Deep learning-based sports action micro-analysis method and system

Abstract

This invention relates to the field of motion analysis technology and discloses a method and system for micro-analysis of sports movements based on deep learning. The system includes a data acquisition module, a preprocessing module, a feature extraction module, a deep learning analysis module, an evaluation and matching module, an optimization feedback module, a report generation module, and an adaptive learning module. The method acquires micro-data of sports movements based on multi-sensor fusion technology, and combines preprocessing algorithms for data denoising, time alignment, and normalization. It performs intelligent analysis of the micro-feature dataset based on a deep learning model, using convolutional neural networks to identify spatial patterns and long short-term memory networks to model temporal dependencies. It optimizes matching between reinforcement learning algorithms and standard motion templates, and combines a federated learning framework to improve the model's generalization ability. This invention improves the reliability, personalization, and practicality of micro-analysis of sports movements.

Description

Technical Field

[0001] This invention relates to the field of motion analysis technology, specifically to a method and system for microscopic analysis of sports motion based on deep learning. Background Technology

[0002] Motion analysis is a method pioneered by the Gilbreths, which studies the operator's subtle movements, records the movement content of body parts with specific symbols, eliminates irrational phenomena, and formulates standard movement sequences. This method includes two levels: process analysis and motion study, and covers a technical system such as motion element analysis, image analysis, and motion economy principles. It is used to optimize work sequence, reduce ineffective movements, and reduce fatigue.

[0003] Currently, due to the various complex factors involved in sports motion analysis, the microscopic data acquired by the multi-sensor acquisition system during deep learning-based microscopic analysis of sports motion cannot guarantee the synchronization and consistency between video sequences, inertial measurement unit data, and biomechanical sensor data in real time. When the acquired data has time alignment errors and environmental noise interference, the quality of the preprocessed standardized motion data will decrease, and the accuracy of feature extraction cannot be guaranteed.

[0004] Therefore, a deep learning-based method and system for microscopic analysis of sports movements is proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based method and system for microscopic analysis of sports movements, which solves the problems mentioned in the background art, such as the inability to guarantee the synchronization and consistency between video sequences, inertial measurement unit data, and biomechanical sensor data in real time, as well as the inability to guarantee the accuracy of feature extraction.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method and system for microscopic analysis of sports movements based on deep learning, wherein the method includes the following steps: S1. Collect microscopic data of sports movements, including video sequences, inertial measurement unit data, and biomechanical sensor data, and generate microscopic data. S2. Preprocess the micro data, including denoising, time alignment and data normalization, to generate standardized motion data; S3. Based on the standardized motion data, extract the micro-features of sports movements, including joint angle changes, muscle activation timing and motion trajectory curvature, and generate a micro-feature dataset. S4. Analyze the micro-feature dataset using a deep learning model, identify spatial patterns through a convolutional neural network, and model temporal dependencies using a long short-term memory network to generate action analysis results. S5. Match the motion analysis results with predefined standard motion templates, calculate similarity scores and deviation indices, and generate preliminary evaluation data; S6. Based on the preliminary evaluation data, apply reinforcement learning algorithms to optimize action parameters and generate action optimization suggestions; S7. Integrate the motion analysis results, preliminary evaluation data, and motion optimization suggestions to construct a micro-analysis report of sports motions and output visual feedback.

[0007] Preferably, the microscopic data of sports movements collected in S1, including video sequences, inertial measurement unit data, and biomechanical sensor data, and the generation of microscopic data includes the following steps: S11. Collect multi-view video data of athletes in training scenarios through a high-speed camera array, and simultaneously record timestamps and spatial coordinates; S12. Use an inertial measurement unit kit to collect data on the athlete's limb acceleration, angular velocity, and orientation, ensuring a sampling frequency greater than 1000 Hz; S13. Collect muscle electrical signals and plantar pressure data through surface electromyography sensors and pressure distribution sensors to generate raw biomechanical data; S14. Integrate video data, inertial data, and biomechanical data into a unified timeline for micro-data streams of sports movements.

