Real-time warning method for abnormal posture of newborns using video analysis

Abstract

The application discloses a new-born abnormal posture real-time early warning method using video analysis, relates to the technical field of image data processing and video recognition, and comprises the following steps: acquiring a real-time video stream of a new-born sleeping area; performing frame-by-frame posture estimation on the real-time video stream to obtain body key point positions of the new-born, and calculating a measurement angle of a target joint according to the body key point positions; and acquiring a detection uncertainty measurement of the body key point positions by a posture estimation model. The application establishes an error propagation model from key point position uncertainty to joint angle measurement uncertainty, obtains a sensitivity coefficient through partial derivative operation, and quantizes the detection uncertainty into an angle confidence interval or a probability distribution. The application effectively overcomes the defect that the same pixel positioning error is significantly amplified in angle calculation due to the short limbs of the new-born, makes the clinical grading determination no longer rely on a single point estimation value, and significantly reduces false positives or missed reports caused by error amplification.

Description

Technical Field

[0001] This invention relates to the field of image data processing and video recognition technology, specifically to a real-time early warning method for abnormal neonatal postures using video analysis. Background Technology

[0002] In recent years, deep learning-based pose estimation techniques have offered new possibilities for non-contact monitoring of abnormal neonatal postures. General models have made significant progress in multimodal data fusion and lightweight deployment, and have initially achieved the extraction of key points of the infant's body from videos and the identification of basic posture categories such as supine and prone positions.

[0003] However, clinical neurodevelopmental assessments for abnormal postures heavily rely on precise quantification of joint angles—for example, an adductor angle below 45° suggests increased muscle tone, and a dorsiflexion angle exceeding 60°–70° indicates abnormality. The angle interval between normal and mild pathology is only 10°–15°, and the accuracy of angle measurement directly determines the reliability of clinical assessments. Current techniques suffer from the following fundamental limitations: The training and evaluation of posture estimation models prioritize keypoint localization accuracy over the clinically required joint angle measurement accuracy. This misalignment leads to a significant error amplification effect in the small-scale anatomy of newborns: the measurement error of joint angles is inversely proportional to the limb length connecting the two keypoints—the shorter the limb, the more severely the same pixel-level localization error is amplified in angle calculation. Newborns have much shorter limbs than adults, and the same pixel-level localization error can lead to angle errors of 5° to 10° or even greater, which is on the same order of magnitude as the 10° to 15° clinical grading interval. Therefore, even a model with keypoint localization accuracy exceeding 90% cannot reliably distinguish the boundaries between normal and mild pathology, or between mild and moderate pathology, in joint angle measurement tasks. Furthermore, existing methods only output a single angle value, providing no quantitative information on measurement uncertainty. The warning logic is based solely on deterministic threshold comparisons, inevitably generating numerous false positives or missed true positives near clinical grading boundaries, failing to meet the reliability requirements of medical-grade neonatal monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time early warning method for abnormal neonatal postures using video analysis, in order to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a real-time early warning method for abnormal neonatal postures using video analysis, comprising: Acquire real-time video streams of the newborn's sleep area; The pose of the real-time video stream is estimated frame by frame to obtain the position of the newborn’s body key points, and the measurement angle of the target joint is calculated based on the position of the body key points. Obtain a measure of the uncertainty in the detection of the body key point positions by the pose estimation model; An error propagation model is established from the uncertainty of key point location to the uncertainty of joint angle measurement. Based on the error propagation model and the detection uncertainty metric, the angle confidence interval or angle probability distribution of the measured angle is calculated. Based on the measured angle and the angle confidence interval, a probability judgment model is used to compare with the preset clinical grading angle threshold, calculate the conditional probability of the current posture relative to each clinical grading level, and obtain the abnormal posture risk level. When the risk level of the abnormal posture exceeds the preset warning threshold, a graded warning message is generated.

[0006] In a preferred embodiment, establishing the error propagation model from the uncertainty of key point location to the uncertainty of joint angle measurement includes: Using the detection uncertainty of each key body point as input, the partial derivative of the position of each key point is calculated using the geometric calculation formula of the joint angle to obtain the sensitivity coefficient of the angle error to the position error of each key point. The formula for calculating the absolute value of the sensitivity coefficient is as follows: ; Where θ is the joint angle. Let be the coordinates of the i-th key point. Let i be the position offset of the i-th key point. The Euclidean distance connecting two key points; Based on the sensitivity coefficient and the detection uncertainty of each key point, the total uncertainty of the target joint angle is calculated according to the error propagation theory. The confidence interval for the measured angle is output as follows: ; in: The total uncertainty is a function of the total uncertainty, calculated by coupling the detection error and sensitivity coefficient of each key point.

