A real-time training assistance system for rotating table tennis balls

Abstract

This invention belongs to the field of table tennis training assistance technology and provides a real-time training assistance system for spinning table tennis. This system simultaneously captures the hand shape and motion characteristics of the ball during the serve and combines motion information with a physical model to achieve spin recognition. Compared with solutions that rely solely on trajectory inference, the recognition accuracy is higher and it can effectively distinguish similar spin types with small differences in motion. It does not require modification of the athlete's commonly used racket, does not change the racket weight or the feel of hitting the ball, avoids the impact of sensor solutions on the athlete's normal training, and has a lower overall deployment cost. Low-latency inference is achieved through edge computing optimization, which can output recognition results and improvement suggestions in real time during training. It supports on-site guidance from coaches and can also help athletes complete motion corrections independently, adapting to the quantitative analysis needs of daily routine training.

Description

Technical Field

[0001] This invention belongs to the field of table tennis training assistance technology, specifically relating to a real-time training assistance system for a rotating table tennis ball. Background Technology

[0002] In table tennis, the type and intensity of spin directly determine the threat level of a serve. Professional coaches typically rely on visual observation combined with experience to judge the spin. This method, dependent on manual judgment, suffers from significant subjective bias and low efficiency, making it unsuitable for quantitative analysis and daily training. With the development of computer vision technology, existing technologies mostly estimate spin type by capturing the ball's trajectory using a racket with sensors or a high-speed camera. However, modifying the racket with sensors is costly, altering its weight and feel, and affecting the athlete's normal serving motion. Furthermore, spin estimation based on the ball's trajectory can only be identified after the ball lands, failing to capture the hand and racket movements at the moment of serve. The accuracy is significantly affected by environmental occlusion and trajectory capture errors, making it difficult to distinguish between similar spin types with minor differences in movement. Therefore, how to automatically and accurately identify the type and intensity of table tennis serve spin based on the movement characteristics during the serve and utilize this information in athlete training is a crucial problem that needs to be solved in the field of table tennis training assistance. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a real-time training assistance system for rotating table tennis balls, thereby resolving the issues in the prior art. The technical solution adopted by this invention is as follows: A real-time training assistance system for spinning table tennis includes: A hand motion analysis model is used to capture changes in hand and upper limb posture during a table tennis serve. The rotation prediction and modeling module is used for 3D trajectory reconstruction, physical model inversion, and multimodal feature fusion. A real-time training assistance module for interactive use by athletes and coaches.

[0004] Furthermore, the hand gesture analysis model includes: A high-precision hand key point detection and optimization module is used to locate 21 hand key points; Temporal modeling and occlusion processing module: Hand key point recovery model based on graph structure reasoning, which infers the occluded parts using visible nodes and historical temporal data; Action feature extraction and classification module: Construct a multi-scale spatiotemporal convolutional network to perform frequency and temporal domain analysis on the motion features of the hand, wrist and forearm, and combine it with a bidirectional recurrent neural network with attention mechanism to achieve automatic recognition and classification of technical actions.

[0005] Furthermore, the high-precision hand key point detection and optimization module is jointly established based on the improved YOLOv8 architecture and the optimized MediaPipe model.

[0006] Furthermore, the rotation prediction and modeling module includes: High-precision 3D trajectory reconstruction: Based on the video stream synchronously acquired by multi-view high-speed cameras, the motion trajectory of the ping-pong ball in 3D space is reconstructed through the principle of binocular parallax and motion estimation algorithm; Rotation inference based on physical models: By combining aerodynamic models and analyzing the curvature changes of the trajectory in the horizontal and vertical directions, the initial rotation vector of the ping-pong ball can be deduced. Multimodal fusion rotation recognition model: By dynamically weighting different features through a cross-modal attention mechanism, it achieves high-precision classification of rotations such as no rotation, upspin, downspin, left-hand rotation, right-hand rotation and their combinations, and performs graded evaluation of rotation intensity.

[0007] Furthermore, in the high-precision three-dimensional trajectory reconstruction, depth sensor-assisted calibration is introduced, point cloud registration technology is used to improve reconstruction accuracy, and an adaptive trajectory smoothing filter is used to eliminate noise and maintain the physical rationality of the trajectory.

