Ice skate arc type control method based on ai vision combined with athlete data

Abstract

This invention relates to the field of ice skate curvature control technology, and more particularly to an ice skate curvature classification control method based on AI vision combined with athlete data. By acquiring multi-view images and inertial data of the athlete during skating, the method analyzes the joint motion trajectory and push-off torque distribution. Temporal dependencies are captured using deep neural network multimodal encoding, and discrete curvature requirement classifications are obtained using manifold learning. Based on the classifications, optimal configuration ranges, and real-time plantar pressure data, a joint optimization solution for blade curvature and contact pressure is searched through variational inference. Control commands are dynamically generated to drive the adjustment mechanism, and the model is updated using biomechanical feedback data. This solution achieves personalized adaptive adjustment of the ice skate curvature, continuously optimizing skating efficiency and athlete performance.

Description

Technical Field

[0001] This invention relates to the field of ice skate curvature control technology, and in particular to an ice skate curvature classification control method based on AI vision combined with athlete data. Background Technology

[0002] Ice skate curvature control is a crucial technical parameter in ice sports such as speed skating and figure skating, directly affecting an athlete's push-off efficiency, skating stability, and turning agility. Currently, ice skate curvature adjustment typically relies on the subjective experience of athletes or coaches, finding a relatively suitable shape through repeated trial runs and manual adjustments to the blade curve. This trial-and-error method requires multiple on-ice experiences, is time-consuming, and struggles to maintain consistency. Some high-end ice skates use preset fixed curvature templates, initially selected based on the athlete's weight, foot type, or skating style; however, once the template is chosen, it cannot be changed during exercise. Additionally, there have been attempts to calculate the optimal curvature using simple mechanical models, such as approximate solutions based on static plantar pressure distribution or peak single push-off force. However, these methods ignore the continuous impact of dynamic changes in joint angles and torques during skating on the required curvature.

[0003] The conventional approach suffers from two major drawbacks. Firstly, the blade curvature is severely disconnected from the athlete's real-time motion. The required blade curvature varies drastically across different phases of skating (such as the start, curves, and acceleration on straightaways), and existing methods cannot adaptively adjust based on dynamic parameters like push-off torque distribution and joint trajectory. This results in the same curvature potentially being either overly efficient or underefficient at different skating stages. Secondly, the degree of individualized matching is low. Athletes exhibit significant differences in their push-off patterns, such as push-off angles and peak torque timing distributions. Traditional methods, relying on static measurements or empirical formulas, struggle to capture these high-level motion characteristics and cannot generate personalized curvature adjustment strategies for each athlete, thus limiting the potential for performance improvement. Summary of the Invention

[0004] This invention provides a method for controlling the arc of ice skates based on AI vision combined with athlete data, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for controlling the arc of an ice skate based on AI vision combined with athlete data, comprising: The system acquires multi-view image sequences and inertial measurement data of athletes gliding on the ice. After spatial registration and temporal alignment of the multi-view image sequences, a three-dimensional posture sequence is extracted. The joint motion trajectory and push-off torque distribution are analyzed by combining the acceleration and angular velocity components in the inertial measurement data. The joint motion trajectory and the push-off torque distribution are encoded using a deep neural network to capture the temporal dependencies within the gliding cycle, construct a latent space representation of the athlete's individual push-off pattern, and map the continuous motion state into discrete arc demand subtypes in the latent space through manifold learning. Based on the arc requirement classification and its corresponding optimal configuration range, combined with the real-time collected plantar pressure distribution data, the joint optimization solution of the cutting edge curvature distribution function and the contact pressure distribution function is searched within the optimal configuration range through variational inference, and control commands for the cutting edge geometry are dynamically generated. The control command is sent to the ice skate curvature adjustment mechanism for adjustment, and the adjusted biomechanical feedback data is collected. The biomechanical feedback data is then used to jointly update the encoding weights of the deep neural network and the hyperparameters of the variational inference.

[0006] After spatial registration and temporal alignment of the multi-view image sequence, a three-dimensional pose sequence is extracted. The joint motion trajectory and push-off torque distribution are then analyzed by combining the acceleration and angular velocity components from the inertial measurement data. Based on polar geometry constraints, cross-view feature point matching is performed on multi-view image sequences. The joint estimation of camera extrinsic matrix and feature point 3D coordinates is optimized by bundle adjustment, and spatial registration results under a unified world coordinate system are constructed. Based on the timestamp information in the inertial measurement data, the feature point sequence in the spatial registration result is interpolated and synchronized, and a three-dimensional attitude sequence is reconstructed from the feature point sequence based on the forward kinematic solution. The joint angle change rate in the three-dimensional posture sequence is fused with the angular velocity component in the inertial measurement data using Kalman filtering to eliminate measurement noise and drift error, thus obtaining the joint motion trajectory. An inverse dynamics solution framework based on the Lagrange dynamics equations is established. The angular acceleration in the joint motion trajectory and the acceleration components in the inertial measurement data are substituted into the motion equations of the inverse dynamics solution framework. The time series of torques borne by each joint during the ice-pushing phase is obtained by recursive calculation of the joint torques, forming the ice-pushing torque distribution.

[0007] The joint angle change rate in the three-dimensional posture sequence is fused with the angular velocity component in the inertial measurement data using Kalman filtering to eliminate measurement noise and drift error, resulting in the joint motion trajectory, which includes: A state-space representation is constructed, defining the rate of change of joint angles as the state vector and the angular velocity components as the observation vector. The state transition matrix and process noise covariance are set according to the biomechanical constraints of human joint motion. Based on the joint angles of the previous frame of the three-dimensional pose sequence, state prediction is performed through the state transition matrix to obtain a prior estimate of the joint angle change rate of the current frame. Calculate the residual between the prior estimate and the angular velocity component, and calculate the Kalman gain based on the process noise covariance and the measurement noise covariance; The prior estimate is corrected using the Kalman gain to obtain the fused posterior estimate of the joint angle change rate, and the joint angle time series is reconstructed by time integration of the posterior estimate of the joint angle change rate. The joint angle time series is smoothed by cubic spline interpolation to eliminate high-frequency noise components and obtain the joint motion trajectory.

[0008] Multimodal feature encoding of the joint motion trajectory and the push-off torque distribution is performed using a deep neural network to capture the temporal dependencies within the gliding cycle, and a latent spatial representation characterizing the athlete's individual push-off pattern is constructed, including: The joint motion trajectory and the distribution of the pushing force are respectively input into two parallel encoder branches. Each encoder branch uses a multi-layer convolutional structure to extract local temporal features and reduces the feature dimension through pooling operations. A cross-modal attention module is introduced into the output layer of the encoder branch to calculate the correlation weight matrix between the joint motion trajectory features and the push-off torque distribution features, and to perform weighted fusion of the feature vectors of the two modalities based on the correlation weight matrix to obtain multimodal fusion features; The multimodal fusion features are input into the recurrent neural network layer, and the dependencies between features at different times within the gliding cycle are captured through the temporal propagation of the hidden states. The global temporal feature vector is extracted at the last time step of the recurrent neural network layer. The global temporal feature vector is input into the encoder of the variational autoencoder, and the global temporal feature vector is mapped to a latent space representation that follows a specific distribution through the reparameterization technique.

