Intelligent auxiliary method and system for equestrian teaching based on motion posture recognition

Abstract

The invention discloses an intelligent auxiliary method and system for equestrian teaching based on motion posture recognition, relates to the technical field of intelligent physical training and man-machine interaction, and aims at the problem of signal mixing caused by rider and horse motion coupling through the following steps: step 1, providing a high-quality data basis by adopting multi-modal data synchronous acquisition and fusion frame sequence generation; 2, performing decoupling and coupling credible coefficient generation by using dual-channel countermeasure coding, realizing high-precision separation of rider and horse signals, and inhibiting horse gait interference; 3, outputting an interpretable attitude deviation curve and an action level label through mapping of a knowledge graph and adaptive normalization of gait residual errors; 4, a feedback channel is dynamically selected based on the posture deviation threshold value and the link state, and a correction prompt is sent to the rider in real time; 5, updating the weight of the model through semi-supervised incremental training after class, and continuously optimizing the performance; the teaching effect is improved, and scientific and efficient support is provided for modern equestrian teaching.

Description

Technical Field

[0001] This invention relates to the field of intelligent sports training and human-computer interaction technology, specifically to an intelligent auxiliary method and system for equestrian teaching based on motion posture recognition. Background Technology

[0002] In modern equestrian teaching scenarios, instructors and students utilize a multimodal perception system to capture rider postures and horse gaits in real time: a combination of visible light and depth cameras enables markerless motion capture, an inertial measurement unit (IMU) tracks rider joints and horse limbs to record acceleration and angular velocity, voice-guided headsets and head-mounted displays simultaneously deliver teaching instructions, and an edge computing platform performs posture estimation and feedback rendering on-site, thus constructing a closed loop of "perception-analysis-feedback." This new teaching method significantly reduces the subjective errors of traditional experience-based judgments and provides a safer and more nuanced training experience for both students and horses.

[0003] A search revealed a system for acquiring and analyzing equestrian rider posture information in patent application CN108447077A. This system includes a motion information acquisition module with sensor nodes installed on various characteristic parts of the rider, a convergence node module for processing the motion information from each node, a wireless data transmission module for communication with a host computer, and a 3D human motion tracking PC interface for reconstructing the rider's posture. This system enables real-time capture and calculation of the rider's three-dimensional posture information through sensor nodes and real-time data-driven operation of a 3D human model via wireless communication. This allows the model to reconstruct the rider's posture during equestrian activities. It can capture and reconstruct rider posture information during equestrian activities, providing professional coaches and riders with high-quality kinematic data to analyze rider deficiencies and ultimately improve rider skill levels.

[0004] However, a natural kinematic coupling exists between rider and horse. When the horse experiences rhythmic fluctuations due to changes in stride length, slight limps, or uneven ground, these changes are transmitted to the rider through the saddle. This results in a mixture of inertial measurement signals and visual skeletal data, containing components from both the rider and the horse. Existing posture recognition algorithms often rely on a single-subject assumption, making it difficult to distinguish between these two types of signals. Consequently, they may misinterpret abnormal horse gait as rider movement deviations. When the system outputs corrective suggestions based on this, trainees may be guided to repeat incorrect force exertion, placing additional load on the horse's back, negatively impacting both teaching effectiveness and animal welfare. This inaccuracy in recognition is becoming a key technical bottleneck hindering the implementation of posture recognition equestrian assistance systems.

[0005] Therefore, this invention provides an intelligent auxiliary method and system for equestrian teaching based on motion posture recognition. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent auxiliary method and system for equestrian teaching based on motion posture recognition. Addressing the signal mixing problem caused by rider-horse motion coupling, the invention employs the following steps: Step 1: Simultaneous acquisition and fusion of multimodal data to generate a high-quality data foundation; Step 2: High-precision separation of rider and horse signals, suppressing horse gait interference, through dual-channel adversarial coding decoupling and coupling reliability coefficient generation; Step 3: Interpretable posture deviation curves and movement level labels are output through knowledge graph mapping and adaptive normalization of gait residuals; Step 4: Dynamically selecting feedback channels based on posture deviation thresholds and link status to provide real-time correction prompts to the rider; Step 5: Post-lesson semi-supervised incremental training to update model weights and continuously optimize performance; This improves teaching effectiveness and solves the technical problems described in the background section.

[0007] (2) Technical solution To achieve the above objectives, the present invention provides the following technical solution: The intelligent auxiliary method for equestrian teaching based on motion posture recognition includes: synchronously collecting visible light-depth dual-stream data, rider inertial measurement unit data and saddle pressure array signals under a unified clock; completing primary clustering and subject label allocation based on the prior of the placement points; and generating a fusion frame sequence with subject labels. During the dual-channel adversarial coding decoupling process, the mutual information difference between rider joint IMU data and saddle pressure array signal, as well as the pressure time-frequency energy dispersion, are integrated in real time to generate a coupling reliability coefficient and dynamically calibrate the rider posture branch weights. The rider's independent posture flow is mapped to the knowledge graph baseline node, the continuous joint deviation vector is calculated by weighting according to the coupling confidence coefficient, and adaptive normalization is performed by combining the gait residual, and the posture deviation curve and action level label are output. Based on the posture deviation curve threshold and the current link status, the system dynamically selects voice, haptic, or augmented reality feedback channels and issues corrective prompts to the rider within a single riding beat. After class, the posture deviation curve, feedback effectiveness labels, and gait residuals are uploaded to the cloud to perform semi-supervised incremental training, and then the updated weight package is sent back to the edge during the link window.

[0008] Furthermore, visible light image data, depth point cloud data, rider inertial measurement unit data, and saddle pressure array data are collected synchronously using a unified clock. Based on prior knowledge of the equipment layout, the rider inertial measurement unit data and saddle pressure array data are assigned main labels to generate rider data and horse data.

[0009] Furthermore, rider and horse skeleton points are extracted from visible light image data and depth point cloud data using a skeleton detection algorithm, and subject labels are assigned.

