Sports action correcting and monitoring system based on image recognition and big data

Abstract

The invention provides a sports action correcting and monitoring system based on image recognition and big data, relates to the technical field of artificial intelligence, and integrates multi-source data, eliminates space-time differences and improves the accuracy and reliability of an analysis result through a data synchronization and fusion module by using a big data processing technology; the biomechanical solving module calculates a lower limb joint torque vector in real time based on a Lagrange equation, and real-time monitoring of the joint torque is achieved; the load optimization module adopts an Adam optimization algorithm, optimizes the stride width and the knee joint buckling angle through a gradient descent method, minimizes the maximum value of the joint torque, reduces the joint load of an athlete, and prevents potential injury; the feedback display module transmits the optimized action parameters to the wearable device in real time, the athlete is guided to adjust the action in real time, and the high efficiency and safety of training are ensured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a sports movement correction and monitoring system based on image recognition and big data. Background Art

[0002] Real-time monitoring of athletes' movement parameters and biomechanical indicators can provide coaches with scientific and objective training basis, assist athletes in optimizing movement patterns and improving their competitive level. In addition, the development of equipment such as inertial measurement units (IMUs) and ground reaction force sensors has made it possible to measure key biomechanical parameters such as lower limb joint torque during training, further promoting the in-depth study of sports science.

[0003] The existing technology has a publication number of CN111738044A, and its name is a campus violence assessment method based on deep learning behavior recognition, which includes collecting surveillance video data distributed throughout the campus, using surveillance cameras at different locations as classification labels, splitting video stream data into different continuous frame groups, inputting, training, and building a three-dimensional convolutional neural network (3D-CNN) campus violence assessment model, and using cross-validation to test the generalization ability of the model; on this basis, the action category of the newly input data of the individual is identified, the safety status of the location is judged, and an alarm is issued for abnormal behavior. In the context of the big data era, it not only ensures the scientificity, efficiency and safety of management, but also provides an effective solution for preventing and controlling campus violence.

[0004] In the indoor training center for sprinting, coaches rely on their naked eyes and single-camera playback to correct stride length and landing point, but they cannot simultaneously know the instantaneous joint load of the athlete's lower limbs, making it difficult to find potential injury points in time;

[0005] Existing systems often lack the ability to synchronously collect and fuse multi-source data, making it difficult to effectively integrate image data, inertial measurement data, and ground reaction force data, which limits the accuracy and reliability of analysis results. In addition, traditional biomechanical analysis methods often have calculation delays when calculating complex parameters such as joint torque in real time, and are unable to provide athletes with immediate action adjustment suggestions, which in turn affects the effect of action optimization;

[0006] The above information disclosed in the above Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention provides a sports movement correction and monitoring system based on image recognition and big data to solve the problems raised in the above background technology.

[0008] The object of the present invention is achieved as follows: A sports action correction and monitoring system based on image recognition and big data, specifically including:

[0009] Data acquisition module: Used to collect real-time multi-view synchronous image streams, three-direction ground reaction force data, and six-axis inertial measurement data of the target athlete during the sprint movement monitoring period;

[0010] Three-dimensional pose estimation module: Used to receive multi-view synchronous image streams and extract the three-dimensional coordinate sequences of multiple lower limb key points of the target athlete;

[0011] Data synchronization and fusion module: Used to fuse the three-dimensional coordinate sequences, three-direction ground reaction force data, and six-axis inertial measurement data to output the initial value vector of sports action correction for multiple lower limb key points;

[0012] Biomechanics solving module: Used to calculate the lower limb joint torque vector of the target athlete during the sprint movement monitoring period based on the Lagrange equation;

[0013] Load optimization module: Used to receive the lower limb joint torque vector of the target athlete during the sprint movement monitoring period, optimize the initial value vector of sports action correction, and output the corrected sports action parameter vector for each lower limb key point;

[0014] Feedback display module: Used to receive the corrected sports action parameter vectors for each lower limb key point and output action adjustment instructions to the wearable display terminal in real time.

[0015] Furthermore, the multi-view synchronous image stream realizes global shutter synchronization through a hardware synchronizer;

[0016] The three-dimensional pose estimation module uses the HRNet-W48 deep learning network as the backbone network and combines a differentiable triangulation algorithm to calculate the three-dimensional coordinates of lower limb key points;

[0017] The three-dimensional pose estimation module includes a camera calibration sub-module for obtaining the internal parameter matrix of three cameras and the external parameter matrix ;

[0018] The three-dimensional pose estimation module further includes a two-dimensional key point detection sub-module for outputting the two-dimensional pixel coordinates of 6 lower limb key points in each frame of image ; where represents six lower limb key points, and c = 1, 2, 3 represents three cameras;

[0019] The three-dimensional pose estimation module includes a differentiable triangulation layer for calculating the three-dimensional coordinates of six lower limb key points based on the camera projection matrix and the two-dimensional key point pixel coordinates 。

[0020] Furthermore, the three-direction ground reaction force data specifically includes:

[0021] Set the three-direction force vector of the three-direction ground reaction force data as ;

[0022] Among them, is the three-direction ground reaction force vector; is the horizontal front-back / left-right component force; is the vertical support force;

[0023] The six-axis inertial measurement data specifically includes:

[0024] Characterize the six-axis inertial measurement data as follows;

[0025]

[0026] Among them, , , are the attitude quaternion, linear acceleration, and angular velocity of the six-axis inertial measurement data at time t, respectively; Used to characterize the direction or orientation of multiple lower limb key points of the target athlete's lower limbs in space, that is, the rotation state of the object relative to the defined coordinate system;

[0027] Attitude quaternion Used to represent the IMU attitude of multiple lower limb key points at time t;

[0028] Linear acceleration Used to represent the three-axis linear acceleration of multiple lower limb key points at time t;

[0029] Angular velocity Used to represent the three-axis angular velocity of multiple lower limb key points at time t.

[0030] Furthermore, the data synchronization and fusion module uses the Kalman filter algorithm to integrate the three-dimensional coordinates , velocity , three-direction ground reaction force and six-axis inertial measurement data to perform data fusion and generate a unified initial value vector for sports action correction ; Specifically:

[0031] The data synchronization and fusion module sets a unified timestamp through the GPS-PPS reference ;

[0032] The data synchronization and fusion module uses the rigid body transformation matrix and the translation vector , map the three-dimensional coordinates of the six lower limb key points to the laboratory coordinate system to achieve a unified spatial reference with the three-direction ground reaction force and the six-axis inertial measurement data ;

[0033] Establish a Kalman filter model: Set the state transition matrix of the Kalman filter and the observation matrix ;

[0034] Using the defined Kalman filter model, input the initial value vector of the sports action correction and the observation vector , and perform the prediction and update steps;

[0035] Set the updated state vector as the initial value vector of the sports action correction at the current moment .