[0009] Based on the aforementioned arc requirement classification and its corresponding optimal configuration range, and combined with real-time collected plantar pressure distribution data, a joint optimal solution for the cutting edge curvature distribution function and the contact pressure distribution function is searched within the optimal configuration range through variational inference. Control commands for dynamically generating the cutting edge geometry include: Based on the parameter configuration table preset by the arc requirement fractal index, obtain the upper and lower bounds of the cutting edge curvature radius corresponding to the current fractal to form the optimal configuration range; The real-time collected plantar pressure distribution data is spatially discretized, dividing the plantar area into the forefoot, midfoot, and heel regions. The pressure peak and pressure center locations of each region are extracted to construct a pressure distribution feature vector. Establish the coupled physical relationship between the cutting edge curvature distribution function and the contact pressure distribution function, express the coupled physical relationship in the form of a probability graph, and iteratively optimize the lower bound of evidence within the optimal configuration interval through variational inference to solve the posterior distribution of the latent variable that maximizes the likelihood of the observed data. The optimal cutting edge curvature distribution parameters are sampled from the latent variable posterior distribution, and the corresponding contact pressure distribution is calculated according to the coupled physical relationship. The cutting edge curvature distribution parameters are then converted into control commands containing curvature adjustment amounts at multiple control points.

[0010] A coupled physical relationship is established between the cutting edge curvature distribution function and the contact pressure distribution function. This coupled physical relationship is expressed as a probabilistic graphical form. The lower bound of evidence is iteratively optimized within the optimal configuration interval through variational inference. The posterior distribution of the latent variables that maximizes the likelihood of the observed data is solved, including: Based on the principle of elastic contact mechanics, the curvature parameter vector of the cutting edge curvature distribution function is mapped to the spatial distribution shape of the contact pressure distribution function through the stress transfer model, and a differential equation expression of the coupled physical relationship is established. The differential equation is transformed into a factor graph structure to form a probability graph. A variational posterior distribution is introduced into the probability graph to approximate the latent variable nodes. The lower bound of evidence is decomposed into the log-likelihood term of the plantar pressure distribution data and the divergence term of the latent variable prior distribution and the variational posterior distribution. Within the optimal configuration interval, the hyperparameters of the variational posterior distribution are iteratively updated through gradient ascent, making the lower bound of evidence monotonically increasing. The iteration is terminated when the increment of the lower bound of evidence is lower than the convergence threshold. The posterior expectation of the latent variables is extracted from the converged variational posterior distribution, and the latent variable posterior distribution that maximizes the likelihood of the observed data is solved.

[0011] A second aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0012] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0013] Based directly on spatiotemporal registration of multi-view images and inertial data, combined with the 3D posture sequence derived from acceleration and angular velocity components, the system can accurately capture the athlete's joint motion trajectory and push-off torque distribution. Deep neural networks encode the multimodal features of joint motion trajectory and push-off torque distribution, capturing the temporal dependencies within the gliding cycle and establishing a latent spatial representation reflecting personalized push-off patterns. By mapping continuous motion states to discrete arc requirement classifications through manifold learning, the system achieves accurate classification and dynamic recognition of blade arc requirements, avoiding the subjective bias of traditional experience-based adjustments and significantly improving the objectivity and adaptability of arc classification.

[0014] Guided by the arc requirement classification and its optimal configuration range, and utilizing real-time collected plantar pressure distribution data, variational inference searches for a joint optimization solution of the blade curvature distribution function and the contact pressure distribution function within the optimal configuration range, dynamically generating control commands for the blade geometry. This process achieves instantaneous matching between the blade arc and the athlete's real-time push-off dynamics, optimizing the contact pressure distribution between the blade and the ice surface and reducing ineffective energy loss during the push-off transition. After the dynamic commands are sent to the blade arc adjustment mechanism, the blade shape can change in real time following the movement state, thereby maximizing the propulsion efficiency of each push-off, reducing gliding resistance, and extending efficient gliding time. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the ice skate arc division control method based on AI vision combined with athlete data, according to an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0018] Figure 1 This is a flowchart illustrating the ice skate arc division control method based on AI vision combined with athlete data, according to an embodiment of the present invention.

[0019] The ice skate radius control method based on AI vision combined with athlete data includes: The system acquires multi-view image sequences and inertial measurement data of athletes gliding on the ice. After spatial registration and temporal alignment of the multi-view image sequences, a three-dimensional posture sequence is extracted. The joint motion trajectory and push-off torque distribution are analyzed by combining the acceleration and angular velocity components in the inertial measurement data. The joint motion trajectory and the push-off torque distribution are encoded using a deep neural network to capture the temporal dependencies within the gliding cycle, construct a latent space representation of the athlete's individual push-off pattern, and map the continuous motion state into discrete arc demand subtypes in the latent space through manifold learning. Based on the arc requirement classification and its corresponding optimal configuration range, combined with the real-time collected plantar pressure distribution data, the joint optimization solution of the cutting edge curvature distribution function and the contact pressure distribution function is searched within the optimal configuration range through variational inference, and control commands for the cutting edge geometry are dynamically generated. The control command is sent to the ice skate curvature adjustment mechanism for adjustment, and the adjusted biomechanical feedback data is collected. The biomechanical feedback data is then used to jointly update the encoding weights of the deep neural network and the hyperparameters of the variational inference.

[0020] In one optional implementation, after spatial registration and temporal alignment of the multi-view image sequence, a three-dimensional pose sequence is extracted, and the joint motion trajectory and push-off torque distribution are analyzed by combining the acceleration and angular velocity components in the inertial measurement data, including: Based on polar geometry constraints, cross-view feature point matching is performed on multi-view image sequences. The joint estimation of camera extrinsic matrix and feature point 3D coordinates is optimized by bundle adjustment, and spatial registration results under a unified world coordinate system are constructed. Based on the timestamp information in the inertial measurement data, the feature point sequence in the spatial registration result is interpolated and synchronized, and a three-dimensional attitude sequence is reconstructed from the feature point sequence based on the forward kinematic solution. The joint angle change rate in the three-dimensional posture sequence is fused with the angular velocity component in the inertial measurement data using Kalman filtering to eliminate measurement noise and drift error, thus obtaining the joint motion trajectory. An inverse dynamics solution framework based on the Lagrange dynamics equations is established. The angular acceleration in the joint motion trajectory and the acceleration components in the inertial measurement data are substituted into the motion equations of the inverse dynamics solution framework. The time series of torques borne by each joint during the ice-pushing phase is obtained by recursive calculation of the joint torques, forming the ice-pushing torque distribution.