[0007] In a preferred embodiment, an adaptive correction step for image distortion is also included: Obtain reference features of known geometric constraints of the scene in the current video frame, the reference features including the rectangular border of the crib or the edge line of the mattress; Based on the projection deformation of the reference feature in the image plane, the shooting posture parameters of the camera relative to the central axis of the newborn's body are inferred. The shooting posture parameters include pitch angle, yaw angle and roll angle. Based on the shooting posture parameters, the perspective distortion mapping relationship between the image plane and the physical space under the current shooting perspective is calculated, and the inverse perspective projection transformation from image coordinates to physical space coordinates is constructed. The coordinates of the key body points are reprojected and corrected using the inverse perspective projection transformation, and the measured angle is recalculated based on the corrected key point coordinates. The recalculated measurement angle is used as the input to the error propagation model.

[0008] In a preferred embodiment, the method further includes a coupled processing based on the error propagation model and the adaptive correction of shooting distortion: Obtain the perspective distortion error influence coefficient of each key body point under the shooting posture parameters; The perspective distortion error influence coefficient and the key point position detection uncertainty are jointly quantified to obtain a comprehensive angle uncertainty measure under the composite error source. The angle confidence interval is updated based on the comprehensive angle uncertainty measure.

[0009] In a preferred embodiment, the step of comparing the measured angle and the angle confidence interval with a preset clinical grading angle threshold using a probability determination model includes: Obtain the newborn's age in months and birth weight parameters, wherein the birth weight parameters include interval markers for three subcategories: normal weight, low birth weight, and macrosomia. Based on the age parameter and the birth weight parameter, retrieve the corresponding clinical grading angle threshold subset from the database, where different age ranges and different birth weight subcategories correspond to different grading angle thresholds; Establish a threshold adjustment function for clinical grading based on differences in age and birth weight; The threshold adjustment function formula is as follows: ; in, The basic angle threshold; Weight correction factor; Age-corrected factor; Birth weight category; Age in months.

[0010] In a preferred embodiment, generating tiered early warning information includes: When the abnormal posture risk level is higher than the first risk threshold but lower than the second risk threshold, and the angle confidence interval crosses the clinical grading angle threshold, it is determined that the current state is at the monitoring boundary, and a prompting message is sent to the user terminal. The prompting message includes the current measurement angle, the angle confidence interval, and suggestions for improving the shooting conditions. When the risk level of the abnormal posture is higher than the second risk threshold, and the angle confidence interval is completely within the clinical abnormality grading interval, it is determined to be a high-confidence abnormal state, and an emergency warning message is sent to the user terminal. The first risk threshold and the second risk threshold are dynamically adjusted based on the birth weight and age parameters, wherein the risk threshold for low birth weight infants is lowered, the risk threshold for macrosomic infants is maintained at the normal baseline level, and the risk threshold for normal weight infants is raised.

[0011] In a preferred embodiment, a respiratory motion analysis step is also included: When the current posture is detected to be a prone posture and the upper limit of the confidence interval of the measured angle exceeds the preset prone angle threshold, the spatiotemporal feature sequence of the infant's chest region is extracted from the real-time video stream; The spatiotemporal feature sequence is subjected to motion amplification processing, which includes constructing a Gaussian pyramid, identifying weakly moving pixels based on features at each level, constructing a Laplacian pyramid, and amplifying and superimposing each level. Frequency domain analysis was performed on the magnified respiratory motion feature sequence to extract respiratory frequency features; The respiratory rate characteristics are compared with a normal respiratory rate range model selected based on age and birth weight parameters. When the respiratory rate is lower than the first respiratory threshold or higher than the second respiratory threshold, a respiratory abnormality warning is generated.

[0012] In a preferred embodiment, an uneasy movement detection step is also included: Extract the temporal posture sequence of the newborn's continuous whole-body movements from the real-time video stream; Based on the temporal posture sequence, the existence of uneasy motion is detected, and an uneasy motion quality score is output. Based on the birth weight parameters, the scoring threshold for judging restlessness and abnormal movement is adaptively adjusted, with the tolerance for judging restlessness and abnormal movement in very low birth weight infants being relaxed, and the threshold for judging restlessness and abnormal movement in macrosomic infants being lowered. When the restlessness movement quality score is lower than the adjusted judgment threshold and continues to exceed the preset time threshold, a neurodevelopmental screening warning is generated.