[0036] Furthermore, the biomechanical solution module is based on the Lagrange equation, combined with the mass-inertia matrix , the Coriolis term and the centrifugal force term , the gravity term , and calculate the lower limb joint torque vector ;

[0037] The biomechanical solution module establishes a 7-degree-of-freedom lower limb biomechanical model, including a hip joint with 3 degrees of freedom, a knee joint with 1 degree of freedom, and an ankle joint with 3 degrees of freedom;

[0038] The biomechanical solution module sets the mass parameters of each segment of the lower limb , where the mass of the ankle segment , the mass of the knee joint segment , the mass of the hip joint segment ; The mass parameter where j in represents the ankle, knee and hip;

[0039] The biomechanical solution module describes the dynamic equations of each segment of the lower limb through the Lagrange equation, combined with the difference L = T - V between the kinetic energy T and the potential energy V;

[0040] Define the lower limb joint torque vector as:

[0041]

[0042] where, is the hip joint torque; is the knee joint torque; is the ankle joint torque.

[0043] Further, the specific logic for outputting the corrected sports action parameter vectors of each lower limb key point includes:

[0044] The load optimization module reads the sequence of lower limb joint torque vectors within a preset time period and calculates the peak value vectors of the torques of each joint ;

[0045] The load optimization module defines a cost function as a function of the maximum absolute value in the lower limb joint torque vector , in the following form:

[0046]

[0047] are adjustable action parameters, representing the stride width and the knee flexion angle respectively;

[0048] The load optimization module uses the Adam optimization algorithm to adjust the action parameters within multiple iterations to minimize the cost function ; and records the optimized action parameter vector as

[0049] Further, the load optimization module calculates the difference in action parameters between the current moment and the previous moment to form a parameter increment vector ;

[0050] Calculate the difference between the optimized action parameter vector and the parameter vector at the previous moment:

[0051]

[0052] wherein, is the optimized action parameter vector; is the action parameter vector at the previous moment;

[0053] The load optimization module transmits the optimized action parameters and the parameter increment vector to the feedback display module for real-time action correction;

[0054] Define the optimized action parameter vector as:

[0055]

[0056] wherein, is the optimized stride width; is the optimized knee flexion angle.

[0057] Further, the feedback display module uses a head-up display (HUD) and vibration feedback to guide the athlete to adjust the stride length and knee flexion angle in real time, forming a closed-loop motion correction mechanism;

[0058] Map the stride increment to the arrow length on the head-up display (HUD), indicating the direction and magnitude of the stride adjustment;

[0059] Map the change in knee flexion angle to the angle scale on the HUD, prompting the angle by which the knee joint needs to be adjusted;

[0060] When the stride increment is reached, activate the vibration motor fixed at the ilium of the athlete, and prompt the athlete to adjust the stride length through vibration.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention synchronously acquires multi-view image streams, ground reaction force data, and six-axis inertial measurement data through the data acquisition module to ensure comprehensive recording of various key parameters during the movement process; The three-dimensional pose estimation module uses advanced image recognition algorithms to accurately extract the three-dimensional coordinates of multiple key points of the lower limbs, providing a basis for detailed motion analysis;

[0062] The data synchronization and fusion module uses big data processing technology to integrate multi-source data, eliminate spatio-temporal differences, and improve the accuracy and reliability of the analysis results; The biomechanics solution module is based on Lagrange's equation and calculates the lower limb joint torque vector in real time to achieve instant monitoring of joint torques;

[0063] The load optimization module uses the Adam optimization algorithm to optimize the stride width and knee flexion angle through the gradient descent method, minimizing the maximum value of joint torque, reducing the joint load of the athlete, and preventing potential injuries;

[0064] The feedback display module transmits the optimized motion parameters to the wearable device in real time to guide the athlete to adjust the motion immediately, ensuring the efficiency and safety of training. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0066] Figure 1 It is a schematic diagram of the overall system module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0068] As shown in the figure

[0069] Embodiment 1:

[0070] Please refer to Figure 1 , the present invention provides a technical solution:

[0071] A sports action correction and monitoring system based on image recognition and big data, specifically including:

[0072] Data acquisition module: used to collect multi-view synchronous image streams, three-way ground reaction force data, and six-axis inertial measurement data of the target athlete during the sprint movement in real time;

[0073] Further explanation: The multi-view synchronous image stream realizes global shutter synchronization through a hardware synchronizer; and the multi-view synchronous image stream is 240 frames per second;

[0074] The specific operation steps to achieve global shutter synchronization are as follows:

[0075] 1.1) Camera arrangement: Arrange three Sony IMX273 cameras on the side of the sprint track in an equilateral triangle arrangement; set a triangular arrangement baseline along the running direction on the side edge of the track, and the baseline length is 4m;

[0076] The optical axes of the three camera lenses all point to the center line of the track, and the horizontal angle is 120°;

[0077] The height of the camera optical center is 1.20m, and the leveling error of the tripod top plate is ≤0.2°;

[0078] The camera is connected to the GPU workstation through a 10GbE optical fiber to provide a dedicated data channel for synchronization configuration;

[0079] 1.2) Synchronization configuration: Each camera is configured with a global shutter, and receives GPS-PPS pulses through a hardware synchronizer to achieve global shutter synchronization. The PPS (Pulse Per Second) pulse synchronization error is ≤0.1 millisecond. The specific operation is as follows:

[0080] Send the GPS-PPS signal to the trigger port of each camera through an 8-way distributor;

[0081] Set the camera trigger mode to "external trigger priority", and the trigger edge is the rising edge;

[0082] Calibration trigger delay: Read the trigger-exposure delay given by the manufacturer and write it into the register;

[0083] Record synchronization error: Measure the difference of the three-channel exposure signals with an oscilloscope and ensure it is ≤ 0.1 ms;

[0084] After synchronization is completed, output a global shutter image stream at 240 fps for the ground force platform installation and alignment.

[0085] 1.3) Data transmission: The camera captures images at a speed of 240 frames per second in RAW 12-bit format and transmits them to the GPU workstation in real time through a 10 GbE network interface. The size of a single-frame data is 1.57 MB.