[0021] After acquiring multi-view image sequences of the athlete gliding, the spatial inconsistency between different cameras needs to be addressed. Based on epipolar geometry constraints, stable feature points such as SIFT or SuperPoint are extracted from image frames from different viewpoints, and cross-viewpoint feature point correspondences are established through fundamental matrix estimation. Specifically, for any two viewpoints, epipolar constraints are used to compress the feature point matching search range from the two-dimensional image plane to the one-dimensional epipolar line, thereby significantly reducing the mismatch rate. After initially establishing the feature point correspondences, bundle adjustment is used to jointly optimize the camera extrinsic matrix and the three-dimensional coordinates of the feature points. The reprojection error is used as the objective function for iterative minimization, ultimately obtaining a high-precision spatial registration result in a unified world coordinate system. During the optimization of the camera extrinsic matrix, the rotation matrix and translation vector of each camera need to be estimated simultaneously to ensure that image information from all viewpoints is uniformly mapped to the same three-dimensional space, eliminating geometric inconsistencies between multiple viewpoints.

[0022] Because multiple cameras and inertial measurement units (IMUs) have their own independent clocks when acquiring data, timestamps between different sensors often have millisecond-level deviations. Based on the timestamp information recorded in the IMU data, cubic spline interpolation is performed on the feature point sequence in the spatial registration result to complete the 3D coordinates of the feature points on the time axis to the same sampling frequency as the IMU data, achieving temporal synchronization. After time alignment, a 3D posture sequence is reconstructed from the feature point sequence based on a forward kinematics solution framework. Specifically, using the 3D coordinates of key skeletal nodes such as the hip, knee, and ankle joints as input, and utilizing predefined bone segment length constraints in the human skeletal model, the 3D rotation angles of each joint are calculated step-by-step through chained coordinate transformations to obtain a complete 3D posture sequence. This posture sequence is expressed as a time series of joint angles, covering motion parameters that directly affect arc control, such as the hip abduction angle, knee flexion-extension angle, and ankle plantar flexion angle during the skating cycle.

[0023] The rate of change of joint angles extracted from 3D posture sequences is inevitably affected by image feature point localization errors in practical applications. While the angular velocity component output by the inertial measurement unit (IMU) has the advantage of high-frequency response, long-term integration introduces drift errors. Fusing these two methods using Kalman filtering can simultaneously utilize the absolute accuracy advantage of visual information and the high-frequency dynamic advantage of IMU. The state vector of the Kalman filter contains the angles and angular velocities of each joint. The prediction step relies on the angular velocity component in the IMU data for state propagation, while the update step uses the joint angles in the 3D posture sequence as observations to correct the predicted state. The process noise covariance matrix reflects the drift characteristics of the IMU, while the observation noise covariance matrix reflects the localization accuracy of visual reconstruction. The appropriate settings of both directly determine the quality of the fusion result. After Kalman filtering fusion, measurement noise and integration drift errors are effectively suppressed, outputting smooth and continuous joint motion trajectories, providing reliable kinematic input for subsequent inverse dynamics analysis.

[0024] Based on obtaining high-quality joint motion trajectories, an inverse dynamics solution framework based on the Lagrange equations is established to analyze the torque distribution borne by each joint during the push-off phase. The Lagrange equations describe the dynamic behavior of the human multibody system using generalized coordinates, and their core form is... ,in For the joint angle generalized coordinate vector, The joint angular velocity vector. The joint angular acceleration vector. The inertia matrix, The matrix of Coriolis force and centrifugal force. The vector of the gravity term. Let be the joint torque vector to be solved.

[0025] The core of solving inverse dynamics lies in: knowing the kinematic quantities , , In reverse, the joint torque required to drive this motion is solved. Joint movement trajectories provide and The time series can be obtained by numerically differentiating the joint angular velocity series. Simultaneously, the acceleration components in the inertial measurement data are used to correct the linear acceleration of the end effector, thereby improving... The estimation accuracy, especially during the high-speed ice-climbing phase, is affected by the large error in angular acceleration obtained solely from visual numerical differentiation. Introducing accelerometer data for fusion correction can effectively improve the accuracy of inverse dynamics solutions.

[0026] In the human skeletal model, the mass, center of mass position, and moment of inertia parameters of each bone segment are individually scaled according to the athlete's weight and height using standard anthropometric proportions to ensure the inertia matrix. The model is constructed to match the actual body shape of athletes. Inverse dynamics calculations employ a recursive Newton-Euler method, progressively transferring force and torque from the distal limbs to the proximal end, resulting in high computational efficiency and suitability for real-time processing. Finally, for each moment of the push-off phase, the torque values ​​of the hip, knee, and ankle joints are output, forming a complete push-off torque time series. The torque time series of each joint are summarized within the push-off cycle to obtain a push-off torque distribution that describes the athlete's individualized push-off force exertion pattern. This distribution not only includes peak and mean torque information but also retains the torque variation characteristics at the beginning, middle, and end of the push-off phases. This provides a dynamic input with rich temporal structure for subsequent multimodal feature encoding in deep neural networks, serving as a crucial data foundation for accurate classification of arc requirements.

[0027] In one optional implementation, the joint angle change rate in the three-dimensional attitude sequence is fused with the angular velocity component in the inertial measurement data using Kalman filtering to eliminate measurement noise and drift error, resulting in the joint motion trajectory, which includes: A state-space representation is constructed, defining the rate of change of joint angles as the state vector and the angular velocity components as the observation vector. The state transition matrix and process noise covariance are set according to the biomechanical constraints of human joint motion. Based on the joint angles of the previous frame of the three-dimensional pose sequence, state prediction is performed through the state transition matrix to obtain a prior estimate of the joint angle change rate of the current frame. Calculate the residual between the prior estimate and the angular velocity component, and calculate the Kalman gain based on the process noise covariance and the measurement noise covariance; The prior estimate is corrected using the Kalman gain to obtain the fused posterior estimate of the joint angle change rate, and the joint angle time series is reconstructed by time integration of the posterior estimate of the joint angle change rate. The joint angle time series is smoothed by cubic spline interpolation to eliminate high-frequency noise components and obtain the joint motion trajectory.

[0028] After obtaining the 3D attitude sequence from multi-view image analysis and the angular velocity components acquired by the inertial measurement unit (IMU), the two types of data need to be accurately fused to eliminate their respective measurement errors. The joint angle change rate obtained from visual analysis is limited by the image frame rate, resulting in insufficient temporal resolution and estimation bias caused by occlusion. Although the angular velocity components from the IMU have high temporal resolution, long-term integration introduces cumulative drift errors. The Kalman filter fusion mechanism can establish an optimal linear unbiased estimate between the two types of signals, complementing each other's error characteristics to obtain a high-precision joint motion trajectory.