[0008] Preferably, the preprocessing of the microscopic data in step S2 includes denoising, time alignment, and data normalization to generate standardized motion data, comprising the following steps: S21. Preprocess the micro data, including performing inter-frame differencing and Gaussian filtering on the video sequence to reduce motion blur, and applying sliding window filtering to denoise the inertial data. S22. A dynamic time warping algorithm is used to align the data from multiple sources of sensors in time to ensure data point synchronization. S23. Scale the data to the [0,1] interval using the min-max normalization method to generate standardized motion data; S24. Store the preprocessed data in tensor format for input into the deep learning model.

[0009] Preferably, in step S3, based on the standardized motion data, the extraction of micro-features of sports movements, including joint angle changes, muscle activation timing, and motion trajectory curvature, to generate a micro-feature dataset includes the following steps: S31. Based on the standardized motion data, the OpenPose algorithm is used to extract the coordinates of key human points from the video data, and the joint angles and relative positions of limbs are calculated. The formula for calculating the joint angle is as follows: ; in, For joint angle, and These are the vector coordinates of adjacent limbs. For vector dot product operation, Let be the magnitude of the vector; S32. Apply wavelet transform to the inertial measurement unit data to extract frequency domain features, including the main frequency band energy and signal entropy; S33. Calculate the muscle activation peak and pressure center trajectory from biomechanical data to generate a temporal feature vector; S34. Merge all features into a multidimensional micro-feature dataset and use principal component analysis for dimensionality reduction.

[0010] Preferably, step S4 involves analyzing the micro-feature dataset using a deep learning model, identifying spatial patterns through a convolutional neural network, and modeling temporal dependencies using a long short-term memory network to generate action analysis results, including the following steps: S41. Construct a deep neural network model, whose input layer receives the micro-feature dataset, and whose hidden layer includes a convolutional neural network layer and a long short-term memory network layer. S42. The model is pre-trained using a sports movement standard database, and the weight parameters are optimized using the backpropagation algorithm. S43. Input the real-time motion data into the trained model and output the motion classification results, anomaly detection scores and time-series prediction values. S44. Highlight key action segments based on attention mechanisms to generate interpretable motion analysis results.

[0011] Preferably, step S5, which involves matching the motion analysis results with a predefined standard motion template to calculate a similarity score and a deviation index, and generating preliminary evaluation data, includes the following steps: S51. Establish a standard action template library, which includes the feature vectors and tolerance thresholds of ideal actions; S52. Calculate the Euclidean distance and dynamic time bending distance between the motion analysis results and the standard template, and generate a similarity matrix; The formula for calculating the Euclidean distance is as follows: ; in, For feature vectors and The Euclidean distance between them and These are real-time action feature vectors and standard template feature vectors, respectively. and Let i be the i-th component of the vector. The dimension of the feature vector S53. Use support vector machines to perform multi-class classification and identify the type of action deviation. S54. Generate preliminary evaluation data including consistency scores and error labels.

[0012] Preferably, step S6, which involves applying a reinforcement learning algorithm to optimize action parameters and generate action optimization suggestions based on the preliminary evaluation data, includes the following steps: S61. Using the preliminary evaluation data as state input, design a reward function based on the magnitude of action improvement; S62. Use the proximal policy optimization algorithm to search for the optimal action parameter adjustment strategy; S63. Generate personalized motion optimization suggestions, including adjustments to joint range of motion and force application timing; S64. Verify the effectiveness of the optimization suggestions through simulation and update it iteratively.

[0013] Preferably, step S7, which integrates the motion analysis results, preliminary evaluation data, and motion optimization suggestions to construct a micro-analysis report of sports motions and outputs visual feedback, includes the following steps: S71. Integrate the motion analysis results, preliminary evaluation data, and motion optimization suggestions into a structured data table; S72. Use data visualization tools to generate heat maps, time-series curves, and 3D animation feedback; S73. Output analysis reports via web interface and mobile application, supporting real-time viewing and historical comparison; S74 integrates voice and text prompts, providing instant corrective guidance.

[0014] Preferably, the method further includes S8, implementing multi-user data collaborative training based on a federated learning framework to ensure data privacy while improving model generalization ability.