[0008] Furthermore, the real-time training assistance module includes: Model optimization and edge computing deployment: The optimized model is deployed on edge computing devices; Low-latency system architecture design: including a front-end and back-end separated real-time analysis system to support high concurrency and low latency requirements; Interaction and feedback mechanism: Based on the user interface, it enables real-time video analysis or offline analysis by uploading.

[0009] Furthermore, the model optimization and edge computing deployment also include: adopting dynamic memory allocation and garbage collection mechanisms to avoid lag or crashes on resource-constrained devices.

[0010] Furthermore, the low-latency system architecture design also includes: the backend uses the Python Flask framework to provide a RESTful API, and integrates OpenCV and TensorFlow / PyTorch for video stream processing and model inference; the frontend is developed based on React and Three.js to realize the visualization of motion capture, rotation parameters, and trajectory prediction results.

[0011] Furthermore, the low-latency system architecture design also includes: implementing real-time data push based on the WebSocket protocol to reduce communication overhead, ensure data synchronization and update, and avoid latency caused by polling.

[0012] Furthermore, the interaction and feedback mechanism also includes generating an assessment report that includes rotation type, intensity level, movement continuity score, and improvement suggestions to assist in guidance and athlete self-correction.

[0013] This invention has the following advantages: It can simultaneously capture the hand shape and motion characteristics of the ball during the serve, and combine motion information with a physical model to achieve spin recognition. Compared with schemes that rely solely on trajectory inference, the recognition accuracy is higher, and it can effectively distinguish similar spin types with small differences in motion. It does not require modification of the athlete's commonly used racket, does not change the racket weight or the feel of hitting the ball, avoids the impact of sensor solutions on the athlete's normal training, and has a lower overall deployment cost. Low-latency inference is achieved through edge computing optimization, which can output recognition results and improvement suggestions in real time during training. It supports on-site guidance from coaches and can also help athletes complete motion corrections independently, adapting to the quantitative analysis needs of daily routine training. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the overall process and technical framework of the present invention. Figure 2 The fusion of spatiotemporal features of hand bones and racket kinematics is used for multimodal motion analysis; Figure 3 This is a schematic diagram of the real-time video analytics backend. Detailed Implementation

[0015] The following will be described in conjunction with embodiments of the present invention. Figures 1-3 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0016] This invention proposes a real-time training assistance system for rotating table tennis balls, comprising: A hand motion analysis model is used to capture changes in hand and upper limb posture during a table tennis serve. like Figure 1 Hand motion analysis models are a fundamental and crucial component, aiming to accurately capture hand and upper limb posture changes during a table tennis serve using improved keypoint detection and post-processing techniques, providing reliable input for subsequent spin and trajectory prediction. Specifically, this includes: (1) High-precision hand key point detection and optimization module: such as Figure 1As shown, this invention employs a multi-level cascaded detection method based on an improved YOLOv8 architecture and an optimized MediaPipe model. Addressing motion blur and complex hand grips in table tennis, it introduces a dedicated dataset, adaptive heatmap mechanism, and anatomical constraints to achieve sub-millimeter accuracy in locating 21 hand keypoints. This invention utilizes a multi-level cascaded architecture of YOLOv8+MediaPipe to solve the challenges of motion blur and finger occlusion caused by high-speed swings. Furthermore, data augmentation techniques simulate different lighting and background conditions to improve model generalization ability, and transfer learning is used to accelerate the training process. During model training, this invention employs a hybrid loss function, combining joint distance error and pose consistency loss to further optimize the stability of keypoint prediction. Simultaneously, for subtle finger movements, a multi-resolution feature pyramid network is introduced to capture multi-level information from global to local perspectives, ensuring detection accuracy even under high-speed motion. To further enhance detection robustness, this invention also integrates depth sensor data for multi-modal fusion, reducing the limitations of single visual information, and dynamically adjusts model parameters through an online learning mechanism to adapt to the personalized hand features of different athletes.

