A smart pick-up racket with an AI vision deviation correction function and a deviation correction method and system thereof

Abstract

This invention discloses a smart Peak racket with AI visual correction function, along with its correction method and system. It relates to the fields of intelligent sports equipment and AI sports training, aiming to solve the core problems of existing smart Peak rackets where high-frequency vibrations during ball impact cause visual acquisition failure and prevent real-time closed-loop correction of the swing kinetic chain. The racket of this invention features a Φ22mm embedded wide-angle camera compartment at the 12 o'clock position on the frame. A cross-axis suspension damping architecture is formed through orthogonally distributed high-damping silicone shock-absorbing pads, achieving a non-rigid suspension connection between the camera compartment and the frame. Simultaneously, a real-time motion coach agent based on deep learning is constructed, forming a complete closed loop of image acquisition, coordinate transformation, motion comparison, and real-time correction suggestions. The swing kinetic chain is reconstructed by fusing visual and IMU data through Kalman filtering, and instantaneous voice correction is achieved based on a preset threshold. This invention fundamentally solves the problem of out-of-focus images, achieves high-precision motion correction, and builds an integrated hardware and software intelligent training system.

Description

Technical Field

[0001] This invention relates to the field of intelligent sports equipment technology, specifically to an intelligent Peak racket with AI visual correction function, as well as its correction method and correction system. Background Technology

[0002] Peakball, a ball sport that combines features of tennis, badminton, and table tennis, has rapidly gained popularity in China in recent years, with a continuously expanding participation base. Users are increasingly demanding improvements in the standardization of their hitting techniques and training efficiency. Against this backdrop, Peak rackets are evolving from traditional purely physical structures towards intelligent designs with instructional feedback functions, becoming a core trend in industry research and development.

[0003] Currently, the development of smart Peak rackets and related sports training programs suffers from three major technological shortcomings, and no existing solution can simultaneously overcome the following problems: First, existing hardware solutions for smart rackets cannot solve the problem of visual data acquisition failure caused by the high-frequency vibrations of Peak ball impact. Current smart rackets that attempt to add cameras to the racket head all use a rigid connection between the camera and the frame. However, the frame vibration frequency at the moment of Peak ball impact is generally above 100Hz. The rigid connection structure cannot decouple the bidirectional mechanical vibration waves from the impact direction and the swing direction. Furthermore, there is no miniaturized camera-bearing structure specifically adapted to the Peak racket's mounting space. This results in severe blurring and motion blur in the camera-captured images. Subsequent AI algorithms cannot extract effective data on racket face deformation and swing posture, ultimately leading to low recognition accuracy and poor correction precision. These solutions remain only in the laboratory testing phase and cannot achieve stable commercial deployment.

[0004] Secondly, existing methods for correcting movement errors cannot achieve real-time closed-loop correction of the entire kinetic chain in the context of pickleball. Existing intelligent training solutions can only provide basic data statistics on swing speed and impact point, failing to capture the complete power chain from the lower limbs to the upper limbs and thus unable to provide professional diagnosis of the swing's power application and consistency. Furthermore, they lack a complete execution loop adapted to the pickleball scenario, and do not set precise error correction threshold trigger rules for the pickleball swing. This results in significant lag in feedback, allowing only post-event video review and failing to provide real-time error correction feedback at the moment of impact. Consequently, user training efficiency is low, and skill improvement is minimal.

[0005] Third, existing technologies have not yet formed an integrated intelligent training system for peakball, encompassing hardware, terminal software, and cloud services. Existing fragmented intelligent racket solutions suffer from extremely poor hardware and software compatibility. They lack a complete implementation system for equipment calibration, hardware self-testing, training process visualization, kinetic chain data review, and algorithm OTA upgrades. This makes it impossible to meet the training needs of users at all levels—beginner, intermediate, and professional—resulting in a very high user threshold and hindering large-scale commercial promotion.

[0006] In summary, currently available technologies lack mature hardware solutions that can fundamentally solve the problem of visual acquisition failure caused by the vibration of the Peak racket during ball impact, as well as a complete method to achieve real-time closed-loop correction of the Peak racket's swing kinetic chain. Furthermore, there is no integrated hardware and software-compatible intelligent training system for Peak rackets, which severely restricts the technological development and commercial application of intelligent Peak rackets. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a novel Peak racket shock absorption structure that fundamentally solves the problem of image defocusing caused by high-frequency vibrations during ball impact, providing a stable and clear image acquisition foundation for AI visual recognition. Construct a real-time motion coaching agent based on deep learning to achieve a complete closed loop from image acquisition, coordinate transformation, motion comparison to real-time correction suggestions, complete the full-dimensional capture and reconstruction of the human force chain, and achieve real-time correction of ball hitting motion with centimeter-level precision; It provides an integrated intelligent training system encompassing hardware, terminal software, and cloud services, catering to the training needs of users at all levels—beginner, intermediate, and professional. It upgrades traditional post-exercise review to real-time teaching feedback at the moment of impact, significantly lowering the learning curve for peakball techniques.

[0008] To achieve the above objectives, the technical solution adopted by this invention is as follows: a smart Peak racket with AI visual correction function, comprising a racket frame, a fiber-reinforced racket face fixed to the inner side of the racket frame, and a hollow handle fixedly connected to the bottom of the racket frame. The handle houses a main control circuit board, a power supply module, and a Bluetooth communication module. A Φ22mm through-type stepped mounting hole is provided at the geometric center position of the top 12 o'clock position of the racket frame. A 22mm embedded wide-angle camera compartment is provided within the stepped mounting hole, and a 12-point visual acquisition module is fixedly installed inside the camera compartment. A high-damping silicone shock-absorbing pad is respectively provided at four orthogonal axial positions (0°, 90°, 180°, and 270°) on the outer wall of the camera compartment. The camera compartment is suspended and non-rigidly connected to the inner wall of the stepped mounting hole through the four high-damping silicone shock-absorbing pads, forming a cross-axis suspended shock-absorbing structure. A 9-axis IMU sensor is also provided inside the handle. The 12-point visual acquisition module and the 9-axis IMU sensor are both electrically connected to the main control circuit board.