[0013] In a preferred embodiment, a method for training the pose estimation model is also included: During the training phase, a newborn posture dataset with annotations of body key points and joint angles was obtained. Construct a multi-task loss function, which includes a key point localization loss term and a joint angle alignment loss term. The calculation of the joint angle alignment loss term is based on the deviation between the joint angles derived from the key points output by the model and the actual joint angles. The joint angle alignment loss term and the key point localization loss term are weighted and combined to form the total loss function for model optimization. The pose estimation model is trained end-to-end based on the total loss function, so that the model can simultaneously optimize the key point localization accuracy and joint angle measurement accuracy during the training process.

[0014] In a preferred embodiment, temporal multi-frame fusion optimization based on angle confidence intervals is also included: Obtain multiple measured angle values ​​of the same target joint and their corresponding angle confidence intervals in multiple consecutive frames of images; The widths of the confidence intervals for each frame are compared; the narrower the confidence interval, the more reliable the angle measurement for that frame. The multiple measured angle values ​​are weighted and fused, where the weight of each frame's measured angle is determined by an inverse proportional function to the width of its confidence interval. The weight coefficient is calculated using the following formula: ; in, For the first Standard deviation of the uncertainty of the frame measurement angle; The fused measurement angle value is output, and the width of the fused angle confidence interval is smaller than the weighted average of the confidence interval widths of each frame, thereby achieving multi-frame temporal fusion noise reduction.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: An error propagation model was established, from the uncertainty of key point location to the uncertainty of joint angle measurement. Sensitivity coefficients were obtained through partial derivative calculations, quantifying the detection uncertainty into angle confidence intervals or probability distributions. This effectively overcomes the defect that the same pixel positioning error is significantly amplified in angle calculation due to the short limbs of newborns, making clinical grading no longer dependent on a single point estimate and significantly reducing false positives or missed detections caused by error amplification.

[0016] The measured angle and its confidence interval are input into a probabilistic judgment model and compared with a preset clinical grading angle threshold. The conditional probability of the current posture relative to each clinical grading level is calculated, and a graded warning is triggered based on the level of abnormal risk. This upgrades the warning logic from deterministic threshold comparison to probabilistic judgment, effectively solving the problem of false alarms or missed alarms caused by measurement uncertainty near the grading boundary, and meeting the reliability requirements of medical-grade neonatal monitoring.

[0017] By incorporating neonatal age and birth weight parameters, a subset of thresholds from a clinical grading perspective is retrieved from the database, and a threshold adjustment function based on the differences in age and birth weight is established. Simultaneously, the risk threshold is dynamically adjusted according to birth weight and age. This allows abnormal posture assessment to adapt to the physiological developmental differences of infants of different ages and weights, avoiding the problem of over-warning for low birth weight infants or under-diagnosing macrosomic infants with a uniform threshold, thus achieving individualized and precise monitoring of abnormal neonatal postures. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0021] Example 1, please refer to Figure 1 As shown in this embodiment, the real-time early warning method for abnormal neonatal postures using video analysis includes: S1. Acquire a real-time video stream of the newborn's sleep area; S2. Perform frame-by-frame pose estimation on the real-time video stream to obtain the positions of the newborn's key body points, and calculate the measurement angle of the target joint based on the positions of the key body points. S3. Obtain the uncertainty measure of the body key point position detection by the posture estimation model; S4. Establish an error propagation model from the uncertainty of key point location to the uncertainty of joint angle measurement. Based on the error propagation model and the detection uncertainty metric, calculate the angle confidence interval or angle probability distribution of the measured angle. S5. Based on the measured angle and the angle confidence interval, a probability judgment model is used to compare with the preset clinical grading angle threshold, and the conditional probability of the current posture relative to the clinical grading state at each level is calculated to obtain the abnormal posture risk level. S6. When the abnormal posture risk level exceeds the preset warning threshold, generate graded warning information.

[0022] In one embodiment, an error propagation model is established from the uncertainty of key point location to the uncertainty of joint angle measurement, specifically including the following sub-steps: Using the detection uncertainty of each key body point as input, the partial derivative of the position of each key point is calculated using the geometric calculation formula of the joint angle to obtain the sensitivity coefficient of the angle error to the position error of each key point.

