement state analysis method, system, equipment and medium based on multi-module collaborative optimization.Method includes: using lightweight real-time model to video stream is human detection and initial key point positioning;Through radial basis function network, initial key point is nonlinear fine-tuning to improve accuracy;From the key point sequence after fine-tuning, extract high-dimensional kinematic feature, and utilize principal component analysis to reduce dimension, obtain core low-dimensional time sequence feature;Finally, through time sequence neural network, feature sequence is modeled, realizes action recognition and posture evaluation.The application passes through the collaborative optimization link of "perception-fine-purification-understanding", effectively solves the problems, such as imbalance of accuracy and speed, feature redundancy, time sequence modeling deficiency and error accumulation, in prior art, provides efficient, accurate, automated analysis solution for wisdom sports training.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of computer vision, artificial intelligence, and sports science, specifically to an intelligent sports state analysis and posture estimation method, system, device, and medium based on multi-module collaborative optimization. Background Technology

[0002] With the rapid development of smart sports and scientific training, the automated and precise analysis of athlete movements using computer vision technology has become a research hotspot in the industry. Accurate motion recognition and posture estimation are the core foundation for evaluating training effectiveness, correcting technical movements, and preventing sports injuries. However, in actual sports monitoring scenarios, achieving full-scene, high-precision, and high-efficiency motion analysis still faces multiple technical challenges.

[0003] Existing technical solutions generally have the following limitations: (1) The contradiction between accuracy and speed: Although deep learning-based pose estimation models (such as HRNet) have high positioning accuracy, they have high computational complexity and are difficult to meet the computing power requirements of real-time monitoring; while lightweight detection models (such as YOLO series) have fast running speed, but in scenarios such as complex poses and target occlusion, the positioning accuracy of key points is often insufficient.

[0004] (2) Risk of feature redundancy and overfitting: High-dimensional features extracted directly from the original image or key point data contain a lot of redundant information and noise. Directly using them for action classification can easily lead to model overfitting and weak generalization ability. (3) Insufficient utilization of temporal information: Most methods focus on the pose estimation of single-frame images and ignore the continuity, rhythm and evolution of actions in the time dimension, which limits the ability to recognize and evaluate complex continuous actions.

[0005] (4) Module isolation and lack of collaborative optimization: Existing solutions mostly adopt a single model or a simple pipeline structure. Each module (such as detection, pose estimation, classification) is optimized independently, and errors are easily accumulated at each level. Furthermore, an end-to-end optimization mechanism from raw pixels to advanced semantic understanding has not been formed.

[0006] Therefore, there is an urgent need for a comprehensive solution that can achieve multi-module collaborative optimization, balance accuracy and efficiency, and make full use of spatiotemporal information to meet the complex analysis needs in smart sports training and testing. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the above-mentioned shortcomings of the prior art. In order to solve the above problems, an intelligent sports state analysis and posture estimation method, system, device and medium based on multi-module collaborative optimization is provided.

[0008] The object of this invention is achieved in the following manner: A method for intelligent sports state analysis and posture estimation based on multi-module collaborative optimization includes the following steps: S1: Based on the target detection model, perform real-time human detection, tracking and preliminary key point localization on the input video stream, and output the target bounding box with tracking ID and the initial key point coordinate sequence; S2: Construct a radial basis function (RBF) network, taking the initial joint coordinate sequence of the current frame and several frames before and after it as input, and outputting the optimized joint coordinate sequence corresponding to the current frame; wherein, the RBF network performs nonlinear filtering and kinematic correction on the initial joint coordinates by fusing temporal neighborhood information; S3: Based on the optimized joint coordinate sequence, extract multidimensional kinematic features, and use a dimensionality reduction method to process the multidimensional kinematic features to obtain a core low-dimensional temporal feature sequence; S4: Construct a temporal modeling network model, using the core low-dimensional temporal feature sequence as input, to perform temporal modeling and pattern recognition, and output action classification results and posture evaluation parameters.