[0009] Furthermore, the stepped mounting hole is a coaxial two-section structure. The inner diameter of the first hole section facing the hitting side of the racket face is 22mm and the hole depth is 8mm. The inner diameter of the second hole section facing the outside of the racket frame is 18mm and the hole depth is 4mm. The connection between the two hole sections forms an annular stepped limiting surface. The outer diameter of the camera compartment is 21.8mm, the inner diameter is 16mm, and the total depth is 12mm. The end of the compartment facing the outside of the racket frame is provided with a limiting flange that cooperates with the annular stepped limiting surface.

[0010] Furthermore, the high-damping silicone shock-absorbing pad is cylindrical with a diameter of 5mm, a thickness of 3mm, and a Shore hardness of 30HA; four orthogonal axial positions on the outer wall of the camera compartment are respectively provided with mounting grooves with an inner diameter of 5mm and a depth of 1.5mm. One end of the high-damping silicone shock-absorbing pad is interference-fitted into the mounting groove, and the other end is interference-fitted with the inner wall of the stepped mounting hole, so that the camera compartment and the frame have no rigid contact.

[0011] Furthermore, the 12-point visual acquisition module is a 4K global shutter wide-angle module with a sampling frame rate of 120fps and a lens field of view of 120°; the sampling frequency of the 9-axis IMU sensor is 1kHz; the Bluetooth communication module is a BLE5.0 low-power Bluetooth module; the main control circuit board has a built-in Flash storage unit for storing deep learning models and standard sweet spot hitting models.

[0012] An AI-based visual kinetic chain correction method based on a smart Peak racket specifically includes the following steps: S1 Image Acquisition: The racket's 12-point visual acquisition module and 9-axis IMU sensor complete dual-modal synchronous data acquisition to obtain high-speed visual flow and inertial navigation data before and after the shot. S2 Coordinate Transformation: After preprocessing the acquired visual images, the pixel coordinates of 12 dynamic points of the human body are extracted in real time and converted into three-dimensional world coordinates with the center of the racket sweet spot as the origin and aligned with the spatial posture of the racket. S3 Action Comparison: The time synchronization of visual data and inertial data is achieved through linear interpolation. Then, the dual-source data is fused by the Kalman filtering algorithm to eliminate vibration noise, reconstruct the swing kinetic chain model, and fit and compare the measured kinetic chain model with the preset standard sweet spot hitting model to calculate the swing deviation value. S4 Real-time Correction Suggestion: The deviation value is judged by a preset correction feedback threshold algorithm. When the deviation value exceeds the trigger threshold, the corresponding real-time correction instruction is generated and broadcast through the voice terminal.

[0013] Furthermore, in step S1, the preset ball-hitting trigger threshold is the instantaneous acceleration peak detected by the IMU ≥ 15g. When the trigger threshold is reached, the 12-point visual acquisition module is activated to acquire 5 frames before and after the ball-hitting trigger moment, for a total of 10 key frames, with a sampling frame rate of 120fps. At the same time, IMU inertial data within 200ms before and after the ball-hitting moment is acquired simultaneously, and the visual images and inertial data are marked with the same high-precision timestamp, with a timestamp accuracy ≤ 100μs.

[0014] Furthermore, in step S2, the 12 extracted human body dynamic points are specifically the left shoulder, left elbow, left wrist, left hip, left knee, left ankle, right shoulder, right elbow, right wrist, right hip, right knee, and right ankle, and the confidence threshold for dynamic point extraction is set to 0.85; in step S3, the reconstructed swing dynamic chain model covers the entire trajectory from backswing, swing, hitting the ball to follow-through.

[0015] Furthermore, in step S4, the triggering rules of the correction feedback threshold algorithm are as follows: voice correction is triggered when the angle deviation between the swing plane and the standard plane is >5°, voice correction is triggered when the racket face torsion angle deviation is >3°, voice correction is triggered when the lateral / longitudinal deviation of the hitting point relative to the center of the sweet spot is >5mm, and voice correction is triggered when the similarity of the swing trajectory of the 12 power points is <85%; the entire process from the triggering of the hit to the user receiving the voice broadcast takes ≤45ms.

[0016] Furthermore, it also includes the execution steps of the supporting mobile terminal APP: smart racket Bluetooth pairing, 12-point camera self-test and calibration, real-time training interface visualization, kinetic chain analysis report generation, and OTA system firmware upgrade; wherein, the kinetic chain analysis report includes total number of training hits, sweet spot rate, average swing speed, stability score, and radar chart visualization data for four dimensions: speed, accuracy, stability, and power.

[0017] An AI-powered visual peak ball intelligent training system includes an intelligent peak ball racket, a mobile terminal APP module, and a cloud data analysis module. The intelligent peak ball racket communicates with the mobile terminal APP module via its own Bluetooth communication module, and the mobile terminal APP module is network-connected to the cloud data analysis module. The mobile terminal APP module provides access to device management, training visualization, and data viewing, while the cloud data analysis module is used for training data statistical analysis, training report generation, and algorithm model iterative optimization.

[0018] The beneficial effects of this invention are as follows: 1. The intelligent Peak racket hardware solution of this invention, through the combination of a 22mm embedded wide-angle camera compartment and a cross-axis floating shock absorption architecture, achieves a non-rigid floating connection between the camera compartment and the racket frame. This effectively decouples the bidirectional mechanical vibration waves from the hitting direction and the swing direction, physically filtering more than 85% of the high-frequency vibrations from the hitting. It ensures that in high-speed hitting scenarios above 60km / h, the image sampling deviation of the visual acquisition module is stably controlled within 0.5mm, completely solving the core problems of image blurring, ghosting, and AI recognition failure caused by vibration in existing rigid connection solutions. At the same time, the miniaturized 22mm camera compartment design perfectly fits the installation space of the Peak racket, without changing the standard size and balance of the racket, and without affecting the normal hitting feel. This breaks through the technical bottleneck of existing solutions that can only remain in the laboratory stage and cannot be commercially implemented.