[0017] (2) Temporal Modeling and Occlusion Processing Module: To address the occlusion problem during fast swings, an adaptive smoothing algorithm based on Kalman filtering and bidirectional temporal consistency constraints is proposed. A keypoint recovery model based on graph structure reasoning is designed, utilizing visible nodes and historical temporal data to infer occluded parts, ensuring the integrity and stability of the keypoint sequence. This model can handle interactive occlusion between the hand, racket, and ball, and enhances the weight of keyframes through a temporal attention mechanism. This invention also integrates a spatiotemporal graph convolutional network to model the motion trajectory of keypoints to capture long-term dependencies, and combines it with a generative adversarial network to generate reasonable occlusion completion sequences, thereby improving the recovery effect in complex scenes. In addition, a self-supervised learning strategy is introduced, using unlabeled video data to pre-train the model, enhancing temporal reasoning capabilities, reducing dependence on a large amount of labeled data, and optimizing the realism of the generated sequences through adversarial training, ensuring low latency and high accuracy in real-time applications.

[0018] (3) Action Feature Extraction and Classification Module: A multi-scale spatiotemporal convolutional network was constructed to perform frequency and temporal domain analysis on the motion features of the hand, wrist, and forearm. Combined with a bidirectional recurrent neural network (BiLSTM+Attention) with an attention mechanism, it automatically identifies and classifies 12 key technical action units, including topspin, backspin, and sidespin. The CNN layer extracts the spatial structural features of the hand posture, the BiLSTM layer captures the dynamic evolution of the swing motion, and the attention mechanism focuses on key frames such as the moment of ball contact, thereby distinguishing subtle differences between similar actions such as topspin, backspin, and sidespin. Furthermore, this invention integrates kinematic features, such as angular velocity and acceleration, to enhance the model's discriminative ability through a feature fusion layer. Cross-entropy loss and center loss are used for joint optimization to improve classification accuracy and robustness. To verify the classification effect, this invention conducted extensive experiments on public datasets and self-built datasets. The results show that the model's recognition rate exceeds 95% in complex scenarios, demonstrating its effectiveness and practicality. Simultaneously, the model also supports incremental learning, continuously optimizing with new data to adapt to technological evolution and changes in the styles of different athletes. Through the collaborative work of the above three sub-modules, the hand motion analysis model can comprehensively and accurately capture the details of the serve motion. Combined with the real-time feedback mechanism, it lays a solid foundation for the rotation and trajectory prediction of the entire system, and ultimately promotes the intelligent development of table tennis training and referee assistance systems.

[0019] The rotation prediction and modeling module is used for 3D trajectory reconstruction, physical model inversion, and multimodal feature fusion. The rotation prediction and modeling module aims to achieve this through 3D trajectory reconstruction, physical model inversion, and multimodal feature fusion, such as... Figure 2 The system accurately identifies the spin type and intensity of serves in real time. Designed for table tennis training and match analysis, it processes high-speed motion data in real time, providing athletes and coaches with precise spin feedback, thereby improving the scientific rigor of technical assessments and tactical planning. Specifically, it includes: (1) High-precision 3D trajectory reconstruction: Based on the video stream synchronously acquired by a multi-view high-speed camera (≥1000fps), the sub-millimeter-level motion trajectory of the ping-pong ball in 3D space is reconstructed through the principle of binocular parallax and motion estimation algorithm. A calibration board is used for camera calibration to ensure the accuracy of the spatial coordinate system, and optical flow is used to compensate for the error caused by motion blur. This invention also introduces depth sensor-assisted calibration, improves the reconstruction accuracy through point cloud registration technology, and develops an adaptive trajectory smoothing filter to eliminate noise and maintain the physical rationality of the trajectory. Specifically, the camera calibration process includes the precise calibration of internal parameters (such as focal length and distortion coefficient) and external parameters (such as position and attitude); the trajectory reconstruction algorithm combines feature point matching and dense optical flow to achieve continuous tracking of the ball's position.