[0010] Furthermore, time stamp alignment technology is used to perform time alignment on visible light image data, depth point cloud data, rider inertial measurement unit data, and saddle pressure array data; Interpolation is used to generate a fused frame sequence that is consistent with the instantaneous movement. Each frame contains a visible light image, a depth image, rider inertial measurement unit data, saddle pressure array data, and a main label.

[0011] Furthermore, by separating rider posture features and horse gait features through a dual-channel adversarial coding decoupling architecture, the mutual information difference between rider inertial measurement unit data and saddle pressure array signal, as well as the pressure time-spectral energy dispersion of saddle pressure array signal, are calculated to generate a coupling reliability coefficient. By using the coupling confidence coefficient to dynamically calibrate the weights of rider posture features, the interference of horse gait feature fluctuations on rider posture feature interpretation is suppressed.

[0012] Furthermore, the rider's independent posture flow is mapped to the baseline node in the equestrian movement knowledge graph; The joint position deviation vector between the rider's independent posture flow and the reference node is calculated by matching the closest standard motion sequence using a dynamic time warping algorithm.

[0013] Furthermore, the joint position deviation vector is weighted and adjusted based on the coupling confidence coefficient, the gait residual in the latent representation of the horse's gait is extracted, the weighted joint position deviation vector is corrected using a nonlinear normalization function, continuous posture deviation curves and movement level labels are generated, and interpretable feedback information is output.

[0014] Furthermore, the rider's posture deviation level is determined based on the average deviation value of the posture deviation curve and the preset threshold range, and the current link status is assessed by real-time monitoring of network latency, bandwidth and packet loss rate. Based on the deviation level and the current link status, the system dynamically selects voice feedback, haptic feedback, or augmented reality feedback channels, generates correction prompts, and sends them to the selected feedback channels within a single riding beat.

[0015] Furthermore, the posture deviation curve, feedback effect label, and gait residual are transmitted to the cloud via an upload operation, and an updated weight package is generated using a semi-supervised incremental training method. During the link window, the updated weight packet is sent back to the edge and applied to the attitude analysis network and the feedback optimization network respectively through weight update operations.

[0016] An intelligent auxiliary system for equestrian teaching based on motion posture recognition, including: The data acquisition and processing module synchronously acquires visible light-depth dual-stream data, rider inertial measurement unit data, and saddle pressure array signals under a unified clock. Based on the prior knowledge of the deployment points, it completes primary clustering and main tag allocation, and generates a fused frame sequence with main tags. The posture calibration module, during the dual-channel adversarial coding decoupling process, integrates the mutual information difference between rider joint IMU data and saddle pressure array signals, as well as the pressure time-frequency energy dispersion, in real time to generate a coupling reliability coefficient and dynamically calibrate the rider posture branch weights. The deviation calculation module maps the rider's independent posture flow to the knowledge graph baseline node, calculates the continuous joint deviation vector by weighting according to the coupling confidence coefficient, and performs adaptive normalization by combining the gait residual, outputting the posture deviation curve and action level label. The feedback correction module dynamically selects voice, haptic, or augmented reality feedback channels based on the posture deviation curve threshold and the current link status, and issues correction prompts to the rider within a single riding beat. The module is optimized and updated. After class, the posture deviation curve, feedback effect label and gait residual are uploaded to the cloud to perform semi-supervised incremental training. Then, the updated weight package is sent back to the edge during the link window.

[0017] (III) Beneficial Effects This invention provides an intelligent auxiliary method and system for equestrian teaching based on motion posture recognition, which has the following beneficial effects: By synchronously acquiring multimodal data under a unified clock, performing primary clustering and time alignment, a fused frame sequence with subject labels is generated, providing a high-quality data foundation for subsequent analysis.

[0018] By employing a dual-channel adversarial coding decoupling technique, and combining mutual information differences and stress-time spectral energy dispersion to generate a coupling reliability coefficient, the rider posture branch weights are dynamically calibrated, effectively suppressing the interference of horse gait fluctuations on rider action interpretation. The two steps work together to achieve effective separation of rider and horse signals for the first time, solving the problem that traditional posture recognition algorithms cannot distinguish between the two actions, and significantly improving the accuracy of posture interpretation.

[0019] By mapping rider's independent posture flow to baseline nodes in a knowledge graph, weighted calculation of continuous joint deviation vectors based on coupling confidence coefficients, and adaptive normalization combined with gait residuals, interpretable posture deviation curves and action level labels are output. This process not only enhances the scientific rigor of posture analysis but also provides intuitive and quantitative evidence for teaching. Compared to traditional methods that only provide vague action evaluations, the dynamic mapping and adaptive normalization of the knowledge graph generate more targeted and interpretable feedback information, providing clear guidance for riders to correct their movements.

[0020] Based on the posture deviation curve threshold and the current link status, the system dynamically selects voice, haptic, or augmented reality feedback channels and issues corrective prompts to the rider within a single riding beat, ensuring low latency and contextual matching of feedback, significantly improving the real-time performance and effectiveness of teaching interaction. The dynamic feedback selection technology optimizes the transmission path according to the actual scenario, avoiding the delay or information mismatch problems of the traditional fixed feedback mode, enabling riders to adjust their posture in a timely manner during riding, thereby enhancing the learning experience and training effect.

[0021] After class, posture deviation curves, feedback effectiveness labels, and gait residuals are uploaded to the cloud for semi-supervised incremental training. Updated weight packets are then fed back to the edge during the link window, allowing the model to continuously adapt as the teaching progresses. Utilizing unlabeled data enables long-term model optimization, reducing the cost of manual annotation and ensuring system stability and high accuracy across different riders, horses, and scenarios. Compared to traditional static models, this incremental training mechanism demonstrates significant innovation, achieving an intelligent upgrade of the teaching system. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the intelligent auxiliary method for equestrian teaching based on motion posture recognition according to the present invention. Figure 2 This is a schematic diagram of the intelligent auxiliary system for equestrian teaching based on motion posture recognition, as described in this invention. Detailed Implementation

[0023] 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.