[0086] Example 2:

[0087] 3D pose estimation module: Used to receive the multi-view synchronized image stream and extract the 3D coordinate sequences of multiple lower limb key points of the target athlete;

[0088] Further explanation: The 3D pose estimation module uses the HRNet-W48 deep learning network as the backbone network and combines a differentiable triangulation algorithm to calculate the 3D coordinates of the lower limb key points;

[0089] The sprint action is fast and the transient force of the lower limbs is obvious. It is necessary to accurately restore the 2D lower limb key points to 3D coordinates from the perspectives of three synchronized cameras;

[0090] HRNet-W48 can maintain the spatial consistency of key points in multi-resolution fusion;

[0091] The differentiable triangulation layer enables the entire network to be trained end-to-end, facilitating the joint optimization of camera parameters and 2D / 3D key points;

[0092] It can converge on the GPU within 0.3 ms / point through 10 iterations of Gauss-Newton, meeting the real-time requirement of 240 Hz;

[0093] The 3D pose estimation module includes a camera calibration sub-module for obtaining the internal parameter matrix of the three cameras and the external parameter matrix ;

[0094] The specific implementation steps are as follows:

[0095] Adopt a 9×6 checkerboard calibration board and place it in 10 different poses on the side of the runway;

[0096] Use the OpenCV-Zhang calibration process to calculate the internal parameter matrix of each camera , and optimize the external parameters through panoramic stitching to obtain an average residual of the three cameras ≤ 0.12 px;

[0097] Store the intrinsic parameter matrix and the extrinsic parameter matrix as a floating-point matrix and solidify and write it into the HRNet-W48 inference script configuration file.

[0098] The three-dimensional pose estimation module further includes a two-dimensional key point detection sub-module for outputting the two-dimensional pixel coordinates of 6 lower limb key points in each frame of image ; where represents six lower limb key points, and c = 1, 2, 3 represent three cameras;

[0099] The specific implementation steps are as follows:

[0100] Image preprocessing: Normalize the channels of the synchronized image streams collected by the three cameras, scale them to the range [0, 1], and adjust the image resolution to 384×288 pixels to adapt to the input requirements of the HRNet-W48 model.

[0101] Key point detection: Load the HRNet-W48 model weight file trained on a 50TB competitive action big data set, perform inference on each frame of image, and extract the two-dimensional pixel coordinates of six lower limb key points.

[0102] Sub-pixel regression: Apply the Gaussian sub-pixel regression method on the output heatmap of the HRNet-W48 model to accurately locate the sub-pixel positions of each key point , and the quantization accuracy reaches 0.01 pixel;

[0103] Data encapsulation: Package the two-dimensional coordinates of six lower limb key points and the corresponding confidence values into a structured data structure Key2D[c][i] for subsequent processing by the triangulation layer.

[0104] The three-dimensional pose estimation module includes a differentiable triangulation layer for calculating the three-dimensional coordinates of six lower limb key points based on the camera projection matrix and the two-dimensional key point pixel coordinates .

[0105] The specific implementation steps are as follows:

[0106] Initialization: For each lower limb key point i, use the two-dimensional pixel coordinates of the three cameras to solve the initial three-dimensional coordinate estimation value through the Direct Linear Transformation (DLT) method ;

[0107] Error function construction:

[0108] Construct the reprojection error function of each lower limb key point at the current time t:

[0109]

[0110] Among them, are the three-dimensional coordinates of the lower limb key points;

[0111] is the perspective projection function of the c-th camera, and are the focal lengths of the camera, and are the principal point coordinates of the camera;

[0112] are the two-dimensional pixel coordinates of the i-th lower limb key point captured by the c-th camera at time t;

[0113] The Gauss-Newton method is used to iteratively optimize the error function The number of iterations is set to 10 times, and the three-dimensional coordinate estimate value is updated each time;

[0114] Adopting the inverse depth parameterization method, the three-dimensional coordinates are represented as , where d is the depth distance to enhance the stability of the optimization process;

[0115] Convergence determination: In each iteration, if the coordinate update amplitude mm, the optimization process is terminated in advance.

[0116] After the optimization is completed, the final three-dimensional coordinates are output, with the unit of millimeters, and the gradient information is fed back to the two-dimensional key point detection sub-module to support the end-to-end network training.

[0117] The three-dimensional pose estimation module includes a joint training and parameter optimization sub-module for simultaneously optimizing the two-dimensional key point detection weights and the parameters of the differentiable triangulation layer;

[0118] The specific implementation steps are as follows:

[0119] The 50TB competitive action big data set is divided into a training set, a validation set, and a test set according to a ratio of 8:1:1. The training set contains 2 million frames of labeled lower limb key point data in the COCO format.

[0120] The total loss function is set as:

[0121]

[0122] Among them, is the mean squared error (MSE) loss of two-dimensional key point detection; L1 distance loss for three-dimensional coordinate prediction; = 1, = 0.5, which is the weight coefficient of the loss function;

[0123] The Adam optimizer is adopted, and the initial learning rate is set to 1×10⁻ 4 , the batch size is 32, and the Cosine Decay strategy is used to adjust the learning rate.

[0124] 60 training epochs are carried out. After each epoch ends, the performance of the model is evaluated using the validation set, and the Mean Per Joint Position Error (MPJPE) is monitored.

[0125] If the MPJPE of the validation set does not improve within 5 consecutive training epochs, the training is terminated early.

[0126] The three-dimensional pose estimation module outputs a sequence of three-dimensional coordinates of six lower limb key points, and synchronizes this sequence with other sensor data according to the time stamp.

[0127] The specific implementation steps are as follows:

[0128] During the inference process, the calculated three-dimensional coordinates are written into the shared memory area.

[0129] At the beginning of each output data frame, a 64-bit time stamp of the GPS-PPS signal source is added, and the time stamp accuracy is 0.05 milliseconds to ensure the time synchronization of the data.

[0130] The three-dimensional coordinate data with the time stamp is pushed to the "Data Synchronization and Fusion Module" through the ZeroMQ protocol, and is aligned with the ground reaction force (GRF) and inertial measurement unit (IMU) data under the unified time reference.

[0131] The beneficial effects of this embodiment are as follows:

[0132] 1. High-precision camera calibration: By using a 9×6 checkerboard calibration board and the OpenCV-Zhang calibration algorithm, high-precision internal and external parameters calibration of the three cameras is achieved, and the reprojection error residual ≤ 0.12 pixels, which significantly improves the accuracy of three-dimensional coordinate calculation.

[0133] 2. Differentiable triangulation algorithm: The differentiable triangulation layer is introduced, so that the three-dimensional pose estimation process can be optimized jointly end-to-end. The camera parameters and key point positions are automatically adjusted using the backpropagation algorithm to achieve higher three-dimensional coordinate accuracy.

[0134] 3. Large-scale data training: The model is trained using a self-built 50TB big data set of competitive action, which contains 2 million frames of labeled lower limb key point data, ensuring the generalization ability and robustness of the model in actual motion scenarios, with the mean per-joint position error (MPJPE) less than 0.9 mm.

[0135] 4. Real-time guarantee: Through the Gauss-Newton method, rapid iterative optimization is carried out on the GPU (10 iterations, convergence time per point ≤ 0.3 ms), ensuring that the system can achieve real-time three-dimensional pose estimation at a high frame rate of 240Hz.