[0009] In step S1, the target detection model is the YOLOv11 model, which is configured to simultaneously perform target detection and preliminary localization of 17 human key points, and to perform correlation tracking of the same target in consecutive frames using the DeepSORT algorithm.

[0010] In step S2, the RBF network uses Gaussian radial basis functions, and its hidden layer centers are initialized using the K-means clustering algorithm. The network uses the mean square error between the key point coordinates in the previous stage and the high-precision labeled data as the loss function, and the network parameters are jointly trained using the backpropagation algorithm.

[0011] In step S3, the multidimensional kinematic features include: spatial coordinates of joints, joint angles, motion velocity, acceleration, and body center of gravity position; the dimensionality reduction method is principal component analysis (PCA), which is used to select the top k principal components whose cumulative variance contribution rate exceeds a preset threshold.

[0012] In step S4, the temporal modeling network is a bidirectional long short-term memory (Bi-LSTM) network that incorporates an attention mechanism. The Bi-LSTM network receives the core low-dimensional temporal feature sequence in a sliding window manner and outputs the result after weighted fusion of features at different time steps through an attention layer.

[0013] It also includes step S5: Based on the action classification and posture evaluation results output in step S4, and using the pre-built sports biomechanics knowledge base, through result analysis and feature association, multi-level matching and reasoning, to generate targeted training improvement suggestions and potential injury risk warnings.

[0014] The sports biomechanics knowledge base includes: a standard movement pattern library, a common error pattern library, an injury risk rule library, and a quantitative assessment and feedback template library; the matching and reasoning process includes at least one of dynamic time warping (DTW) template matching, threshold-based rule matching, and generative rule reasoning.

[0015] An intelligent sports activity state analysis system includes: The image processing module is used to receive video streams and output the initial joint coordinate sequence of the human target with tracking ID; The key point optimization module, connected to the image processing module, is used to receive multiple frames of the initial joint point coordinate sequence, and perform nonlinear mapping based on the radial basis function network to output a high-precision joint point coordinate sequence. The feature engineering module, connected to the key point optimization module, is used to extract multidimensional kinematic features from the high-precision joint coordinate sequence and perform dimensionality reduction using principal component analysis to output a low-dimensional temporal feature sequence. The temporal analysis module, connected to the feature engineering module, is used to receive the low-dimensional temporal feature sequence, perform temporal modeling based on the long short-term memory network, and output action classification results and posture evaluation parameters.

[0016] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0018] The beneficial effects of this invention are as follows: This invention achieves the optimal balance between accuracy and efficiency: It uses a target detection model (YOLOv11) as the front-end detector, which provides reliable initial attitude estimation while ensuring real-time performance; Subsequently, it performs local fine-tuning through a lightweight RBF network, avoiding the high computational cost brought by a single ultra-large model, and achieving high-precision attitude output under limited computing power.

[0019] Optimizing feature quality and improving generalization ability: High-dimensional kinematic features are reduced in dimensionality using PCA, effectively eliminating redundant noise and extracting essential features strongly correlated with motion states. This not only reduces the input dimensionality and training difficulty of subsequent LSTM models but also significantly improves the model's adaptability and generalization performance to different individuals and environments.

[0020] Deep integration of spatiotemporal information: The innovative introduction of LSTM with attention mechanism into the action analysis process enables the model to fully learn and understand the temporal evolution of actions, thereby enabling more accurate identification and stage decomposition evaluation of continuous and complex technical actions (such as the entire process of long jump, including approach run, take-off, flight, and landing).

[0021] Modular collaborative design for strict error propagation control: This invention constructs a progressive pipeline of "detection-fine-tuning-dimensionality reduction-recognition-empowerment," with each module performing its specific function while complementing each other: YOLOv11 enables rapid localization, RBF performs fine-tuning, PCA extracts core features, and LSTM grasps temporal patterns. This design supports independent optimization and upgrades of each module, while RBF compensates for detection errors and PCA purifies input data, effectively controlling the accumulation and propagation of errors, significantly enhancing the overall robustness and operational stability of the system.