[0024] Please see Figure 1 This invention provides an intelligent auxiliary method for equestrian teaching based on motion posture recognition, including: Step 1: Synchronously acquire visible light image data, depth point cloud data, rider inertial measurement unit (IMU) data, and saddle pressure array data using a unified clock. Based on prior equipment layout, assign subject labels to the rider IMU data and saddle pressure array data to generate rider and horse data. Simultaneously, extract rider and horse skeleton points from the visible light image data and depth point cloud data using a skeleton detection algorithm and assign subject labels. Use timestamp alignment technology to time-align the visible light image data, depth point cloud data, rider IMU data, and saddle pressure array data. Through interpolation processing, generate a fused frame sequence with consistent instantaneous movement. Each frame contains a visible light image, depth image, rider IMU data, saddle pressure array data, and subject labels.

[0025] Step one includes the following: Step 101: Data Acquisition and Synchronization During the data acquisition and synchronization phase, multiple devices are used to capture the movement information of riders and horses, and this information is kept consistent over time.

[0026] Specifically, a combined system of visible light and depth cameras records two-dimensional image data of the rider and three-dimensional depth point cloud data of both the rider and horse. Simultaneously, inertial measurement unit sensors are deployed at key joints such as the rider's shoulders, elbows, hips, and knees to collect three-axis acceleration and angular velocity data. Furthermore, a pressure sensor array is installed under the saddle to measure pressure distribution data, reflecting the horse's gait characteristics. To achieve data synchronization, all devices employ a unified clock control mechanism, operate at the same sampling frequency, and add a timestamp to each frame of data. The timestamp alignment process relies on network time protocols or hardware trigger signals to keep the time deviation between data streams generated by different devices within milliseconds. Time consistency of multi-source data is a prerequisite for subsequent analysis, avoiding information confusion caused by time misalignment; synchronous acquisition provides a reliable time reference for the integration of multimodal data, improving the accuracy of subsequent processing.

[0027] Step 102: Primary Clustering and Main Label Allocation In the initial clustering and main label allocation stage, the collected data are classified and labeled according to prior knowledge of equipment layout.

[0028] First, based on the preset positions of the inertial measurement unit sensors at the rider's key joints, the triaxial acceleration and angular velocity data are explicitly associated with the rider and labeled as "rider data." Second, the pressure distribution data generated by the pressure sensor array under the saddle is associated with the horse's gait characteristics and labeled as "horse data." Next, using deep learning-based skeletal detection algorithms, such as OpenPose or MediaPipe, key skeletal points of the rider and horse are extracted from visible light images and depth point cloud data, labeled as "rider skeletal points" and "horse skeletal points," respectively. Based on this, a clear subject label, "rider" or "horse," is assigned to each type of data point, generating a preliminary dataset containing these labels. Implementing primary clustering and subject label assignment, clearly defining data attribution, distinguishes the source of movement between the rider and horse, avoiding signal mixing. This classification and labeling method provides a clear basis for subsequent separation of motion signals, improving the targeting of data processing.

[0029] Step 103: Time Alignment and Data Fusion During the time alignment and data fusion phase, data streams from different sources are time-calibrated and integrated into a unified sequence. First, using the timestamps of each data frame as a reference, interpolation is used to adjust all data streams, ensuring that visible light image data, depth point cloud data, inertial measurement unit sensor data, and pressure sensor array data correspond to the same instant of action at any given time point. Specifically, interpolation involves estimating the data values ​​for missing time points based on the chronological order of the timestamps, ensuring the continuity and consistency of the data along the timeline.

[0030] After time alignment, these multimodal data are integrated into a fused frame sequence. Each frame contains a visible light image, a depth image, rider inertial measurement unit data, saddle pressure sensor array data, and the corresponding subject label. Time alignment is crucial for the comprehensive analysis of multimodal data, preventing motion matching errors caused by time discrepancies. Data fusion integrates scattered information into a cohesive whole for comprehensive analysis. The fused frame sequence provides a complete and consistent data structure for subsequent signal decoupling and motion recognition, improving processing efficiency and result reliability.

[0031] In practice, the process involves three stages: data acquisition and synchronization, initial segmentation and subject label allocation, and time alignment and data fusion, generating a fused frame sequence with subject labels. This process first uses multiple devices to collect rider and horse motion information and achieves synchronization through timestamp alignment; then, the data is classified and labeled according to device placement and skeletal detection algorithms; finally, interpolation and integration generate a unified fused frame sequence. The synchronization, clear classification, and complete integration of multimodal data are crucial for rider and horse motion analysis, effectively distinguishing rider and horse motion signals, reducing the influence of signal confounding, and significantly improving the accuracy of the analysis results.

[0032] Step 2: Separate rider posture features and horse gait features using a dual-channel adversarial coding decoupling architecture. Calculate the mutual information difference between the rider inertial measurement unit data and the saddle pressure array signal, as well as the pressure time-spectral energy dispersion of the saddle pressure array signal. Generate a coupling reliability coefficient. Use this coefficient to dynamically calibrate the weights of the rider posture features and suppress the interference of horse gait feature fluctuations on the interpretation of rider posture features.

[0033] Step two includes the following: Step 201: Dual-channel adversarial coding decoupling In the dual-channel adversarial coding decoupling stage, an adversarial learning framework is used to separate the motion signals of the rider and the horse.

[0034] Specifically, two independent encoding channels are constructed: the first channel processes the rider's inertial measurement unit data (including acceleration and angular velocity) and visual skeleton data (joint positions), extracting the rider's independent posture features through the encoder; the second channel processes the saddle pressure array signal (pressure distribution) and the horse's visual skeleton data (limb and hoof positions), extracting the horse's gait features. After the encoder transforms the input data into latent feature representations of the rider and horse, a discriminator is introduced to attempt to distinguish between these two types of features.

[0035] The encoder is optimized to make it difficult for the discriminator to distinguish between rider and horse features, thus achieving feature decoupling between the two. An adversarial learning framework is employed, which effectively separates signal features from different sources, reducing mutual interference between rider and horse motion signals. The decoupled rider posture features more accurately reflect the rider's movements, reducing the impact of horse gait fluctuations on rider feature extraction.

[0036] Step 202: Mutual Information Difference Calculation During the mutual information difference calculation phase, the degree of coupling between rider motion signals and horse motion signals is assessed.

