Somatosensory interaction method and system applied to physical training

Abstract

The invention provides a somatosensory interaction method and system applied to physical training, and relates to the field of biological information processing, wearable sensor technology and artificial intelligence, and the method comprises the steps: obtaining real-time physiological load data of an athlete; acquiring technical action posture data of the athlete under the real-time physiological load data; on the basis of the physiological load data and the technical action posture data, constructing or updating a prediction model representing dynamic relevance of the physiological load data and the technical action posture data in real time; using the prediction model to predict a specific error deformation of the imminent technical action attitude data based on the change trend of the current physiological load data; and before the specific error deformation occurs, generating and outputting an early warning correction instruction. Physiological load data and technical action attitude data are acquired in real time, and a dynamic association prediction model of the two data is constructed, so that the technical problem that only hysteresis and reactivity error correction can be performed in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of bio-information processing, wearable sensor technology and artificial intelligence, and in particular, to a predictive somatosensory interaction method and system based on physiological and kinematic data fusion applied to sports training. BACKGROUND

[0002] In competitive sports training, the quality of a player's technical action execution is a core factor determining their competitive performance. Athletes (e.g., shooting, gymnastics, golf, etc.) need to consolidate muscle memory through thousands of repeated training to ensure the stability, accuracy and consistency of the action.

[0003] However, in the process of long-time and high-intensity training, athletes inevitably enter a state of physiological fatigue. Physiological fatigue (e.g., heart rate fluctuations, muscle fatigue) will directly lead to a decline in the athlete's neural control ability, and in turn cause minor deformations or errors in technical actions (kinematic postures). If these erroneous deformations cannot be discovered and corrected in time, not only will the quality of a single training be reduced, but also the training effect will stagnate or even regress, and the risk of sports injuries may be increased.

[0004] Currently, action correction in sports training mainly relies on the naked eye observation of coaches or post-training analysis through optical motion capture (CV) systems. The limitations of these methods are: Lagging: Whether it is coach observation or video review, it is a reactive correction after the error action occurs, and cannot intervene at the moment of error occurrence.

[0005] Superficiality: Existing monitoring methods can at most tell the athlete that they are already tired (i.e., state monitoring). They cannot establish a deep causal relationship between physiological fatigue and the impending action deformation.

[0006] Technical limitations: The use of optical capture (CV) alone is easily disturbed by obstructions (such as other athletes or equipment), while the use of inertial measurement units (IMU) alone will produce cumulative errors due to long-time integration.

[0007] In summary, the existing sports training methods lack a mechanism that can predict the erroneous deformation of an athlete's technical action based on their current physiological state changes before the deformation occurs, and output real-time correction instructions. Therefore, how to dynamically associate the physiological load data of an athlete with the technical action posture data to achieve predictive and immediate error correction is a technical problem that needs to be solved in the field at present. SUMMARY

[0008] Therefore, the present application aims to provide a somatosensory interaction method and system applied to sports training, which aims to solve the technical problem that the prior art mentioned in the background art can only conduct reactive monitoring or macroscopic early warning, and cannot predict specific motion deformation in real time based on physiological state and correct in advance.

[0009] In a first aspect, the present application discloses a somatosensory interaction method applied to sports training, comprising: acquiring real-time physiological load data of an athlete; acquiring technical motion posture data of the athlete under the real-time physiological load data; based on the physiological load data and the technical motion posture data, constructing or updating a prediction model representing the dynamic correlation between the two in real time; using the prediction model, predicting a specific error deformation of the technical motion posture data that is about to occur based on the trend of the current physiological load data; generating and outputting a warning correction instruction before the specific error deformation occurs.

[0010] Optionally, the acquisition of the real-time physiological load data comprises: acquiring electrocardiogram (ECG) signal data; and / or, acquiring surface electromyography (sEMG) signal data.

[0011] Optionally, the acquisition of the technical motion posture data comprises: acquiring posture information of the athlete using an inertial measurement unit (IMU); and / or, acquiring video image data of the athlete using an optical motion capture (CV) system.

[0012] Optionally, the acquisition of the technical motion posture data further comprises: fusing the posture information acquired by the inertial measurement unit (IMU) and the video image data acquired by the optical motion capture (CV) system to correct the cumulative error of the inertial measurement unit (IMU) using the unobstructed characteristics of the inertial measurement unit (IMU) and the low-frequency high-precision characteristics of the optical motion capture (CV) system.