[0015] Preferably, the system includes: The data acquisition module uses a multi-sensor fusion unit to synchronously acquire video, inertial, and biomechanical data, generates time-consistent raw data through a timestamp alignment unit, and outputs the raw motion stream through a data buffer unit. The preprocessing module receives the raw action stream, performs filtering and outlier removal through the denoising unit, synchronizes multi-source data using the alignment unit, and outputs standardized action data through the normalization unit. The feature extraction module receives the standardized motion data, calculates joint motion parameters through the key point detection unit, extracts signal features using the frequency domain analysis unit, and generates a micro-feature dataset through the feature fusion unit. The deep learning analysis module receives the micro-feature dataset, identifies spatial features through convolutional neural network units, models temporal dynamics using recurrent neural network units, and generates action analysis results through model inference units. The evaluation matching module receives the action analysis results, performs similarity calculation through the standard template library unit, identifies error patterns using the deviation analysis unit, and outputs preliminary evaluation data through the scoring unit. The optimization feedback module receives the preliminary evaluation data, searches for optimization strategies through the reinforcement learning unit, verifies the effectiveness of the suggestions using the simulation unit, and provides real-time correction instructions through the feedback generation unit. The report generation module integrates the action analysis results, evaluation data, and optimization feedback, generates charts and animations through the visualization unit, and outputs a comprehensive analysis report using the report compilation unit. The adaptive learning module receives historical user data, analyzes the athlete's ability baseline through the profile building unit, identifies training weaknesses through the weakness identification unit, and dynamically optimizes analysis parameters through the planning adjustment unit.

[0016] Compared with existing technologies, this invention provides a method and system for microscopic analysis of sports movements based on deep learning, which has the following beneficial effects: 1. In this invention, when performing microscopic data analysis of sports movements, the multi-sensor fusion unit in the data acquisition module synchronously acquires video sequences, inertial measurement unit data, and biomechanical sensor data. The preprocessing module performs data denoising, time alignment, and normalization, which can ensure the consistency and integrity of microscopic data in real time, avoid feature extraction errors caused by asynchronous acquisition and noise interference, ensure the accuracy of standardized movement data, and further improve the reliability of microscopic analysis of sports movements.

[0017] 2. In this invention, when performing micro-feature analysis of sports movements, key micro-features such as joint angle changes and muscle activation sequence are extracted through the feature extraction module, and real-time matching is performed using a deep learning analysis module combined with standard movement templates. This can automatically identify abnormal movement segments, enabling the system to reduce the deviation of movement analysis results from the actual movement state. Furthermore, when an anomaly is detected, preliminary evaluation data is generated through the evaluation matching module to ensure the accuracy of movement evaluation.

[0018] 3. In this invention, when performing sports movement optimization feedback, the reinforcement learning unit in the optimization feedback module generates personalized movement optimization suggestions based on individual differences, and uses a federated learning framework to achieve multi-user data collaborative optimization. This can automatically adapt the optimal training plan for different athletes, reduce the problem of insufficient universality of movement optimization suggestions, and further improve the personalization and practicality of sports movement micro-analysis. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the architecture of the deep learning-based sports motion micro-analysis system of the present invention; Figure 2 This is a flowchart illustrating the steps of the deep learning-based microscopic analysis method for sports movements according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1-2 The specific implementation of the deep learning-based micro-analysis method and system for sports movements is as follows, and the method includes the following steps: S1. Collect microscopic data of sports movements, including video sequences, inertial measurement unit data, and biomechanical sensor data, and generate microscopic data. S2. Preprocess the micro data, including denoising, time alignment and data normalization, to generate standardized motion data; S3. Based on standardized motion data, extract micro-features of sports movements, including joint angle changes, muscle activation timing and motion trajectory curvature, and generate a micro-feature dataset. S4. Analyze the micro-feature dataset using a deep learning model, identify spatial patterns through a convolutional neural network, and model temporal dependencies using a long short-term memory network to generate action analysis results. S5. Match the motion analysis results with predefined standard motion templates, calculate similarity scores and deviation indices, and generate preliminary evaluation data; S6. Based on the preliminary evaluation data, apply reinforcement learning algorithms to optimize action parameters and generate action optimization suggestions; S7. Integrate motion analysis results, preliminary assessment data, and motion optimization suggestions to construct a micro-analysis report of sports motions and output visual feedback.