[0037] First, temporal correlation analysis is performed on rider inertial measurement unit (IMU) data and saddle pressure array signals within a fixed time window to calculate the actual mutual information, reflecting the strength of the correlation between the two. Next, baseline mutual information is generated through a permutation test: the time series of rider IMU data is randomly shuffled and re-paired with the saddle pressure array signals, and the mutual information is calculated repeatedly to obtain the mean and distribution fluctuation of the mutual information when there is no correlation. Then, the difference between the actual mutual information and the baseline mutual information is calculated, and the result is normalized to keep it within a reasonable range. The smaller the difference in mutual information, the weaker the coupling between the rider's motion signal and the horse's motion signal; conversely, the greater the difference, the more significant the coupling interference. By quantifying the coupling strength between rider and horse signals, an objective basis can be provided for subsequent calibration of rider posture characteristics. This quantification method can accurately reflect the degree of mutual influence between signals and improve the accuracy of posture analysis.

[0038] Step 203: Calculation of spectral energy dispersion under pressure In the stage of calculating the spectral energy dispersion under pressure, the frequency domain characteristics of the saddle pressure array signal are analyzed to detect gait anomalies in horses. First, wavelet decomposition is performed on the saddle pressure array signal, dividing it into time-frequency spectra across multiple frequency bands, and the energy distribution of each band is extracted. Next, the spectral entropy of the energy distribution in each band is calculated. The spectral entropy is obtained by taking the logarithm of the normalized energy proportion and then weighted summing, which characterizes the degree of dispersion of the energy distribution. Then, the calculated spectral entropy is compared with the individual baseline spectral entropy under normal gait, and the relative distance between the two is calculated, reflecting the change in energy dispersion caused by gait fluctuations. A larger relative distance indicates a more significant gait anomaly in the horse. Calculating the spectral energy dispersion under pressure, frequency domain analysis can capture subtle features of gait fluctuations, compensating for the shortcomings of time domain analysis. By introducing frequency domain features, the impact of horse gait on rider posture characteristics can be more comprehensively assessed, improving the sensitivity of anomaly detection.

[0039] Step 204: Generation of Coupling Credibility Coefficients In the coupling confidence coefficient generation stage, the mutual information difference and the frequency spectrum energy dispersion under pressure are input into the fuzzy logic system. Through predefined fuzzy rules and member functions, a coupling confidence coefficient is generated, with a range limited to 0 to 1. The fuzzy rules determine the strength of coupling interference based on the magnitude of the mutual information difference and the frequency spectrum energy dispersion under pressure: when both the mutual information difference and energy dispersion are small, the coupling confidence coefficient is close to 1, indicating weak coupling interference; when either value is large, the coupling confidence coefficient is close to 0, indicating strong coupling interference. By combining the mutual information difference in the time domain and the energy dispersion in the frequency domain, the confidence of rider posture characteristics can be assessed more accurately. The fuzzy logic system can effectively handle uncertainties and nonlinear relationships, providing a smooth and reasonable basis for weight adjustment.

[0040] During the dynamic calibration of rider posture branch weights, the weights of rider posture features are adjusted based on the coupling confidence coefficient.

[0041] Specifically, the coupling confidence coefficient is multiplied by the latent feature representation of the rider posture channel to generate weighted rider posture features. When the coupling confidence coefficient is low, the weight of the rider posture features is reduced, thereby suppressing the interference of horse gait fluctuations on rider movement interpretation. Dynamically calibrating the rider posture branch weights and adjusting the importance of features according to the real-time coupling degree ensures the accuracy of rider posture interpretation. This adaptive adjustment mechanism maintains the robustness of posture recognition in different scenarios and improves the reliability of the equestrian teaching assistance system.

[0042] In practice, this system employs five stages: dual-channel adversarial coding decoupling, mutual information difference calculation, stress-time spectral energy dispersion calculation, coupling credibility coefficient generation, and dynamic calibration of rider posture branch weights. This process achieves precise decoupling of rider and horse motion signals and calibration of rider posture features. First, an adversarial learning framework is used to separate rider and horse features. Then, the coupling degree between the two is quantified using time-domain mutual information difference and frequency-domain energy dispersion. Next, a coupling credibility coefficient is generated based on a fuzzy logic system. Finally, the weights of rider posture features are dynamically adjusted according to the coupling credibility coefficient. By combining time-domain and frequency-domain analysis methods with the integrated application of adversarial learning and fuzzy logic, the system effectively suppresses the interference of horse gait fluctuations on rider posture interpretation, significantly improving the accuracy and robustness of posture recognition and providing more reliable technical support for equestrian teaching.

[0043] Step 3: Map the rider's independent posture flow to the reference node in the equestrian movement knowledge graph. Use the dynamic time warping algorithm to match the closest standard movement sequence, calculate the joint position deviation vector between the rider's independent posture flow and the reference node, and adjust the joint position deviation vector by weighting according to the coupling confidence coefficient. Extract the gait residual from the horse's gait latent representation, correct the weighted joint position deviation vector using a nonlinear normalization function, generate continuous posture deviation curves and movement level labels, and output interpretable feedback information.

[0044] Step three includes the following: Step 301: Mapping the Knowledge Graph Baseline Nodes In the knowledge graph baseline node mapping stage, an equestrian movement knowledge graph is first constructed, which contains baseline nodes for standard rider movements. Each baseline node corresponds to a keyframe of a standard movement, recording the rider's joint angles and positional features in that movement, such as shoulder angle and hip position. Next, the rider's independent posture stream obtained in step two (i.e., the weighted time series of rider posture features) is matched with the baseline nodes in the equestrian movement knowledge graph.

[0045] The specific matching process employs a dynamic time warping algorithm. It determines the closest baseline node sequence by comparing the similarity between the rider's independent posture flow time series and the standard motion sequence. Similarity calculation is based on the cumulative distance of joint features in the time series, and the optimal matching path is found by adjusting the time axis alignment. Using predefined features of the standard motion as a reference, it accurately identifies the corresponding standard of the rider's actual movements, facilitating subsequent deviation analysis. The dynamic time warping algorithm can adapt to differences in movement speed and rhythm, ensuring the accuracy of the matching results and thus improving the reliability of posture evaluation.