[0136] 5. Multi-source data synchronization: High-precision time synchronization (time alignment error ≤ 0.05 ms) of three-dimensional coordinate data with GRF and IMU data is achieved through GPS-PPS signals, improving the accuracy of multi-source data fusion analysis, which is different from the existing technologies that only analyze a single data source.

[0137] Example Two:

[0138] Further explanation: The three-direction ground reaction force data specifically includes:

[0139] The sampling frequency of the three-direction ground reaction force data is 1kHz;

[0140] Set the three-direction force vector of the three-direction ground reaction force data as ;

[0141] Among them, is the three-direction ground reaction force vector; is the horizontal front-back / left-right component force; is the vertical support force;

[0142] The specific operation steps are as follows:

[0143] Select an AMTI OR6-7 force measurement platform with a range of ±10kN and a sensitivity of 2µV / N;

[0144] Embed the force measurement platform in the starting area of the runway, with the platform surface flush with the runway surface layer, and the installation error controlled within ±0.5mm;

[0145] Use four-point bolt fastening for the force platform to ensure that its verticality error is less than 0.05°;

[0146] Configure the platform data acquisition system and set the sampling frequency to 1kHz;

[0147] Set a fourth-order Butterworth low-pass filter with a cut-off frequency of 100Hz to filter the original data, and the output three-direction force vector is (unit: Newton);

[0148] Structurally store the collected data packets, attach a sampling timestamp to each data frame, and later call it for the data fusion module.

[0149] Connect the GPS-PPS pulse to the SYNC_IN port of the force platform controller through a shielded cable;

[0150] With the help of the GPS-PPS signal, achieve time alignment between the platform data frame and the image acquisition module, ensuring that the time alignment error is less than 0.05 ms;

[0151] Write a unified time stamp in the data stream and record the data frame acquisition time for subsequent multi-module data fusion.

[0152] The ground reaction force acquisition module provides mechanical data synchronized with the output of the three-dimensional attitude estimation module. Through the data synchronization and fusion module, accurately correlate the GRF data with the three-dimensional coordinate sequence of the lower limb key points, providing necessary mechanical information for the biomechanics solution module.

[0153] Further explanation: The six-axis inertial measurement data specifically includes:

[0154] Characterize the six-axis inertial measurement data as follows;

[0155]

[0156] Among them, , , are the attitude quaternion, linear acceleration, and angular velocity of the six-axis inertial measurement data at time t, respectively; Used to characterize the direction or orientation of the lower limbs of multiple lower limb key points of the target athlete in space, that is, the rotation state of the object relative to the defined coordinate system;

[0157] Attitude quaternion Used to represent the IMU attitude of multiple lower limb key points at time t;

[0158] Linear acceleration Used to represent the three-axis linear acceleration (m / s²) of multiple lower limb key points at time t;

[0159] Angular velocity Used to represent the three-axis angular velocity (degrees / second) of multiple lower limb key points at time t;

[0160] The specific operation steps are as follows:

[0161] 2.1) Layout and installation: Select Bosch BMI270 IMU, and each unit is configured with a range of ±16 g and ±2000 ° / s;

[0162] Fix the IMU on the athlete's waist (near the L4 vertebral body level area) and the anterior edges of the left and right tibias, and use a 3D printed shell (weight < 25g) and a combination of medical-grade double-sided tape + Velcro for fixation to ensure an anti-vibration ability with an anti-vibration frequency greater than 200Hz;

[0163] Check the installation angles of each IMU to ensure that the installation directions of all devices are consistent with the human body coordinates, and the installation error does not exceed 1°;

[0164] 2.2) Attitude solution and data acquisition:

[0165] Set the IMU to SPI slave mode and configure the external clock input, and use the locked GPS-PPS pulse (1kHz) as the external clock;

[0166] Use the Mahony filter, set the parameter β = 0.05, perform real-time attitude solution on the original acceleration and angular velocity signals, and calculate and output the quaternion ;

[0167] At the same time, collect the linear acceleration and angular velocity values, and encapsulate them into a data frame structure ;

[0168] Each data frame contains an accurate timestamp, which is determined by the synchronization result of the internal counter of the device and the external GPS-PPS, and the sampling rate is maintained at 1kHz.

[0169] 2.3) Time alignment:

[0170] Input the Firefly synchronization pulse signal into each IMU through the GPIO interface to achieve phase locking with the GPS-PPS signal, and ensure that the internal and external clocks of all IMUs are aligned;

[0171] Use an oscilloscope to detect the time drift between IMU data frames to ensure that the time drift is less than 0.2ms / hour;

[0172] Unify the timestamps in the data frames of each IMU with the images and force platforms, and perform subsequent spatio-temporal correlation processing through the data synchronization and fusion module.

[0173] The six-axis inertial measurement data obtained by the inertial attitude acquisition module, the data of the three-dimensional attitude estimation module and the ground reaction force acquisition module are precisely time-aligned and spatially correlated through the data synchronization and fusion module, providing complete lower limb motion dynamic information for the biomechanics solution module.

[0174] Example three:

[0175] Data Synchronization and Fusion Module: It is used to fuse the three-dimensional coordinate sequence, three-direction ground reaction force data, and six-axis inertial measurement data to output the initial vector of sports action correction for multiple lower limb key points;

[0176] Further Explanation: The data synchronization and fusion module uses the Kalman filtering algorithm to integrate the three-dimensional coordinates , speed , three-direction ground reaction force and six-axis inertial measurement data to perform data fusion and generate a unified initial vector of sports action correction ; Specifically:

[0177] The data synchronization and fusion module sets a unified timestamp through the GPS-PPS reference , and interpolates the three-dimensional coordinates of six lower limb key points to 1kHz to ensure that the time alignment error is less than 0.05 milliseconds; Specifically:

[0178] Use the GPS-PPS (Global Positioning System - Pulse Per Second) signal as the unified time reference to set the unified timestamp for the entire system .

[0179] Align the original timestamp of the three-dimensional coordinate sequence of six lower limb key points with the unified timestamp to ensure that the time reference of all data sources is consistent.

[0180] For the data, the five-point difference method is used to calculate the speed and interpolate it to a sampling rate of 1kHz, and the specific calculation formula is as follows:

[0181]

[0182] where, is the three-dimensional coordinate of the i-th lower limb key point at time t; is the speed vector of the i-th lower limb key point; seconds is the sampling interval;

[0183] Use a high-precision timing device to measure the time alignment error after interpolation to ensure that the error value is less than

[0184] The data synchronization and fusion module uses the rigid body transformation matrix and the translation vector to transform the three-dimensional coordinates of six lower limb key points Map to the laboratory coordinate system to achieve unification with the three - dimensional ground reaction force and six - axis inertial measurement data for a unified spatial reference;

[0185] The specific operation steps are as follows:

[0186] Obtain rigid - body transformation parameters: According to the camera calibration results, determine the rigid - body transformation matrix and the translation vector as follows:

[0187]

[0188] where, is the three - dimensional coordinate of the i - th lower - limb key point in the laboratory coordinate system; is the rotation matrix; is the translation vector.