[0022] The S1 collects microscopic data of sports movements, including video sequences, inertial measurement unit data, and biomechanical sensor data, and generates microscopic data through the following steps: S11. Collect multi-view video data of athletes in training scenarios through a high-speed camera array, and simultaneously record timestamps and spatial coordinates; S12. Use an inertial measurement unit kit to collect data on the athlete's limb acceleration, angular velocity, and orientation, ensuring a sampling frequency greater than 1000 Hz; S13. Collect muscle electrical signals and plantar pressure data through surface electromyography sensors and pressure distribution sensors to generate raw biomechanical data; S14. Integrate video data, inertial data, and biomechanical data into a unified timeline for micro-data streams of sports movements; The specific operations for multi-sensor data time alignment include: First, obtaining the timestamps of the raw data from each sensor. and reference time base ,in Based on the timestamps of the high-speed camera, the maximum permissible time offset is then calculated. Set to twice the sampling interval, and then, according to the calibration coefficient... Apply the time alignment formula: ; in, For the aligned timestamps, The timestamp of the raw sensor data. For reference time base, This is the maximum allowable time offset. This formula serves as a calibration coefficient, ensuring synchronization of multi-source data and improving the accuracy of data fusion. Finally, the aligned timestamps Data fusion is achieved to generate micro data streams with consistent time. This step ensures synchronization of multi-source data and avoids feature extraction errors caused by time deviations.

[0023] In S2, micro-data is preprocessed, including denoising, time alignment, and data normalization, to generate standardized motion data. This process includes the following steps: S21. Preprocess the micro data, including performing inter-frame differencing and Gaussian filtering on the video sequence to reduce motion blur, and applying sliding window filtering to denoise the inertial data. S22. A dynamic time warping algorithm is used to align the time of multi-source sensor data. This algorithm achieves synchronization by calculating the cumulative cost matrix between data sequences and backtracking to find the minimum cost path, thus ensuring data point synchronization. S23. Scale the data to the [0,1] interval using the min-max normalization method to generate standardized motion data; The specific steps include: First, extracting the minimum value from the preprocessed data sequence. and maximum value ,in and Obtained through sliding window statistics, then for each data point Apply the normalization formula: ; in, These are the normalized data values. The original data values, and These are the minimum and maximum values ​​of the data sequence, respectively. Finally, normalize the results. The data is stored in tensor format for use in subsequent models. This step addresses the influence of units and improves data stability. S24. Store the preprocessed data in tensor format for input into the deep learning model.

[0024] S3 extracts micro-features of sports movements based on standardized motion data, including joint angle changes, muscle activation timing, and motion trajectory curvature, generating a micro-feature dataset through the following steps: S31. Based on standardized motion data, load a pre-trained OpenPose convolutional neural network model, input video frames to extract the coordinates of 18 human body key points, and calculate joint angles and relative limb positions. The formula for calculating the joint angle is as follows: ; in, For joint angle, and These are the vector coordinates of adjacent limbs. For vector dot product operation, Given the magnitude of the vector, this formula quantifies the degree of joint bending, enhancing the accuracy of feature extraction. S32. Apply a 5-level discrete wavelet transform based on the db4 wavelet basis to the inertial measurement unit data to extract frequency domain features, including the main frequency band energy and signal entropy. S33. Calculate the muscle activation peak and pressure center trajectory from biomechanical data to generate a temporal feature vector; S34. Merge all features into a multidimensional micro-feature dataset and use principal component analysis for dimensionality reduction.