[0029] When constructing the state-space representation, the rate of change of joint angles is defined as the state vector. subscript This represents the discrete time step index. The dimension of the state vector corresponds to the number of degrees of freedom of the tracked joints. For key lower limb joints of ice skaters (including the three-dimensional rotational degrees of freedom of the hip, knee, and ankle joints), the state vector typically contains multiple components. The angular velocity component output by the inertial measurement unit is defined as the observation vector. State transition matrix Based on the biomechanical constraints of human joint movement, the state transition matrix is ​​designed to approximate an identity matrix, as joint angular velocities exhibit strong continuity between adjacent frames. Damping coefficients related to the gliding period are introduced near the diagonal elements to reflect the natural decay characteristics of joint angular velocities during the push-off phase. (Process noise covariance matrix...) Each element is initialized based on the biomechanical range of motion and intensity of movement of different joints of the athlete. The process noise variance corresponding to the hip joint is relatively large, while the process noise variance corresponding to the ankle joint is relatively small, in order to match the actual motion characteristics of each joint during the ice pushing process.

[0030] Based on the joint angles of the previous frame of the 3D pose sequence Through the state transition matrix Perform state prediction to obtain a prior estimate of the joint angle change rate in the current frame. The calculation method is as follows Meanwhile, the prior estimated covariance matrix pass To carry out dissemination and updates, among which Let be the posterior covariance matrix of the previous time step. This prediction step relies only on the state information and system dynamics model of the previous frame, and has not yet incorporated the inertial measurement data of the current frame. Therefore, the resulting estimate is a priori.

[0031] Calculate prior estimates Angular velocity observation vector output by the inertial measurement unit The residuals between ,Right now The observation matrix This describes the linear mapping relationship from the state vector to the observation space. Since the rate of change of joint angles directly corresponds to the angular velocity components of the inertial measurement unit in a physical sense, Typically, it is taken as the identity matrix or a rotation matrix after coordinate system alignment. Measurement noise covariance matrix. The Kalman gain, determined based on the device specifications and actual calibration results of the inertial measurement unit, reflects the white noise level of the angular velocity sensor. The calculation is as follows The Kalman gain tends to be larger when process noise is high and measurement noise is low, meaning that more trust is placed in inertial measurement data; conversely, it relies more on state prediction results. This adaptive weighting mechanism ensures that the fusion results maintain optimal estimation performance in different motion phases.

[0032] Using Kalman gain The prior estimate is corrected to obtain the posterior estimate of the fused joint angle change rate. The calculation method is as follows The posterior covariance matrix is ​​updated synchronously to... ,in It is the identity matrix. Posterior estimate. This represents the optimal joint angular velocity estimate for the current frame after fusing visual and inertial data, effectively suppressing drift error and high-frequency noise. The posterior estimation sequence obtained frame by frame is numerically integrated along the time axis, using the initial joint angle as the starting point of integration, to reconstruct the joint angle time series. Integration step size and sampling time interval Consistency. During the integration process, an absolute angle correction from the 3D pose sequence is introduced every fixed number of frames to suppress any residual low-frequency drift during integration and ensure the long-term stability of the joint angle time series.

[0033] The reconstructed joint angle time series Cubic spline interpolation is used for smoothing. Cubic spline interpolation constructs cubic polynomial segments between adjacent data nodes, requiring each segment to satisfy the conditions of continuous function values, continuous first derivative, and continuous second derivative at the nodes. This effectively eliminates high-frequency noise components introduced by integral errors and sampling quantization while preserving the overall trend of joint angle changes. The spacing between interpolation nodes is determined based on the actual sampling frequency and the characteristic frequency of the gliding motion: the push-off period for gliding on ice is typically between 0.8 and 1.5 seconds, and the main frequency of joint angle changes is below 10 Hz. Therefore, the node spacing of cubic spline interpolation is set so that the bandwidth of the smoothed curve is slightly higher than 10 Hz, filtering out noise components above this frequency while preserving effective motion information. After smoothing, the output joint motion trajectory has continuous differentiability in the time dimension and can be directly used as input for subsequent multimodal feature encoding, providing high-quality kinematic feature sequences for deep neural networks.

[0034] In practical engineering implementation, the initialization of the Kalman filter has a significant impact on the fusion effect, and the initial state vector... Take the mean of the joint angular velocity estimates during the stationary phase before the start of gliding, and the initial posterior covariance matrix. Set as a diagonal matrix, with diagonal elements taking values ​​consistent with the measurement variance of each joint in a static state. Process noise covariance matrix. With measurement noise covariance matrix Maximum likelihood estimation can be performed using offline calibration data, or the adaptive Kalman filter mechanism can be dynamically adjusted based on residual statistical characteristics during online operation to adapt to changes in noise characteristics caused by individual differences among athletes and variations in ice surface conditions. The entire fusion process completes two steps—prediction and update—at each sampling time. The computational complexity is proportional to the square of the number of joint degrees of freedom, and it can meet the latency requirements for real-time processing on an embedded computing platform.

[0035] In one optional implementation, multimodal feature encoding of the joint motion trajectory and the push-off torque distribution is performed using a deep neural network to capture the temporal dependencies within the gliding cycle, and a latent spatial representation characterizing the athlete's individual push-off pattern is constructed, including: The joint motion trajectory and the distribution of the pushing force are respectively input into two parallel encoder branches. Each encoder branch uses a multi-layer convolutional structure to extract local temporal features and reduces the feature dimension through pooling operations. A cross-modal attention module is introduced into the output layer of the encoder branch to calculate the correlation weight matrix between the joint motion trajectory features and the push-off torque distribution features, and to perform weighted fusion of the feature vectors of the two modalities based on the correlation weight matrix to obtain multimodal fusion features; The multimodal fusion features are input into the recurrent neural network layer, and the dependencies between features at different times within the gliding cycle are captured through the temporal propagation of the hidden states. The global temporal feature vector is extracted at the last time step of the recurrent neural network layer. The global temporal feature vector is input into the encoder of the variational autoencoder, and the global temporal feature vector is mapped to a latent space representation that follows a specific distribution through the reparameterization technique.