[0013] Optionally, the prediction model comprises: a support vector machine (SVM) model; or, a radial basis function (RBF) neural network model; or, a recurrent neural network (RNN) model.

[0014] Optionally, the process further includes, before building or updating the prediction model in real time: Feature extraction is performed on the physiological load data and the technical movement posture data; Based on an optimal fusion coefficient vector, feature-level fusion is performed on the features of the physiological load data and the technical action posture data to generate a fused feature vector.

[0015] Optionally, the warning correction instruction includes: Haptic interaction commands; or, Visual warning instructions; or, Voice prompts and commands.

[0016] Optionally, motion-sensing interaction methods applied to sports training also include: When one or more other individuals are present in the training area, obtain the athlete's historical specific error deformation prediction sequence; Obtain historical spatial interaction feature data between the athlete and one or more other individuals that corresponds in time to the historical specific error deformation prediction sequence; Based on the historical specific error deformation prediction sequence and the historical spatial interaction feature data, an attribution model is constructed to characterize the influence of the spatial interaction feature data on the probability of occurrence of the specific error deformation. And using the attribution model, an optimal spatial interaction configuration instruction is generated to reduce the probability of future occurrence of the specific erroneous deformation.

[0017] Secondly, this application discloses a motion-sensing interactive system for sports training, comprising: The physiological data acquisition module is used to acquire real-time physiological load data of athletes; The posture data acquisition module is used to acquire the athlete's technical movement posture data under the real-time physiological load data; The dynamic modeling module is used to construct or update a predictive model that characterizes the dynamic correlation between the physiological load data and the technical action posture data in real time. The predictive analysis module is used to predict a specific error deformation of the technical movement posture data that is about to occur, based on the trend of the current physiological load data, using the predictive model. The instruction generation module is used to generate and output a warning correction instruction before the specific error variation occurs.

[0018] Optionally, the motion-sensing interactive system applied to sports training also includes: The historical data acquisition module is used to acquire the athlete's historical specific error deformation prediction sequence when one or more other individuals are present in the training venue; The spatial interaction analysis module is used to acquire historical spatial interaction feature data between the athlete and one or more other individuals that corresponds in time to the historical specific error deformation prediction sequence. The attribution modeling module is used to construct an attribution model based on the historical specific error deformation prediction sequence and the historical spatial interaction feature data to characterize the influence of the spatial interaction feature data on the probability of occurrence of the specific error deformation. The spatial configuration optimization module is used to generate an optimal spatial interaction configuration instruction that aims to reduce the future probability of the specific error deformation by utilizing the attribution model.

[0019] Thirdly, this application also discloses an electronic device, comprising: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the haptic interaction method described in any of the first aspects above.

[0020] Fourthly, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the motion-sensing interaction method described in any of the first aspects above.

[0021] The somatosensory interaction method and system disclosed in this application acquire physiological load data and technical movement posture data in real time, and construct a dynamic correlation prediction model between the two. This aims to solve the technical problem that existing technologies can only perform delayed and reactive error correction. This invention can predict specific movement deformities based on changes in physiological state before erroneous movements occur and output early warning corrective instructions. Because this invention can output corrective instructions before specific erroneous deformities occur, it provides an immediate intervention capability that existing delayed error correction technologies lack, improving the immediacy, targeting, and effectiveness of training, helping athletes form correct muscle memory, and preventing sports injuries.

[0022] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a motion-sensing interaction method applied to sports training, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of a motion-sensing interactive system applied to sports training, as described in an embodiment of the present invention. Detailed Implementation

[0025] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0026] As described in the background section, existing sports training error correction mechanisms suffer from lag, superficiality (monitoring only the state rather than predicting specific errors), and technical limitations (such as CV occlusion and IMU cumulative error), making it impossible to effectively intervene before athletes' movements become distorted due to physiological fatigue.

[0027] To this end, the present invention provides a predictive motion-sensing interaction scheme that can correlate athletes’ physiological state with their kinematic performance in real time and dynamically, accurately capture impending errors and deformations, and achieve timely error correction to prevent problems before they occur.

[0028] Reference Figure 1 As shown in the figure, an embodiment of the present invention discloses a motion-sensing interaction method applied to sports training, comprising: Step S11: Obtain the athlete's real-time physiological load data.