[0023] The absolute value of the sensitivity coefficient is calculated using the following formula: ; Where θ represents the measured angle of the target joint (in degrees). This represents the position coordinates (in pixels) of the i-th key point. The value represents the tiny offset applied to the keypoint position (in pixels), and L represents the Euclidean distance between the two keypoints (in pixels). This formula describes the inverse relationship between the sensitivity of the angle error to the keypoint position error and the limb length: the shorter the limb, the greater the angle change caused by the same pixel offset.

[0024] In some implementations, the sensitivity coefficient can be calculated using numerical differentiation methods, that is, keeping the positions of other key points unchanged, and applying a small perturbation Δ to the position of a single key point. Calculate the resulting angular change Δθ, and then use Δθ / Δ The ratio approximates the partial derivative. In an optional implementation, the angle θ with respect to the keypoint coordinates can also be directly derived using analytical geometry methods. , The partial derivative expression of ) is obtained, thus yielding a more accurate sensitivity coefficient.

[0025] Based on the sensitivity coefficient and the detection uncertainty of each key point, the total uncertainty of the target joint angle is calculated according to the error propagation theory.

[0026] In some implementations, the first-order error propagation formula can be used: ; in, This represents the standard deviation (in pixels) of the detection uncertainty at the location of the i-th keypoint output by the pose estimation model. This represents the uncertainty in joint angle measurement (in degrees) obtained from the propagation of keypoint positioning errors. The formula is based on the error propagation theory of first-order Taylor expansion, which weights the uncertainties of each keypoint using sensitivity coefficients to synthesize the total uncertainty in angle measurement.

[0027] It should be noted that, and Although both represent the positional error of key points, they have different meanings: It is a tiny perturbation artificially applied when calculating the sensitivity coefficient, used to approximate the partial derivative; while This is the standard deviation of the detection uncertainty in the actual output of the attitude estimation model, used to calculate the total uncertainty of angle measurement. Both are in pixels, and a mapping relationship is established through a sensitivity coefficient.

[0028] In an optional implementation, when there are many key points and the sensitivity coefficients vary significantly, a weighted error propagation model can be used to assign higher weights to key points with larger sensitivity coefficients, so as to more accurately reflect the contribution of each key point error to the final angle measurement.

[0029] The confidence interval for the measured angle is output as follows: ; Wherein, δ is a function of the total uncertainty, which is calculated by coupling the detection error of each key point with the sensitivity coefficient.

[0030] In some implementations, a confidence level of 95% can be set, resulting in a confidence interval of ±1.96δ (or approximately ±2δ). Alternatively, a complete angular probability distribution can be output instead of a single confidence interval. For example, a Monte Carlo sampling method can be used to simulate the uncertainty propagation of keypoint locations, thereby obtaining the probability density function of the joint angle. This approach better captures the probability characteristics of non-Gaussian distributions, providing richer uncertainty information for subsequent probability determination models.

[0031] By establishing an error propagation model from the uncertainty of key point location to the uncertainty of joint angle measurement, this embodiment transforms the inherent detection error of the posture estimation model into quantifiable angle measurement uncertainty, effectively solving the error amplification effect problem caused by the small limb length of newborns in the background art. The sensitivity coefficient establishes a quantitative relationship between key point positioning error and angle measurement error, enabling the detection uncertainty of each key point to be reasonably propagated to the final joint angle measurement result. The confidence interval or angle probability distribution provides a reliable uncertainty quantification basis for subsequent probability determination, so that clinical grading determination no longer relies on a single point estimate, but can fully consider the impact of measurement uncertainty on the determination result, thereby significantly reducing the probability of false positives or missed reports.

[0032] In one embodiment, the method further includes an adaptive correction step for image distortion, which specifically includes the following sub-steps: Reference features are obtained from the known geometric constraints of the scene in the current video frame. These reference features include the rectangular border of the crib or the edge lines of the mattress. In some implementations, the reference features can be automatically extracted from the video frame using image edge detection algorithms (such as the Canny operator or Hough transform).

[0033] In an alternative implementation, a more precise reference feature can be provided using a preset crib marking pattern (such as a checkerboard or AprilTag marker). When using the crib's rectangular border as a reference feature, the degree of projection distortion in the image plane can be calculated based on the coordinates of the four corners of the rectangle; when using the mattress edge line as a reference feature, perspective distortion can be assessed based on the curvature of the edge line.

