A coupling relationship index generation method based on a 3D human posture image model

Abstract

This invention discloses a health assessment method based on a 3D human posture image model, belonging to the field of computer vision and pattern recognition technology. This invention constructs a health assessment method based on three-dimensional human posture temporal data, eliminating the need for electrodes attached to the subject's body surface or additional sensors. Data acquisition can be completed simply by collecting image sequences in a natural, awake state using a regular camera. Compared to traditional polysomnography, this significantly reduces the examination threshold and the burden on the subject, making it more suitable for repeated use in non-specialized sleep laboratory settings such as homes and community clinics. Through three-dimensional human posture estimation and a virtual chest surface model, this invention can extract volume or cross-sectional area changes related to breathing without contact, thereby obtaining the first signal of respiratory rhythm and avoiding the problem of environmental audio being susceptible to noise and multi-source interference.

Description

Technical Field

[0001] This invention relates to the field of computer vision and pattern recognition technology, and in particular to a method for generating coupling relationship indicators based on a 3D human pose image model. Background Technology

[0002] Currently, the clinical diagnosis of obstructive sleep apnea (OSA) usually relies on polysomnography (PSG) monitoring. This method requires attaching multiple physiological sensors to the subject's body surface and completing overnight recording in a sleep laboratory equipped with specialized equipment and technicians. Although PSG is considered the gold standard for OSA diagnosis, its examination process is complex and costly, which can affect subject compliance and comfort. Furthermore, a single overnight monitoring session is insufficient to reflect the fluctuations in sleep-disordered breathing over multiple nights or long-term follow-up, thus limiting its large-scale application in primary healthcare institutions, community screening, and long-term health management scenarios.

[0003] To alleviate these problems, several contactless or low-intensity alternative assessment solutions have emerged in recent years. In the realm of wearable devices, products such as smartwatches, wristbands, and chest straps integrate biosensors like photoplethysmography (PPG) and accelerometers. By analyzing signals such as heart rate variability, body movement, and blood oxygen saturation, they indirectly infer the risk of nocturnal respiratory events or obstructive sleep apnea (OSA). Regarding environmental perception, studies have utilized microphones placed beside the bed to collect snoring and breathing sounds, analyzing acoustic characteristics to estimate the apnea-hypopnea index (AHI) or grade OSA severity. Furthermore, computer vision-based contactless monitoring methods are being explored, using ordinary visible light or infrared cameras to track body movements, chest and abdominal movements, or changes in overall body contour during sleep to estimate respiratory rate, sleep posture, or assist in identifying abnormal events. These solutions, to some extent, reduce the reliance on dedicated spaces and numerous leads for monitoring, improving the accessibility and comfort of sleep-related assessments.

[0004] However, existing alternatives still have limitations; wearable device-based methods have not completely escaped the constraints of direct contact with human skin, and some subjects may remove them at night due to discomfort or wear them improperly, leading to a decline in data quality or monitoring interruption; snoring analysis based on ambient audio is easily affected by background noise, snoring of others in the same room, etc., and it is difficult to accurately distinguish the snoring of a target individual in a multi-sound source environment; existing monocular visual monitoring methods mostly focus on overall body movement or periodic chest and abdominal movements during sleep, and have relatively limited exploration of the static or quasi-static postural characteristics of the human body, especially lacking quantitative means for the continuous and subtle neck and shoulder postural compensations generated in the waking state to maintain upper airway patency.

[0005] Some studies have shown that OSA patients differ from healthy individuals in terms of craniofacial structure and cervical / head-neck posture, and there may be a correlation between upper airway morphology and head-neck posture. However, current technologies lack a technical solution to use three-dimensional human posture image models to quantify and analyze the stability and movement patterns of the head and neck of subjects in a natural sitting or relaxed state during the day, and to assess the risk of upper airway dysfunction or sleep apnea. This has led to insufficient utilization of the aforementioned potential physiological signals to some extent. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] This invention provides a method for generating coupling relationship indicators based on a 3D human posture image model, which addresses the problem that existing non-contact vision solutions do not adequately utilize the correlation information between respiratory rhythm and head and neck posture stability in a conscious state.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] This invention provides a method for generating coupling relationship indicators based on a 3D human pose image model, comprising:

[0010] Step S1: Collect image sequences of the target user in a natural, awake state;

[0011] Step S2: Process the image sequence to generate corresponding three-dimensional human pose time-series data;

[0012] Step S3: Based on the three-dimensional human posture time-series data, extract three-dimensional motion data of at least one upper airway associated region;

[0013] Step S4: Based on the three-dimensional motion data of the upper airway associated region, generate a first signal characterizing respiratory rhythm and a second signal characterizing posture stability.