[0029] In one specific embodiment, step S11 can be achieved using a wearable biosensor worn by the athlete. The physiological load data may specifically include: Electrocardiogram (ECG) signal data. This can be acquired, for example, through a smart chest strap or flexible electrodes worn by athletes. R-wave peak identification is performed on the ECG signal to obtain the RR interval (RRI) sequence. Further time-domain and frequency-domain analysis is then performed to extract heart rate variability (HRV) related features, such as low-frequency power (LF), high-frequency power (HF), and the ratio of LF to HF power (LF / HF). An abnormally high LF / HF ratio typically indicates sympathetic nervous system excitation and increased psychological stress.

[0030] Surface electromyography (sEMG) signal data. This is acquired, for example, via electrode pads attached to key muscle groups in an athlete (e.g., the deltoid muscle in a shooter's arm, or forearm muscles), or via a sensor array integrated into a wearable armband. The sEMG signals are processed to extract time-domain features such as integrated electromyography (iEMG) values ​​and root mean square (RMS) values, as well as frequency-domain features such as mean power frequency (MPF) and median frequency (MF). A significant decrease in MPF ​​or MF is a recognized indicator of localized muscle fatigue.

[0031] Those skilled in the art will understand that the physiological load data may also include other data that can characterize physiological state, such as skin conductance (GSR), electroencephalography (EEG), or blood oxygen saturation (SpO2), and the present invention is not limited to ECG and sEMG.

[0032] Step S12: Obtain the athlete's technical movement posture data under the real-time physiological load data.

[0033] Step S12 is collected synchronously with step S11 to ensure that the acquired posture data and physiological data correspond strictly in time, which is the basis for building dynamic correlation.

[0034] In one specific embodiment, step S12 can be implemented in one or more of the following ways: An inertial measurement unit (IMU) is used to acquire the athlete's posture information. For example, IMU sensors (including gyroscopes, accelerometers, and magnetometers) are attached to key limbs of the athlete (such as the wrist, elbow, and shoulder). The pitch, yaw, and roll angles of these limbs are calculated in real time using attitude calculation algorithms (such as complementary filtering or Kalman filtering), forming a kinematic data stream. The advantages of IMUs are their high sampling frequency (e.g., 200Hz) and insensitivity to ambient light and occlusion. Their disadvantage is that the integration of the gyroscope leads to accumulated errors, causing attitude angle drift after prolonged operation.

[0035] Video image data of the athlete is acquired using an optical motion capture (CV) system. For example, one or more high-speed cameras (marked or unmarked) are aimed at the athlete. For unmarked systems, 2D or 3D coordinates of the athlete's keypoints are extracted in real time using computer vision algorithms (such as deep learning-based pose estimation algorithms, such as OpenPose and MediaPipe) to form pose data. The advantage of CV systems is their low-frequency, high-precision operation; that is, they do not accumulate drift. Their disadvantages include: (a) a relatively low sampling rate (e.g., 30-60Hz); (b) susceptibility to occlusion problems, such as self-occlusion by the athlete or occlusion by other objects; and (c) sensitivity to lighting conditions.

[0036] Furthermore, in a preferred embodiment, in order to overcome the limitations of a single sensor (such as the aforementioned cumulative error and the aforementioned occlusion problem), step S12 may specifically include: The attitude information (high frequency, with drift, unobstructed) acquired by the IMU and the video image data (low frequency, no drift, with occlusion) acquired by the CV are fused. The technical advantage of this fusion scheme (e.g., implemented through Extended Kalman Filter (EKF) or multi-sensor fusion algorithms) lies in its utilization of the IMU's unobstructed and high-frequency characteristics, while simultaneously leveraging the CV's low-frequency, high-precision characteristics (i.e., the CV has no accumulated drift), to correct the IMU's accumulated error in real time. Through this multimodal data fusion, a long-term stable, high-frequency (e.g., 200Hz) and high-precision attitude data stream can be obtained.

[0037] Step S13: Based on the physiological load data and the technical action posture data, construct or update a prediction model that characterizes the dynamic correlation between the two in real time.

[0038] This is the core technical point of the present invention. The model aims to learn and quantify the nonlinear mapping relationship between what kind of physiological load (from S11) and what kind of changing trend corresponds to what kind of technical movement posture (from S12) and what kind of erroneous deformation.

[0039] The model can be trained offline in the cloud using a large amount of historical data and updated or fine-tuned in real time locally (e.g., on an edge computing device).

[0040] In a preferred embodiment, before performing step S13 (building the model), in order to improve the training efficiency and prediction accuracy of the model, the method may further include a data preprocessing step: Feature extraction: Feature extraction is performed on the raw physiological data (such as ECG waveforms) obtained in S11 and the raw posture data (such as IMU quaternions) obtained in S12 to obtain physiological feature vectors and posture feature vectors.