[0019] 2. The AI ​​visual kinetic chain correction method corresponding to this invention constructs a complete execution closed loop of "image acquisition - coordinate transformation - action comparison - real-time correction suggestion". Through millisecond-level dynamic trajectory capture of 12 power points of the human body, it can complete the reconstruction of the complete force generation kinetic chain from the lower limbs, core to upper limbs and racket, realizing professional diagnosis of the standardization and consistency of the swing action. It breaks through the limitation of existing solutions that can only count basic data such as swing speed and hitting point. At the same time, through a targeted correction feedback threshold algorithm, it achieves an instantaneous response of ≤45ms from the trigger of the hit to the user receiving voice feedback. The correction accuracy reaches the centimeter level, which completely solves the problems of feedback lag, only being able to review after the fact and low training efficiency of existing solutions. It truly realizes the function of "real-time AI coach" and significantly reduces the learning threshold of pickleball technology.

[0020] 3. The AI ​​visual peak ball intelligent training system corresponding to this invention establishes a complete three-level architecture of "intelligent peak ball racket hardware - mobile terminal APP module - cloud data analysis module," achieving deep hardware and software adaptation. It covers the entire process from device Bluetooth pairing, camera self-test calibration, real-time training visualization, power chain analysis report generation to OTA algorithm firmware upgrades. It eliminates the need for complex professional debugging by users and can adapt to the training needs of users at all levels, from beginners to advanced professionals. It completely solves the problems of poor hardware and software compatibility, high usage threshold, and inability to be commercially scaled in existing solutions. Simultaneously, the cloud platform can continuously iterate and optimize the algorithm model based on anonymized training data, significantly extending the product's lifecycle and enhancing its commercial value and market competitiveness. Attached Figure Description

[0021] Figure 1 This is a system architecture block diagram of the real-time action coaching agent based on deep learning described in this invention. Figure 2 This is a flowchart illustrating the logical flow of the AI ​​error correction "identification-analysis-feedback" closed loop described in this invention. Figure 3 This is a timing diagram illustrating the entire process of the system described in this invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings. This embodiment is implemented based on the technical solution of the present invention, providing detailed implementation methods, specific operating procedures, and complete technical parameters. Those skilled in the art can completely reproduce the technical solution of the present invention based on the following content. The scope of protection of the present invention is not limited to the following embodiments.

[0023] This invention discloses an AI-powered visual Peak racket with a cross-axis suspension damping structure and its kinetic chain correction system. The core of this solution consists of two main components that can be implemented independently yet work together: first, an intelligent Peak racket hardware structure with a 22mm embedded wide-angle camera housing and a cross-axis suspension damping architecture; and second, a real-time motion coaching agent based on deep learning. The following embodiments will provide a complete and practical detailed description of the hardware structure, system methods, supporting terminal implementation, and the entire workflow.

[0024] Example 1: AI Vision Peak Racket with Cross-Axis Suspension Shock Absorption Structure

[0025] The smart Peak racket provided in this embodiment is dedicated hardware for smart Peak rackets, including a one-piece molded frame, a fiber-reinforced racket face, and a hollow handle. The racket face is fixed to the inside of the frame using a hot-pressing process to form the hitting surface. The handle is fixed to the bottom of the frame at the 6 o'clock position using threads and epoxy adhesive. The overall design conforms to the USAPA Peak racket official size standard, with a total frame width of 220mm, a total racket face length of 400mm, and a handle length of 140mm. Each racket's main control MCU is pre-programmed with a globally unique device identification code (SN code) for authorization binding and activation with the accompanying terminal software, ensuring genuine authorization of the device and system.

[0026] 1.1 22mm Embedded Wide-Angle Camera Compartment and Cross-Axis Suspension Vibration Damping Architecture (Core Technical Features)

[0027] At the geometric center of the top of the racket frame at the 12 o'clock position, a Φ22mm through-type stepped mounting hole is made. The stepped mounting hole has a coaxial two-section structure: the inner diameter of the first section facing the hitting side of the racket face is 22mm and the hole depth is 8mm; the inner diameter of the second section facing the outside of the racket frame is 18mm and the hole depth is 4mm. The connection between the two sections forms an annular stepped limiting surface facing the outside of the racket frame, which is used to limit the axial displacement of the camera chamber and prevent the chamber from falling out due to the vibration of hitting the ball.

[0028] The stepped mounting holes contain a 22mm embedded wide-angle camera compartment, which serves as a dedicated support structure for the racket's 12-point visual module. The specific parameters of the camera compartment are as follows: Chamber specifications: outer diameter 21.8mm (with a 0.2mm gap reserved with the inner wall of the first hole section), inner diameter 16mm, total depth of the chamber 12mm; Material selection: Made of 6061-T6 lightweight aluminum alloy, precision-carved by CNC, with a chamber wall thickness of 2.9mm and a single chamber weight of ≤3.2g, to avoid affecting the balance and weight distribution of the racket; Structural Design: The end of the chamber facing the impact surface is the lens acquisition port, and the other end is a closed end with a limiting flange. The limiting flange has an outer diameter of 21.5mm and mates with the limiting surface of the stepped mounting hole to limit the axial displacement of the chamber towards the impact surface. Along the circumferential direction of the outer wall of the chamber, at four orthogonal axial positions of 0°, 90°, 180°, and 270°, there is a cylindrical mounting groove with an inner diameter of 5mm and a depth of 1.5mm. The central axes of the four grooves are orthogonal to each other, forming a cross-shaped shock-absorbing mounting position. Optical configuration: A 4K global shutter wide-angle visual acquisition module (12-point visual module) is fixedly installed inside the chamber. The lens focal length is 3.6mm, the field of view is 120°, and the lens optical axis is parallel to the normal of the shooting face. It can completely cover the entire area of ​​the shooting face and the full range of motion of 12 dynamic points of the human body on the side of the shooting hand, with no blind spots.