[0046] Step 302: Calculation of continuous joint deviation vector In the continuous joint deviation vector calculation stage, for each time point in the rider's independent posture flow, the joint position features are compared with the joint position features in the corresponding reference node determined in the knowledge graph reference node mapping stage, and the difference between the two is calculated.

[0047] Specifically, if the shoulder position in the rider's independent posture flow differs from the shoulder position of the reference node, the direction and magnitude of this difference are recorded as a vector, forming a deviation vector. Next, the deviation vector is weighted and adjusted using the coupling confidence coefficient (ranging from 0 to 1) generated in step two. The coupling confidence coefficient represents the degree to which the horse's gait influences the rider's posture; a smaller value indicates a greater influence from the horse's gait, thus reducing the weight of the deviation vector accordingly; conversely, a larger value preserves the original magnitude of the deviation vector. The weighting adjustment process is achieved by multiplying the deviation vector by the coupling confidence coefficient. Calculating the deviation vector for continuous joints quantifies the specific differences between the rider's movement and the standard movement, providing accurate data support for teaching feedback. By introducing the coupling confidence coefficient for weighting, the interference of abnormal horse gait on rider posture assessment can be effectively reduced, improving the accuracy of deviation calculation.

[0048] Step 303: Adaptive normalization of gait residuals In the adaptive normalization stage of gait residuals, the latent representation of the horse's gait is first extracted from step two. The gait characteristics are analyzed and compared with predefined standard gait characteristics. The difference between the two is calculated, called the gait residual. For example, if the horse's hip height deviates from the expected value of the standard gait, the deviation is recorded as a residual. Next, a nonlinear normalization function is designed to adjust the weighted deviation vector obtained in the continuous joint deviation vector calculation stage according to the magnitude of the gait residual. Specifically, when the gait residual is large, the nonlinear normalization function proportionally reduces the amplitude of the deviation vector; when the gait residual is small, the amplitude of the deviation vector remains basically unchanged. The adjusted deviation vector reflects the rider's posture deviation after eliminating the influence of abnormal horse gait. Adaptive normalization of gait residuals, by correcting the interference of horse gait fluctuations on rider posture assessment, enables the assessment results to more accurately reflect the quality of the rider's movements. This adaptive mechanism ensures the stability of the system under different horse gait conditions and enhances the robustness of posture assessment.

[0049] Step 304: Generation of Posture Deviation Curve and Motion Level Labels In the posture deviation curve and motion level label generation stage, the deviation vectors obtained in the gait residual adaptive normalization stage are concatenated in chronological order to form a continuous posture deviation curve, which describes the rider's motion changing over time. Then, based on the magnitude of the posture deviation curve and a preset threshold range, corresponding motion level labels are generated.

[0050] The specific classification method is as follows: if the deviation range is between 0 and 5 degrees, it is marked as "Excellent"; if it is between 5 and 10 degrees, it is marked as "Good"; if it is between 10 and 15 degrees, it is marked as "Pass"; and if it is greater than 15 degrees, it is marked as "Fail". Furthermore, by combining the results of the knowledge graph baseline node mapping stage and the normalized deviation vector, interpretable feedback information is generated, such as "right knee flexion less than 5 degrees" or "left shoulder elevation deviation 10 degrees". The generation of posture deviation curves and movement level labels presents the quality of rider movements in an intuitive and quantitative way, facilitating understanding and improvement by both instructors and trainees. This method not only provides an overall level of movement assessment but also gives detailed feedback on specific deviation areas, effectively improving the relevance and practicality of teaching.

[0051] In practice, step three involves four stages: mapping reference nodes from the knowledge graph, calculating continuous joint deviation vectors, adaptive normalization of gait residuals, and generating posture deviation curves and action level labels. This process achieves precise quantitative evaluation of rider movements. First, using an equestrian movement knowledge graph and a dynamic time warping algorithm, the rider's independent posture flow is matched with standard movements. Then, by comparing joint position features and combining them with a coupling confidence coefficient, the deviation vector is calculated and adjusted. Next, based on gait residuals and a nonlinear normalization function, the deviation vector is further corrected. Finally, posture deviation curves and action level labels are generated, providing intuitive feedback. By comprehensively considering standard movement references, the influence of horse gait, and deviation correction, this method accurately reflects the rider's true skill level. It significantly improves the accuracy and adaptability of posture evaluation, providing reliable technical support for equestrian teaching, while the generated feedback information has high practical value.

[0052] Step 4: Determine the rider's posture deviation level based on the average deviation value of the posture deviation curve and the preset threshold range. Assess the current link status by monitoring network latency, bandwidth and packet loss rate in real time. Dynamically select voice feedback, haptic feedback or augmented reality feedback channel according to the deviation level and the current link status, generate correction prompts and send them to the selected feedback channel within a single riding beat.

[0053] Step four includes the following: Step 401: Attitude Deviation Curve Threshold Analysis In the posture deviation curve threshold analysis stage, a threshold range for the posture deviation curve is first preset to classify the degree of rider posture deviation. Specifically, a deviation angle between 0 and 5 degrees is defined as "normal," between 5 and 10 degrees as "slight deviation," between 10 and 15 degrees as "moderate deviation," and greater than 15 degrees as "severe deviation." These threshold ranges are set based on the analysis results of the impact of rider joint angle deviation on movement standardization. Next, the posture deviation curve generated in step three is analyzed in real time to calculate the average deviation value within the current time window (set as the most recent 5 seconds). The calculation method is to add up all values ​​of the posture deviation curve within the most recent 5 seconds and divide by the length of the time window to obtain the average deviation value. Then, the calculated average deviation value is compared with the preset threshold range to determine the current rider posture deviation level. For example, when the average deviation value exceeds 5 degrees, it is judged as "slight deviation" or a higher level, and the subsequent feedback mechanism is triggered. By quantifying and classifying the degree of rider posture deviation, we can provide an objective basis for selecting subsequent feedback channels. This analysis method can achieve differentiated treatment based on the specific severity of the deviation, ensuring that feedback measures match actual needs and improving the accuracy and practicality of feedback.