[0036] Joint motion trajectories and push-off moment distributions are treated as two independent input signals, fed into two parallel encoder branches for feature extraction. The joint motion trajectory branch receives the multi-joint angle time series after Kalman filtering smoothing, while the push-off moment distribution branch receives the joint moment time series obtained from inverse dynamics analysis. Each encoder branch employs a multi-layer one-dimensional convolutional structure, with the convolution kernel sliding along the time axis to capture local temporal patterns. Taking the joint motion trajectory branch as an example, suppose the input feature map is at the... The number of channels in a convolutional layer is The kernel size is Then the length of this layer is Perform a causal convolution with a stride of 1 on the input sequence, and keep the output length constant. The sequence is then processed with batch normalization and ReLU activation to stabilize gradient propagation and introduce nonlinear expressive power. Max pooling is inserted after every two convolutional layers, with a pooling window size of 2, compressing the sequence length to half its original size, thereby gradually reducing the feature dimensionality and expanding the receptive field. The ice-push moment distribution branch uses the same number of convolutional layers and pooling strategy, but the number of channels in each layer can be independently configured according to the dimension of the moment signal to adapt to the difference in feature complexity between the two signals. After multiple convolutions and pooling, the two branches output dimensions of... and The local temporal feature tensor, where Output feature dimensions for the joint trajectory branch. Output feature dimensions for the torque distribution branch.

[0037] A cross-modal attention module is introduced into the output layer of both encoder branches to explicitly model the interdependence between joint motion trajectory features and push-off torque distribution features. Let the feature matrix output by the joint motion trajectory branch be... The characteristic matrix output by the ice-pushing moment distribution branch is ,in The number of time steps after pooling. The cross-modal attention modules are respectively... and As the source of queries and key-value pairs, the two feature paths are mapped to a unified attention space through linear projection, with a projection dimension of [missing value]. Relevance weight matrix It is calculated through the scaled dot product attention mechanism, i.e. ,in and These are the projection matrices for the query and the key, respectively. This is a scaling factor used to prevent the gradient from vanishing due to excessively large dot product values. Based on the obtained correlation weight matrix, the value vector of the characteristic ice-climbing moment distribution is... ( We perform weighted aggregation on the value projection matrix to obtain the cross-modal context vector. Then will Features of joint movement trajectory The concatenation is performed along the channel dimension, and then a linear transformation layer is applied to compress the concatenated dimensions to a unified fusion feature dimension. Thus, the multimodal fusion feature matrix is ​​obtained. This fusion process preserves the local temporal features of both modalities and explicitly captures the dynamic influence of torque distribution on joint motion trajectories through attention weights, effectively avoiding the information aliasing problem caused by simple splicing or weighted summation.

[0038] Multimodal fusion feature matrix The data is fed into a recurrent neural network layer to capture long-range dependencies across time steps within the gliding cycle. The recurrent neural network layer employs a multi-layer bidirectional long short-term memory (Bi-LSTM) structure, with each layer containing two LSTM units (forward and backward). The forward unit processes data in chronological order. Each row is processed in reverse chronological order by the backward unit, thus enabling the hidden state at each time step to simultaneously integrate historical and future information. Let the th... Layer Bi-LSTM at time steps The forward hidden state is The backward hidden state is Then the complete hidden state at that time step is ,in For the hidden layer dimension of a unidirectional LSTM, For time step index, This is used for layer indexing. Dropout regularization is added between adjacent Bi-LSTM layers, with a dropout rate set to 0.3 to suppress overfitting. This is done at the last time step of the last Bi-LSTM layer. At this point, the concatenated vector of the forward and backward hidden states is extracted as the global temporal feature vector. ,in denoted as the total number of layers in the Bi-LSTM. This vector, through temporal integration of features from all moments throughout the gliding cycle, carries a global temporal representation of the athlete's complete skating pattern, providing a sufficient information foundation for subsequent latent spatial mapping.

[0039] global temporal feature vector The input to the encoding end of the variational autoencoder is mapped to a latent space representation following a specific distribution using a reparameterization technique. The encoding end consists of two parallel fully connected layers, each outputting a mean vector of the latent variables. With log-variance vector ,in The potential spatial dimension. Both the mean vector and the log-variance vector are transformed linearly from... The calculated logarithmic variance is used to ensure numerical stability. The reparameterization process introduces... Irrelevant random noise vector The latent space representation vector is obtained through The calculation yielded, where , This represents element-wise multiplication. for A 3D identity matrix. This reparameterization operation allows the gradient to be passed through... Backpropagation is used to calculate the mean and variance, thus supporting end-to-end training while maintaining random sampling characteristics. The loss function during training consists of two parts: reconstruction loss and KL divergence regularization term. The KL divergence term constrains the latent space distribution to approach the standard normal prior, preventing discontinuous regions in the latent space and ensuring the smoothness of continuous motion state mapping in subsequent manifold learning. Throughout the multimodal feature encoding process, from parallel convolution branches to cross-modal attention fusion, and then through Bi-LSTM temporal integration and variational coding, the parameters at each stage are jointly optimized through end-to-end backpropagation, enabling the latent space representation to fully capture the deep structural features of the athlete's individualized skating patterns.

[0040] In one optional implementation, based on the arc requirement classification and its corresponding optimal configuration interval, and combined with real-time collected plantar pressure distribution data, a joint optimization solution of the cutting edge curvature distribution function and the contact pressure distribution function is searched within the optimal configuration interval through variational inference, and control commands for dynamically generating the cutting edge geometry are generated, including: Based on the parameter configuration table preset by the arc requirement fractal index, obtain the upper and lower bounds of the cutting edge curvature radius corresponding to the current fractal to form the optimal configuration range; The real-time collected plantar pressure distribution data is spatially discretized, dividing the plantar area into the forefoot, midfoot, and heel regions. The pressure peak and pressure center locations of each region are extracted to construct a pressure distribution feature vector. Establish the coupled physical relationship between the cutting edge curvature distribution function and the contact pressure distribution function, express the coupled physical relationship in the form of a probability graph, and iteratively optimize the lower bound of evidence within the optimal configuration interval through variational inference to solve the posterior distribution of the latent variable that maximizes the likelihood of the observed data. The optimal cutting edge curvature distribution parameters are sampled from the latent variable posterior distribution, and the corresponding contact pressure distribution is calculated according to the coupled physical relationship. The cutting edge curvature distribution parameters are then converted into control commands containing curvature adjustment amounts at multiple control points.

[0041] Based on the parameter configuration table preset by the arc requirement classification index, the upper and lower bounds of the cutting edge curvature radius corresponding to the current classification are obtained, thus forming the optimal configuration range. This parameter configuration table is derived from statistical analysis of a large amount of athlete experimental data during the offline phase, storing reasonable ranges of curvature radii for each arc requirement classification category. Let the category index of the current arc requirement classification be... The corresponding lower bound of the cutting edge radius of curvature is The upper boundary is The optimal configuration interval is defined as follows: In practical applications, the range widths corresponding to different curvature types vary: for skaters who use a strong push-off technique, the curvature radius range is usually smaller to ensure that the contact pressure between the blade and the ice surface meets the high thrust requirements; for skaters who maintain their position, the range is generally larger to reduce friction loss and extend the gliding distance. The index lookup process of the parameter configuration table is completed within milliseconds, ensuring that the real-time performance of the overall control flow is not affected.