[0034] Based on the projection deformation of the reference features in the image plane, the shooting posture parameters of the camera relative to the newborn's body central axis are inferred. These shooting posture parameters include pitch angle, yaw angle, and roll angle. In some implementations, an extended form of the Direct Linear Transform (DLT) algorithm can be used. By establishing a mapping equation system between the spatial geometric relationship of the reference features and their image projection, the extrinsic parameter matrix of the camera can be obtained, and then the pitch angle, yaw angle, and roll angle can be decomposed. In optional implementations, the Perspective-n-Point (PnP) algorithm can also be used to infer the shooting posture parameters. When the pitch angle exceeds ±15°, the yaw angle exceeds ±20°, or the roll angle exceeds ±10°, it can be determined that there is a significant tilt in the current shooting angle, at which point the necessity for perspective distortion correction increases significantly.

[0035] Based on the shooting posture parameters, the perspective distortion mapping relationship between the image plane and physical space under the current shooting viewpoint is calculated, and the inverse perspective projection transformation from image coordinates to physical space coordinates is constructed. In some implementations, a mathematical relationship between the pinhole camera model and the perspective projection transformation can be established. Let the transformation matrix from the image coordinate system to the camera coordinate system be K, the rotation matrix from the camera coordinate system to the world coordinate system be R, and the translation vector be t. Then, the inverse perspective projection transformation can be expressed as: ; In an optional implementation, when the camera's angle of view is tilted, resulting in different degrees of distortion in different areas of the field of view, a segmented perspective distortion model can be used to apply different distortion correction coefficients to different areas of the image in order to achieve more accurate spatial coordinate recovery.

[0036] The coordinates of the key body points are reprojected and corrected using the inverse perspective projection transformation, and the measured angles are recalculated based on the corrected key point coordinates. In some implementations, the key point coordinates (u, v) in the original image coordinate system can be mapped to three-dimensional coordinates (X, Y, Z) in the physical space coordinate system through the inverse perspective projection transformation, and then the joint angles are calculated in the physical space coordinate system to eliminate the angle measurement deviation caused by perspective projection. In an optional implementation, a two-dimensional approximate correction method can also be used, that is, the key point coordinates are geometrically corrected in the image plane according to the shooting posture parameters, which is suitable for scenarios where the camera is approximately parallel to the plane of the newborn's body and the degree of distortion is small.

[0037] The recalculated measurement angle is used as input to the error propagation model. In some implementations, the corrected measurement angle can more accurately reflect the actual degree of flexion of the neonatal joint, enabling subsequent probabilistic judgment models to perform clinical grading assessments based on more reliable measurement data. In an optional implementation, when the difference in angle before and after fluoroscopic distortion correction exceeds a preset threshold (e.g., 5°), a shooting condition optimization prompt can be triggered, reminding the user to adjust the camera position or angle to obtain more accurate measurement results.

[0038] By introducing an adaptive correction step for shooting distortion, this embodiment effectively solves the perspective distortion problem caused by the tilt of the camera's shooting angle. When the camera is tilted relative to the newborn's body plane by pitch, yaw, or roll angles, the projection of joint angles onto the image plane will produce significant distortion, directly affecting measurement accuracy. By acquiring scene geometric constraints such as the crib frame or mattress edge, this embodiment can automatically estimate shooting posture parameters without manual calibration, and then construct an inverse perspective projection transformation to achieve geometric correction of key point coordinates. The corrected measurement angle is closer to the true angle of the newborn's joints in physical space, providing more reliable input data for subsequent error propagation models and probability determination.

[0039] In one embodiment, the method further includes a coupled processing based on the error propagation model and the adaptive correction of shooting distortion, specifically including the following sub-steps: Obtain the perspective distortion error influence coefficients for each key body point under the aforementioned shooting posture parameters. In some implementations, the sensitivity of the projection offset of each key point in the image plane relative to its spatial position can be calculated based on the specific values ​​of the camera's pitch angle, yaw angle, and roll angle, thus obtaining the perspective distortion error influence coefficients.

[0040] In an optional implementation, key points located at the edge of the image typically have a larger perspective distortion error influence coefficient than key points in the center of the image; when a key point is located far from the central axis of the newborn's body in the image, its perspective distortion error influence coefficient also increases accordingly.

[0041] The perspective distortion error influence coefficient and the key point position detection uncertainty are jointly quantified to obtain a comprehensive angular uncertainty measure under the composite error source.