[0054] Step 402: Current Link Status Assessment During the current link status assessment phase, network parameters between the edge computing platform and feedback devices (including voice headsets, haptic devices, and head-mounted displays) are collected in real time, specifically including network latency, bandwidth, and packet loss rate. These network parameters are directly measured through the communication module of the edge computing platform. Next, based on predefined evaluation criteria, the network status is categorized into three levels: "Good," "Medium," and "Poor." The specific evaluation criteria are as follows: network latency less than 50 milliseconds is defined as "Good," 50 to 100 milliseconds as "Medium," and greater than 100 milliseconds as "Poor"; bandwidth greater than 10 megabits per second is defined as "Good," 5 to 10 megabits per second as "Medium," and less than 5 megabits per second as "Poor"; packet loss rate less than 1% is defined as "Good," 1% to 5% as "Medium," and greater than 5% as "Poor."

[0055] In the comprehensive evaluation, the worst of the three parameters—network latency, bandwidth, and packet loss rate—is taken as the final grade of the current link status. For example, if the network latency is 30 milliseconds (good), the bandwidth is 7 megabits per second (medium), and the packet loss rate is 6% (poor), then the current link status is judged as "poor." The current link status assessment is crucial because network status directly affects the transmission quality and timeliness of feedback information. Quantitative evaluation provides a reliable basis for selecting feedback channels. This evaluation method reflects real-time network performance, ensuring that the selection of subsequent feedback channels is adapted to network conditions, and improving the stability and reliability of feedback.

[0056] Step 403: Dynamic Selection of Feedback Channels In the dynamic selection phase of feedback channels, the characteristics of each feedback channel are first clarified: voice feedback transmits specific instructions through voice headsets, is sensitive to network latency (requiring network latency of less than 100 milliseconds), and has low bandwidth requirements (approximately 0.1 megabits per second); haptic feedback transmits simple signals through haptic devices, has extremely low network latency requirements (less than 50 milliseconds), and extremely low bandwidth requirements (approximately 0.01 megabits per second); augmented reality feedback provides visual guidance through head-mounted displays, and has higher requirements for network latency and bandwidth (requiring network latency of less than 50 milliseconds and bandwidth greater than 10 megabits per second). Next, based on the deviation level determined in the posture deviation curve threshold analysis phase and the link status determined in the current link status assessment phase, a suitable feedback channel is dynamically selected.

[0057] Specifically: when the link status is "good" and the deviation level is "severe deviation" (average deviation greater than 15 degrees), augmented reality feedback is selected to provide detailed visual correction guidance; when the link status is "medium" and the deviation level is "moderate deviation" (average deviation between 10 and 15 degrees), voice feedback is selected to convey specific action adjustment instructions; when the link status is "poor" or the deviation level is "slight deviation" (average deviation between 5 and 10 degrees), haptic feedback is selected to provide simple prompts for riders to adjust their posture. Furthermore, the deviation level and link status are reassessed every second, and the selected feedback channel is updated to ensure the selection matches the current situation. The dynamic selection of feedback channels, where different channels are applicable to different network conditions and deviation levels, optimizes the delivery of feedback. This method allows for flexible adaptation to different network environments and rider postures, improving the relevance and efficiency of teaching interaction.

[0058] Step 404: Generating and Sending Correction Notices In the correction prompt generation and transmission phase, specific correction prompts are first generated based on the posture deviation curve and movement level label output in step three. For example, if the posture deviation curve shows that the rider's right knee joint deviation is +8 degrees, the prompt "Please increase the right knee flexion angle by 8 degrees" is generated. The prompt content is generated based on the predefined mapping rules between knowledge graph baseline nodes and deviation vectors. Next, it is ensured that the correction prompt generation and transmission are completed within a single riding beat (one gait cycle for a horse, approximately 1 to 2 seconds).

[0059] Specifically, the total latency from detecting a deviation to sending a prompt must be less than one gait cycle to ensure real-time feedback. The transmission latency is determined by the processing speed of the edge computing platform and the network latency of the selected feedback channel. Finally, the generated corrective prompt is sent to the selected feedback channel via the edge computing platform, using a low-latency communication protocol (such as WebRTC) to minimize transmission latency. For example, if voice feedback is selected, the prompt is sent as an audio stream; if augmented reality feedback is selected, it is sent as rendered image data. Timely and accurate feedback generation and transmission of corrective prompts are crucial for improving rider skills and protecting horse welfare. This approach ensures the real-time and targeted nature of the feedback, effectively improving the quality and practical effect of teaching interaction.

[0060] In practice, this system employs four stages: posture deviation curve threshold analysis, current link status assessment, dynamic selection of feedback channels, and generation and transmission of correction prompts. This achieves a low-latency, context-appropriate real-time feedback mechanism for equestrian teaching. First, the average deviation value of the posture deviation curve is calculated and the deviation level is determined, quantifying the degree of rider posture deviation. Then, the current link status is assessed by real-time monitoring of network latency, bandwidth, and packet loss rate. Next, voice feedback, haptic feedback, or augmented reality feedback channels are dynamically selected based on the deviation level and link status. Finally, specific correction prompts are generated and sent within a single riding beat. By comprehensively analyzing rider posture deviation and network conditions, the system provides riders with the most context-appropriate feedback method, ensuring timeliness and accuracy. This method significantly improves the flexibility and adaptability of teaching interaction, effectively enhancing the efficiency of rider posture correction and playing a positive role in protecting horse welfare.

[0061] Step 5: After training, upload the posture deviation curve, feedback effectiveness labels, and gait residuals to the cloud, perform semi-supervised incremental training, generate updated weight packages, and transmit them back to the edge during the link window to update the weights of the posture analysis network and the feedback optimization network. The specific technical features can be summarized as follows: Upload the posture deviation curve, feedback effectiveness labels, and gait residuals to the cloud; generate updated weight packages using semi-supervised incremental training; transmit the updated weight packages back to the edge during the link window; and apply the updated weights to the posture analysis network and the feedback optimization network respectively.