[0029] The camera compartment and the racket frame are connected by a cross-axis suspension shock absorption structure, which is a non-rigid connection and the core innovative solution of this invention to solve the problem of blurry images caused by ball impact. The specific implementation of the shock absorption structure is as follows: Shock absorption unit: It uses 4 identical high-damping silicone shock absorption pads. The shock absorption pads are cylindrical with a diameter of 5mm and a thickness of 3mm. The Shore hardness is 30HA, the elongation at break is ≥600%, the resilience is ≥55%, and it can withstand the working temperature of -20℃ to 80℃, making it suitable for outdoor training scenarios. Installation method: The 1.5mm thick section of each shock-absorbing pad is interference-fitted into the mounting groove on the outer wall of the camera chamber, and the remaining 1.5mm thick section protrudes from the outer wall of the chamber. The protruding ends of the four shock-absorbing pads are interference-fitted with the inner wall of the first hole of the stepped mounting hole. After installation, the outer wall and the closed end of the camera chamber do not have any rigid contact with the frame. They are completely suspended and fixed in the mounting hole by the four orthogonally distributed silicone shock-absorbing pads, forming a dual-axis decoupled shock absorption structure with X-axis (0° / 180° corresponding to the hitting direction) and Y-axis (90° / 270° corresponding to the swing direction). Technical effect: This architecture can effectively absorb and decouple the mechanical waves along the X and Y axes at the moment of impact, physically filter more than 85% of high-frequency vibrations, and ensure that the sampling deviation of the 4K vision module is controlled within 0.5mm when hitting the ball at high speed (above 60km / h), completely solving the pain point of AI recognition of out-of-focus.

[0030] 1.2 Onboard Hardware and Circuit Layout

[0031] The handle has a hollow ABS engineering plastic shell, with a physical power switch at the bottom. It integrates complete main control and sensing hardware, as shown in the following layout: Main control circuit board: The ESP32-S3 dual-core microcontroller is used as the main control core and is soldered to the fixed bracket in the middle of the handle. The onboard Flash storage capacity is 16MB, which is used to store deep learning models, standard hitting models, local hitting data, system programs and device unique SN code. 9-axis IMU sensor: It adopts BMI088 high dynamic inertial measurement unit, which is soldered on the main control circuit board and coaxially aligned with the central axis of the racket. The sensor has a sampling frequency of 1kHz and can simultaneously collect three-axis acceleration, three-axis angular velocity and three-axis geomagnetic data during the swing process to calculate the real-time spatial attitude P(x,y,z) of the racket. Power supply module: It adopts a 3.7V 500mAh high-density polymer lithium battery, which is fixed at the end of the handle and is charged through a Type-C interface. When fully charged, it can support the system to work continuously for ≥8 hours and standby time for ≥30 days. Communication and output module: Onboard BLE5.0 low-power Bluetooth module with a maximum transmission distance of 10m and a one-way transmission latency of ≤20ms, compatible with mobile terminal devices with Bluetooth 5.0 and above; The end of the handle has a built-in 8Ω 1W miniature voice speaker, which also supports outputting voice correction commands via Bluetooth connection to Bluetooth headsets and smart sports watches.

[0032] 1.3 Structural Performance Verification of this Embodiment

[0033] Through vibration table simulation and actual hitting tests, the cross-axis suspension damping architecture of this embodiment can physically filter 86.2% of the vibration amplitude under the high-frequency vibration conditions of peak ball hitting at 100Hz~500Hz. In the high-speed hitting scenario of 65km / h, the spatial deviation of the image sampling of the 12-point vision module is stably controlled within 0.32mm, without any image defocus or ghosting issues. It fully meets the requirements of AI algorithms for high-definition and stable image acquisition, providing a hardware foundation for centimeter-level accuracy in ball hitting correction.

[0034] Example 2: Real-time motion coaching agent based on deep learning

[0035] This embodiment is based on the smart Peak racket hardware of Embodiment 1, and constructs a real-time motion coaching agent based on deep learning. For the overall system architecture, please refer to [link / reference needed]. Figure 1 The core process includes: image acquisition -> coordinate transformation -> motion comparison -> real-time correction suggestions. The system is divided into three layers: the perception layer and processing layer at the racket end, and the application layer between the terminal and the cloud. The specific implementation methods of each layer are as follows: 2.1 Perception Layer: Dual-modal synchronous data acquisition The perception layer is deployed locally on the smart Peak racket. Its core function is to achieve time-synchronous acquisition of visual signals from the 12-point vision module and inertial signals from the IMU, completing the front-end input of multi-source heterogeneous data. The specific implementation steps are as follows: 1. Trigger Mechanism Settings: The system presets the ball-hitting trigger threshold to be the instantaneous acceleration peak detected by the IMU ≥ 15g. When the acceleration peak is below this threshold, the system is in a low-power standby state, with only the IMU continuously collecting data at a frequency of 1kHz, and the 12-point vision module is in a sleep state. When the acceleration peak reaches the trigger threshold, the vision module is immediately woken up and the ball-hitting event is locked.

[0036] 2. Image Acquisition: After the 12-point vision module is awakened, it immediately acquires 5 frames before and after the ball-hitting trigger moment, for a total of 10 key frames, with a sampling frame rate of 120fps. Each frame is stamped with a high-precision timestamp with a timestamp accuracy of ≤100μs. After acquisition, the 10 key frames are transmitted to the processing layer, and the remaining non-key frames are discarded to reduce computing power consumption.