[0020] (2) Rotation Inference Based on Physical Model: Combining aerodynamic models (gravity, drag, Magnus force), a rotation calculation algorithm based on trajectory inversion (Gray-Box model) was developed. By analyzing the curvature changes of the trajectory in the horizontal and vertical directions, the initial rotation vector (angular velocity magnitude and axial direction) of the ping-pong ball is inverted. The model also considers factors such as table elasticity, ball material, and air density, and improves the inversion accuracy through iterative optimization, and introduces uncertainty estimation to assess the reliability of the prediction. This invention further extends the model by incorporating surface friction coefficient and racket rubber characteristic parameters, making the rotation inference closer to the actual physical process. At the same time, Bayesian optimization method is used to automatically adjust parameters to minimize prediction errors. For example, the Magnus force model calculates the lateral force based on the cross product of rotational velocity and linear velocity, thereby explaining the curvature of the trajectory; iterative optimization uses the Levenberg-Marquardt algorithm, combined with experimental data to fit the model parameters; uncertainty estimation is based on Monte Carlo simulation, quantifying the confidence interval of the inversion results to ensure the robustness of the system in complex environments. This invention extends the dual-path architecture into a "visual semantic-physical inversion" multimodal fusion framework—not only integrating the dual-path estimation of racket posture, but also incorporating semantic features such as hand motion sequences and hitting point positions output by the hand shape analysis model into a unified decision space. Through cross-modal attention mechanisms, information complementarity is achieved, significantly improving the recognition stability of complex spins (especially combined spins).

[0021] (3) Multimodal Fusion Spin Recognition Model: An integrated decision network was designed, which integrates physical features (rotation vector, curvature, etc.) obtained from trajectory inversion and semantic features (hand action sequence, racket posture, hitting point position, etc.) obtained from hand shape analysis model. The model dynamically weights different features through a cross-modal attention mechanism to achieve high-precision classification of no-spin, topspin, backspin, left-side spin, right-side spin and their combinations, and performs graded evaluation of spin intensity. For example, for strong backspin, the model will simultaneously capture the rapid drop of the trajectory and the downward friction action of the hand, and output a confidence score. This invention also introduces a graph neural network to model the relationships between multimodal features in order to capture the complex interactions between features, and combines multiple base classifiers through ensemble learning to improve the generalization ability and anti-interference ability of the overall system. Specifically, the cross-modal attention mechanism calculates the correlation weights between physical and semantic features, enabling the model to focus on key information; the graph neural network represents features as nodes and edges as dependencies between features, aggregating information through graph convolutional layers; ensemble learning uses random forests and gradient boosting trees as base classifiers, outputting the final classification result through voting or weighted averaging, ensuring the system's stability under noisy data or abnormal conditions. Temporal stability is enhanced by improving the YOLOv8-MediaPipe network by combining Kalman filtering and Local Refinement Network (LRN), while an edge computing feedback system (latency <0.2s) is deployed to meet real-time training requirements. This invention achieves three breakthroughs: first, the modal dimension is expanded from "visual + mechanical" to four modalities: "visual + IMU + force table + electromyography," resulting in more comprehensive feature representation; second, the introduction of Transformer to replace the early CNN-LSTM backbone significantly enhances long-range dependency modeling capabilities; and third, the expert knowledge embedding mechanism transforms coaching experience into learnable rule constraints, enabling the model to maintain physical interpretability even with 1000+ samples. Experiments show that the upgraded fusion model improves the accuracy of spin type recognition from 90% to 94.5% and the accuracy of intensity classification from 85% to 87.5%, while maintaining real-time performance with an end-to-end latency of ≤0.2 seconds. This fully verifies the continuous optimization potential of the multimodal fusion technology route in intelligent analysis of table tennis. Furthermore, the model training used a large-scale labeled dataset covering serve samples from different athletes, rackets, and environments to enhance generalization performance.