[0014] Step S5: Calculate the coupling relationship index between the first signal and the second signal;

[0015] Step S6: Based on the coupling relationship index, assess the health risks associated with obstructive sleep apnea for the target user.

[0016] As a preferred embodiment of the coupling relationship index generation method based on a 3D human posture image model described in this invention, the three-dimensional human posture temporal data includes a three-dimensional coordinate sequence representing multiple key points of the spinal segment, shoulder, and head.

[0017] As a preferred embodiment of the coupling relationship index generation method based on a 3D human posture image model described in this invention, the extraction of three-dimensional motion data of at least one upper airway associated region includes:

[0018] Based on the key points of the upper torso in the three-dimensional human posture time-series data, a three-dimensional spatial reference frame is defined.

[0019] Within the reference frame, a curved surface model covering the chest area and driven by preset control points is constructed as a virtual region;

[0020] The algorithm is optimized to match the projection of the surface model in the image sequence with the upper torso outline of the target user.

[0021] During the matching process, the changes in the volume or cross-sectional area parameters of the virtual region over time are tracked as three-dimensional motion data of the upper airway associated region.

[0022] As a preferred embodiment of the coupling relationship index generation method based on a 3D human posture image model described in this invention, wherein: the generation of the first signal representing respiratory rhythm includes:

[0023] The changes in the volume or cross-sectional area parameters of the virtual region over time are used to construct an original time series.

[0024] The original time series is subjected to frequency domain filtering or bandpass filtering to extract waveform components that match the respiratory rate;

[0025] The waveform component is used as the first signal characterizing the respiratory rhythm.

[0026] As a preferred embodiment of the coupling relationship index generation method based on a 3D human pose image model described in this invention, wherein: the generation of the second signal characterizing pose stability includes:

[0027] Based on the three-dimensional human posture time-series data, the relative position of the head relative to the torso, the relative posture angle, or a combination of both are calculated to obtain a posture sequence representing the change of head and neck posture over time.

[0028] The posture sequence is smoothed or detrended to highlight slow changes or slight swaying.

[0029] The processed attitude sequence is used as a second signal to characterize attitude stability.

[0030] As a preferred embodiment of the coupling relationship index generation method based on a 3D human pose image model according to the present invention, wherein: calculating the coupling relationship index between the first signal and the second signal includes:

[0031] Within a sliding time window, the regularity features of the first signal are calculated, wherein the regularity features of the first signal include at least one feature value obtained by sample entropy calculation.

[0032] Within the sliding time window, the energy characteristics or complexity characteristics of the second signal are calculated, and the energy characteristics or complexity characteristics include at least one feature value obtained by spectral entropy calculation or time-domain complexity measurement calculation.

[0033] The regularity feature sequence obtained along time is subjected to statistical correlation analysis or mutual information analysis with the energy feature or complexity feature sequence to obtain a value as the coupling relationship index.

[0034] As a preferred embodiment of the coupling relationship index generation method based on a 3D human pose image model described in this invention, the statistical correlation analysis or mutual information analysis includes calculating at least one of Pearson correlation coefficient, Spearman rank correlation coefficient, or maximum information coefficient.