[0041] Feature-level fusion: Due to the differences in dimensions, scale, and importance between physiological features and posture features, feature-level fusion can be performed based on an optimal fusion coefficient vector. This optimal fusion coefficient vector can be obtained through iterative optimization during offline training (e.g., using particle swarm optimization or genetic algorithms), with the goal of maximizing the accuracy of subsequent prediction models.

[0042] Generate a fusion feature vector: The extracted physiological and posture features are weighted and combined according to the optimal fusion coefficient vector to generate a unified, high-information-density fusion feature vector. This fusion feature vector is then used as input to the prediction model in step S13.

[0043] In one specific embodiment, the prediction model can select different machine learning models based on the complexity and real-time requirements of the application scenario: Support Vector Machine (SVM) model: suitable for simple classification tasks with low feature dimensionality and high real-time requirements (e.g., binary classification prediction: whether it will be deformed or not).

[0044] Radial basis function (RBF) neural network models are suitable for handling more complex nonlinear relationships, such as classifying damage risk (high, medium, and low risk).

[0045] Recurrent Neural Network (RNN) model (or its variants such as Long Short-Term Memory Network (LSTM) or Gated Recurrent Unit (GRU): This is a preferred embodiment. Since both physiological data (S11) and posture data (S12) are time-series data, the RNN / LSTM model is extremely suitable for learning the dynamic relationships between such time-series data. It can capture trends of change, rather than just static instantaneous values.

[0046] Step S14: Using the prediction model, based on the current trend of the physiological load data, predict a specific error deformation of the technical movement posture data that is about to occur.

[0047] This step is the key difference between this invention and existing technologies. This step provides a specific, implementable application scenario (e.g., shooting training) to illustrate this: Scene: A shooting athlete is taking a long time to aim his gun.

[0048] Data Acquisition (S11, S12): The system acquires sEMG and HRV data of the arm holding the gun in real time, as well as wrist posture data acquired through IMU.

[0049] Model (S13): The system is equipped with a pre-trained LSTM model. This model has learned from historical data that when an athlete's sEMG median frequency (MPF) continues to decline (trend) over the past 5 seconds and falls below the fatigue threshold A, while their HRV LF / HF ratio (stress) spikes instantaneously (trend) and exceeds the threshold B—this combination of physiological trends (fatigue + stress) is a precursor to loss of control.

[0050] Prediction (S14): At second T, the system detects the aforementioned physiological data trend (MPF continues to decline and LF / HF spikes instantaneously). The LSTM model immediately (at T+0.1 seconds) outputs a high-probability prediction: within the next 0.5 seconds (i.e. before T+0.6 seconds), a specific error distortion will occur in the athlete's wrist posture data (e.g., involuntary wrist drop due to muscle fatigue and tension or tremor when pulling the trigger).

[0051] Step S15: Before the specific error deformation occurs, generate and output a warning correction instruction.

[0052] Continuing with the above scenario: After making a prediction at T+0.1 seconds, the system must issue a warning within a 12-second interval (i.e., within 0.5 seconds) before the actual wrist drop occurs (T+0.6 seconds). The system immediately generates and outputs the following command: Tactile interaction command (preferred): A short, forceful vibration is emitted via a vibrator (part of a wearable device) worn on the athlete's wrist. This tactile feedback is a non-invasive, high-response command that instantly draws the athlete's attention back to control of the wrist, thereby proactively counteracting impending erroneous movements.

[0053] Visual warning instruction: If the athlete is wearing smart glasses, a red LED will flash on the edge of the lens as a warning.

[0054] Voice prompts: A short prompt or keyword "wrist" is played through the bone conduction headphones worn by the athlete.

[0055] The embodiments of the present invention may further include optimization steps S16 to S19 for multi-person training environments. This solution aims to address a previously overlooked technical problem: the deformation of athletes' technical movements is not only affected by their own physiological load (S11), but may also be affected by the spatial position and psychological pressure brought by other individuals (such as coaches and other athletes) in the training venue.

[0056] Step S16: When one or more other individuals are present in the training area, obtain the athlete's historical specific error deformation prediction sequence.