[0022] With a wide range of applications and outstanding practical value, this method forms a complete technology stack, from raw video input to advanced semantic output (action category, posture score, risk warning), which can be directly applied to various smart sports scenarios such as gym training guidance, physical fitness testing in sports classes, and rehabilitation training monitoring. It has extremely high practical value and industrialization potential. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of an intelligent sports activity state analysis method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the collaborative operation of the image processing module and the key point optimization module in one embodiment of the present invention; Figure 3 This is a schematic diagram of the feature dimensionality reduction process performed by principal component analysis in one embodiment of the present invention; Figure 4 This is a schematic diagram of a bidirectional long short-term memory network structure with an attention mechanism in one embodiment of the present invention; Figure 5 This is a performance comparison chart of an embodiment of the present invention and the prior art in terms of key point positioning accuracy and motion recognition accuracy. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same technical meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0026] like Figures 1-5 As shown, this invention provides an intelligent sports state analysis and posture estimation method based on multi-module collaborative optimization, including the following steps: S1: Based on the target detection model, perform real-time human detection, tracking and preliminary key point localization on the input video stream, and output the target bounding box with tracking ID and the initial key point coordinate sequence; S2: Construct a Radial Basis Function (RBF) network, taking the initial joint coordinate sequence of the current frame and several frames before and after it as input, and outputting the optimized joint coordinate sequence corresponding to the current frame; wherein, the RBF network performs nonlinear filtering and kinematic correction on the initial joint coordinates by fusing temporal neighborhood information; (the Radial Basis Function (RBF) network uses a "multi-frame input, single-frame output" approach). The mechanism works as follows: Centered on the current frame, it takes the initial keypoint coordinates of several frames before and after it, and merges them into a spatiotemporal vector representing the local motion context, which is then input into the network. The input is a multi-frame sequence, including the previous n frames, the current frame, and the next n frames, where each frame is the initial keypoint coordinate sequence output from the previous stage. After flattening, the multi-frame sequence becomes a single input vector, and the output is the optimized current frame coordinate sequence. Internally, the network uses a nonlinear model constructed from radial basis functions to learn how to infer more reasonable and smoother keypoint coordinates for the current frame from this noisy temporal observation data. Essentially, it uses information within the temporal neighborhood to perform intelligent filtering and kinematic rationality correction on the observation results of the current frame, thereby outputting a more stable and regular keypoint trajectory. S3: Based on the optimized joint coordinate sequence, extract multidimensional kinematic features, and use a dimensionality reduction method to process the multidimensional kinematic features to obtain a core low-dimensional temporal feature sequence; S4: Construct a temporal modeling network model, using the core low-dimensional temporal feature sequence as input, to perform temporal modeling and pattern recognition, and output action classification results and posture evaluation parameters.

[0027] In step S1, the target detection model is the YOLOv11 model, which is configured to simultaneously perform target detection and preliminary localization of 17 human key points, and to perform correlation tracking of the same target in consecutive frames using the DeepSORT algorithm.

[0028] In step S2, the RBF network uses Gaussian radial basis functions, and its hidden layer centers are initialized using the K-means clustering algorithm. The network uses the mean square error between the key point coordinates in the previous stage and the high-precision labeled data as the loss function, and the network parameters are jointly trained using the backpropagation algorithm.

[0029] In step S3, the multidimensional kinematic features include: spatial coordinates of joints, joint angles, motion velocity, acceleration, and body center of gravity position; the dimensionality reduction method is principal component analysis (PCA), which is used to select the top k principal components whose cumulative variance contribution rate exceeds a preset threshold.

[0030] In step S4, the temporal modeling network is a bidirectional long short-term memory (Bi-LSTM) network that incorporates an attention mechanism. The Bi-LSTM network receives the core low-dimensional temporal feature sequence in a sliding window manner and outputs the result after weighted fusion of features at different time steps through an attention layer.