[0035] The beneficial effects of this invention are as follows: This invention constructs a health assessment method based on three-dimensional human posture time-series data. It eliminates the need for electrodes to be attached to the subject's body surface or for additional sensors to be worn. Data acquisition can be completed simply by collecting image sequences in a natural, awake state using a regular camera. Compared to traditional polysomnography, this significantly reduces the examination threshold and the burden on the subject, making it more suitable for repeated use in non-specialized sleep laboratory settings such as homes and community clinics. Through three-dimensional human posture estimation and a virtual surface model of the chest, this invention can extract volume or cross-sectional area changes related to breathing without contact, thereby obtaining the first signal of respiratory rhythm and avoiding the problem of environmental audio being susceptible to noise and interference from multiple sound sources. Simultaneously, by calculating the relative position and posture angle of the head relative to the torso, a second signal of posture stability is constructed, transforming the originally difficult-to-quantify head and neck compensatory posture behavior into a dynamic indicator that can be tracked over time. Furthermore, within a sliding time window, methods such as sample entropy and spectral entropy are used to measure the regularity of respiratory rhythm and the energy or complexity characteristics of posture signals. Coupling indices between these two types of signals are obtained through correlation or mutual information analysis. This allows the model to not only focus on the regularity of breathing but also characterize whether coordinated head and neck adjustments occur during respiratory abnormalities, resulting in higher sensitivity and richer information in the early screening of upper airway dysfunction, especially obstructive sleep apnea. By inputting the coupling indices into a risk assessment model trained on labeled samples, risk levels or probabilities can be output and automatic warnings can be given. This provides individuals with a low-cost, multi-night repeatable risk monitoring method, which is beneficial for large-scale pre-screening in community populations and provides quantitative evidence for whether further PSG examinations should be performed. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0037] Figure 1 This is a flowchart illustrating the coupling relationship index generation method based on a 3D human pose image model in this embodiment. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0039] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0040] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0041] This application proposes a method for generating coupling relationship indicators based on a 3D human pose image model, combining... Figure 1 As shown, the method includes:

[0042] Step S1: Collect image sequences of the target user in a natural, awake state;

[0043] In this embodiment, the image sequence can be obtained by continuously capturing images of a target user in a seated or relaxed standing posture using a fixedly installed camera. The camera is preferably positioned one to two meters away from the target user, at a height roughly level with the chest and head, ensuring that the spine, shoulders, and head remain within the field of view throughout the acquisition process. Specifically, the target user should remain naturally awake during acquisition, without consciously controlling their breathing or performing any specific movements; a relaxed, everyday posture, such as sitting in a chair with back support, is sufficient to minimize significant body movement. The image resolution can be set to a common range of 640 x 480 pixels to 1280 x 720 pixels by default. The frame rate is preferably around 25 frames per second, adjustable between 15 and 30 frames per second to balance temporal resolution and storage overhead. The default acquisition duration is three to ten minutes, but can be extended as needed to improve statistical stability, provided it is at least 60 seconds. The start and end times of acquisition can be recorded by software in a data file for subsequent alignment with the timeline of the three-dimensional posture sequence. Optionally, it is preferable to maintain relatively stable indoor lighting in the acquisition environment to avoid strong backlight and large-area obstruction. When the target user is detected to leave the field of view for a short time or be completely obscured by a foreground object, the image frame of the corresponding time period can be marked as an invalid frame in this embodiment. This part of the data can be ignored in the subsequent processing stage or the time axis can be filled in by interpolation to ensure that the continuous image sequence has a clear and effective interval division in the engineering implementation.

[0044] Step S2: Process the image sequence to generate corresponding three-dimensional human pose time-series data;

[0045] Specifically, 3D human pose temporal data can be obtained by first performing human detection and 2D skeleton keypoint detection on each frame of the image, and then calling existing 3D pose estimation algorithms to revert the 2D keypoints to 3D coordinates in the camera coordinate system. The resulting coordinates can be uniformly represented in a right-handed coordinate system with the camera optical center as the origin, the horizontal axis as the horizontal axis, the vertical axis as the vertical axis, and the direction pointing to the target user as the depth axis. The depth unit can be meters or millimeters. On the time axis, the sampling interval of the 3D human pose temporal data is consistent with the image frame rate. The keypoint sequences of adjacent frames are stored aligned according to the image timestamps, thereby ensuring temporal synchronization with subsequent breathing and posture signals. In this embodiment, to reduce the impact of single-frame detection noise on the 3D trajectory, a first-order or second-order low-pass filter can be applied to the keypoint coordinates in the time dimension. The filter cutoff frequency is preferably set in the range of 1 Hz to 2 Hz to retain low-frequency components related to slow changes in breathing and posture. Optionally, when keypoint detection fails or the confidence score is lower than a preset threshold in a certain frame, the three-dimensional coordinates of the previous frame can be directly used or the interpolation results of several frames before and after can be used as the estimated value of the frame, and the frame can be marked as an interpolated frame in the data so as to distinguish between the original data and the interpolated data in subsequent quality control or anomaly handling.