[0057] To achieve this step and subsequent steps S17 to S19, the system first requires a tool capable of simultaneously acquiring the athlete's own posture and the spatial positions of all individuals within the playing field. This embodiment preferably utilizes the optical motion capture (CV) system described in step S12. While performing its primary task (i.e., acquiring video image data of the athlete to support fusion), this CV system's inherent technical capabilities enable it to identify and track the two-dimensional or three-dimensional coordinates of all other individuals within its field of view in real time using mature computer vision target detection and tracking algorithms (e.g., algorithms based on YOLO or DeepSORT).

[0058] Simultaneously, a central data processing unit (e.g., processor and memory) is configured as the data synchronization hub. When data frames from the motion capture (CV) system and data packets from the athlete's wearable IMU (S12) arrive, this unit adds a unified, high-precision system timestamp (e.g., a microsecond-level server clock) to all data streams. This timestamp synchronization mechanism is the cornerstone for subsequent time-correspondence analysis.

[0059] Under this system architecture, the specific implementation of obtaining the historical specific error variation prediction sequence is as follows: During the athlete's long-term training, the system continuously runs the prediction model in steps S13 and S14. Whenever the output probability of this model (e.g., the aforementioned LSTM model) exceeds a preset high confidence threshold (e.g., 80% probability), that is, when the system predicts that the athlete will soon (e.g., within the next 0.5 seconds) experience a specific error variation (such as wrist tremor during shooting), the system will not only execute the warning in S15, but also write this event as an entry into a historical log database. This entry is a data point in the prediction sequence, and it contains at least two key technical parameters: one is the predicted error type, and the other is the precise system timestamp of the predicted event. Through multiple training iterations, the system constructs a complete historical specific error variation prediction sequence containing hundreds or thousands of entries. This sequence is the target variable (i.e., the Y value) required for the subsequent step S18 to construct the attribution model.

[0060] Step S17: Obtain the historical spatial interaction feature data between the athlete and one or more other individuals that corresponds in time to the historical specific error deformation prediction sequence.

[0061] This step prepares the input variables (i.e., X values) for the attribution model in step S18. This can be achieved by the system performing a historical backtracking data processing task. The system iterates through each record in the historical error-specific deformity prediction sequence obtained in S16. For each timestamp in the sequence, the system uses that timestamp to query the synchronized historical data log and extracts the raw spatial coordinate data (from the CV system) of all other individuals at that time point (and typically within a short window before that time point, such as 2 seconds).

[0062] Simply acquiring raw coordinate data is insufficient; this invention requires acquiring feature data. Therefore, the system executes a feature engineering process to transform the raw, variable coordinate data into a fixed set of highly informative spatial interaction features.

[0063] In a specific, feasible scenario, these characteristic parameters may include: The first feature is the nearest individual distance. This parameter is obtained by the system calculating the instantaneous straight-line distances between the athlete and all other individuals on the field within the stated time window, identifying the smallest distance value (i.e., the distance to the nearest interferer), and calculating the average of this distance value within the time window. This average parameter (e.g., the average nearest distance is 1.5 meters) is one feature.

[0064] The second feature is local density. This parameter is obtained by defining a virtual radius (e.g., two meters) around the athlete's coordinates, and then calculating in real time how many other individuals enter this virtual radius area within that time window, and taking the average value (e.g., an average density of 2.3 individuals).

[0065] The third characteristic is visual field intrusion. This parameter is obtained by combining the athlete's head orientation (i.e., gaze direction) calculated from the athlete's head IMU (from S12) with the coordinates of other individuals obtained by the CV system to dynamically calculate how many other individuals are within the athlete's primary visual field cone area (e.g., within a 60-degree angle in front) within the stated time window. Its characteristic value can be the total duration or average number of other individuals remaining in the visual field.

[0066] The fourth feature is relative velocity. This parameter is obtained by the CV system calculating the velocity vectors of other individuals based on the changes in coordinates over time. The system calculates the velocity vector difference between the athlete and the nearest individual to quantify whether a disruptor is stationary or approaching or passing through at high speed.

[0067] Through the aforementioned feature engineering process, the system matches each erroneous prediction in S16 with a set of time-corresponding, quantified historical spatial interaction feature data.

[0068] Step S18: Based on the historical specific error deformation prediction sequence and the historical spatial interaction feature data, construct an attribution model to characterize the influence of the spatial interaction feature data on the probability of occurrence of the specific error deformation.