[0025] S4 uses deep learning models to analyze micro-feature datasets, identifies spatial patterns through convolutional neural networks, and models temporal dependencies using long short-term memory networks to generate action analysis results, including the following steps: S41. Construct a deep neural network model whose input layer receives a dataset of micro-features and whose hidden layer includes two convolutional neural networks and one long short-term memory network containing 128 units. S42. The model is pre-trained using a sports movement standard database, with cross-entropy loss function and Adam optimizer, and weight parameters are optimized through backpropagation algorithm; The specific steps include: First, loading the pre-training dataset and inputting the micro-feature dataset into the model; then, calculating the input weighted sum z for each neuron and applying the Sigmoid activation function formula: ; in, For the output of the Sigmoid function, The input to the neuron is weighted and summed. in, The weights are obtained by multiplying the weight matrix by the input features and adding a bias term. Finally, the weights are adjusted using the backpropagation algorithm to minimize the loss function and enhance the model's nonlinear fitting ability. S43. Input the real-time motion data into the trained model and output the motion classification results, anomaly detection scores and time-series prediction values. S44. Highlight key action segments based on attention mechanisms to generate interpretable motion analysis results.

[0026] In S5, the motion analysis results are matched with predefined standard motion templates to calculate similarity scores and deviation indices, generating preliminary evaluation data. This includes the following steps: S51. Establish a standard action template library, which includes the feature vectors and tolerance thresholds of ideal actions; S52. Calculate the Euclidean distance and dynamic time bending distance between the motion analysis results and the standard template. The dynamic time bending distance is calculated using classical boundary conditions, continuity constraints, and monotonic constraints to generate a similarity matrix. The formula for calculating the Euclidean distance is as follows: ; in, For feature vectors and The Euclidean distance between them and These are real-time action feature vectors and standard template feature vectors, respectively. and Let i be the i-th component of the vector. Given the dimension of the feature vector, this formula quantifies action similarity, improving the objectivity of the evaluation. S53. Use support vector machines to perform multi-class classification and identify the type of action deviation. S54. Generate preliminary evaluation data including consistency scores and error labels.

[0027] In S6, based on the preliminary evaluation data, reinforcement learning algorithms are applied to optimize action parameters and generate action optimization suggestions, including the following steps: S61. Using the preliminary evaluation data as the state input, design a reward function based on the magnitude of action improvement, specifically defined as the negative value of the Euclidean distance between the current action parameter vector and the target vector; The specific steps include: First, analyzing the motion bias values ​​in the preliminary assessment data; then, setting the discount factor γ to 0.9 and calculating the cumulative discount reward formula for time step t: ; in, For time step Cumulative discount rewards, As a discount factor, The immediate reward for the k-th step in the future; in, Based on the real-time calculation of the improvement in movement, finally, To optimize strategies and balance short-term and long-term returns; S62. Use the proximal policy optimization algorithm to search for the optimal action parameter adjustment policy. This algorithm updates the policy network parameters in each iteration through importance sampling and pruning mechanism to stabilize the training process. S63. Generate personalized motion optimization suggestions, including adjustments to joint range of motion and force application timing; S64. Verify the effectiveness of the optimization suggestions through simulation and update it iteratively.

[0028] S7 integrates motion analysis results, preliminary assessment data, and motion optimization suggestions to construct a micro-analysis report of sports movements and outputs visual feedback, including the following steps: S71. Integrate motion analysis results, preliminary evaluation data, and motion optimization suggestions into a structured data table; S72. Use data visualization tools to generate heat maps, time-series curves, and 3D animation feedback; The specific operations include: First, extracting real-time action parameter vectors from the structured data table. and ideal motion parameter vector Then, set the tolerance threshold. The formula for calculating the consistency score is 0.1: ; in, To score for consistency of movement, and These are the real-time motion parameter vector and the ideal motion parameter vector, respectively. This is the tolerance threshold; The norm is calculated using Euclidean distance; finally, the scores are... Mapping to a heatmap color gradient generates visual feedback; this step quantifies action deviations and enhances report clarity. S73. Output analysis reports via web interface and mobile application, supporting real-time viewing and historical comparison; S74 integrates voice and text prompts, providing instant corrective guidance.

[0029] The method also includes S8, which implements multi-user collaborative training based on a federated learning framework. Specifically, it adopts a federated averaging algorithm, which aggregates model weight updates from each client on a central server without exchanging the original data, thus ensuring data privacy while improving the model's generalization ability.