[0189] Coordinate transformation: Map the three - dimensional coordinates of the six interpolated lower - limb key points to the laboratory coordinate system through the above rigid - body transformation to obtain the three - dimensional coordinates under the unified spatial reference;

[0190] Spatial alignment verification: Use the known spatial calibration points in the laboratory to verify the spatial alignment accuracy of, and ensure that the spatial error is within millimeters;

[0191] Set the initial value vector for sports action correction as follows:

[0192]

[0193] where, is the three - dimensional coordinate of the i - th lower - limb key point at time t;

[0194] is the velocity of the i - th lower - limb key point;

[0195] is the three - dimensional ground reaction force vector;

[0196] is the six - axis inertial measurement data vector.

[0197] Establish a Kalman filter model: Set the state - transition matrix and the observation matrix of the Kalman filter; as follows:

[0198]

[0199] Among them, is the original observation vector; is the state transition matrix for predicting the current state; is the observation matrix for combining current observation data; is the process noise, assuming it follows a zero-mean Gaussian distribution.

[0200] Kalman filter processing: Using the defined Kalman filter model, input the initial value vector of sports action correction and the observation vector , and execute the prediction and update steps; specifically as follows:

[0201]

[0202]

[0203] Among them, is the predicted state vector; is the predicted error covariance matrix; is the process noise covariance matrix.

[0204] The update steps are as follows:

[0205]

[0206]

[0207]

[0208] Among them, is the Kalman gain; is the observation matrix; is the observation noise covariance matrix; is the identity matrix;

[0209] Generate the initial value vector: Set the updated state vector as the initial value vector of sports action correction at the current moment ;

[0210] Write the initial value vector of sports action correction into the circular buffer and transmit it to the subsequent processing module through the high-speed data channel.

[0211] The data synchronization and fusion module outputs the unified motion state vector , which contains the initial value vector of sports action correction , with a sampling frequency of 1 kHz, and write it into the circular buffer for subsequent module calls.

[0212] The specific operation steps are as follows:

[0213] Set the unified motion state vector as:

[0214]

[0215] Write sequentially into a circular buffer with a length of 2 seconds to ensure the continuity and real-time nature of the data.

[0216] Configure the data interface so that the subsequent biomechanics solution module can read in real time from the circular buffer to complete the subsequent processing and analysis of the data.

[0217] The beneficial effects of this embodiment are as follows:

[0218] High-precision time alignment: Use the GPS-PPS reference to set a unified timestamp, and combine the five-point difference method for interpolation to 1 kHz to align the real-time error, improving the accuracy of multi-source data synchronization.

[0219] Unified spatial reference conversion: Through the rigid body transformation matrix and translation vector, accurately map the three-dimensional coordinates of six lower limb key points to the laboratory coordinate system to ensure spatial consistency, which is different from the problem of inconsistent multi-source data spatial references in the prior art.

[0220] Integrated Kalman filter algorithm: Introduce the Kalman filter algorithm, combine the state transition matrix and the observation matrix, and achieve efficient fusion of coordinate, velocity, ground reaction force, and inertial measurement data to generate a comprehensive initial value vector for sports action correction, improving the accuracy and real-time nature of data fusion.

[0221] Example 4:

[0222] Biomechanics solution module: Used to calculate the lower limb joint torque vector of the target athlete during the sprint movement based on the Lagrangian equation;

[0223] The said biomechanics solution module is based on the Lagrangian equation and combines the mass-inertia matrix Coriolis terms and centrifugal force terms Gravitational terms to calculate the lower limb joint torque vector ;

[0224] Further explanation: 3.1) The said biomechanics solution module establishes a 7-degree-of-freedom lower limb biomechanics model, including a hip joint with 3 degrees of freedom, a knee joint with 1 degree of freedom, and an ankle joint with 3 degrees of freedom;

[0225] Select a 7-degree-of-freedom lower limb biomechanics model to describe the movement and mechanical characteristics of the hip, knee, and ankle joints. Specifically, it includes:

[0226] Hip joint: 3 degrees of freedom, representing flexion / extension, abduction / adduction, and rotation respectively;

[0227] Knee joint: 1 degree of freedom, representing flexion / extension;

[0228] Ankle joint: 3 degrees of freedom, representing flexion / extension, inversion / eversion, and rotation respectively;

[0229] Define a local coordinate system for each joint to ensure that the motion relationships of all parts of the model are consistent with the actual human anatomy; the origin of the coordinate system is set at the joint center, and the axis directions conform to the right-hand rule;

[0230] Set the connection parameters between the joints, including joint length, angle limits, and motion ranges, to simulate the dynamic characteristics of the lower limbs of actual athletes.

[0231] 3.2) The biomechanical solution module sets the mass parameters of each segment of the lower limb , where the mass of the ankle segment , the mass of the knee segment , the mass of the hip segment ; the mass parameter where j represents ankle, knee, and hip;

[0232] The specific operation steps are as follows: In this embodiment, based on the Dempster scale and combined with the tester's weight of 75 kg, we obtain , and ;

[0233] Based on the mass parameters of each segment, calculate the mass-inertia matrix to describe the inertial characteristics of the lower limb joint system; the specific calculation formula of the inertia matrix is as follows:

[0234]

[0235] where, represents the inertial coupling term of the system in the generalized coordinate ;

[0236] Through simulation tests, verify the accuracy of the mass-inertia matrix to ensure that it can truly reflect the dynamic characteristics of the lower limb joint system.

[0237] 3.3) The biomechanical solution module uses the Lagrange equation, combined with the difference L = T - V between kinetic energy T and potential energy V, to describe the dynamic equations of each segment of the lower limb;

[0238]

[0239] The specific operation steps are as follows:

[0240] Define the kinetic energy \(T\) and potential energy \(V\): The kinetic energy ; The potential energy ;

[0241] Using the Lagrangian equation \(L = T - V\), establish the dynamic equation of the lower limb joint system:

[0242]

[0243] where, is the generalized coordinate of the lower limb joint, 7-dimensional (hip × 3, knee × 1, ankle × 3);

[0244] is the generalized velocity of the lower limb joint; is the generalized acceleration of the lower limb joint;

[0245] is the torque vector of the lower limb joint, with the unit of Newton-meter (N·m);

[0246] is the plantar Jacobian matrix, which describes the relationship between the plantar and the joint;

[0247] is the three-dimensional ground reaction force vector.