[0062] Step five includes the following: Step 501: Data Upload and Preprocessing During the data upload and preprocessing phase, key data is first collected and uploaded to the cloud after each equestrian lesson. Specifically, the uploaded data includes posture deviation curves, feedback effectiveness labels, and gait residuals.

[0063] The posture deviation curve, derived from the calculation results in step three, represents the continuous joint deviation between the rider's movement and the standard movement, reflecting the rider's movement characteristics. Feedback effectiveness labels record the effect of the rider's adjustments after receiving feedback in step four, categorized as "effective" and "ineffective," manually labeled by the coach or generated by an automated evaluation system. Gait residuals, also derived from step three, represent the difference between the horse's gait and the standard gait, used to analyze the potential impact of horse movement on the rider's posture. Next, these data undergo preprocessing: for the posture deviation curve, a temporal segmentation method is used, extracting key movement segments using a fixed time window (e.g., 10 frames per second) to ensure the data granularity is suitable for subsequent model training; for the feedback effectiveness labels, they are encoded in binary form, where "effective" is converted to a value of 1 and "ineffective" to a value of 0, for use in the subsequent supervised learning process; for the gait residuals, normalization is implemented using a min-max normalization method, calculating the difference between the gait residual and the minimum and maximum values ​​in the current lesson, then dividing by the difference between the maximum and minimum values ​​to scale the data to a range of 0 to 1.

[0064] By standardizing and structuring the data, the quality and consistency of the uploaded data are ensured, enabling the model to effectively identify and utilize it. The preprocessed data has a standardized format and uniform scale, which can improve the efficiency and accuracy of model training and provide high-quality input data for subsequent semi-supervised incremental training.

[0065] Step 502, Semi-supervised incremental training In the semi-supervised incremental training phase, the cloud utilizes preprocessed data to optimize the deep learning model, which consists of two parts: a pose analysis network and a feedback optimization network. The pose analysis network receives pose deviation curves and normalized gait residuals to extract rider pose features; the feedback optimization network receives pose features and feedback effectiveness labels to optimize the generation of feedback strategies. The training process employs a semi-supervised learning method, divided into supervised training and unsupervised training.

[0066] In supervised training, the effective input labels and the feedback probability predicted by the feedback optimization network are optimized using a cross-entropy loss function. Specifically, this involves calculating the logarithmic loss between the true label and the predicted probability to measure the model's prediction accuracy and adjust model parameters. In unsupervised training, the input posture deviation curve and normalized gait residuals are used to reconstruct the input data using an autoencoder. The loss function is a weighted combination of mean squared errors, meaning that the difference between the reconstructed data and the original data is calculated, and the weights of the gait residuals are adjusted according to a preset balance coefficient to optimize the model's ability to represent the data. Furthermore, the training process employs an incremental training strategy, fine-tuning the model's weights based on existing model weights, updating only parameters relevant to new data each time. The learning rate is scheduled using cosine annealing, dynamically adjusting the learning rate through a cosine function to gradually and smoothly decrease during training. Semi-supervised incremental training, by combining a small amount of labeled data with a large amount of unlabeled data, can fully utilize existing resources and improve the model's adaptability to rider and horse movement characteristics. This training method can effectively cope with dynamic data changes, maintain model accuracy, reduce reliance on manual annotation, and enhance the system's adaptability.

[0067] Step 503, Weight Update and Backhaul During the weight update and backhaul phase, after training is completed, the optimized model weights are backhauled from the cloud to the edge to ensure that the teaching equipment and the cloud model remain synchronized.

[0068] Specifically, the process begins by extracting the weight changes from the trained model and generating compressed weight packets that only contain the differences compared to the previous version, reducing the amount of data transmitted. Next, the network status between the cloud and the edge is monitored, and transmission is performed during periods of low load (e.g., nighttime) to ensure that the transmission does not interfere with real-time teaching activities. An incremental update strategy is employed during transmission, transmitting only the differential weights, and an automatic retransmission request protocol is used to ensure reliable transmission by checking data integrity and retransmitting lost portions when necessary. Upon receiving the weight packets, the edge device decompresses them and loads them into the pose analysis network and feedback optimization network. During loading, a linear interpolation method is used to smoothly transition between the old and new weights, gradually increasing the proportion of the new weights (from 0 to 1) to fuse the old and new weights, ensuring the smoothness of model updates. Weight updates and backhaul, by synchronizing the cloud training results to the edge device, ensure that the teaching equipment can use the latest optimized model, improving its real-time performance. This mechanism continuously improves model performance without interrupting the teaching process, enhancing system stability and long-term adaptability, and providing technical support for continuous optimization of riders and horses.

[0069] In practice, the intelligent auxiliary method for equestrian teaching achieves continuous optimization and improved adaptability of the model through three stages: data upload and preprocessing, semi-supervised incremental training, and weight update and feedback. In the data upload and preprocessing stage, posture deviation curves, feedback effectiveness labels, and gait residuals are collected and normalized to ensure high-quality and consistent training data. In the semi-supervised incremental training stage, a combination of supervised and unsupervised learning is used to optimize the posture analysis network and feedback optimization network, enabling the model to more accurately identify the movement characteristics of riders and horses. In the weight update and feedback stage, the optimized model weights are synchronized to the edge to ensure the real-time performance and accuracy of the teaching equipment. Utilizing a post-lesson data-driven model optimization method effectively adapts to the dynamic changes in rider and horse movement characteristics, maintaining the system's high precision and adaptability. This method not only improves the accuracy of posture recognition and the effectiveness of feedback strategies but also reduces the need for manual intervention, providing a reliable technical guarantee for the long-term stable operation of equestrian teaching.