[0030] The system includes: The data acquisition module uses a multi-sensor fusion unit to synchronously acquire video, inertial, and biomechanical data, generates time-consistent raw data through a timestamp alignment unit, and outputs the raw motion stream through a data buffer unit. The preprocessing module receives the raw motion stream, performs filtering and outlier removal through the denoising unit, synchronizes multi-source data using the alignment unit, and outputs standardized motion data through the normalization unit. The feature extraction module receives standardized motion data, calculates joint motion parameters through the key point detection unit, extracts signal features using the frequency domain analysis unit, and generates a micro-feature dataset through the feature fusion unit. The deep learning analysis module receives a dataset of micro-features, identifies spatial features through convolutional neural network units, models temporal dynamics using recurrent neural network units, and generates action analysis results through the model inference unit. The evaluation matching module receives the action analysis results, performs similarity calculations through the standard template library unit, identifies error patterns using the deviation analysis unit, and outputs preliminary evaluation data through the scoring unit. The optimization feedback module receives preliminary evaluation data, searches for optimization strategies through reinforcement learning units, verifies the effectiveness of the suggestions using simulation units, and provides real-time correction instructions through the feedback generation unit. The report generation module integrates action analysis results, evaluation data, and optimization feedback, generates charts and animations through the visualization unit, and outputs a comprehensive analysis report using the report compilation unit. The adaptive learning module receives historical user data, analyzes the athlete's ability baseline through the profile building unit, identifies training weaknesses through the weakness identification unit, and dynamically optimizes analysis parameters through the planning adjustment unit.

[0031] The operational steps of the deep learning-based sports motion micro-analysis method and system are as follows: I. Synchronous Acquisition of Multi-Source Data: The acquisition of microscopic data on sports movements is achieved through multi-sensor fusion technology. First, a high-speed camera array is used to capture multi-view video sequences of athletes, and timestamps and spatial coordinates are recorded simultaneously. At the same time, limb motion parameters are collected through an inertial measurement unit kit, and muscle activation signals are obtained by combining surface electromyography (EMG) sensors. All sensor data are synchronously fused through a timestamp alignment unit to form a microscopic data stream of sports movements with a unified timeline, providing a complete data foundation for subsequent analysis.

[0032] II. Data Preprocessing and Standardization: A multi-level processing flow is adopted to ensure data quality. Inter-frame differencing and Gaussian filtering are applied to video data to solve motion blur. Sliding window filtering is applied to inertial data for noise reduction. Dynamic time warping algorithm is used to achieve time alignment of multi-source data. Then, the min-max normalization method is used to scale the data to a standard range, and finally standardized motion data is generated to prepare for feature extraction.

[0033] III. Multi-dimensional Feature Extraction: Multi-level feature mining is performed based on standardized motion data. The OpenPose algorithm is used to extract the coordinates of key points on the human body and calculate joint angles and motion trajectories. Wavelet transform is performed on inertial data to extract frequency domain features. Muscle activation time sequence is analyzed from biomechanical data. All features are fused into a micro-feature dataset after dimensionality reduction by principal component analysis to characterize motion characteristics.

[0034] IV. Deep Learning Model Analysis: A composite neural network model is constructed for action analysis. The input layer of the model receives a dataset of micro-features, extracts spatial features through a convolutional neural network, models temporal dynamics through a long short-term memory network, focuses on key action segments based on an attention mechanism, and outputs action classification results and anomaly detection scores to generate interpretable action analysis results.

[0035] V. Intelligent Assessment and Matching: A standard action template library is established for comparison. The Euclidean distance and dynamic time curvature distance between real-time action features and standard templates are calculated. Action deviation classification is performed using support vector machines to generate preliminary evaluation data containing consistency scores and error labels, providing a basis for optimization.

[0036] VI. Adaptive Optimization Feedback: Personalized adjustments are achieved based on a reinforcement learning framework. Preliminary evaluation data is used as state input, a reward function is designed to evaluate the effect of action improvement, the optimal parameter adjustment strategy is searched through a proximal policy optimization algorithm, targeted action optimization suggestions are generated, and continuous iteration is verified through simulation.