[0248] Integrate the mass-inertia matrix , the Coriolis term and centrifugal force term , and the gravity term into the equations:

[0249]

[0250] Through simulation tests, verify the accuracy described by the Lagrangian equation, and fine-tune the equation parameters according to the actual motion data to ensure the authenticity and accuracy of the dynamic model.

[0251] The biomechanical solution module adopts the improved Newmark-β method, sets the parameters , for numerical integration, with a step size milliseconds, and discretely solves the equations;

[0252] According to the initial value vector of sports action correction provided by the data synchronization and fusion module, set the generalized coordinate , velocity and acceleration of the lower limb joint;

[0253] Discretize the dynamic equation and apply the improved Newmark-β method to iteratively solve the lower limb joint torque :

[0254]

[0255]

[0256] Among them, it is gradually updated through an iterative method , and , until the system of equations reaches the convergence condition.

[0257] In each iteration, if N·m, it is considered that the joint torque at the current moment has converged, and the iteration is stopped.

[0258] The biomechanical solution module outputs the lower limb joint torque vector , with the unit of Newton-meter (N·m), and transmits it to the load optimization module;

[0259] Define the lower limb joint torque vector as:

[0260]

[0261] Among them, is the hip joint torque; is the knee joint torque; is the ankle joint torque.

[0262] Through the high-speed data interface, the calculated joint torque vector is transmitted to the load optimization module to ensure the transmission delay milliseconds.

[0263] Store the joint torque vector in a structured data format (such as CSV or binary format) for real-time storage, which is convenient for subsequent analysis and processing.

[0264] The beneficial effects of this embodiment are as follows:

[0265] High-degree-of-freedom biomechanical model: A 7-degree-of-freedom lower limb biomechanical model is adopted, covering the multi-degree-of-freedom movements of the hip, knee, and ankle joints, which is more accurate and comprehensive than the existing single-degree-of-freedom models.

[0266] Accurate mass parameter setting: Based on the Dempster ratio table, the mass parameters of each segment of the lower limb are set, combined with the actual weight of the tester, to ensure the authenticity of the model mass distribution and improve the accuracy of the dynamic equation calculation.

[0267] Embodiment 5:

[0268] Load Optimization Module: It is used to receive the lower limb joint torque vectors of a target athlete during a sprint exercise, optimize the initial value vector of sports action correction, and output the corrected sports action parameter vectors for each lower limb key point;

[0269] Further Explanation: The specific logic of outputting the corrected sports action parameter vectors for each lower limb key point includes:

[0270] 4.1) The load optimization module reads the sequence of lower limb joint torque vectors within a preset time period and calculates the peak value vectors of each joint torque ;

[0271] In this embodiment, the preset time period is within 200 milliseconds;

[0272] 4.1) The specific operation steps are as follows:

[0273] Set the rolling window length to 200 milliseconds, the sampling frequency to 1 kHz, and the window size to 200 time frames.

[0274] Receive the sequence of lower limb joint torque vectors from the biomechanics solution module .

[0275] At the current time t, extract the sequence of lower limb joint torque vectors in the time range from t - 200 milliseconds to t:

[0276]

[0277] Within the extracted rolling window, calculate the maximum absolute values of the hip joint, knee joint, and ankle joint torques respectively:

[0278]

[0279] Among them, , , ;

[0280] Store the calculated peak value vectors in the memory buffer for subsequent steps; ensure that the rolling window slides 1 millisecond each time to achieve real-time processing.

[0281] 4.2) The load optimization module defines a cost function as a function of the maximum absolute value in the lower limb joint torque vector , and the form is as follows:

[0282]

[0283] The specific operation steps are as follows:

[0284] are adjustable motion parameters, representing the stride width and the knee flexion angle respectively;

[0285] In this embodiment, the adjustable range of the motion parameters is set as follows:

[0286] Stride width The range is [0.5, 1.5] meters; the knee flexion angle The range is [0°, 120°]; set the initial motion parameter vector as the starting point of the optimization algorithm.

[0287] 4.3) The load optimization module adopts the Adam optimization algorithm to adjust the motion parameters within multiple iterations to minimize the cost function ; and denote the optimized motion parameter vector as ;

[0288] Further explanation: The learning rate of the Adam optimization algorithm in this embodiment , the first moment parameter

[0289] Select the finite difference method as the gradient estimation method of the cost function for the motion parameters as follows:

[0290]

[0291] Among them, and are calculated by the finite difference method respectively;

[0292] Perform 40 optimization iterations, and each iteration executes the following steps:

[0293] Forward propagation: Based on the motion parameters at the current moment t, adjust the stride width and the knee flexion angle, and re-collect and calculate the lower limb joint torque vector ;

[0294] Calculate the cost function .

[0295] Estimate the gradient by the finite difference method.

[0296] Apply the Adam optimizer to update the motion parameters according to the gradient:

[0297]

[0298] If , terminate the iteration in advance.

[0299] After each iteration, record the optimized action parameters and the corresponding cost function values , for subsequent analysis and verification.

[0300] 4.4) The load optimization module calculates the difference in action parameters between the current moment and the previous moment to form a parameter increment vector ;

[0301] The specific calculation steps are as follows:

[0302] Calculate the difference between the optimized action parameter vector and the parameter vector at the previous moment:

[0303]

[0304] Among them, is the optimized action parameter vector; is the action parameter vector at the previous moment;

[0305] Set the threshold of the parameter increment vector ; to ensure the smoothness of action parameter adjustment; in this embodiment ;

[0306]

[0307] Store the calculated parameter increment vector in the memory buffer for use in subsequent steps. It should be noted that the threshold of this embodiment is determined by the expert group using methods such as the fuzzy analytic hierarchy process (FAHP);

[0308] 4.5) The load optimization module transfers the optimized action parameters and the parameter increment vector to the feedback display module for real-time action correction; this embodiment ensures that the calculation delay does not exceed 4 milliseconds.

[0309] Define the optimized action parameter vector as:

[0310]

[0311] Among them, is the optimized step width; is the optimized knee flexion angle.

[0312] Through the high-speed data interface, transfer the optimized action parameter vector and the parameter increment vector Transmit to the feedback display module to ensure that the transmission delay does not exceed 4 milliseconds;

[0313] The beneficial effects of this embodiment are as follows:

[0314] Innovation in the definition of the cost function: Taking the maximum absolute value of the lower limb joint torque vector as the cost function, directly optimizing for the most unfavorable situation of joint load, ensuring the overall reduction of joint torque, and enhancing the pertinence and effectiveness of the optimization effect.