[0042] In some implementations, the angular error introduced by perspective distortion can be treated as an additional source of uncertainty, and its sum of squares and square roots can be used in a joint calculation with the inherent detection uncertainty of the pose estimation model. ; Among them, δ-detection is the angular uncertainty caused by detection uncertainty, and δ-perspective is the angular uncertainty caused by perspective distortion.

[0043] In an optional implementation, when any one of the pitch angle, yaw angle, or roll angle exceeds a preset safety threshold, the perspective distortion error influence coefficient can be weighted and amplified to reflect a greater risk of systematic error.

[0044] The angle confidence interval is updated based on the comprehensive angle uncertainty measure. In some implementations, when there is a significant difference between the original confidence interval before perspective distortion correction and the corrected confidence interval (e.g., overlap rate less than 80%), the corrected confidence interval is used as the final output. When perspective distortion correction cannot completely eliminate the error introduced by the shooting angle, the comprehensive angle uncertainty measure can more comprehensively reflect the reliability of the measurement results.

[0045] In an optional implementation, the union of the original confidence interval and the corrected confidence interval can be retained as the final confidence interval estimate, thereby ensuring the safety of clinical judgment under conservative boundary assessment.

[0046] By coupling the error propagation model with adaptive correction for image distortion, this embodiment achieves unified quantification of composite error sources. Traditional error propagation models only consider the detection uncertainty of the pose estimation model. However, in real-world applications, changes in the camera's shooting angle introduce additional perspective distortion errors. These two error sources have different physical causes and statistical characteristics. By obtaining the perspective distortion error influence coefficient and performing joint quantification, this embodiment can superimpose the influence of perspective distortion on detection uncertainty, obtaining a more comprehensive measure of overall angle uncertainty. The updated confidence interval fully reflects the contributions of all known error sources during the measurement process, enabling clinical grading decisions to make more reliable decisions based on a full consideration of measurement uncertainty.

[0047] In one embodiment, the step of comparing the measured angle and the angle confidence interval with a preset clinical grading angle threshold using a probability determination model specifically includes the following sub-steps: Obtain the newborn's age in months and birth weight parameters, wherein the birth weight parameters include interval markers for three subcategories: normal weight, low birth weight, and macrosomia.

[0048] In some implementations, the age parameter can be automatically obtained through the hospital information system or manually entered by the user, and the unit of the age parameter is the actual number of months after birth.

[0049] In an optional implementation, the birth weight parameter can be obtained from newborn birth records or periodic physical examination data, and its classification criteria refer to the definition: normal birth weight is 2500g to 4000g, low birth weight is less than 2500g (of which very low birth weight is less than 1500g and extremely low birth weight is less than 1000g), and macrosomia is more than 4000g. When the birth weight parameter is very low birth weight or extremely low birth weight, a stricter judgment standard should be adopted in the subsequent threshold adjustment to adapt to the special characteristics of infant physical development.

[0050] Based on the age parameter and the birth weight parameter, a subset of corresponding clinical grading angle thresholds is retrieved from the database, wherein different age ranges and different birth weight subcategories correspond to different grading angle thresholds.

[0051] In some implementations, a clinical grading threshold matrix can be pre-established, which includes different age ranges (such as 0-1 month, 1-3 months, 3-6 months, 6-12 months) and different birth weight categories (normal weight, low birth weight, macrosomia).

[0052] In an optional implementation, when no perfectly matching combination of age range and weight category exists in the database, an interpolation method (such as linear interpolation or spline interpolation) can be used to calculate the corresponding clinical grading angle threshold. When both the age parameter and birth weight parameter are within the normal range, the standard clinical grading angle threshold is used; when either parameter deviates from the normal range, an adjusted threshold set is retrieved.

[0053] A threshold adjustment function based on clinical grading differences in gestational age and birth weight is established. The formula for the threshold adjustment function is as follows: ; in: The base angle threshold is α(W); the weight correction factor is β(M); the age correction factor is W; the birth weight category is W; and the age is M.

[0054] In some implementations, The baseline angle thresholds can be obtained by querying a pre-built clinical grading angle threshold table, which corresponds to the standard judgment boundary for newborns of normal age (3-6 months) and normal weight (2500g-4000g).

[0055] The value of α(W) can be set as follows: α(W) = 1.0 for normal weight category, α(W) = 0.85-0.95 for low birth weight category, and α(W) = 1.05-1.15 for macrosomic infant category. The specific value of the correction factor can be determined based on clinical research data.