[0022] The real-time training assistance module is deployed as an interactive system that athletes and coaches can use directly, such as... Figure 3 As shown. The system design emphasizes real-time performance, interactivity, and usability to meet the immediate feedback requirements in training scenarios, while also considering ease of use and scalability; specifically including: (1) Model Optimization and Edge Computing Deployment: This invention utilizes model pruning and quantization techniques (such as converting FP32 models to INT8) to compress YOLOv8, LSTM, and other models, significantly reducing computational load and storage requirements. Specifically, pruning removes redundant parameters, and quantization reduces model size while maintaining controllable accuracy loss, thereby improving inference speed while maintaining high accuracy. These optimized models have been successfully deployed on edge computing devices such as the NVIDIA Jetson series, ensuring localized real-time analysis capabilities, reducing reliance on cloud services, and enhancing data privacy. Simultaneously, this invention optimizes memory management strategies, employing dynamic memory allocation and garbage collection mechanisms to avoid stuttering or crashes on resource-constrained devices. The dynamic model loading function allows the system to automatically select the optimal model version based on hardware configuration, improving compatibility and flexibility. This invention also tested performance under different power consumption modes, balancing accuracy and speed through an adaptive scheduling algorithm. For example, real-time performance is prioritized in low-power mode, while analysis accuracy is improved in high-performance mode, ensuring stable operation for several hours in battery-powered scenarios, and supporting hot-swapping and remote monitoring functions.

[0023] (2) Low-latency system architecture design: This invention designs a real-time analysis system with front-end and back-end separation to support high concurrency and low latency requirements. The back-end uses the Python Flask framework to provide a RESTful API and integrates libraries such as OpenCV, TensorFlow / PyTorch for video stream processing and model inference. Through asynchronous processing and GPU acceleration, this invention optimizes the decoding, inference, and result return process of video frames, ensuring that the end-to-end analysis latency is controlled within 0.2 seconds, meeting the requirements of real-time feedback, and supporting parallel processing of multiple video streams. This architecture design draws on the engineering experience of the intelligent training system for table tennis trajectory routes based on the Eagle Eye system—this system addresses the challenge of real-time synchronization of multi-source heterogeneous data in high-speed motion scenarios (ball speed > 20m / s), and also adopts asynchronous processing and GPU acceleration strategies to reduce latency, realizing a complete link from data acquisition to closed-loop feedback of decision-making. The front-end is developed based on React and Three.js, realizing the visualization of motion capture, rotation parameters, and trajectory prediction results. The user interface supports 3D animation playback and slow-motion viewing, and allows interactive adjustment of the view angle for multi-angle observation of action details. It also provides color coding and annotation tools to enhance visualization. The system also integrates [unspecified features]. Furthermore, it employs load balancing technologies such as Nginx reverse proxy and containerized deployment to handle high concurrency access, ensuring a smooth experience for multiple users simultaneously. It maintains stable connections even under network fluctuations and features failover and automatic scaling capabilities.

[0024] (3) Interaction and Feedback Mechanism: This invention has developed an intuitive user interface that supports real-time video analysis or offline analysis via upload. The interface design is simple and clear, suitable for quick operation on the training site. The system can automatically generate an evaluation report containing rotation type, intensity level, movement continuity score, and improvement suggestions to assist coaches in providing precise guidance and athletes in self-correction. The report content is based on algorithm analysis, providing quantitative indicators and qualitative suggestions, such as angular velocity analysis and posture correction tips for rotational movements, and incorporates sports science principles to enhance professionalism. The report can be exported as PDF or video format for easy archiving and sharing, supporting team collaboration and remote guidance. At the same time, it integrates a voice broadcast function to provide immediate verbal feedback during training, such as "insufficient rotation angle" or "good movement rhythm," enhancing the interactive experience and supporting multilingual options to accommodate international users. This invention also adds a personalized settings module, allowing users to customize feedback thresholds and report templates to adapt to different training stages and individual differences, such as setting sensitivity parameters and preferred display methods. Through the historical data comparison function, the system can visualize training progress and help users track performance optimization, such as generating trend charts and progress reports. It also supports data export and integration with third-party analysis tools to facilitate long-term training planning.