[0031] It also includes step S5: Based on the action classification and posture evaluation results output in step S4, and using the pre-built sports biomechanics knowledge base, through result analysis and feature association, multi-level matching and reasoning, to generate targeted training improvement suggestions and potential injury risk warnings.

[0032] The sports biomechanics knowledge base includes: a standard movement pattern library, a common error pattern library, an injury risk rule library, and a quantitative assessment and feedback template library; the matching and reasoning process includes at least one of dynamic time warping (DTW) template matching, threshold-based rule matching, and generative rule reasoning.

[0033] An intelligent sports activity state analysis system includes: The image processing module is used to receive video streams and output the initial joint coordinate sequence of the human target with tracking ID; The key point optimization module, connected to the image processing module, is used to receive multiple frames of the initial joint point coordinate sequence, and perform nonlinear mapping based on the radial basis function network to output a high-precision joint point coordinate sequence. The feature engineering module, connected to the key point optimization module, is used to extract multidimensional kinematic features from the high-precision joint coordinate sequence and perform dimensionality reduction using principal component analysis to output a low-dimensional temporal feature sequence. The temporal analysis module, connected to the feature engineering module, is used to receive the low-dimensional temporal feature sequence, perform temporal modeling based on the long short-term memory network, and output action classification results and posture evaluation parameters.

[0034] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method.

[0035] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0036] To overcome the shortcomings of isolated modules and error accumulation in existing technologies, this invention pioneers a progressive and collaborative technology chain of "efficient detection—local fine-tuning—feature purification—temporal understanding—knowledge empowerment." This architecture deeply integrates core algorithm modules with a sports science knowledge base, forming a complete and closed-loop optimization pipeline.

[0037] I. Systemic Architectural Innovation: A Technology Chain of Multi-Stage Collaborative Optimization To overcome the shortcomings of isolated modules and error accumulation in existing technologies, this invention pioneers a progressive and collaborative technology chain of "efficient detection—local fine-tuning—feature purification—temporal understanding—knowledge empowerment." This architecture deeply integrates core algorithm modules with a sports science knowledge base, forming a complete and closed-loop optimization pipeline.

[0038] The core process is as follows: High-efficiency detection: A highly optimized YOLOv11 model is used for real-time human target detection and initial key point localization, providing stable and fast input for subsequent processes.

[0039] Local Fine-tuning: A radial basis function (RBF) network is creatively introduced to perform residual compensation and local deformation correction on the initial keypoint coordinates. This module is specifically designed for complex situations such as occlusion and blurring, utilizing its powerful local nonlinear interpolation capabilities to optimize keypoint accuracy, replacing traditional graph convolutional networks or heatmap regression methods.

[0040] Feature Refinement: Rich kinematic features (joint angles, velocities, accelerations, etc.) are calculated from the finely tuned keypoint sequence. Principal Component Analysis (PCA) is innovatively introduced to reduce the dimensionality of high-dimensional features. This step not only reduces computational overhead, but more importantly, filters out noise and individual differences, extracting the core low-dimensional feature vectors that represent the essence of motion.

[0041] Temporal understanding: The refined feature input is combined with a bidirectional LSTM model using an attention mechanism for in-depth temporal modeling and analysis. PCA preprocessing enables the temporal model to focus more on the regular patterns of actions, thereby significantly improving the accuracy of action recognition, stage segmentation, and preliminary evaluation.

[0042] Knowledge empowerment: The end of the technology chain does not simply output classifications or scores, but is deeply integrated with the sports biomechanics knowledge base, transforming the precise and refined time-series data generated in the first three steps into concrete diagnosis, assessment and guidance.

[0043] Synergistic effect: Each module is designed and compensated for the output characteristics of the previous stage (such as RBF to fine-tune the detection results and PCA to purify the input of the time series model), realizing step-by-step error control and layer-by-layer information purification. In the end, the overall system performance (accuracy, speed, robustness, interpretability) is greater than the simple sum of the independent performance of each module, which is "1+1>2".