[0046] Step S3: Based on the three-dimensional human posture time series data, extract the three-dimensional motion data of at least one upper airway associated region;

[0047] Step S4: Based on the three-dimensional motion data of the upper airway associated region, generate a first signal characterizing respiratory rhythm and a second signal characterizing postural stability.

[0048] Step S5: Calculate the coupling relationship index between the first signal and the second signal;

[0049] Step S6: Based on coupling relationship indicators, assess the health risks associated with obstructive sleep apnea for the target user;

[0050] In one embodiment, the three-dimensional human pose temporal data includes a sequence of three-dimensional coordinates representing multiple key points of the spinal segment, shoulder, and head.

[0051] In one embodiment, extracting three-dimensional motion data of at least one upper airway associated region includes:

[0052] Based on the key points of the upper torso in the three-dimensional human posture temporal data, a three-dimensional spatial reference frame is defined.

[0053] Within the reference frame, a surface model covering the chest area and driven by preset control points is constructed as a virtual region;

[0054] The algorithm is optimized to match the projection of the curved model in the image sequence with the upper torso outline of the target user.

[0055] During the matching process, the changes in the volume or cross-sectional area parameters of the virtual region over time are tracked as three-dimensional motion data of the upper airway associated region.

[0056] Furthermore, the surface model can be implemented using spline surfaces or other continuously differentiable parametric surfaces. The preset control points are preferably arranged in a two-dimensional regular grid within the projected rectangle of the chest region. The number of control points can be set to a grid size of four x four to eight x eight to achieve a trade-off between fitting accuracy and computational complexity. The volume parameters can be calculated by selecting several cross-sectional planes approximately parallel to the ground within a three-dimensional spatial reference frame. The intersection lengths or cross-sectional profiles of the virtual surface and each cross-sectional plane are approximated as local volumes through numerical integration, and then accumulated across all cross-sections to obtain the volume change over time. The cross-sectional area parameters can be calculated by pixel counting or polygon area solving on the projection of the virtual region on the reference plane to obtain a sequence of cross-sectional area changes over time. In this embodiment, the matching of the curved surface model with the upper part of the torso can be achieved by minimizing the distance or overlap error between the virtual contour and the torso edge in the image. The number of iterations can be set to between ten and fifty by default, and the specific value can be adjusted according to the processor performance and real-time requirements. When the matching residual of a certain frame exceeds the preset upper limit, the volume parameter or cross-sectional area parameter corresponding to that frame can be marked as an outlier. When constructing three-dimensional motion data in the future, the mean or interpolation of several frames before and after can be used to replace it, so as to avoid the failure of a single frame fitting causing a significant jump in the respiratory signal.

[0057] In one embodiment, generating a first signal characterizing respiratory rhythm includes:

[0058] The changes in the volume or cross-sectional area parameters of the virtual region over time are used to construct the original time series.

[0059] Frequency domain filtering or bandpass filtering is applied to the original time series to extract waveform components that match the respiratory rate;

[0060] The waveform component is used as the first signal to characterize the respiratory rhythm;

[0061] For example, frequency domain filtering or bandpass filtering can be implemented using finite impulse response (FIR) or infinite impulse response (IR) filters. The respiratory rate of adults at rest is typically in the range of 0.1 Hz to 0.6 Hz. Therefore, the lower and upper limits of the respiratory bandwidth can be preset to adjustable ranges of 0.05 Hz to 0.1 Hz and 0.8 Hz to 1 Hz, respectively. The default bandwidth can be selected from 0.1 Hz to 0.6 Hz to cover most populations. The filter order can be set according to the sampling frequency and the required transition bandwidth. When the sampling frequency of the first signal is approximately 25 samples per second, the order of the bandpass filter is preferably between 20 and 100 to obtain sufficient frequency selectivity while controlling computational overhead and phase delay. After obtaining the bandpass signal, this embodiment can perform amplitude normalization processing on the waveform components to limit the numerical range of the first signal to the interval between negative one and one, which facilitates the numerical stability of subsequent sample entropy and correlation calculations. Optionally, when the number of effective respiratory peaks detected within a certain time window is lower than a preset lower limit, it can be determined that the respiratory rhythm estimation within that time window is unreliable. In the subsequent calculation of coupling relationship indicators, this time window can be skipped or the estimated value of a neighboring window can be used for interpolation, thereby avoiding the disproportionate impact of extreme abnormal segments on the overall evaluation results.