[0037] 3. Inertial signal acquisition: After the ball-hitting event is triggered, the IMU synchronously acquires all inertial navigation data within 200ms before and after the ball-hitting event, including three-axis acceleration a(x,y,z), three-axis angular velocity ω(x,y,z), and three-axis geomagnetic data m(x,y,z). Each set of data is stamped with a timestamp of the same origin as the visual frame and synchronously transmitted to the processing layer.

[0038] 2.2 Processing Layer: Multi-sensor Fusion and Powertrain Restructuring

[0039] The processing layer is deployed in the main control MCU of the smart Peak racket. Its core function is to denoise and fuse the dual-modal data to reconstruct a complete impact kinetic chain model. The specific implementation steps are as follows: Step 1: Video stream preprocessing and face deformation extraction 1. Lens distortion correction: Call the pre-stored camera intrinsic parameter matrix and distortion coefficients to perform distortion correction on the captured key frame images, correct the radial and tangential distortion of the lens, and restore the true geometric dimensions of the shooting surface; 2. IMU-linked image stabilization correction: Based on the synchronously acquired IMU data, the vibration offsets offset_x and offset_y of the X and Y axes at the moment of impact are calculated, a 2×3 affine transformation matrix is ​​constructed, and the corrected image is translated and rotated to compensate for the image jitter caused by high-frequency vibration, so as to obtain a clear image after stabilization. 3. Racket face deformation and hitting point extraction: The stabilized image is preprocessed with grayscale and Gaussian blur. The Canny operator is used to extract the racket face contour and the contact area with the pickle. The subpixel edge detection algorithm is used to calculate the microscopic deformation displacement Δd of the racket face at the moment of impact, including the lateral and longitudinal offset of the hitting point relative to the center of the sweet spot of the racket face, as well as the torsion angle of the racket face. The center of the sweet spot of the racket face is pre-calibrated as the geometric center of the racket face and stored in the system in advance.

[0040] Step 2: Multi-sensor data time synchronization

[0041] For IMU inertial data with a sampling frequency of 1kHz, a linear interpolation algorithm is used for resampling. The inertial data is interpolated to the visual frame time node of 120fps to ensure that each set of visual frame data corresponds to a set of synchronized inertial data. The time synchronization error is controlled within 1ms, eliminating the time difference of dual-modal data and completing the preprocessing before the fusion of multi-source heterogeneous data.

[0042] Step 3: Real-time extraction of coordinates of 12 dynamic points of the human body and millisecond-level dynamic trajectory capture

[0043] 1. Deep learning model deployment: A lightweight human pose estimation model (such as the OpenPose variant based on MobileNetV2) is pre-trained and deployed in the Flash of the main control MCU. The model is quantized in INT8 format, with an inference speed of ≥30fps, and can achieve millisecond-level dynamic trajectory capture. 2. Real-time extraction of coordinates of 12 dynamic points on the human body: The model performs inference on the keyframe images after image stabilization and extracts the coordinates of 12 key dynamic points on the side of the holding hand in real time, including: left shoulder, left elbow, left wrist, left hip, left knee, left ankle, right shoulder, right elbow, right wrist, right hip, right knee, and right ankle; each dynamic point outputs pixel-level coordinates (u,v) and confidence level, with the confidence level threshold set to 0.85, and only dynamic point data with confidence levels higher than the threshold are retained; 3. Coordinate Transformation: The extracted pixel-level coordinates (u,v) of the 12 dynamic points are converted into three-dimensional world coordinates (X,Y,Z) with the center of the sweet spot of the racket as the origin through the pre-stored camera extrinsic parameter matrix (the position and attitude of the camera relative to the racket), so as to achieve the alignment of the dynamic point coordinates with the racket's spatial attitude in a unified coordinate system.

[0044] Step 4: Kalman filter fusion and powertrain reconstruction

[0045] 1. Establish the system state equations: Define the system state vector as the three-dimensional position, velocity, and acceleration of 12 dynamic points, and the three-dimensional spatial position, velocity, acceleration, and angular velocity of the racket; 2. Establish dual-source observation equations: Use the 3D coordinates of the 12 dynamic points extracted by vision as the first observation vector $Z_1$, and use the acceleration and angular velocity data collected by the IMU as the second observation vector $Z_2$ to construct the corresponding observation matrix; 3. Perform Kalman filter iteration: Prediction steps: Based on the posterior state vector and system state equation of the previous time step, predict the prior state vector and prior covariance matrix of the current time step. Update steps: Combine dual-source observation data, calculate the Kalman gain, correct the prior state vector, obtain the posterior state vector and posterior covariance matrix at the current time, and eliminate random noise caused by high-frequency vibrations. 4. Kinetic Chain Model Reconstruction: Based on the filtered smooth state data, the Kalman filtering algorithm is used to reconstruct the entire swing trajectory from backswing, swing, hitting the ball to follow-through, completing the construction of a complete hitting kinetic chain model and restoring the power transmission trajectory of 12 power points and the changes in racket spatial posture during the power generation process.

[0046] 2.3 Application Layer: Action Comparison, Real-time Correction, and Implementation via Supporting Terminals

[0047] The application layer is divided into a local real-time racket correction module, a supporting mobile terminal APP module, and a cloud data analysis module, which fully realizes the entire process from real-time training to data review. The specific implementation method is as follows: 2.3.1 Local Real-Time Correction Module This module is deployed on the racket's main control MCU and its core function is to implement second-level voice correction. It performs intelligent diagnosis of the point of force application, racket face angle, and swing speed. The specific implementation is as follows: 1. Pre-storage of Standard Sweet Spot Shot Model: 1200 sets of labeled standard sweet spot shots from ATP professional peakball players are pre-collected, including the full trajectory of 12 dynamic points, the position of the hitting point, racket face deformation parameters, and the posture data at the moment of hitting. Through Gaussian mixture model clustering optimization, a general standard sweet spot shot model is obtained and stored in the Flash memory of the main control board. At the same time, the system supports downloading customized standard models at three levels—beginner, intermediate, and professional—through the accompanying APP to adapt to users of different skill levels.