[0044] II. Core Value Module: Sports Biomechanics Knowledge Base This knowledge base is key to transforming domain expert knowledge into a computable and reasonable rule system, and is central to achieving the leap from "seeing actions" to "understanding actions." It adopts a hybrid "rule-data" architecture; taking basketball shooting as an example, the structure is as follows: Module Content and Functions Example (basketball shooting) Standard Action Pattern Library Define the spatiotemporal standards for ideal motion. The ideal range for wrist flexion angle during the "release phase" is specified as 45°-60°. Common error mode library Summarize typical errors and their quantitative characteristics. Define error mode "E01_Elbow abduction too large": The angle between the upper arm and the torso is >110° when the ball is released. Damage Risk Rule Base Establish rules for associating error patterns with potential damage. Rule R01: If E01 occurs frequently and is accompanied by a shoulder abduction angle >90°, then "shoulder impingement risk (intermediate)" is triggered. Quantitative Assessment and Feedback Template Library Provides scoring criteria, textual and visual feedback templates. Text template: "You have been detected to have a {{E01}} problem. We recommend that you perform {{wall shooting simulation training}}." The core matching algorithm of this knowledge base includes: 1. Dynamic Time Warping (DTW): Used to compare the low-dimensional temporal feature sequences extracted by S3 with ideal action templates in the knowledge base. DTW can handle actions at different speeds and calculate the "distance" or similarity score between them.

[0045] 2. Rule-Based Matching: The pose evaluation score output by S4 is compared with the specific kinematic parameters extracted by S3 (such as maximum joint angle and asymmetry index).

[0046] Example rule: IF knee valgus angle > 15 degrees AND landing impact force > 2 times body weight THEN trigger "ACL high risk" alert.

[0047] The knowledge base reasoning engine includes: 1. Production Rule Engine: This is the "brain" of the knowledge base. It contains a series of "IF-THEN" rules that connect matched posture deviations with potential causes, improvement suggestions, and risk levels.

[0048] 2. Case-Based Reasoning (CBR): When encountering complex or rare posture patterns, the system will search for the most similar solved cases in the historical case library, and use their diagnoses and suggested solutions to make adaptive modifications before outputting the results.

[0049] 3. Lightweight Bayesian Networks: Used to handle uncertainty. For example, a posture deviation may be caused by multiple factors. Bayesian networks can calculate the posterior probability of various potential causes based on observed features (such as abnormal combinations of angles of multiple joints) and give the most probable cause inference.

[0050] The knowledge base uses S4 results to generate suggestions and warnings in the following steps: 1. Result analysis and feature association: The system receives the output of S4 (e.g., action classification = "running", posture evaluation score = 72 / 100), and at the same time, retrieves the core kinematic parameters (e.g., stride length, ground contact time, pelvic tilt angle, knee flexion angle curve, etc.) calculated by S3 for this running video segment.

[0051] 2. Multi-level matching and reasoning: First, motion template matching is performed (e.g., comparing the joint angle curve of the current runner with the template of "elite runner" in the knowledge base using DTW to find that the runner has excessive pelvic rotation and knee stiffness upon landing). Then, rules are triggered (e.g., Rule A: IF knee flexion angle < 20 degrees during landing THEN poor cushioning capacity and high impact force; Rule B: IF pelvic rotation amplitude > X degrees AND low core stability assessment score THEN insufficient core strength leading to low efficiency). Finally, causal reasoning is performed (e.g., the rule engine combines the two triggered rules to infer that the main problem is: "Insufficient core stability leads to poor pelvic control, and at the same time, lower limb flexibility may be limited, which together cause poor cushioning capacity and energy leakage."