[0069] This step is the core of the invention's solution to the psycho-spatial interference problem. The system first uses the historical specific error deformation prediction sequence (i.e., the probability value of the predicted error) obtained in step S16 as the target variable (Y value) for training, and the historical spatial interaction feature data (i.e., features such as nearest distance, local density, and field of view intrusion) obtained in step S17 as the input feature variable (X value) for training. The system loads this historical pairing dataset into a central processing unit (e.g., a processor) to construct the attribution model.

[0070] The attribution model constructed in this embodiment is a composite structure, consisting of a high-performance prediction model and supporting attribution analysis tools, to ensure that the model can not only predict but also explain and confirm the impact. In the specific technical solution, a tree-based ensemble learning algorithm, such as Gradient Boosting Decision Tree (GBDT), is preferably used as the cornerstone of the prediction model. The process of building this GBDT model includes the following specific training and parameter tuning steps: First, model instantiation and hyperparameter definition are performed. The model is instantiated as a regressor, with its objective function set to minimize the difference between the predicted probability and the actual error probability. Before training, key hyperparameters must be defined, such as the learning rate, the number of estimators (n_estimators, i.e., the number of trees built), and the maximum depth (max_depth). The learning rate controls the contribution of each new tree to the final prediction result and is usually set to a small value (e.g., 0.01 or 0.1) to prevent overfitting. The number of estimators defines the scale of the model ensemble and needs to be adjusted according to the size of the dataset (e.g., 500 to 1000 trees). The maximum depth limits the complexity of each tree to prevent the model from becoming overly reliant on noise in the training data.

[0071] Secondly, model training is performed. During training, the system employs an iterative optimization strategy. In each iteration, the model builds a new, simple decision tree. This new tree does not directly predict the original target variable (error probability), but rather fits the residual (or more precisely, the negative gradient) between the current model's prediction and the true error probability. Through this iterative fitting of residuals, the model can progressively and robustly optimize its predictive performance. In practice, the system typically uses techniques such as cross-validation and grid search (GridSearchCV) to systematically search and optimize the aforementioned hyperparameters, ensuring that the model maintains the highest predictive accuracy even on unseen data.

[0072] Finally, an attribution analysis tool is constructed. Once the GBDT prediction model is trained, the system needs to transform it into an interpretable attribution model. This embodiment preferably uses the SHAP (SHapley Additive Explanations) tool. The SHAP analyzer (i.e., the attribution model referred to in this invention) is loaded and utilizes the trained GBDT model. The core of SHAP is based on cooperative game theory principles; it can inversely decompose each prediction of the GBDT model, calculating the independent contribution of each input spatial interaction feature (e.g., nearest individual distance) to the final error probability prediction. This contribution value is called the Shapley value. By calculating and analyzing this Shapley value, the system can accurately characterize the extent to which a certain spatial feature (e.g., an increase in the number of individuals intruding into the field of vision) numerically affects the probability of a specific error variation in the athlete's technical movement (e.g., increasing the risk probability by 15%). This combination of the GBDT prediction model and the SHAP analyzer fully realizes the construction of the highly interpretable and scientifically rigorous attribution model required in this step, thus providing a solid technical basis for the subsequent generation of intervention instructions.

[0073] Step S19: Using the attribution model, generate an optimal spatial interaction configuration instruction designed to reduce the future probability of the specific error deformation.

[0074] This step is the final execution link in the entire method chain, realizing a closed loop from historical data analysis to real-time environmental intervention. Its purpose is to physically avoid risks caused by spatial factors by guiding athletes to change their positions. The process of this implementation is a real-time optimization solution loop, characterized by high systematicity, reproducibility, and extremely low time latency.

[0075] First, real-time assessment of intervention conditions is the starting point for instruction generation. During training, the system uses the feature engineering module in step S17 to collect the latest position and motion status of all individuals (including athletes and other individuals) within the field at high frequency (e.g., once every 0.1 seconds), calculating a set of current spatial interaction feature data (i.e., X_current) representing the current environmental risk. The system immediately inputs this X_current into the prediction part of the attribution model trained in step S18 (i.e., the GBDT prediction model) to obtain a real-time current error probability (P_current). The system pre-sets an intervention threshold (P_intervene, e.g., 75%), which is usually determined by offline analysis of the balance between risk and false alarms in historical data. If P_current is below this threshold, the system determines that the risk is acceptable and remains silent; once P_current exceeds this threshold, the system immediately confirms that the current spatial layout poses an excessive risk of interference to the athletes and must initiate the instruction generation process.