[0048] 2. Motion Comparison: The measured kinetic chain model and the deformation displacement Δd of the hitting point output by the processing layer are compared with the preset standard sweet spot hitting model: The similarity deviation of the swing trajectory of 12 kinetic points is calculated using the Dynamic Time Warping (DTW) algorithm, the deviation value ε of the hitting point relative to the center of the sweet spot is calculated using Euclidean distance, and the angular deviation θ between the swing plane and the standard plane, the racket face torsion angle deviation Φ, and the instantaneous swing speed are calculated at the same time. 3. Correction Feedback Threshold Algorithm: The system presets a specific correction trigger threshold, as follows: A swing plane angle deviation θ > 5° triggers speech correction. If the racket face torsion angle deviation Φ>3°, voice correction is triggered. If the lateral / vertical deviation of the impact point is greater than 5mm, voice correction will be triggered. If the similarity of the swing trajectory of the 12 dynamic points is less than 85%, voice correction is triggered. 4. Real-time Correction Suggestions: Based on the calculated deviation type, the system matches corresponding personalized correction commands and transmits them in real-time to Bluetooth headsets, the built-in speaker on the handle, or the accompanying terminal APP via the BLE5.0 Bluetooth module. The voice output latency is ≤45ms, achieving user-perceptible second-level voice correction and providing instant feedback after the shot. When the swing plane angle deviation is θ=7°, the output is: "Correction reminder: The swing plane is tilted by 7 degrees. Please press down the racket head and use the core rotation to generate power to keep the swing plane parallel to the ground." When the hitting point is 8mm to the right, the output is: "Correction reminder: The hitting point is 8mm to the right. Please adjust your stance and aim at the center of the sweet spot." When the hitting point is too low, the output is: "Hitting point is too low, please strengthen wrist lift"; When all deviations are within the threshold, the output is: "Perfect action, sweet spot shot, maintain shot depth".

[0049] For the complete closed-loop logic of the AI ​​correction mechanism in this invention, including "identification-analysis-feedback", please refer to [link / reference]. Figure 2 It fully reproduced the execution flow of the system's core code and clarified the judgment rules for the correction trigger threshold.

[0050] 2.3.2 Supporting Mobile Terminal APP Implementation Module

[0051] This module is a companion mobile application that supports smartphones, tablets, smartwatches, and other terminal devices, providing users with a visual interface for operation and data viewing. The specific implementation is as follows: 1. Installation and Environment Configuration Software distribution: Users can obtain the APP installation package and complete the installation by scanning the official authorization QR code on the back of the product warranty card; Permission configuration: During the installation and first launch of the APP, the app needs to request and obtain "Bluetooth connection", "camera access" and "location information" permissions from the user to ensure the normal execution of device pairing, data transmission and function upgrades; Account and Device Activation: After completing account registration, users need to enter the unique device identification code (SN code) built into the racket upon first login to complete the authorization binding of the device and account and activate all system functions.

[0052] 2. Smart racket Bluetooth pairing function

[0053] Operation path: APP homepage → Settings → Racket connection; Implementation logic: After the user turns on the physical power switch at the bottom of the racket handle, the APP scans for nearby devices via Bluetooth. Once the smart racket with the device name "ProsperousFuture_P01" is found, the user can click to complete the pairing and connection. After successful pairing, the device status and system version information are automatically synchronized.

[0054] 3. 12-point camera self-test and calibration function

[0055] Operation path: Device connection successful → System self-test; Implementation logic: After the user triggers the self-test, the system automatically completes the zero-point calibration of the 12-point vision module, the initial state detection of the cross-axis suspension damping module, and the gyroscope calibration of the IMU sensor. If the self-test passes, the APP interface displays a simulated real-time image of the racket's 12-point position, with a green crosshair in the center, indicating "Self-test passed". If the self-test fails, a fault prompt will automatically pop up, guiding the user to check whether there are foreign objects blocking the silicone pad in the camera compartment or whether there are stains obstructing the lens glass.

[0056] 4. Real-time training and correction mode

[0057] Operation path: APP homepage → Start training; Display and Interaction Logic: The APP's real-time training interface is divided into three core areas. The upper left corner displays the current swing speed (unit: km / h) in real time. The central area displays a simulated 3D model of the racket and a schematic diagram of the racket face. After hitting the ball, the actual hitting position is marked on the schematic diagram of the racket face with a red dot. A "Stop Training" control button is set at the bottom. During the training process, the APP synchronously receives the correction voice commands from the racket end and synchronously completes the voice broadcast and interface pop-up prompts.

[0058] 5. Function for generating and viewing power chain analysis reports

[0059] Operation path: Training completed → View report; Key metrics in the report include: total number of hits, sweet spot percentage (percentage of hits to the optimal power zone), average swing speed, and stability score (image stabilization score calculated based on cross-axis damping frequency). The scores for the four core dimensions of "speed," "accuracy," "stability," and "power" are visualized using a radar chart. The report also allows viewing historical training reports to compare technical improvement trends.

[0060] 6. OTA system upgrade function

[0061] Users can check the latest correction algorithm package and system firmware on the APP's Settings → Version Update page, and complete the system update of the racket and APP via OTA wireless upgrade, without the need for additional hardware operations.

[0062] 2.3.3 Cloud Data Analysis Module

[0063] The raw data and analysis results of each user's shot will be synchronously uploaded to the cloud training platform via the accompanying app. The specific functions of the cloud platform are as follows: 1. Data Statistical Analysis: Perform multi-dimensional statistics on users' historical training data, including core indicators such as total number of hits, sweet spot hit rate, swing stability, power generation continuity, and power transmission efficiency of 12 power points; 2. Training Report Generation: Standardized AI teaching and training reports are generated weekly / monthly, including the user's technical improvement, analysis of technical shortcomings, and targeted training suggestions; 3. Model Iteration and Optimization: Based on massive user training data, the cloud platform continuously iterates and optimizes the deep learning pose estimation model and the standard hitting model. It also supports users to upload their own hitting videos and customize their own personal standard motion model.