[0056] In an optional implementation, the value of β(M) can be set as follows: β(M) = 1.10-1.20 for 0-1 months of age, β(M) = 1.0 for 3-6 months of age, and β(M) = 0.90-1.00 for 6-12 months of age, to reflect the differences in joint range of motion and muscle tone development at different ages. When the age parameter and / or birth weight parameter are at the critical boundary, the adjustment function can adopt a smoothly transitioning sigmoid function form to avoid judgment oscillations caused by threshold abrupt changes.

[0057] It should be noted that β(M) is used to adjust the clinical grading angle threshold (i.e., the angle boundary value used to determine normal / mild / moderate pathology), while the first risk threshold and the second risk threshold in claim 6 are used to adjust the warning trigger sensitivity. The two are independent of each other: the larger β(M), the stricter the grading angle threshold (more easily judged as abnormal); the lower the risk threshold, the more sensitive the warning trigger. For example, β(M) = 1.10 for 0-1 month old infants indicates a 10% increase in the grading angle threshold, meaning a stricter judgment of abnormal postures; simultaneously, a 10%-20% decrease in the first risk threshold indicates a relaxed warning trigger condition. This independent bidirectional adjustment mechanism allows the system to control the judgment criteria and warning sensitivity separately, improving clinical adaptability.

[0058] By introducing differentiated threshold adjustments based on age and birth weight parameters, this embodiment achieves personalized clinical grading for newborns with different physiological characteristics. Newborn joint mobility and muscle tone development exhibit significant age-related and individual variability; the clinical significance of the same measurement angle may differ significantly under different ages and weights. Traditional fixed threshold methods cannot accommodate these physiological developmental differences, easily leading to increased false positive rates for low-age or low-weight infants and increased underreporting rates for high-age or high-weight infants. By establishing a threshold function that is dually adjusted for age and birth weight, this embodiment can dynamically modify the clinical grading criteria based on the specific physiological characteristics of each newborn, making the identification of abnormal postures more consistent with the principles of personalized medicine and significantly improving the clinical applicability and accuracy of the early warning system.

[0059] In one specific implementation, the probabilistic decision model can employ Bayesian inference. Assume the measured angle θ follows a normal distribution N( , σ 2 ),in Let σ be the angle measurement value, and σ be the angle uncertainty (output by the error propagation model). Then, the current posture belongs to a certain clinical grading interval (such as the normal interval [0, 1 ... ], mild pathological range , ], moderate pathological range [ , The probability of ]) is: P(normal) = Φ(( - ) / σ)-Φ((0- ) / σ); P(mild) = Φ(( - ) / σ)-Φ(( - ) / σ); P(moderate) = Φ(( - ) / σ)-Φ(( - ) / σ); Where Φ(·) is the cumulative distribution function of the standard normal distribution. The abnormal attitude risk level is the level with the highest probability, and this probability value is output as the confidence level. When the maximum probability value is lower than a preset threshold (e.g., 0.6), it is determined to be a monitoring boundary state.

[0060] In one embodiment, generating graded early warning information specifically includes the following graded determination logic: When the abnormal posture risk level is higher than the first risk threshold but lower than the second risk threshold, and the angle confidence interval crosses the clinical grading angle threshold, it is determined that the current state is at the monitoring boundary, and a prompting message is sent to the user terminal. The prompting message includes the current measurement angle, the angle confidence interval, and suggestions for improving the shooting conditions.

[0061] In some implementations, the triggering condition for the monitoring boundary state can be specifically described as follows: both P(normal|θ) and P(abnormal|θ) are in the medium probability range (e.g., 30%-70%), and the confidence interval width exceeds the preset narrow threshold (e.g., 10°), indicating that the current measurement has high uncertainty but is close to the classification boundary.

[0062] In an optional implementation, when a monitoring boundary condition is determined, in addition to sending a prompt message, a high-frequency sampling mode (e.g., increasing from 15 frames per second to 30 frames per second) can be activated to accumulate more measurement data in a short time for subsequent confidence enhancement analysis. Suggestions for improving shooting conditions may include specific guidance measures such as adjusting the camera position to face the newborn's midline, reducing changes in ambient light, and removing unnecessary obstructions.