[0025] The present invention also includes: The construction of professional databases and knowledge graphs will create a continuously iterative digital technology knowledge base to support in-depth research and application of table tennis serving techniques, including: (1) Multi-source data acquisition and standardization: By deploying a multi-angle high-speed camera system (frame rate up to 1000fps), an inertial measurement unit (IMU), and a three-dimensional force table, this invention acquired multimodal data of athletes of different levels (national level 2 and above) performing various serving techniques in a standard laboratory environment. This included high-definition video sequences, six-degree-of-freedom kinematic parameters, and ground reaction force dynamics data. All acquired data underwent rigorous preprocessing, such as time synchronization, noise filtering, and coordinate system unification. Subsequently, precise frame-level manual annotation was performed, covering the coordinates of 21 key points of the hand, the angles of major joints, racket trajectory, spin type (such as topspin, backspin, and sidespin), and the true value of intensity (quantified in terms of rotational speed and arc). Based on this, this invention constructed a "Table Tennis Serving Multimodal Database (TTSpinDataset)" with more than 500 valid samples. Each sample is accompanied by metadata (such as athlete ID, technique type, and environmental conditions). The annotation process is cross-checked by a team of experienced professional coaches, and a consistency evaluation protocol is used to ensure data reliability. All data is stored in structured formats (such as JSON and HDF5) to facilitate subsequent machine learning analysis. To further enhance the representativeness of the database, this invention expands the data collection scope to include athlete samples from different age groups (teenagers to adults), genders, and training backgrounds (professional and amateur). Simultaneously, a semi-automated annotation tool (based on computer vision algorithms) is introduced to assist the manual process, significantly improving annotation efficiency and reducing subjective bias, laying a solid foundation for subsequent model training.

[0026] (2) Construction of Hand Shape-Rotation Association Knowledge Graph: Through in-depth mining and analysis of the database, the present invention systematically extracts fine-grained features such as different hand shape movements (e.g., grip method, finger fine-tuning), wrist change trajectory (speed, acceleration curve), and racket angle (tilt angle, azimuth angle), and uses statistical methods and machine learning models (e.g., decision trees and cluster analysis) to reveal the quantitative correspondence between these features and the final rotation type and intensity, thus forming a preliminary "hand shape-rotation" mapping knowledge graph. This graph is stored in the form of a graph database (e.g., Neo4j), where nodes represent action units or rotation categories, and edges represent associations (e.g., "cause" "influence"), supporting complex queries (e.g., SPARQL language) and logical reasoning. For example, it can dynamically retrieve typical hand posture sequences corresponding to a specific rotation or reverse deduce the action combination to achieve the target rotation. This invention also introduces semantic relationship modeling, which associates action units with rotation effects through probabilistic graphical models (such as Bayesian networks), quantifies uncertainty and simulates dynamic generation processes, so that the knowledge graph can not only describe static mapping rules, but also predict the impact of action changes on rotation effects, providing coaches with deeper training insights and personalized feedback.

[0027] (3) Knowledge Base Application and Continuous Update Mechanism: Based on the knowledge graph, this invention developed an interactive visualization technology teaching platform that can generate standard action templates (including 3D animation and keyframe decomposition) and support comparative analysis of athletes' actual performance and ideal models. This platform construction inherits and expands upon the previous research results on the training effect and influence mechanism of table tennis competitive ability based on virtual reality technology. This research, through the synergistic application of algorithm optimization and multi-source data acquisition technology, initially verified the effectiveness of the virtual training environment in improving competitive ability and explored the intrinsic influence mechanism of technical action visualization and autonomous learning, providing a methodological foundation and technical reserves for the subsequent advancement of this invention's innovation research. On this basis, this invention further integrates the algorithm-driven 3D animation generation system with the knowledge graph to form a more mature standard action modeling and personalized feedback framework: on the one hand, it uses the action-rotation association rules in the knowledge graph to drive the accurate generation of 3D animation, realizing the transformation from abstract parameters to intuitive demonstration; on the other hand, through real-time comparison of athletes' actual data and standard templates, it constructs a closed-loop autonomous learning system of "collection-analysis-visualization-feedback", significantly improving the immersion and pertinence of technical teaching. Meanwhile, this invention designs a continuous learning mechanism based on online and incremental learning. It automatically updates the knowledge base and prediction model by integrating newly acquired data (such as from sensor streams) in real time, and utilizes reinforcement learning to optimize parameters, thereby improving its adaptability to new motion variations and environmental conditions (such as different racket materials). For example, when a new athlete joins, the system can quickly adapt to their personal style using a small number of samples, and continuously optimize prediction accuracy through a closed-loop feedback loop (based on training effect evaluation). This invention also establishes a strict version control and audit log system to record each knowledge update (such as adding new relationships or revising rules), ensuring traceability throughout the entire process. Furthermore, it plans to open up partially anonymized datasets and toolchains through an API interface to promote collaboration, verification, and sharing within the academic community, and drive the long-term development of table tennis technology research.