[0076] Secondly, the local search and simulation evaluation of the optimal position is the computational core of instruction generation. Since the intervention object of this invention is limited to the athlete's own position, the problem is defined as a local spatial optimization problem, rather than a large-scale scheduling problem controlling all individuals. To find the lowest-risk position, the system utilizes a local grid search technique, a stable and engineering-easy-to-implement optimization method. The system virtually defines a finite search area (e.g., 1 meter in each direction) on a two-dimensional plane, centered on the athlete's current position, and evenly distributes several candidate movement points within this area (e.g., one every 0.5 meters, for a total of 25 points). The system then enters a high-speed assumption simulation loop: for each candidate movement point, the system assumes the athlete has moved to that point. After the assumed movement, the system keeps the current positions of all other individuals in the field unchanged and recalculates a set of hypothetical spatial interaction feature vectors. The system feeds this hypothetical feature vector into the prediction part of the attribution model to obtain a hypothesis error probability (P_test). The system repeats this simulation and prediction process for all 25 candidate points, ultimately obtaining a set containing 25 different risk values.

[0077] Finally, the system determines the instruction and delivers the haptic feedback. After obtaining the risk assessment results for all candidate positions, the system selects the candidate movement point from these 25 probability values ​​that minimizes the probability of error in the hypothesis. This point is determined as the optimal physical position with the lowest risk for the athlete under the current interference environment. The system then calculates the displacement vector difference between this optimal physical position and the athlete's current position and transforms this vector difference into an executable optimal spatial interaction configuration instruction designed to guide the athlete's displacement. The output of this instruction must follow the form of a warning correction instruction. For example, the system can generate a voice prompt instruction, playing a short and directional instruction (e.g., take a step to the left or move back half a meter) through the athlete's bone conduction headphones; or it can use spatial audio technology to play a slight, spatially directional cue in the direction of the optimal position, using the athlete's hearing to guide their subconscious movement. Through this real-time feedback, the athlete is guided away from the high-risk standing area, thereby physically eliminating or reducing the potential psychological stress and movement distortion risks caused by external spatial factors (such as too close a distance or a fast-moving interferer).

[0078] Reference Figure 2 The present invention also provides a motion-sensing interactive system for sports training, the system comprising: Physiological data acquisition module 1 is used to perform the function of acquiring physiological load data (such as ECG, sEMG1) as described in the aforementioned step S11; The attitude data acquisition module 2 is used to perform the function of acquiring technical action attitude data (such as IMU6, CV2 or their fusion) as described in the aforementioned step S12; The dynamic modeling module 3 is used to perform the functions described in step S13 above, which are to build or update prediction models (such as SVM4, RNN2) based on physiological and posture data. This module may also include the function of performing feature extraction and fusion. Predictive analysis module 4 is used to perform the function described in step S14 above, which is to predict specific error deformations in real time using the model; The instruction generation module 5 is used to perform the function described in step S15 above, which is to generate and output warning correction instructions (such as tactile or visual instructions) before an error occurs.

[0079] It should be noted that, Figure 2 The system shown has modules that are functionally limited, and the specific ways in which these functions are implemented have been described in detail in the above method embodiments, and will not be repeated here.

[0080] Furthermore, motion-sensing interactive systems applied to sports training also include: The historical data acquisition module is used to acquire the athlete's historical specific error deformation prediction sequence when one or more other individuals are present in the training venue; The spatial interaction analysis module is used to acquire historical spatial interaction feature data between the athlete and one or more other individuals that corresponds in time to the historical specific error deformation prediction sequence. The attribution modeling module is used to construct an attribution model based on the historical specific error deformation prediction sequence and the historical spatial interaction feature data to characterize the influence of the spatial interaction feature data on the probability of occurrence of the specific error deformation. The spatial configuration optimization module is used to generate an optimal spatial interaction configuration instruction that aims to reduce the future probability of the specific error deformation by utilizing the attribution model.

[0081] This application also discloses an electronic device. Specifically, the electronic device may include: at least one processor (e.g., CPU, GPU, or NPU dedicated to AI computing), at least one memory, a power supply, a communication interface (e.g., a Bluetooth or Wi-Fi module for receiving sensor data), an input / output interface (e.g., for connecting to a tactile vibrator), and a communication bus.

[0082] The memory is used to store computer programs (e.g., operating system, computer program) and data (e.g., physiological data of S11, posture data of S12, and prediction model parameters of S13); the processor is used to load and execute the computer programs stored in the memory to implement all or part of the steps of the somatosensory interaction method disclosed in any of the foregoing embodiments.