[0062] In one embodiment, generating a second signal characterizing attitude stability includes:

[0063] Based on three-dimensional human posture time-series data, the relative position of the head relative to the torso, the relative posture angle, or a combination of both are calculated to obtain a posture sequence representing the change of head and neck posture over time.

[0064] Smoothing or detrending of the posture sequence highlights slow changes or subtle oscillations;

[0065] The processed attitude sequence is used as a second signal characterizing attitude stability.

[0066] Similarly, the second signal for posture stability can be obtained by converting the three-dimensional keypoint coordinates of the head and torso into a parameter sequence reflecting the relative posture. The relative position can be represented by the three-dimensional displacement vector of the head's center of mass relative to the torso's center of mass, and the relative posture angle can be represented by the rotation angle of the head's local coordinate system relative to the torso's local coordinate system and its change over time. Smoothing can be achieved using moving average filtering or low-pass filtering, with the time window length preferably between 0.5 and 2 seconds. The filter cutoff frequency can be set in the range of 0.5 Hz to 1 Hz, so that the second signal mainly reflects slow swaying and posture adjustments without excessively amplifying instantaneous jitter. Detrending can be achieved by subtracting the linear fitting result or the overall mean of the entire posture sequence, causing the second signal to fluctuate slightly around zero, which is beneficial for alignment with the first signal on a numerical scale. Optionally, when it is detected that the head key points are severely missing or occluded in several consecutive frames, making it impossible to stably estimate the head and neck posture, this embodiment can adopt an interpolation method based on the previous and next valid frames to recover some data; if the continuous missing time exceeds a preset threshold, such as three seconds, the entire time period is marked as an invalid interval of the posture signal, and the data of this time period is not used in the subsequent calculation of coupling indicators and health risk assessment.

[0067] In one embodiment, calculating the coupling relationship index between the first signal and the second signal includes:

[0068] Within a sliding time window, the regularity features of the first signal are calculated, and the regularity features of the first signal include at least one feature value obtained by sample entropy calculation.

[0069] Within the sliding time window, the energy characteristics or complexity characteristics of the second signal are calculated. The energy characteristics or complexity characteristics include at least one feature value obtained through spectral entropy calculation or time-domain complexity measurement calculation.

[0070] Statistical correlation analysis or mutual information analysis is performed on the regularity feature sequence obtained along the time and the energy feature or complexity feature sequence to obtain a value as an index of coupling relationship.

[0071] In one implementation, the regularity feature calculation step of the first signal includes:

[0072] Step S51, in the Within a time window, the first signal sequence representing the respiratory rhythm is normalized to zero mean and unit variance:

[0073] ,

[0074] in, Indicates the first Samples after window normalization Represents the discrete sequence of the first signal. Indicates the first The starting sample index of the window in the entire sequence. Indicates the sample index within the window. Indicates the window length (number of samples). Indicates the first Mean within the window Indicates the first Standard deviation within the window Represents the window index (a non-negative integer starting from zero);

[0075] Step S52: Estimate the respiratory fundamental frequency based on the main peak of the power spectrum within the window, and set the window length covering several cycles accordingly.

[0076] ,

[0077] ,

[0078] in, Indicates the respiratory fundamental frequency. Indicates the search range for the respiratory frequency band. Indicates the first Power spectrum estimation within the window Indicates the window length. The coefficient representing the number of respiratory cycles contained in each window (dimensionless). Indicates the sampling frequency. This represents the sliding step size (number of samples). This represents the window overlap ratio coefficient (dimensionless). Indicates rounding up. Indicates rounding down. and These represent the minimum and maximum candidate respiratory rates, respectively.