[0315] Real-time optimization and low-latency transmission: By strictly controlling the delay of optimization calculation and data transmission, ensuring that the full-process delay does not exceed 4 milliseconds, real-time action correction is achieved. Compared with the relatively high latency in the prior art, the real-time performance and response speed of the system are significantly improved.

[0316] Rolling window and peak extraction mechanism: Introducing a 200-millisecond rolling window processing and peak vector calculation method to ensure that the optimization target is based on the recent motion state, enhancing the dynamic adaptability and accuracy of the optimization process.

[0317] Smoothing control of the parameter increment vector: By setting the parameter increment threshold, avoiding excessive parameter adjustment, ensuring the smoothness of the action optimization process and the naturalness of human actions, different from the possible mutation optimization problems in the prior art.

[0318] The following description is made for this embodiment:

[0319] The sprint load optimization experiment was carried out on a standard track and field training ground. Six national-level young sprint athletes were selected as experimental subjects (numbers: Zhang San, Li Si, Wang Wu, Zhao Liu, Sun Qi, Zhou Ba). Each person had completed a 30m sprint warm-up at maximum effort. Before the experiment, the height, weight, and lower limb length of each athlete were measured, and the average weight was recorded as 75 kg and the lower limb length as 0.92 m. The instruments used in the experiment included: a three-dimensional motion capture system (sampling rate 1 kHz, positioning error <0.5 mm), a ground reaction force sensing platform (sampling rate 1 kHz, error <5 N), a six-axis inertial measurement unit (sampling rate 1 kHz, drift <0.1° / s), and a computing platform (Intel i7, 16GB RAM). The experiment was carried out under the conditions of a temperature of 22°C and a humidity of 50%, with uniform illumination.

[0320] The experimental steps are as follows:

[0321] First, start the data synchronization and fusion module to obtain the three-dimensional coordinate sequence of the lower limb key points, joint velocity, ground reaction force, and inertial data of each athlete during the sprint, and generate the initial value vector of sports action correction through Kalman filtering Subsequently, it is input into the biomechanical solution module to calculate the lower limb joint moment vector by the Lagrangian method Record the initial maximum joint moments: Hip joint: 120–140 N·m, knee joint: 220–245 N·m, ankle joint: 105–120 N·m.

[0322] On this basis, the load optimization module is called, and its learning rate , first moment parameter reads and calculates the peak vector within the rolling window .

[0323] The optimization objective function takes the maximum value of the three-joint peak, and performs 40 iterations through the Adam algorithm. Each iteration re-simulates the sprint gait and recalculates the new , with the stride width and the knee flexion angle as adjustable parameters. The finite difference method estimates the gradient, and the convergence threshold is set to ;

[0324] Based on the initial parameters , it is iterated to the optimal motion parameters : The stride width is adjusted to the range of 0.92–1.05 m, and the knee flexion angle is adjusted to the range of 38–52°; the corresponding maximum joint moments after optimization decrease significantly: the hip joint decreases to 102–118 N·m, the knee joint decreases to 190–215 N·m, and the ankle joint decreases to 90–105 N·m. The calculation delay for each athlete in the entire optimization process is ≤4 ms, and the parameter increment vector is output in real time, and the results are sent to the coach's terminal and the athlete's wearable display through wireless communication.

[0325] The above implementation process fully verifies the stability and real-time performance of the load optimization module in the real sprint scenario. Through Quantitative data display: The average maximum joint moment is reduced by 15.8%, and the optimized gait parameters are more in line with the human mechanics characteristics, thus effectively reducing the risk of sports injuries, with significant innovation and practical value.

[0326] The data table of the embodiment is as follows:

[0327]

[0328] The above table data clearly shows the optimization effect of the load optimization module in actual sprint sports: The stride and flexion angle parameters are reasonably adjusted, and the peak values of the three-joint moments all decrease significantly, with an average reduction rate of about 15%.

[0329] Example 6:

[0330] Feedback display module: It is used to receive the sports action parameter vectors after correcting the key points of each lower limb, and output action adjustment instructions to the wearable display terminal in real time.

[0331] Further explanation: The feedback display module guides the athlete to adjust the stride and knee flexion angle in real time through the head-up display (HUD) and vibration feedback, forming a closed-loop action correction mechanism;

[0332] Map the stride increment to the arrow length on the head-up display (HUD), indicating the direction and amplitude of stride adjustment;

[0333] Map the change in knee flexion angle to the angle scale on the HUD, prompting the angle that the knee joint needs to be adjusted;

[0334] Send the parameter increment vector to the HUD glasses worn by the athlete through the UDP protocol to ensure the real-time and stability of data transmission, and maintain a rendering frequency of 90Hz;

[0335] When the stride increment , start the vibration motor fixed at the ilium of the athlete, and prompt the athlete to adjust the stride through vibration; the vibration motor in this example is 100Hz; the number of trigger pulses ; and in this example It is adaptively adjusted by the expert group according to the actual experimental needs, such as using the fuzzy analytic hierarchy process (FAHP), etc., without limitation;

[0336] Subscribe to the receipt signal of the HUD, and detect the sending and receiving status of data packets in real time.

[0337] If a data packet loss is detected, the system will automatically retransmit within 2 frames to ensure the integrity of the feedback instruction.

[0338] The overall closed-loop delay of the system is controlled within 42 milliseconds to meet the real-time action correction requirements.

[0339] The feedback display module passes the action parameter increment vector generated by the load optimization module to the athlete in real time through the HUD and vibration feedback devices. Through the dual feedback of vision and touch, the athlete can immediately understand and adjust the stride width and knee flexion angle, realize the precise correction and optimization of the action, and form a complete closed-loop system.

[0340] It should be noted that: All calculation formulas in this application document adopt regression analysis including but not limited to machine learning algorithms to deeply analyze the collected relevant parameters, identify their natural trends and interrelationships. Using professional software, such as the Scikit-learn library in Python or the R language, a mathematical model matching the data is automatically generated. Then, the performance of the model is objectively evaluated through methods such as cross-validation, and combined with continuous feedback and optimization to ensure that the created formula truly reflects the internal laws of the data, thereby ensuring its effectiveness and accuracy. In all calculation formulas of this application, the parameters in each formula are processed by dimensionless normalization within a consistent range to ensure that different physical quantities are compared on the same scale; the technical means of dimensionless normalization include but are not limited to Min-Max Normalization and Z-Score standardization;

[0341] The technical solution of the present invention can be embodied in the form of a software product in essence or in the part that contributes to the prior art. This computer software product can be stored in a computer-readable storage medium, such as a floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disc of a computer, etc., including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of various embodiments of the present invention.

[0342] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus or device), or used in combination with these instruction execution systems, apparatus or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus or device.