[0064] 2.4 System Anomaly Self-Check and Fault Tolerance Mechanism

[0065] This system has a built-in end-to-end fault detection and fault-tolerance mechanism, and corresponding solutions are set up for frequently used faults. The specific implementation is as follows: 1. AI-powered defocusing anomaly handling: The system monitors the sharpness score of the visual image in real time. When the image sharpness is lower than the preset threshold, it automatically triggers a defocusing anomaly prompt. The system guides the user through the APP to check whether the cross suspension damping module is loose and whether there are stains on the 12 o'clock lens glass. It also automatically performs secondary lens calibration to restore normal acquisition effect. 2. Data latency anomaly handling: The system monitors Bluetooth transmission latency in real time. When the latency exceeds a preset threshold, it automatically prompts the user to check if the phone's Bluetooth version is not lower than 5.0, and guides the user to close unnecessary background applications, free up terminal memory, and reduce transmission latency. 3. Handling of shock absorption module malfunctions: The system uses IMU to detect the attenuation frequency of the impact vibration in real time. When the vibration attenuation efficiency is lower than the preset threshold, it will automatically prompt the user to check whether the silicone shock absorption pad is aging or falling off, ensuring that the hardware shock absorption effect meets the system requirements.

[0066] Example 3: System Full-Process Working Example

[0067] This embodiment fully describes the entire workflow of the present invention from device power-on to completion of training and review. For the millisecond-level timing nodes of the entire process, please refer to [link / reference needed]. Figure 3 The details are as follows: 1. System Power-On and App Activation & Pairing: The user presses and holds the physical power switch at the bottom of the racket handle for 3 seconds to power on the device. The main control MCU sequentially completes hardware self-tests of the 4K vision module, 9-axis IMU sensor, and Bluetooth module. The user scans the product warranty card QR code to install the accompanying app, completes account registration, enters the racket's serial number to activate the device, and then enters the [Racket Connection] page in the app to search for and connect to the racket device named "ProsperousFuture_P01" to complete Bluetooth pairing.

[0068] 2. System Self-Check and Calibration: After successful pairing, the user triggers the [System Self-Check] in the APP. The system automatically completes the initial zero-point calibration of the cross-shaped shock absorber module, IMU gyroscope calibration, and lens distortion correction. The APP interface displays a green crosshair and prompts "Self-Check Passed". The system enters the training state. The user selects the advanced standard sweet spot hitting model according to their own skill level. The model is synchronized to the racket's local storage.

[0069] 3. Ball-hitting event triggering and data acquisition: When the user enters the "Start Training" page of the APP and starts the real-time training mode, when the racket is swung to hit the ball, the IMU detects in real time that the instantaneous acceleration peak reaches 18g, which exceeds the trigger threshold of 15g. The 12-point vision module is immediately activated to lock the key images of 10 frames before and after the ball hit. At the same time, the IMU synchronously collects the inertial data of the corresponding timestamp. The dual-modal data is synchronously transmitted to the processing layer of the main control MCU.

[0070] 4. Data Processing and Kinetic Chain Reconstruction: The processing layer first performs distortion correction and IMU image stabilization compensation on the visual image, extracts the pixel coordinates of 12 dynamic points of the human body and converts them into three-dimensional world coordinates; then, it performs time synchronization interpolation on the IMU data, fuses the dual-modal data through the Kalman filtering algorithm, eliminates vibration noise, and reconstructs the complete swing kinetic chain model of this shot.

[0071] 5. Motion Comparison and Real-time Correction: The system fits and compares the measured data with the standard model, and calculates that the deviation of the swing plane angle is θ=6°, which exceeds the trigger threshold of 5°. Then, it generates a corresponding correction voice command, which is transmitted in real time to the Bluetooth headset and APP worn by the user via Bluetooth. Within 40ms after hitting the ball, the user receives voice feedback: "Correction reminder: The swing plane is deviated by 6 degrees. Please press down the racket head, use the core rotation of the waist to generate power, and keep the swing plane parallel to the ground." At the same time, the APP interface simultaneously marks the deviation data of this hit.

[0072] 6. Training Review and Report Generation: After completing training, users can click the "Stop Training" button at the bottom of the app. The system will automatically summarize all the data from this training session and generate a kinetic chain analysis report. The report will display the scores for four dimensions—speed, accuracy, stability, and strength—using a radar chart. It will also calculate key indicators such as sweet spot rate, average swing speed, and stability score. The training data will be uploaded to the cloud platform, where periodic training reports and improvement suggestions will be generated for the user.

[0073] 7. System Standby and Maintenance: After training, the system returns to a low-power standby state, waiting for the next shot to trigger; users can check for system version updates and complete OTA upgrades in the APP, or trigger system self-checks to troubleshoot hardware and system abnormalities.

[0074] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart Peak racket with AI visual correction function, comprising a frame, a fiber-reinforced racket face fixed to the inner side of the frame, and a hollow handle fixedly connected to the bottom of the frame, wherein the handle houses a main control circuit board, a power supply module, and a Bluetooth communication module, characterized in that, A Φ22mm through-type stepped mounting hole is provided at the geometric center position of the top 12 o'clock direction of the frame. A 22mm embedded wide-angle camera compartment is set in the stepped mounting hole, and a 12-point visual acquisition module is fixedly installed inside the camera compartment. A high-damping silicone shock-absorbing pad is set at the four orthogonal axial positions of 0°, 90°, 180° and 270° on the outer wall of the camera compartment. The camera compartment is suspended and non-rigidly connected to the inner wall of the stepped mounting hole through the four high-damping silicone shock-absorbing pads, forming a cross-axis suspended shock-absorbing structure. A 9-axis IMU sensor is also set inside the handle. The 12-point visual acquisition module and the 9-axis IMU sensor are both electrically connected to the main control circuit board.