[0052] 3. Personalized Suggestions and Warnings: Training improvement suggestions include immediate feedback (e.g., pay attention to core engagement and pelvic stability; try shortening stride length and increasing cadence) and long-term training plans (e.g., adding the following to weekly training: 1) plank variations; 2) hamstring and calf flexibility exercises; 3) jumping squats). These suggestions are precisely retrieved from the "Training Intervention Library" and matched to the user's fitness level. Injury risk warnings include immediate alerts (e.g., if insufficient knee cushioning is detected, pay close attention to knee sensations during this training session; stop immediately if discomfort occurs) and long-term risk assessment reports (e.g., your running pattern indicates a high risk of stress fractures (tibia) and patellar tendinitis; gait correction and consultation with a physical therapist are recommended). The risk level is calculated and labeled (e.g., low / medium / high) based on the severity of the deviation and the user's personal injury history.

[0053] III. Collaborative Workflow between Technology Chain and Knowledge Base Scenario: Analyzing a basketball player's shooting video.

[0054] Data generation (S1-S4): The system processes video using a technical chain of YOLOv11 → RBF fine-tuning → PCA purification → bidirectional LSTM modeling.

[0055] Output: Action classification results (e.g., 15 "standard shots", 5 "abnormal elbow abduction"), raw posture assessment scores, and detailed and accurate joint angle and velocity timing curves for each shot.

[0056] Knowledge Matching and Reasoning (S5): Pattern matching: The knowledge base receives time-series data and automatically matches 5 shots that match the "E01_Elbow abduction too large" error pattern.

[0057] Rule Triggering and Risk Assessment: Based on the frequency of the error pattern (5 / 20=25%) and the associated characteristics (shoulder angle), the injury risk rule R01 is triggered to generate a comprehensive judgment of "shoulder impingement risk - intermediate".

[0058] Intelligent feedback generation (final output): Training improvement suggestions: Combining the error pattern library and feedback templates, generate personalized correction suggestions: "It was detected that you have an excessive elbow abduction problem when shooting. It is recommended to conduct 'wall shooting simulation training'...".

[0059] Damage Risk Warning: Based on the risk assessment results, output warning information: There is a risk of shoulder impingement (intermediate)... We recommend incorporating 'elastic band shoulder external rotation strengthening' training. Visual analysis: It can generate skeletal overlay animations of user actions and standard movements, highlighting deviation joints.

[0060] This invention achieves substantial multi-dimensional innovation through the integration of interdisciplinary technologies (computer vision, numerical analysis, deep learning, and sports biomechanics) and a problem-oriented architectural design: Architectural Innovation: A collaborative optimization closed-loop technology chain was proposed, which systematically solved the problems of error accumulation and information redundancy.

[0061] Algorithm innovation: Original designs were introduced in the keypoint fine-tuning (RBF network) and feature engineering (PCA preprocessing) stages.

[0062] Application Innovation: It pioneered a model that deeply integrates a computable knowledge base into the end of the technology chain, achieving: Interpretable output: Transform coordinate data into diagnostics with significance in sports science.

[0063] Personalized and proactive guidance: Providing targeted corrective solutions and early warnings for damage prevention.

[0064] Closed-loop optimization ecosystem: Supports a closed-loop training optimization process of "evaluation-feedback-re-evaluation".

[0065] End-to-end value: It provides a full-stack, automated solution from raw video to advanced semantic feedback (motion scoring, correction suggestions, risk warnings). The YOLOv11 front-end ensures real-time performance, and the modular design ensures scalability, enabling its widespread application in various scenarios such as smart sports teaching, fitness guidance, and rehabilitation monitoring.

[0066] In summary, this solution not only revolutionizes the technical approach to motion analysis, but also achieves a true leap from "perception" to "cognition" and then to "decision support" through deep collaboration between the technology chain and the knowledge base, providing a complete technical solution that is high-precision, real-time, interpretable, and practical for the field of smart sports.