[0079] Step S53, in the Within the window, the length is constructed based on the embedding dimension and the time delay. The template vector is used, and the matching count is performed using the infinity norm:

[0080] ,

[0081] ,

[0082] ,

[0083] in, Indicates index The starting length is template, and Indicates the starting index of the template. Indicates time delay (number of samples). Indicates the template component index. and They respectively represent the dimensions of and The matching ratio below, This represents the matching tolerance (dimensionless, as it has been normalized). Indicates the available length is The number of templates, Indicates the available length is The number of templates, Indicates an indicator function, This represents the infinite norm of the difference between the components;

[0084] Step S54, for the first The logarithmic ratio of the sample entropy of the window is used to obtain a measure of regularity, and these samples are concatenated into a time series:

[0085] ,

[0086] in, Indicates the first Window sample entropy, This indicates the smallest positive number whose denominator is zero. Represents the natural logarithm. Indicating regularity feature sequences in the index The value at;

[0087] Step S55: Correlate the embedded parameters with the respiratory fundamental frequency to balance time resolution and robustness.

[0088]

[0089] in, Indicates time delay. Indicates the sampling frequency. Indicates the respiratory fundamental frequency. Indicates the matching tolerance. This represents a dimensionless tolerance coefficient. Indicates the embedding dimension. This represents the threshold coefficient for grading, on the right-hand side of the grading condition. This indicates the number of samples per period;

[0090] exist The internal calculation can be performed using weighted Welch spectral estimation. Before step S51, a light detrending process can be performed to reduce baseline drift. If the values ​​are abnormally low or high, the restrictions can be relaxed accordingly. To maintain a sufficient number of cycles within the window;

[0091] Specifically, the above implementation method uses a sliding window as a carrier to generate a time-varying sample entropy sequence to measure the regularity of the first signal; firstly, standardization is performed within each window to make the matching statistics insensitive to amplitude scaling; then, the fundamental frequency is estimated using the main peak of the power spectrum; and the window length and step size are set on a scale of several periods to balance statistical stability and time resolution; the sample entropy is compared... and The matching ratio of the dimensional template is a logarithmic ratio. The smaller the value, the stronger the repeatability and the more regular the waveform. The embedding parameters adopt an adaptive setting related to the fundamental frequency: the delay is taken as one-quarter of the period to reduce component redundancy, the short window uses low dimension to reduce sparse counting bias, the long window increases the dimension to enhance the resolution of subtle perturbations, and the tolerance is set with a dimensionless coefficient to facilitate reuse across subjects and scenarios. The obtained regularity feature sequence can be aligned with the attitude-side energy or complexity features on the same time axis, providing a stable input for the calculation of coupling indexes and serving subsequent risk assessment related to obstructive sleep apnea.

[0092] In this embodiment, the sampling frequency of the discrete sequence of the first signal can be consistent with the image acquisition frame rate. When the frame rate is between 15 and 30 frames per second, the duration of the sliding time window can be automatically adjusted according to the estimated respiratory fundamental frequency, so that each time window contains approximately 3-8 complete respiratory cycles, ensuring that the sample entropy estimation has a sufficient number of samples while taking into account temporal resolution. The window overlap ratio can be set between 50% and 80%, and the corresponding sliding step size is determined by multiplying the window length by the overlap ratio, thereby forming a smooth transition between adjacent windows. The amplitude tolerance used for matching in the sample entropy calculation can be set in the range of 0.1-0.3 according to the normalized signal amplitude, to prevent the tolerance from being too small, resulting in insufficient matching times, or too large, resulting in different waveforms being regarded as the same; to avoid extremely small positive numbers with a denominator of zero, values ​​between 10 to the power of -6 and 10 to the power of -4 can be selected to improve numerical stability without significantly affecting the entropy value. Optionally, when the number of valid samples within a certain time window is lower than a certain proportion of the expected total number of samples due to imputation or missing data, such as lower than 70%, in this embodiment, the regularity feature of the window may not be calculated. Instead, the entropy value of the adjacent time window may be used for linear interpolation or the window may be directly marked as missing, thereby avoiding the distortion of the statistical distribution of the coupling relationship index by a small number of severely missing data windows.