[0028] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, alterations, or substitutions made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A real-time training assistance system for rotating table tennis balls, characterized in that, include: A hand motion analysis model is used to capture changes in hand and upper limb posture during a table tennis serve. The rotation prediction and modeling module is used for 3D trajectory reconstruction, physical model inversion, and multimodal feature fusion. A real-time training assistance module for interactive use by athletes and coaches.

2. The real-time training assistance system for rotating table tennis according to claim 1, characterized in that, The hand gesture analysis model includes: A high-precision hand key point detection and optimization module is used to locate 21 hand key points; Temporal modeling and occlusion processing module: Hand key point recovery model based on graph structure reasoning, which infers the occluded parts using visible nodes and historical temporal data; Action feature extraction and classification module: Construct a multi-scale spatiotemporal convolutional network to perform frequency and temporal domain analysis on the motion features of the hand, wrist and forearm, and combine it with a bidirectional recurrent neural network with attention mechanism to achieve automatic recognition and classification of technical actions.

3. The real-time training assistance system for rotating table tennis according to claim 2, characterized in that, The high-precision hand key point detection and optimization module is built on an improved YOLOv8 architecture and an optimized MediaPipe model.

4. The real-time training assistance system for rotating table tennis according to claim 1, characterized in that, The rotation prediction and modeling module includes: High-precision 3D trajectory reconstruction: Based on the video stream synchronously acquired by multi-view high-speed cameras, the motion trajectory of the ping-pong ball in 3D space is reconstructed through the principle of binocular parallax and motion estimation algorithm; Rotation inference based on physical models: By combining aerodynamic models and analyzing the curvature changes of the trajectory in the horizontal and vertical directions, the initial rotation vector of the ping-pong ball can be deduced. Multimodal fusion rotation recognition model: By dynamically weighting different features through a cross-modal attention mechanism, it achieves high-precision classification of rotations such as no rotation, upspin, downspin, left-hand rotation, right-hand rotation and their combinations, and performs graded evaluation of rotation intensity.

5. The real-time training assistance system for rotating table tennis according to claim 4, characterized in that, In the high-precision 3D trajectory reconstruction, a depth sensor-assisted calibration is introduced, the reconstruction accuracy is improved through point cloud registration technology, and an adaptive trajectory smoothing filter is used to eliminate noise and maintain the physical rationality of the trajectory.

6. The real-time training assistance system for rotating table tennis according to claim 1, characterized in that, The real-time training assistance module includes: Model optimization and edge computing deployment: The optimized model is deployed on edge computing devices; Low-latency system architecture design: including a front-end and back-end separated real-time analysis system to support high concurrency and low latency requirements; Interaction and feedback mechanism: Based on the user interface, it enables real-time video analysis or offline analysis by uploading.

7. The real-time training assistance system for rotating table tennis according to claim 6, characterized in that, The model optimization and edge computing deployment also include: adopting dynamic memory allocation and garbage collection mechanisms to avoid lag or crashes on resource-constrained devices.

8. The real-time training assistance system for rotating table tennis according to claim 6, characterized in that, The low-latency system architecture design also includes: the backend uses the Python Flask framework to provide a RESTful API, and integrates OpenCV and TensorFlow / PyTorch for video stream processing and model inference; the frontend is developed based on React and Three.js to realize the visualization of motion capture, rotation parameters, and trajectory prediction results.

9. A real-time training assistance system for rotating table tennis according to claim 6, characterized in that, The low-latency system architecture design also includes: implementing real-time data push based on the WebSocket protocol to reduce communication overhead, ensure data synchronization and update, and avoid latency caused by polling.

10. A real-time training assistance system for rotating table tennis according to claim 6, characterized in that, The interaction and feedback mechanism also includes generating an assessment report that includes rotation type, intensity level, movement continuity score, and improvement suggestions to assist in guidance and athlete self-correction.

Images