[0067] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several changes and improvements without departing from the overall concept of the present invention, and these should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent sports state analysis and posture estimation based on multi-module collaborative optimization, characterized in that: Includes the following steps: S1: Based on the target detection model, perform real-time human detection, tracking and preliminary key point localization on the input video stream, and output the target bounding box with tracking ID and the initial key point coordinate sequence; S2: Construct a radial basis function (RBF) network, taking the initial joint coordinate sequence of the current frame and several frames before and after it as input, and outputting the optimized joint coordinate sequence corresponding to the current frame; wherein, the RBF network performs nonlinear filtering and kinematic correction on the initial joint coordinates by fusing temporal neighborhood information. S3: Based on the optimized joint coordinate sequence, extract multidimensional kinematic features, and use a dimensionality reduction method to process the multidimensional kinematic features to obtain a core low-dimensional temporal feature sequence; S4: Construct a temporal modeling network model, using the core low-dimensional temporal feature sequence as input, to perform temporal modeling and pattern recognition, and output action classification results and posture evaluation parameters.

2. The intelligent sports state analysis and posture estimation method based on multi-module collaborative optimization according to claim 1, characterized in that: In step S1, the target detection model is the YOLOv11 model, which is configured to simultaneously perform target detection and preliminary localization of 17 human key points, and to perform correlation tracking of the same target in consecutive frames using the DeepSORT algorithm.

3. The intelligent sports state analysis and posture estimation method based on multi-module collaborative optimization according to claim 1, characterized in that: In step S2, the RBF network uses Gaussian radial basis functions, and its hidden layer centers are initialized using the K-means clustering algorithm. The network uses the mean square error between the key point coordinates in the previous stage and the high-precision labeled data as the loss function, and the network parameters are jointly trained using the backpropagation algorithm.

4. The intelligent sports state analysis and posture estimation method based on multi-module collaborative optimization according to claim 1, characterized in that: In step S3, the multidimensional kinematic features include: spatial coordinates of joints, joint angles, motion velocity, acceleration, and body center of gravity position; the dimensionality reduction method is principal component analysis (PCA), which is used to select the top k principal components whose cumulative variance contribution rate exceeds a preset threshold.

5. The intelligent sports state analysis and posture estimation method based on multi-module collaborative optimization according to claim 1, characterized in that: In step S4, the temporal modeling network is a bidirectional long short-term memory (Bi-LSTM) network that incorporates an attention mechanism. The Bi-LSTM network receives the core low-dimensional temporal feature sequence in a sliding window manner and outputs the result after weighted fusion of features at different time steps through an attention layer.

6. The intelligent sports state analysis and posture estimation method based on multi-module collaborative optimization according to claim 1, characterized in that: It also includes step S5: Based on the motion classification and posture assessment results output in step S4, and using the pre-built sports biomechanics knowledge base, targeted training improvement suggestions and potential injury risk warnings are generated through result analysis, feature association, multi-level matching and reasoning.

7. The intelligent sports state analysis and posture estimation method based on multi-module collaborative optimization according to claim 6, characterized in that: The sports biomechanics knowledge base includes: a standard movement pattern library, a common error pattern library, an injury risk rule library, and a quantitative assessment and feedback template library; the matching and reasoning process includes at least one of dynamic time warping (DTW) template matching, threshold-based rule matching, and generative rule reasoning.

8. An intelligent sports activity status analysis system, characterized in that, include: The image processing module is used to receive video streams and output the initial joint coordinate sequence of the human target with tracking ID; The key point optimization module, connected to the image processing module, is used to receive multiple frames of the initial joint point coordinate sequence, and perform nonlinear mapping based on the radial basis function network to output a high-precision joint point coordinate sequence. The feature engineering module, connected to the key point optimization module, is used to extract multidimensional kinematic features from the high-precision joint coordinate sequence and perform dimensionality reduction using principal component analysis to output a low-dimensional temporal feature sequence. The temporal analysis module, connected to the feature engineering module, is used to receive the low-dimensional temporal feature sequence, perform temporal modeling based on the long short-term memory network, and output action classification results and posture evaluation parameters.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images