[0093] In one embodiment, statistical correlation analysis or mutual information analysis includes calculating at least one of Pearson correlation coefficient, Spearman rank correlation coefficient, or maximum information coefficient;

[0094] In one embodiment, assessing health status related to motor control includes:

[0095] Based on the distribution of variance contribution rate, the motion differentiation entropy is calculated as a differentiation index to characterize the motion coordination of the target joint chain.

[0096] The motor differentiation entropy is compared with at least one reference threshold, or the motor differentiation entropy is input into a pre-trained motor control health assessment model, and the assessment results are output to distinguish between normal motor control states and abnormal motor control states.

[0097] The steps for calculating the motion differentiation entropy include: variance contribution rate, probability distribution construction, motion differentiation entropy calculation and normalization.

[0098] In one implementation, the motion differentiation entropy is calculated using the following steps:

[0099] Step S71, based on the obtained Sequence of each motion cooperative component and its variance eigenvalues The variance contribution rate is denoted as:

[0100] ,

[0101] in, Indicates the first The variance contribution rate of each cooperating component Indicates the first The variance eigenvalues ​​of each cooperating component. Indicates the number of cooperative components ( ), Indicates the summation index;

[0102] Step S72, to avoid zero probability and over-concentration, for Prior smoothing and temperature shaping are performed to obtain the distribution. :

[0103] ,

[0104] in, The variance contribution rate after smoothing is represented by , and the Dirichlet smoothing strength is represented by (dimensionless). ), surface Show the first The probability of each component. Represents the temperature coefficient (dimensionless). To soften the weighting, Strengthen the head, Flattening the distribution);

[0105] Step S73, with The degree of unevenness characterizes the differentiation degree, and the motion differentiation entropy and its normalized form are defined as follows:

[0106] ,

[0107] ,

[0108] in, Represents the unnormalized motion differentiation entropy (natural logarithm base). Represents the normalized motion differentiation entropy, with a value range of... , Represents the natural logarithm;

[0109] Step S74, when multiple action cycles are used to obtain When calculating the statistics, you can first calculate the average on the periodic set before proceeding to step S72:

[0110] ,

[0111] in, Indicates cross The mean characteristic value of each period, Indicates the first Characteristic values ​​of each period, Indicates the number of periods included in the statistics; in computer implementation, when... In the event of a very small amount of underflow, it can be Internal numerical truncation threshold ,Will by Substitute; where, This section represents the lower limit constant used for numerical stability;

[0112] Pick – The interval is used to suppress zero probability. Available Internal fine-tuning to match participant differences. From the cumulative variance coverage ratio to – Freeze the interval to avoid comparison bias caused by changing dimensions across samples;

[0113] Specifically, the above implementation scheme starts with the variance contribution rate of the collaborative components. First, it formalizes the relative proportion of each component to the overall motion variation, and then maps this proportion to a probability distribution to construct an information entropy measure. The smoothing term is used to avoid zero probability under limited samples, and the temperature coefficient is used to adjust the concentration of probability, so that the measure can respond to the concentration phenomenon of the dominant component and will not be overly sensitive when there are long-tail components. The motion differentiation entropy calculated subsequently characterizes the degree of balance of component participation: when a few components contribute more, the probability distribution is peaked and the entropy decreases, indicating a low degree of collaborative differentiation. When multiple components participate together and their contributions are close, the distribution tends to be flat and the entropy increases, indicating a high degree of differentiation.