[0083] Furthermore, this application also discloses a computer-readable storage medium. This medium (e.g., RAM, ROM, hard disk, optical disk) is used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed haptic interaction method.

[0084] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A somatosensory interaction method applied to sports training, characterized in that, The method comprises: acquiring real-time physiological load data of an athlete; acquiring technical action posture data of the athlete under the real-time physiological load data; based on the physiological load data and the technical action posture data, constructing or updating a prediction model in real time, which represents the dynamic correlation between the two; using the prediction model, predicting a specific error deformation of the technical action posture data that is about to occur based on the trend of the current physiological load data; generating and outputting a pre-warning correction instruction before the specific error deformation occurs.

2. The somatosensory interaction method for sports training according to claim 1, wherein, The acquisition of the real-time physiological load data comprises: acquiring electrocardiogram (ECG) signal data; and / or acquiring surface electromyography (sEMG) signal data. The acquisition of the technical action posture data comprises:

3. The somatosensory interaction method for sports training according to claim 1, wherein, acquiring posture information of the athlete using an inertial measurement unit (IMU); and / or acquiring video image data of the athlete using an optical motion capture (CV) system. The acquisition of the technical action posture data further comprises: fusing the posture information acquired by the inertial measurement unit (IMU) and the video image data acquired by the optical motion capture (CV) system to correct the cumulative error of the inertial measurement unit (IMU) using the unobstructed characteristics of the inertial measurement unit (IMU) and the low-frequency high-precision characteristics of the optical motion capture (CV) system.

4. The somatosensory interaction method for sports training according to claim 3, wherein, The prediction model comprises: a support vector machine (SVM) model; 5. The somatosensory interaction method for sports training according to claim 1, wherein, or a radial basis function (RBF) neural network model; or a recurrent neural network (RNN) model. Before constructing or updating the prediction model in real time, the method further comprises: extracting features from the physiological load data and the technical action posture data; performing feature-level fusion on the features of the physiological load data and the technical action posture data based on an optimal fusion coefficient vector to generate a fusion feature vector. The pre-warning correction instruction comprises:

6. The somatosensory interaction method for sports training according to any one of claims 1 to 3, wherein, a tactile interaction instruction; or a visual pre-warning instruction; or a voice prompt instruction.

7. The somatosensory interaction method for sports training according to claim 1, wherein, The method further comprises: when one or more other individuals are present in the training venue, acquiring a historical specific error deformation prediction sequence of the athlete; acquiring historical spatial interaction feature data between the athlete and the one or more other individuals corresponding in time to the historical specific error deformation prediction sequence; based on the historical specific error deformation prediction sequence and the historical spatial interaction feature data, constructing an attribution model for representing the influence of the spatial interaction feature data on the occurrence probability of the specific error deformation; and using the attribution model, generating an optimal spatial interaction configuration instruction aimed at reducing the future occurrence probability of the specific error deformation. The method comprises: 8.The somatosensory interaction method for sports training of claim 1, wherein, a physiological data acquisition module for acquiring real-time physiological load data of an athlete; a posture data acquisition module for acquiring technical action posture data of the athlete under the real-time physiological load data; a dynamic modeling module for constructing or updating a prediction model in real time based on the physiological load data and the technical action posture data, which represents the dynamic correlation between the two; ​ ​ 9. A somatosensory interaction system applied to sports training, characterized in that, ​ ​ ​ ​ a predictive analysis module configured to predict, using the prediction model, a specific erroneous deformation of the technical movement gesture data based on a variation trend of the current physiological load data; an instruction generation module configured to generate and output a pre-warning correction instruction before the specific erroneous deformation occurs.

10. The somatosensory interaction system for sports training according to claim 9, wherein, Further comprising: a historical data acquisition module configured to acquire a historical specific erroneous deformation prediction sequence of the athlete when one or more other individuals are present in a training field; a spatial interaction analysis module configured to acquire historical spatial interaction feature data between the athlete and the one or more other individuals corresponding to the historical specific erroneous deformation prediction sequence in time; an attribution modeling module configured to construct an attribution model based on the historical specific erroneous deformation prediction sequence and the historical spatial interaction feature data, the attribution model being configured to represent an influence of the spatial interaction feature data on a probability of occurrence of the specific erroneous deformation; a spatial configuration optimization module configured to generate an optimal spatial interaction configuration instruction aiming to reduce a future probability of occurrence of the specific erroneous deformation using the attribution model.

Images