[0070] See also Figure 2 This invention provides an intelligent auxiliary system for equestrian teaching based on motion posture recognition, comprising: The data acquisition and processing module synchronously acquires visible light-depth dual-stream data, rider inertial measurement unit data, and saddle pressure array signals under a unified clock. Based on the prior knowledge of the deployment points, it completes primary clustering and main tag allocation, and generates a fused frame sequence with main tags. The posture calibration module, during the dual-channel adversarial coding decoupling process, integrates the mutual information difference between rider joint IMU data and saddle pressure array signals, as well as the pressure time-frequency energy dispersion, in real time to generate a coupling reliability coefficient and dynamically calibrate the rider posture branch weights. The deviation calculation module maps the rider's independent posture flow to the knowledge graph baseline node, calculates the continuous joint deviation vector by weighting according to the coupling confidence coefficient, and performs adaptive normalization by combining the gait residual, outputting the posture deviation curve and action level label. The feedback correction module dynamically selects voice, haptic, or augmented reality feedback channels based on the posture deviation curve threshold and the current link status, and issues correction prompts to the rider within a single riding beat. The module is optimized and updated. After class, the posture deviation curve, feedback effect label and gait residual are uploaded to the cloud to perform semi-supervised incremental training. Then, the updated weight package is sent back to the edge during the link window.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0072] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0073] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent auxiliary method for equestrian teaching based on motion posture recognition, characterized in that: include, Visible light-depth dual-stream data, rider inertial measurement unit data and saddle pressure array signals are synchronously acquired under a unified clock. Based on the prior of the deployment points, primary clustering and main tag allocation are completed to generate a fused frame sequence with main tags. During the dual-channel adversarial coding decoupling process, the mutual information difference between rider joint IMU data and saddle pressure array signal, as well as the pressure time-frequency energy dispersion, are integrated in real time to generate a coupling reliability coefficient and dynamically calibrate the rider posture branch weights. The rider's independent posture flow is mapped to the knowledge graph baseline node, the continuous joint deviation vector is calculated by weighting according to the coupling confidence coefficient, and adaptive normalization is performed by combining the gait residual, and the posture deviation curve and action level label are output. Based on the posture deviation curve threshold and the current link status, the system dynamically selects voice, haptic or augmented reality feedback channels and issues corrective prompts to the rider within a single riding beat. After class, the posture deviation curve, feedback effectiveness labels, and gait residuals are uploaded to the cloud to perform semi-supervised incremental training, and then the updated weight package is sent back to the edge during the link window.

2. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 1, characterized in that: Visible light image data, depth point cloud data, rider inertial measurement unit data, and saddle pressure array data are collected synchronously using a unified clock. Based on prior knowledge of the equipment layout, the rider inertial measurement unit data and saddle pressure array data are assigned main labels to generate rider data and horse data.

3. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 2, characterized in that: The rider and horse skeleton points are extracted from visible light image data and depth point cloud data using a skeleton detection algorithm, and subject labels are assigned.

4. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 3, characterized in that: Time-stamp alignment technology was used to align visible light image data, depth point cloud data, rider inertial measurement unit data, and saddle pressure array data. Interpolation is used to generate a fused frame sequence that is consistent with the instantaneous movement. Each frame contains a visible light image, a depth image, rider inertial measurement unit data, saddle pressure array data, and a main label.

5. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 4, characterized in that: By separating rider posture features and horse gait features through a dual-channel adversarial coding decoupling architecture, the mutual information difference between rider inertial measurement unit data and saddle pressure array signal, as well as the pressure time-spectral energy dispersion of saddle pressure array signal, are calculated to generate a coupling reliability coefficient. By using the coupling confidence coefficient to dynamically calibrate the weights of rider posture features, the interference of horse gait feature fluctuations on rider posture feature interpretation is suppressed.

6. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 5, characterized in that: Map the rider’s independent posture flow to the baseline node in the equestrian movement knowledge graph; The joint position deviation vector between the rider's independent posture flow and the reference node is calculated by matching the closest standard motion sequence using a dynamic time warping algorithm.

7. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 6, characterized in that: The joint position deviation vector is weighted and adjusted based on the coupling confidence coefficient. The gait residual in the latent representation of horse gait is extracted. The weighted joint position deviation vector is corrected using a nonlinear normalization function. Continuous posture deviation curves and movement level labels are generated, and interpretable feedback information is output.

8. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 7, characterized in that: The rider's posture deviation level is determined based on the average deviation value of the posture deviation curve and the preset threshold range. The current link status is assessed by real-time monitoring of network latency, bandwidth and packet loss rate. Based on the deviation level and the current link status, the system dynamically selects voice feedback, haptic feedback, or augmented reality feedback channels, generates correction prompts, and sends them to the selected feedback channels within a single riding beat.

9. The intelligent auxiliary method for equestrian teaching based on motion posture recognition according to claim 8, characterized in that: The posture deviation curve, feedback effectiveness label, and gait residual are transmitted to the cloud via an upload operation, and an updated weight package is generated using a semi-supervised incremental training method. During the link window, the updated weight packet is sent back to the edge and applied to the attitude analysis network and the feedback optimization network respectively through weight update operations.

10. An intelligent auxiliary system for equestrian teaching based on motion posture recognition, characterized in that: include, The data acquisition and processing module synchronously acquires visible light-depth dual-stream data, rider inertial measurement unit data, and saddle pressure array signals under a unified clock. Based on the prior knowledge of the deployment points, it completes primary clustering and main tag allocation, and generates a fused frame sequence with main tags. The posture calibration module, during the dual-channel adversarial coding decoupling process, integrates the mutual information difference between rider joint IMU data and saddle pressure array signals, as well as the pressure time-frequency energy dispersion, in real time to generate a coupling reliability coefficient and dynamically calibrate the rider posture branch weights. The deviation calculation module maps the rider's independent posture flow to the knowledge graph baseline node, calculates the continuous joint deviation vector by weighting according to the coupling confidence coefficient, and performs adaptive normalization by combining the gait residual, outputting the posture deviation curve and action level label. The feedback correction module dynamically selects voice, haptic, or augmented reality feedback channels based on the posture deviation curve threshold and the current link status, and issues correction prompts to the rider within a single riding beat. The module is optimized and updated. After class, the posture deviation curve, feedback effect label and gait residual are uploaded to the cloud to perform semi-supervised incremental training. Then, the updated weight package is sent back to the edge during the link window.

Images