[0042] When spatially discretizing the real-time acquired plantar pressure distribution data, the plantar region is divided into three functional zones: the forefoot zone, the midfoot zone, and the heel zone. A pressure sensor array is used to acquire the pressure values ​​at each measurement point on the plantar surface. The sensor array is assumed to contain [number of sensors]. There are [number] measurement points, and the coordinates of each measurement point are [coordinates]. The corresponding pressure sampling value is ,in After classifying all measurement points into three regions based on anatomical boundaries, the peak pressure and pressure center location within each region are extracted. For a given region... ( Peak pressure is defined as the maximum pressure value at all measurement points within the zone. The location of the pressure center was calculated using a weighted average, with the horizontal axis being... The vertical axis is The pressure peak values ​​and pressure center coordinates of the three zones are concatenated to construct a pressure distribution feature vector. This vector contains nine components, corresponding to the two-dimensional coordinates of the peak pressure and pressure center in each of the three zones. The pressure distribution feature vector serves as a core representation of the observed data in subsequent variational inference, directly affecting the solution for the posterior distribution of the latent variables.

[0043] Establishing the coupled physical relationship between the blade edge curvature distribution function and the contact pressure distribution function is the theoretical foundation for joint optimization solutions. The contact pressure distribution of the blade edge on the ice surface is jointly determined by the blade edge's geometric curvature and the athlete's body weight load: in the local contact area, the smaller the radius of curvature, the smaller the unit contact area, and the higher the contact pressure; the larger the radius of curvature, the larger the contact area, and the lower the pressure. Let the curvature distribution function along the blade edge arc length be... ,in Let the coordinates be arc length. The total arc length of the cutting edge; the corresponding contact pressure distribution function is: According to Hertzian contact mechanics, the local contact pressure and curvature satisfy a nonlinear coupling relationship, which can be expressed as: ,in The normal load applied to the skater's blade. This is a nonlinear operator determined by the contact mechanics equations. The above coupling relationship is expressed as a probabilistic graphical form: using the edge curvature distribution parameters... As latent variables, the pressure distribution feature vector For the observed variables, establish a generative model. The prior distribution From the optimal configuration range Determined, conditional likelihood By coupling physical relationships Establish.

[0044] By using variational inference to iteratively optimize the lower bound of evidence (ELBO) within the optimal configuration interval, the posterior distribution of latent variables that maximizes the likelihood of the observed data is solved. A variational distribution is introduced. Approximate True Posterior ,in Let these be variational parameters. The lower bound of evidence is defined as... The first term is the log-likelihood of the reconstruction expectation, and the second term is the KL divergence penalty term between the variational distribution and the prior distribution. During the iterative optimization process, the gradient ascent method is used to... Update the parameters, using reparameterization techniques in each iteration. Monte Carlo sampling is performed to estimate the gradient. To ensure that the optimization results strictly fall within the optimal configuration interval, a truncation constraint is imposed on the support domain of the variational distribution. Set as a truncated normal distribution, the cutoff boundaries correspond to... and The iterative process continues until the change in the lower bound of evidence between two consecutive iterations is less than a preset convergence threshold. The process terminates when the maximum number of iterations is reached. The entire variational inference process runs online on the embedded computing unit, and the inference cycle is synchronized with the athlete's gliding gait cycle to ensure the timeliness of control commands.

[0045] From the converged latent variable posterior distribution Mid-sampling, taking the posterior mean As the optimal cutting edge curvature distribution parameter. Based on the coupled physical relationship Calculate the corresponding contact pressure distribution Verify its correlation with the characteristic vector of plantar pressure distribution. Consistency. The optimal cutting edge curvature distribution parameters... Converted into a control command containing multiple control point curvature adjustment values: uniformly set along the cutting edge arc length direction. There are [number] control points, with control point indices as follows: , No. Target curvature value at each control point Depend on Interpolation yields the curvature adjustment amount, defined as the difference between the target curvature and the current measured curvature. The difference, that is Control commands are sent to the skate blade curvature adjustment mechanism in the form of structured data packets. These packets contain the position coordinates of each control point, the target curvature value, and the adjustment amount. The adjustment mechanism drives the corresponding fine-tuning actuators to correct the curvature based on the adjustment amount. After the control commands are generated, the kinematic state and pressure data at the current moment are recorded simultaneously as input for subsequent biomechanical feedback updates, forming a closed-loop adaptive control process.

[0046] In one optional implementation, a coupled physical relationship is established between the cutting edge curvature distribution function and the contact pressure distribution function. This coupled physical relationship is represented as a probabilistic graphical form. Variational inference is used to iteratively optimize the lower bound of evidence within the optimal configuration interval. The posterior distribution of the latent variables that maximizes the likelihood of the observed data is solved, including: Based on the principle of elastic contact mechanics, the curvature parameter vector of the cutting edge curvature distribution function is mapped to the spatial distribution shape of the contact pressure distribution function through the stress transfer model, and a differential equation expression of the coupled physical relationship is established. The differential equation is transformed into a factor graph structure to form a probability graph. A variational posterior distribution is introduced into the probability graph to approximate the latent variable nodes. The lower bound of evidence is decomposed into the log-likelihood term of the plantar pressure distribution data and the divergence term of the latent variable prior distribution and the variational posterior distribution. Within the optimal configuration interval, the hyperparameters of the variational posterior distribution are iteratively updated through gradient ascent, making the lower bound of evidence monotonically increasing. The iteration is terminated when the increment of the lower bound of evidence is lower than the convergence threshold. The posterior expectation of the latent variables is extracted from the converged variational posterior distribution, and the latent variable posterior distribution that maximizes the likelihood of the observed data is solved.

[0047] Based on the principles of elastic contact mechanics, the contact behavior between the ice skate blade and the ice surface can be described by Hertz contact theory and its extended forms. The curvature distribution along the arc length of the blade is not uniform; the curvature value at different locations determines the degree of local stress concentration when in contact with the ice surface. By introducing the blade curvature parameter vector into the stress transfer model, a set of differential equations describing the coupling relationship between curvature and contact pressure can be obtained. Specifically, let the blade arc length coordinates be... The curvature distribution function along the cutting edge direction is The contact pressure distribution function is The two are related through the fundamental equation of elasticity: the local contact pressure is affected by the curvature at that point, the elastic modulus of the cutting edge material, the elastic response of the ice surface, and the normal load. The common constraints. In the continuous domain, this coupling relationship can be expressed as a mixed differential equation containing the second derivative of curvature and the integral of pressure, with boundary conditions given by the free end constraints at both ends of the cutting edge and the equilibrium condition of the normal load. Discretizing this differential equation, the continuous arc length coordinates... exist The area is evenly divided into By establishing control points, the continuous equations are transformed into a system of finite-dimensional linear equations, providing a structured physical constraint basis for subsequent probabilistic graphical modeling.