2. The intelligent Peak racket with AI visual correction function according to claim 1, characterized in that, The stepped mounting hole has a coaxial two-section structure. The first hole section facing the hitting side of the racket face has an inner diameter of 22mm and a depth of 8mm. The second hole section facing the outside of the racket frame has an inner diameter of 18mm and a depth of 4mm. The connection between the two hole sections forms an annular stepped limiting surface. The camera compartment has an outer diameter of 21.8mm, an inner diameter of 16mm, and a total depth of 12mm. The end of the compartment facing the outside of the racket frame is provided with a limiting flange that mates with the annular stepped limiting surface.

3. The intelligent Peak racket with AI visual correction function according to claim 1, characterized in that, The high-damping silicone shock-absorbing pad is cylindrical with a diameter of 5mm, a thickness of 3mm, and a Shore hardness of 30HA. The outer wall of the camera compartment has four orthogonal axial positions with an inner diameter of 5mm and a depth of 1.5mm. One end of the high-damping silicone shock-absorbing pad is interference-fitted into the mounting groove, and the other end is interference-fitted with the inner wall of the stepped mounting hole, so that the camera compartment and the frame have no rigid contact.

4. The intelligent Peak racket with AI visual correction function according to claim 1, characterized in that, The 12-point visual acquisition module is a 4K global shutter wide-angle module with a sampling frame rate of 120fps and a lens field of view of 120°; the sampling frequency of the 9-axis IMU sensor is 1kHz; the Bluetooth communication module is a BLE5.0 low-power Bluetooth module; the main control circuit board has a built-in Flash storage unit for storing deep learning models and standard sweet spot hitting models.

5. An AI-based visual kinetic chain correction method based on a smart Peak racket, characterized in that, Based on the execution of the smart Peak racket according to any one of claims 1-4, the core calculation is completed by the racket's main control MCU, specifically including the following steps: S1 Image Acquisition: The racket's 12-point visual acquisition module and 9-axis IMU sensor complete dual-modal synchronous data acquisition to obtain high-speed visual flow and inertial navigation data before and after the shot. S2 Coordinate Transformation: After preprocessing the acquired visual images, the pixel coordinates of 12 dynamic points of the human body are extracted in real time and converted into three-dimensional world coordinates with the center of the racket sweet spot as the origin and aligned with the spatial posture of the racket. S3 Action Comparison: The time synchronization of visual data and inertial data is achieved through linear interpolation. Then, the dual-source data is fused by the Kalman filtering algorithm to eliminate vibration noise, reconstruct the swing kinetic chain model, and fit and compare the measured kinetic chain model with the preset standard sweet spot hitting model to calculate the swing deviation value. S4 Real-time Correction Suggestion: The deviation value is judged by a preset correction feedback threshold algorithm. When the deviation value exceeds the trigger threshold, the corresponding real-time correction instruction is generated and broadcast through the voice terminal.

6. The AI ​​visual kinetic chain correction method based on a smart Peak racket according to claim 5, characterized in that, In step S1, the preset ball-hitting trigger threshold is the instantaneous acceleration peak detected by the IMU ≥ 15g. When the trigger threshold is reached, the 12-point visual acquisition module is activated to acquire 5 frames before and after the ball-hitting trigger moment, for a total of 10 key frames, with a sampling frame rate of 120fps. At the same time, the IMU inertial data within 200ms before and after the ball-hitting moment is acquired simultaneously, and the visual images and inertial data are marked with the same high-precision timestamp, with a timestamp accuracy ≤ 100μs.

7. The AI ​​visual kinetic chain correction method based on a smart Peak racket according to claim 5, characterized in that, In step S2, the 12 extracted human body dynamic points are specifically the left shoulder, left elbow, left wrist, left hip, left knee, left ankle, right shoulder, right elbow, right wrist, right hip, right knee, and right ankle. The confidence threshold for dynamic point extraction is set to 0.

85. In step S3, the reconstructed swing dynamic chain model covers the entire trajectory from backswing, swing, hitting the ball to follow-through.

8. The AI ​​visual kinetic chain correction method based on a smart Peak racket according to claim 5, characterized in that, In step S4, the triggering rules of the correction feedback threshold algorithm are as follows: voice correction is triggered when the angle deviation between the swing plane and the standard plane is >5°, voice correction is triggered when the racket face torsion angle deviation is >3°, voice correction is triggered when the lateral / longitudinal deviation of the hitting point relative to the center of the sweet spot is >5mm, and voice correction is triggered when the similarity of the swing trajectory of the 12 power points is <85%; the entire process from the hitting trigger to the user receiving the voice broadcast takes ≤45ms.

9. The AI ​​visual kinetic chain correction method based on a smart Peak racket according to claim 5, characterized in that, It also includes the execution steps of the supporting mobile terminal APP: smart racket Bluetooth pairing, 12-point camera self-test and calibration, real-time training interface visualization, kinetic chain analysis report generation, and OTA system firmware upgrade; wherein, the kinetic chain analysis report includes total number of training hits, sweet spot rate, average swing speed, stability score, and radar chart visualization data for four dimensions: speed, accuracy, stability, and power.

10. An AI-powered visual pickle ball intelligent training system, characterized in that, The system includes a smart Peak racket, a mobile terminal APP module, and a cloud data analysis module. The smart Peak racket is the same as described in any one of claims 1 to 4. The smart Peak racket communicates with the mobile terminal APP module via its own Bluetooth communication module, and the mobile terminal APP module is network-connected to the cloud data analysis module. The smart Peak racket is used to execute the AI ​​visual kinetic chain correction method described in any one of claims 5 to 9. The mobile terminal APP module provides access for device management, training visualization, and data viewing. The cloud data analysis module is used for training data statistical analysis, training report generation, and algorithm model iterative optimization.

Images