[0114] To facilitate comparisons across subjects, tasks, or devices, entropy values ​​are calculated as follows: Normalization to The range is expanded to avoid confusion of dimensions due to different component numbers; in multi-period scenarios, by first aggregating the feature values ​​and then calculating the distribution, the impact of occasional action fluctuations can be reduced, forming a stable evaluation index and providing consistent input for threshold comparison or model discrimination;

[0115] Furthermore, the number of motion co-factors can be determined by the cumulative variance coverage threshold. After the eigenvalue sequence of the dimensionality reduction analysis is sorted in descending order, the number of components corresponding to the first time the cumulative variance contribution rate reaches 90% to 95% can be taken as the number of effective components. Subsequent calculations of motion differentiation entropy are only performed on these effective components to both retain the main motion information and avoid introducing too many noise components. The intensity parameter used for prior smoothing can be a value between 10 to the power of -6 and 10 to the power of -3 to suppress the occurrence of zero probability in the case of finite samples and ensure that the probability distribution is always bounded. The temperature coefficient used for temperature shaping can be set in the range of 0.7 to 1.5. When the temperature coefficient is close to 1, the probability distribution is basically consistent with the original variance contribution rate. When the temperature coefficient is less than 1, the motion differentiation entropy is more sensitive to a few dominant co-factors. When the temperature coefficient is greater than 1, the motion differentiation entropy emphasizes the balance of multiple components. In engineering implementation, when the number of effective collaborative components obtained by dimensionality reduction analysis is less than two, or when the variance eigenvalues ​​of some components are found to be close to zero and fluctuate drastically between different periods, in this embodiment, the motion differentiation entropy index of that segment of data can be temporarily not calculated, and the corresponding action period can be marked as an invalid period. Only the motion control-related health assessment results are statistically analyzed on the effective period to avoid abnormal values ​​caused by division by zero or logarithmic operations affecting the overall judgment.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0117] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for generating coupling relationship indicators based on a 3D human pose image model, characterized in that, include: Step S1: Collect image sequences of the target user in a natural, awake state; The image sequence is obtained by continuously shooting the target user in a sitting or relaxed standing position using a fixedly installed camera. During the acquisition, the target user does not need to consciously control their breathing rhythm or perform any specific actions. Step S2: Process the image sequence to generate corresponding three-dimensional human pose time-series data; Step S3: Based on the three-dimensional human posture time-series data, extract three-dimensional motion data of at least one upper airway associated region; Step S4: Based on the three-dimensional motion data of the upper airway associated region, generate a first signal characterizing respiratory rhythm and a second signal characterizing posture stability. Step S5: Calculate the coupling relationship index between the first signal and the second signal; The three-dimensional human posture temporal data includes a three-dimensional coordinate sequence representing multiple key points of the spinal segment, shoulder, and head; The extraction of three-dimensional motion data for at least one upper airway-related region includes: Based on the key points of the upper torso in the three-dimensional human posture time-series data, a three-dimensional spatial reference frame is defined. Within the reference frame, a curved surface model covering the chest area and driven by preset control points is constructed as a virtual region; The algorithm is optimized to match the projection of the surface model in the image sequence with the upper torso outline of the target user. During the matching process, the changes in the volume or cross-sectional area parameters of the virtual region over time are tracked as three-dimensional motion data of the upper airway associated region; The first signal that generates the respiratory rhythm includes: The changes in the volume or cross-sectional area parameters of the virtual region over time are used to construct an original time series. The original time series is subjected to frequency domain filtering or bandpass filtering to extract waveform components that match the respiratory rate; The waveform component is used as the first signal characterizing the respiratory rhythm; The second signal representing attitude stability includes: Based on the three-dimensional human posture time-series data, the relative position of the head relative to the torso, the relative posture angle, or a combination of both are calculated to obtain a posture sequence representing the change of head and neck posture over time. The posture sequence is smoothed or detrended to highlight slow changes or slight swaying. The processed attitude sequence is used as a second signal characterizing attitude stability. The calculation of the coupling relationship index between the first signal and the second signal includes: Within a sliding time window, the regularity features of the first signal are calculated, wherein the regularity features of the first signal include at least one feature value obtained by sample entropy calculation. Within the sliding time window, the energy characteristics or complexity characteristics of the second signal are calculated, and the energy characteristics or complexity characteristics include at least one feature value obtained by spectral entropy calculation or time-domain complexity measurement calculation. The regularity feature sequence obtained along time is subjected to statistical correlation analysis or mutual information analysis with the energy feature or complexity feature sequence to obtain a value as the coupling relationship index. 2.The method of claim 1, wherein, The statistical correlation analysis or mutual information analysis includes calculating at least one of the Pearson correlation coefficient, Spearman rank correlation coefficient, or maximum information coefficient.

Images