[0048] When the above discretized coupled physical relationship is transformed into a factor graph structure, the curvature parameter at each control point The contact pressure is modeled as a latent variable node. The discrete values ​​at each control point are modeled as observed variable nodes, while the coupling relationship described by the stress transfer model is encoded as factor nodes connecting the latent variable nodes and the observed variable nodes. Let the latent variable curvature distribution parameter vector be... Its prior distribution is categorized according to the radian requirement. The corresponding optimal configuration range Configure the settings and use a truncated Gaussian distribution. Prior constraints are applied to ensure that the curvature parameter always falls within a physically reasonable range during the optimization process. The plantar pressure distribution data is then statistically analyzed by region to obtain the pressure distribution feature vector. This vector, as observed data, enters the likelihood calculation stage of the factor graph. The introduction of the factor graph structure allows complex joint posterior distributions to be decomposed through the product of local factors, thus providing operational structural support for variational inference.

[0049] When introducing a variational posterior distribution in a probabilistic graphical form, a mean-field approximation strategy is adopted to approximate the latent variables. The true posterior distribution is approximately a family of parameterized distributions, whose variational parameters are denoted as . Under the mean-field assumption, the latent curvature variables at each control point are considered independent, thus decomposing the inference problem of the high-dimensional joint posterior distribution into an optimization problem of several low-dimensional marginal distributions. Lower bound of evidence. It is broken down into two components: one is observation data. The log-likelihood expectation under the variational posterior distribution reflects the degree of fit of the variational posterior to the observed data; secondly, it is the negative value of the KL divergence between the latent variable prior distribution and the variational posterior distribution, which plays a regularization role and prevents the variational posterior from deviating excessively from the prior constraints. The sum of the two terms constitutes a complete expression for the lower bound of evidence. Maximizing this lower bound is equivalent to minimizing the KL divergence between the variational posterior and the true posterior in an information-theoretical sense, thereby achieving a high-quality approximation of the latent variable posterior distribution.

[0050] When performing gradient ascent iterations within the optimal configuration interval, the variational parameters... Regarding the lower bound of evidence Calculate the gradient and update along the gradient direction. During gradient calculation, the gradient of the log-likelihood term is backpropagated through the Jacobian matrix of the stress transfer model, while the gradient of the KL divergence term is directly calculated using the analytical form of the variational posterior distribution family. To ensure the numerical stability of the iteration process, an adaptive learning rate strategy is used to control the step size of each update, avoiding non-monotonic fluctuations in the lower bound of evidence due to excessively large step sizes. After each iteration, the increment of the current lower bound of evidence relative to the previous iteration value is calculated. When the absolute value of this increment is lower than a preset convergence threshold, the lower bound is considered closed. When convergence is reached, the optimization loop is terminated. The convergence determination also introduces a maximum upper limit on the number of iterations as a fallback mechanism to prevent iteration from failing to terminate due to numerical precision issues in extreme cases.

[0051] After iterative convergence, latent variables are extracted from the variational posterior distribution. The posterior expected value is used to obtain the optimal cutting edge curvature distribution parameters. This posterior expectation integrates physical prior constraints, observational data, and the configuration range defined by the arc requirement classification, representing the optimal blade geometry configuration under the current individualized athlete's push-off pattern and real-time plantar pressure distribution conditions. By substituting the stress transfer model, the optimal contact pressure distribution can be simultaneously reconstructed. The solution verifies whether the obtained curvature distribution meets the requirements of contact pressure uniformity and load balance at the physical level. If the restored contact pressure distribution shows local stress concentration exceeding the limit, a penalty term is applied to the variational posterior parameter at the control point, and the local iteration is re-executed until the contact pressure distribution meets the preset uniformity index.

[0052] In practical engineering implementation, the above variational inference process is executed in a pipelined manner on an embedded computing platform, triggering a complete inference process after each gliding cycle. The plantar pressure sensor array acquires data at high frequency from each measurement point. Pressure sampling value at the location The peak pressure of each region was obtained through regional statistics. and pressure center coordinates , This is then concatenated into a pressure distribution feature vector. Input factor graph. Radius demand classification. The optimal configuration range is provided in real time by the learning results of the latent spatial manifold of the deep neural network. The parameters are retrieved from a pre-stored fractal configuration table and used as prior constraint boundaries for variational inference. The entire inference process, from triggering to outputting the optimal curvature parameters... The latency must meet the response requirements of real-time control to ensure that control commands can be issued and executed by the blade curvature adjustment mechanism before the start of the next gliding cycle.

[0053] In the quality evaluation stage of the joint optimization solution, the obtained solution will be... The calculation compares the results with historical gliding cycles to determine the magnitude of change in the curvature parameter vector between adjacent cycles. If the magnitude of change exceeds a preset threshold, an anomaly detection process is triggered to investigate sensor data quality issues or classification errors. (The last sentence appears to be incomplete and possibly refers to a separate process.) Converted into target curvature values ​​at each control point And combined with the current measured curvature value Calculate curvature adjustment amount Finally, it is encapsulated into control commands and sent to the ice skate curvature adjustment mechanism, completing a complete closed loop from variational inference to physical execution.

[0054] A second aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0055] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0056] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for controlling the curvature classification of ice skates based on AI vision combined with athlete data, characterized in that, include: The system acquires multi-view image sequences and inertial measurement data of athletes gliding on the ice. After spatial registration and temporal alignment of the multi-view image sequences, a three-dimensional posture sequence is extracted. The joint motion trajectory and push-off torque distribution are analyzed by combining the acceleration and angular velocity components in the inertial measurement data. The joint motion trajectory and the push-off torque distribution are encoded using a deep neural network to capture the temporal dependencies within the gliding cycle, construct a latent space representation of the athlete's individual push-off pattern, and map the continuous motion state into discrete arc demand subtypes in the latent space through manifold learning. Based on the arc requirement classification and its corresponding optimal configuration range, combined with the real-time collected plantar pressure distribution data, the joint optimization solution of the cutting edge curvature distribution function and the contact pressure distribution function is searched within the optimal configuration range through variational inference, and control commands for the cutting edge geometry are dynamically generated. The control command is sent to the ice skate curvature adjustment mechanism for adjustment, and the adjusted biomechanical feedback data is collected. The biomechanical feedback data is then used to jointly update the encoding weights of the deep neural network and the hyperparameters of the variational inference.