[0343] The description of the above embodiments is only used to help understand the method and its core idea of the present invention. It should be noted that for those of ordinary skill in the art of this technology, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A sports movement correction and monitoring system based on image recognition and big data, characterized in that: Specifically include: Data acquisition module: used to collect multi-view synchronized image streams, three-dimensional ground reaction force data and six-axis inertial measurement data of the target athlete performing sprinting during the monitoring period in real time; 3D posture estimation module: used to receive multi-view synchronized image streams and extract the 3D coordinate sequence of multiple lower limb key points of the target athlete; Data synchronization and fusion module: used to fuse the three-dimensional coordinate sequence, three-dimensional ground reaction force data, and six-axis inertial measurement data to output the initial value vector of sports action correction for multiple key points of the lower limbs; Biomechanics solution module: used to calculate the torque vector of the lower limb joints of the target athlete during sprinting during the monitoring period based on the Lagrange equation; Load optimization module: used to receive the lower limb joint torque vector of the target athlete during the sprinting during the monitoring period, optimize the initial value vector of the sports action correction, and output the corrected sports action parameter vector of each lower limb key point; Feedback display module: used to receive the corrected sports movement parameter vectors of each lower limb key point, and output movement adjustment instructions to the wearable display terminal in real time.

2. The sports movement correction and monitoring system based on image recognition and big data according to claim 1, characterized in that: The multi-view synchronized image streams are synchronized with the global shutter via a hardware synchronizer; The 3D posture estimation module uses the HRNet-W48 deep learning network as the backbone network and combines a differentiable triangulation algorithm to calculate the 3D coordinates of the key points of the lower limbs; The 3D posture estimation module includes a camera calibration submodule for obtaining the intrinsic parameter matrices of the three cameras. With the external parameter matrix ; The 3D posture estimation module further includes a 2D key point detection submodule for outputting the 2D pixel coordinates of 6 lower limb key points in each frame image. ;in represents six lower limb key points, c=1,2,3 represents three cameras; The 3D posture estimation module includes a differentiable triangulation layer for calculating the 3D coordinates of the six lower limb key points based on the camera projection matrix and the 2D key point pixel coordinates. .

3. The sports movement correction and monitoring system based on image recognition and big data according to claim 2, characterized in that: The three-directional ground reaction force data specifically include: The three-way force vector of the three-way ground reaction force data is set as ; in, is the three-way ground reaction force vector; It is the horizontal front-back / left-right component; is the vertical support force; The six-axis inertial measurement data specifically includes: The six-axis inertial measurement data is characterized as follows; ； in, , , They are the attitude quaternion, linear acceleration and angular velocity of the six-axis inertial measurement data at time t; It is used to characterize the direction or orientation of the lower limbs of multiple key points of the target athlete in space, that is, the rotation state of the object relative to the defined coordinate system; Attitude Quaternion Used to represent the IMU posture of multiple lower limb key points at time t; Linear acceleration Used to represent the three-axis acceleration of multiple key points of the lower limbs at time t; Angular velocity Used to represent the three-axis angular velocity of multiple lower limb key points at time t.

4. The sports movement correction and monitoring system based on image recognition and big data according to claim 3 is characterized in that: The data synchronization and fusion module uses the Kalman filter algorithm to integrate the three-dimensional coordinates of the six lower limb key points ,speed , three-way ground reaction force And six-axis inertial measurement data Perform data fusion to generate a unified initial value vector for sports action correction Specific: The data synchronization and fusion module sets a unified timestamp through the GPS-PPS benchmark ; The data synchronization and fusion module uses the rigid body transformation matrix and translation vectors , the three-dimensional coordinates of the six lower limb key points Mapped to the laboratory coordinate system to achieve three-dimensional ground reaction force and six-axis inertial measurement data A unified spatial benchmark; Establishing Kalman filter model: Setting the state transfer matrix of Kalman filter and the observation matrix ; Using the defined Kalman filter model, input the initial value vector of sports action correction and the observation vector , perform prediction and update steps; Set the updated state vector as the initial value vector of the sports action correction at the current moment .

5. The sports movement correction and monitoring system based on image recognition and big data according to claim 4 is characterized in that: The biomechanics solution module is based on the Lagrange equation, combined with the mass-inertia matrix , Coriolis term and centrifugal force term , Gravity , calculate the lower limb joint torque vector ; The biomechanical solution module establishes a 7-degree-of-freedom lower limb biomechanical model, including a 3-degree-of-freedom hip joint, a 1-degree-of-freedom knee joint, and a 3-degree-of-freedom ankle joint; The biomechanical solution module sets the mass parameters of each segment of the lower limb , of which the ankle mass , knee joint quality , hip joint quality ; Quality parameters The "j" in the letter represents the ankle, knee and hip. The biomechanical solution module describes the dynamic equations of each segment of the lower limb by combining the Lagrange equation with the difference L=TV between kinetic energy T and potential energy V; The lower limb joint torque vector is defined as: ； in, is the hip joint torque; is the knee joint torque; is the ankle joint torque.

6. The sports movement correction and monitoring system based on image recognition and big data according to claim 5 is characterized in that: The specific logic of outputting the corrected sports action parameter vectors of each lower limb key point includes: The load optimization module reads the lower limb joint torque vector sequence within a preset time period and calculates the peak vector of each joint torque ; The load optimization module defines a cost function is the lower limb joint torque vector The function of the maximum absolute value in is as follows: ； are adjustable movement parameters, representing stride width and knee flexion angle respectively; The load optimization module uses the Adam optimization algorithm to adjust the action parameters in multiple iterations. , to minimize the cost function ; and the optimized action parameter vector is recorded as .

7. The sports movement correction and monitoring system based on image recognition and big data according to claim 6 is characterized in that: The load optimization module calculates the difference between the action parameters at the current moment and the previous moment to form a parameter increment vector ; Calculate the difference between the optimized action parameter vector and the parameter vector at the previous moment: ； in, is the optimized action parameter vector; is the action parameter vector of the previous moment; The load optimization module optimizes the action parameters and parameter increment vector Transmitted to the feedback display module for real-time action correction; Define the optimized action parameter vector for: ； in, is the optimized stride width; The optimized knee flexion angle.

8. The sports movement correction and monitoring system based on image recognition and big data according to claim 7 is characterized in that: The feedback display module guides the athlete to adjust the stride and knee flexion angle in real time through the head-up display HUD and vibration feedback, forming a closed-loop action correction mechanism; Increment the stride length Mapped to the length of the arrow on the head-up display (HUD), indicating the direction and magnitude of the stride adjustment; Change in knee flexion angle Mapped as an angle scale on the HUD, indicating the angle that the knee joint needs to be adjusted; When the stride increment When the athlete steps on the saddle, the vibration motor fixed to the athlete's ilium is activated, and the athlete is prompted to adjust his stride through vibration.

Images