[0037] VII. Visual Report Generation: The integrated analysis results generate comprehensive feedback, merging motion analysis results, evaluation data, and optimization suggestions into structured data, which is presented in a visual format such as heatmaps and time-series curves. Combined with voice and text prompts, analysis reports are output to achieve real-time monitoring and guidance of training effects.

[0038] VIII. System Adaptive Learning: Continuous optimization is achieved based on the federated learning framework. By constructing athlete ability profiles through user historical data, weak points in training are identified, and analysis parameters and training plans are dynamically adjusted to ensure that the system can adapt to the personalized needs of different athletes and continuously improve the accuracy of analysis.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0040] 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 alterations 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 claims and their equivalents.

Claims

1. A deep learning-based method for microscopic analysis of sports movements, characterized in that, The method includes the following steps: S1. Collect microscopic data of sports movements, including video sequences, inertial measurement unit data, and biomechanical sensor data, and generate microscopic data. S2. Preprocess the micro data, including denoising, time alignment and data normalization, to generate standardized motion data; S3. Based on the standardized motion data, extract the micro-features of sports movements, including joint angle changes, muscle activation timing and motion trajectory curvature, and generate a micro-feature dataset. S4. Analyze the micro-feature dataset using a deep learning model, identify spatial patterns through a convolutional neural network, and model temporal dependencies using a long short-term memory network to generate action analysis results. S5. Match the motion analysis results with predefined standard motion templates, calculate similarity scores and deviation indices, and generate preliminary evaluation data; S6. Based on the preliminary evaluation data, apply reinforcement learning algorithms to optimize action parameters and generate action optimization suggestions; S7. Integrate the motion analysis results, preliminary evaluation data, and motion optimization suggestions to construct a micro-analysis report of sports motions and output visual feedback.

2. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, The process of collecting microscopic data of sports movements in S1, including video sequences, inertial measurement unit data, and biomechanical sensor data, and generating microscopic data includes the following steps: S11. Collect multi-view video data of athletes in training scenarios through a high-speed camera array, and simultaneously record timestamps and spatial coordinates; S12. Use an inertial measurement unit kit to collect data on the athlete's limb acceleration, angular velocity, and orientation, ensuring a sampling frequency greater than 1000 Hz; S13. Collect muscle electrical signals and plantar pressure data through surface electromyography sensors and pressure distribution sensors to generate raw biomechanical data; S14. Integrate video data, inertial data, and biomechanical data into a unified timeline for micro-data streams of sports movements.

3. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, The preprocessing of the microscopic data in step S2 includes denoising, time alignment, and data normalization to generate standardized motion data, comprising the following steps: S21. Preprocess the micro data, including performing inter-frame differencing and Gaussian filtering on the video sequence to reduce motion blur, and applying sliding window filtering to denoise the inertial data. S22. A dynamic time warping algorithm is used to align the data from multiple sources of sensors in time to ensure data point synchronization. S23. Scale the data to the [0,1] interval using the min-max normalization method to generate standardized motion data; S24. Store the preprocessed data in tensor format for input into the deep learning model.

4. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, In step S3, based on the standardized motion data, micro-features of sports movements are extracted, including joint angle changes, muscle activation timing, and motion trajectory curvature. The process of generating a micro-feature dataset includes the following steps: S31. Based on the standardized motion data, the OpenPose algorithm is used to extract the coordinates of key human points from the video data, and the joint angles and relative positions of limbs are calculated. The formula for calculating the joint angle is as follows: ； in, For joint angle, and These are the vector coordinates of adjacent limbs. For vector dot product operation, Let be the magnitude of the vector; S32. Apply wavelet transform to the inertial measurement unit data to extract frequency domain features, including the main frequency band energy and signal entropy; S33. Calculate the muscle activation peak and pressure center trajectory from biomechanical data to generate a temporal feature vector; S34. Merge all features into a multidimensional micro-feature dataset and use principal component analysis for dimensionality reduction.

5. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, In step S4, a deep learning model is used to analyze the micro-feature dataset. Spatial patterns are identified through a convolutional neural network, and temporal dependencies are modeled using a long short-term memory network to generate action analysis results. This includes the following steps: S41. Construct a deep neural network model, whose input layer receives the micro-feature dataset, and whose hidden layer includes a convolutional neural network layer and a long short-term memory network layer. S42. The model is pre-trained using a sports movement standard database, and the weight parameters are optimized using the backpropagation algorithm. S43. Input the real-time motion data into the trained model and output the motion classification results, anomaly detection scores and time-series prediction values. S44. Highlight key action segments based on attention mechanisms to generate interpretable motion analysis results.

6. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, In step S5, the motion analysis results are matched with predefined standard motion templates to calculate similarity scores and deviation indices, generating preliminary evaluation data. This includes the following steps: S51. Establish a standard action template library, which includes the feature vectors and tolerance thresholds of ideal actions; S52. Calculate the Euclidean distance and dynamic time bending distance between the motion analysis results and the standard template, and generate a similarity matrix; The formula for calculating the Euclidean distance is as follows: ； in, For feature vectors and The Euclidean distance between them and These are real-time action feature vectors and standard template feature vectors, respectively. and Let i be the i-th component of the vector. The dimension of the feature vector S53. Use support vector machines to perform multi-class classification and identify the type of action deviation. S54. Generate preliminary evaluation data including consistency scores and error labels.

7. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, In step S6, based on the preliminary evaluation data, the application of a reinforcement learning algorithm to optimize action parameters and generate action optimization suggestions includes the following steps: S61. Using the preliminary evaluation data as state input, design a reward function based on the magnitude of action improvement; S62. Use the proximal policy optimization algorithm to search for the optimal action parameter adjustment strategy; S63. Generate personalized motion optimization suggestions, including adjustments to joint range of motion and force application timing; S64. Verify the effectiveness of the optimization suggestions through simulation and update it iteratively.

8. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, The S7 process integrates the motion analysis results, preliminary evaluation data, and motion optimization suggestions to construct a micro-analysis report of sports motions and outputs visual feedback, including the following steps: S71. Integrate the motion analysis results, preliminary evaluation data, and motion optimization suggestions into a structured data table; S72. Use data visualization tools to generate heat maps, time-series curves, and 3D animation feedback; S73. Output analysis reports via web interface and mobile application, supporting real-time viewing and historical comparison; S74 integrates voice and text prompts, providing instant corrective guidance.

9. The deep learning-based microscopic analysis method for sports movements according to claim 1, characterized in that, The method also includes S8, which implements multi-user data collaborative training based on a federated learning framework, ensuring data privacy while improving the model's generalization ability.

10. A deep learning-based sports motion micro-analysis system, used to implement the deep learning-based sports motion micro-analysis method according to any one of claims 1-9, characterized in that, The system includes: The data acquisition module uses a multi-sensor fusion unit to synchronously acquire video, inertial, and biomechanical data, generates time-consistent raw data through a timestamp alignment unit, and outputs the raw motion stream through a data buffer unit. The preprocessing module receives the raw action stream, performs filtering and outlier removal through the denoising unit, synchronizes multi-source data using the alignment unit, and outputs standardized action data through the normalization unit. The feature extraction module receives the standardized motion data, calculates joint motion parameters through the key point detection unit, extracts signal features using the frequency domain analysis unit, and generates a micro-feature dataset through the feature fusion unit. The deep learning analysis module receives the micro-feature dataset, identifies spatial features through convolutional neural network units, models temporal dynamics using recurrent neural network units, and generates action analysis results through model inference units. The evaluation matching module receives the action analysis results, performs similarity calculation through the standard template library unit, identifies error patterns using the deviation analysis unit, and outputs preliminary evaluation data through the scoring unit. The optimization feedback module receives the preliminary evaluation data, searches for optimization strategies through the reinforcement learning unit, verifies the effectiveness of the suggestions using the simulation unit, and provides real-time correction instructions through the feedback generation unit. The report generation module integrates the action analysis results, evaluation data, and optimization feedback, generates charts and animations through the visualization unit, and outputs a comprehensive analysis report using the report compilation unit. The adaptive learning module receives historical user data, analyzes the athlete's ability baseline through the profile building unit, identifies training weaknesses through the weakness identification unit, and dynamically optimizes analysis parameters through the planning adjustment unit.

Images