2. The method of claim 1, wherein, After spatial registration and temporal alignment of the multi-view image sequence, a three-dimensional pose sequence is extracted. The joint motion trajectory and push-off torque distribution are then analyzed by combining the acceleration and angular velocity components from the inertial measurement data. Based on polar geometry constraints, cross-view feature point matching is performed on multi-view image sequences. The joint estimation of camera extrinsic matrix and feature point 3D coordinates is optimized by bundle adjustment, and spatial registration results under a unified world coordinate system are constructed. Based on the timestamp information in the inertial measurement data, the feature point sequence in the spatial registration result is interpolated and synchronized, and a three-dimensional attitude sequence is reconstructed from the feature point sequence based on the forward kinematic solution. The joint angle change rate in the three-dimensional posture sequence is fused with the angular velocity component in the inertial measurement data using Kalman filtering to eliminate measurement noise and drift error, thus obtaining the joint motion trajectory. An inverse dynamics solution framework based on the Lagrange dynamics equations is established. The angular acceleration in the joint motion trajectory and the acceleration components in the inertial measurement data are substituted into the motion equations of the inverse dynamics solution framework. The time series of torques borne by each joint during the ice-pushing phase is obtained by recursive calculation of the joint torques, forming the ice-pushing torque distribution.

3. The method according to claim 2, characterized in that, The joint angle change rate in the three-dimensional posture sequence is fused with the angular velocity component in the inertial measurement data using Kalman filtering to eliminate measurement noise and drift error, resulting in the joint motion trajectory, which includes: A state-space representation is constructed, defining the rate of change of joint angles as the state vector and the angular velocity components as the observation vector. The state transition matrix and process noise covariance are set according to the biomechanical constraints of human joint motion. Based on the joint angles of the previous frame of the three-dimensional pose sequence, state prediction is performed through the state transition matrix to obtain a prior estimate of the joint angle change rate of the current frame. Calculate the residual between the prior estimate and the angular velocity component, and calculate the Kalman gain based on the process noise covariance and the measurement noise covariance; The prior estimate is corrected using the Kalman gain to obtain the fused posterior estimate of the joint angle change rate, and the joint angle time series is reconstructed by time integration of the posterior estimate of the joint angle change rate. The joint angle time series is smoothed by cubic spline interpolation to eliminate high-frequency noise components and obtain the joint motion trajectory.

4. The method according to claim 1, characterized in that, Multimodal feature encoding of the joint motion trajectory and the push-off torque distribution is performed using a deep neural network to capture the temporal dependencies within the gliding cycle, and a latent spatial representation characterizing the athlete's individual push-off pattern is constructed, including: The joint motion trajectory and the distribution of the pushing force are respectively input into two parallel encoder branches. Each encoder branch uses a multi-layer convolutional structure to extract local temporal features and reduces the feature dimension through pooling operations. A cross-modal attention module is introduced into the output layer of the encoder branch to calculate the correlation weight matrix between the joint motion trajectory features and the push-off torque distribution features, and to perform weighted fusion of the feature vectors of the two modalities based on the correlation weight matrix to obtain multimodal fusion features; The multimodal fusion features are input into the recurrent neural network layer, and the dependencies between features at different times within the gliding cycle are captured through the temporal propagation of the hidden states. The global temporal feature vector is extracted at the last time step of the recurrent neural network layer. The global temporal feature vector is input into the encoder of the variational autoencoder, and the global temporal feature vector is mapped to a latent space representation that follows a specific distribution through the reparameterization technique.

5. The method according to claim 1, characterized in that, Based on the aforementioned arc requirement classification and its corresponding optimal configuration range, and combined with real-time collected plantar pressure distribution data, a joint optimal solution for the cutting edge curvature distribution function and the contact pressure distribution function is searched within the optimal configuration range through variational inference. Control commands for dynamically generating the cutting edge geometry include: Based on the parameter configuration table preset by the arc requirement fractal index, obtain the upper and lower bounds of the cutting edge curvature radius corresponding to the current fractal to form the optimal configuration range; The real-time collected plantar pressure distribution data is spatially discretized, dividing the plantar area into the forefoot, midfoot, and heel regions. The pressure peak and pressure center locations of each region are extracted to construct a pressure distribution feature vector. Establish the coupled physical relationship between the cutting edge curvature distribution function and the contact pressure distribution function, express the coupled physical relationship in the form of a probability graph, and iteratively optimize the lower bound of evidence within the optimal configuration interval through variational inference to solve the posterior distribution of the latent variable that maximizes the likelihood of the observed data. The optimal cutting edge curvature distribution parameters are sampled from the latent variable posterior distribution, and the corresponding contact pressure distribution is calculated according to the coupled physical relationship. The cutting edge curvature distribution parameters are then converted into control commands containing curvature adjustment amounts at multiple control points.

6. The method according to claim 5, characterized in that, A coupled physical relationship is established between the cutting edge curvature distribution function and the contact pressure distribution function. This coupled physical relationship is expressed as a probabilistic graphical form. The lower bound of evidence is iteratively optimized within the optimal configuration interval through variational inference. The posterior distribution of the latent variables that maximizes the likelihood of the observed data is solved, including: Based on the principle of elastic contact mechanics, the curvature parameter vector of the cutting edge curvature distribution function is mapped to the spatial distribution shape of the contact pressure distribution function through the stress transfer model, and a differential equation expression of the coupled physical relationship is established. The differential equation is transformed into a factor graph structure to form a probability graph. A variational posterior distribution is introduced into the probability graph to approximate the latent variable nodes. The lower bound of evidence is decomposed into the log-likelihood term of the plantar pressure distribution data and the divergence term of the latent variable prior distribution and the variational posterior distribution. Within the optimal configuration interval, the hyperparameters of the variational posterior distribution are iteratively updated through gradient ascent, making the lower bound of evidence monotonically increasing. The iteration is terminated when the increment of the lower bound of evidence is lower than the convergence threshold. The posterior expectation of the latent variables is extracted from the converged variational posterior distribution, and the latent variable posterior distribution that maximizes the likelihood of the observed data is solved.

7. An ice skate curvature parting control system based on AI vision combined with athlete data, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data acquisition unit is used to acquire multi-view image sequences and inertial measurement data of athletes during ice skating. After spatial registration and temporal alignment of the multi-view image sequences, a three-dimensional posture sequence is extracted. The joint motion trajectory and push-off torque distribution are analyzed by combining the acceleration and angular velocity components in the inertial measurement data. The feature classification unit is used to encode the joint motion trajectory and the push-off torque distribution through a deep neural network, capture the temporal dependency within the gliding cycle, construct a latent space representation of the athlete's individual push-off pattern, and map the continuous motion state into discrete arc demand classification in the latent space through manifold learning. The optimization control unit is used to search for the joint optimization solution of the cutting edge curvature distribution function and the contact pressure distribution function within the optimal configuration range based on the arc requirement classification and its corresponding optimal configuration range, combined with the real-time collected plantar pressure distribution data, and dynamically generate control commands for the cutting edge geometry. The adjustment and update unit is used to send the control command to the ice skate curvature adjustment mechanism for adjustment, and to collect the adjusted biomechanical feedback data, and to use the biomechanical feedback data to jointly update the encoding weights of the deep neural network and the hyperparameters of the variational inference.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images