[0063] When the risk level of the abnormal posture is higher than the second risk threshold, and the confidence interval of the angle is completely within the clinical abnormality grading interval, it is determined to be a high-confidence abnormal state, and an emergency warning message is sent to the user terminal. In some implementations, the determination condition of a high-confidence abnormal state can be specifically expressed as: P(abnormal|θ) is higher than the second risk threshold (e.g., 85%), and the upper and lower bounds of the confidence interval [θ-δ, θ+δ] are both within the abnormal grading angle range, indicating that even considering measurement uncertainty, the current posture still belongs to the clinical abnormality category with a high probability.

[0064] In an optional implementation, the emergency warning information may include an abnormality type identifier (such as adductor angle abnormality, foot dorsiflexion angle abnormality, etc.), the current measurement angle and confidence interval, the specific value exceeding the threshold, and detailed information such as suggested clinical intervention measures. When a high-confidence abnormality is determined, the system can simultaneously push the warning to preset medical monitoring terminals to shorten the delay time of clinical response.

[0065] The first risk threshold and the second risk threshold are dynamically adjusted based on the birth weight and age parameters, wherein the risk threshold for low birth weight infants is lowered, the risk threshold for macrosomic infants is maintained at the normal baseline level, and the risk threshold for normal weight infants is raised.

[0066] In some implementations, when the birth weight parameter is low birth weight or very low birth weight, the first risk threshold and the second risk threshold can be reduced by 5%-15% from the normal baseline to accommodate the physiological characteristics of low birth weight infants with low or high muscle tone.

[0067] In an optional implementation, when the age parameter is in the 0-1 month range, the first risk threshold can be appropriately reduced by 10%-20%, because newborns at this stage have weak voluntary motor ability and may maintain a single posture for a long time, requiring a more sensitive monitoring strategy; when the age parameter is in the 6-12 month range and the birth weight is macrosomic, the second risk threshold can be appropriately increased by 5%-10% to reduce false alarms caused by normal movement due to the increased activity of the infant.

[0068] By establishing a tiered early warning mechanism, this embodiment achieves differentiated responses to abnormal postures at different risk levels. The determination of monitoring boundary states utilizes the key feature that the angle confidence interval crosses the clinical grading threshold. When the confidence interval simultaneously covers both the normal and abnormal ranges, it indicates that the current measurement data is in a judgment ambiguity zone. In this case, it is not advisable to issue an emergency warning directly, but attention is still required. Suggestions for improving shooting conditions in the prompting information can help users obtain more accurate data in subsequent measurements, helping to eliminate uncertainty in subsequent judgments. The determination of high-confidence abnormal states requires both the abnormal risk level and the confidence interval to be met simultaneously, ensuring the reliability of the warning while avoiding misjudgments that may arise from relying on a single indicator. By introducing birth weight and age parameters to dynamically adjust the risk threshold, the warning sensitivity can adapt to individual differences in newborns with different physiological characteristics, significantly reducing the underreporting rate of low-birth-weight infants and the false alarm rate of normal-birth-weight infants.

[0069] In one embodiment, a respiratory motion analysis step is also included, specifically comprising the following sub-steps: When the current posture is detected to be a prone posture and the upper limit of the confidence interval of the measured angle exceeds the preset prone angle threshold (e.g., 60°), the spatiotemporal feature sequence of the infant's chest region is extracted from the real-time video stream.

[0070] In a preferred embodiment, the preset prone angle threshold can be set to 60°. This threshold is based on clinical studies showing that when an infant's prone angle exceeds 60°, the probability of their mouth and nose being completely covered by the bed surface increases significantly, raising the risk of suffocation. However, this threshold can be dynamically adjusted according to the infant's age and weight: for younger infants or those with lower weight, the threshold can be appropriately lowered to 50°, while for older infants or those with higher weight, it can be increased to 70°.

[0071] In some implementations, the bounding box of the infant's chest region can be determined based on the location of key body points, and then an image sequence of the chest region can be cropped from the real-time video stream as input data for subsequent analysis.

[0072] In an optional implementation, when the newborn is in a prone position, the chest area may be obstructed to varying degrees relative to the camera. In this case, a temporal median filtering method can be used to select the frame with the least degree of obstruction from multiple consecutive frames as the analysis sample. When the prone angle exceeds 60° and the duration exceeds a preset duration (e.g., 2 minutes), a necessity assessment for respiratory motion analysis is triggered.

[0073] The spatiotemporal feature sequence is subjected to motion amplification processing, which includes constructing a Gaussian pyramid, identifying weakly moving pixels based on features at each level, constructing a Laplacian pyramid, and amplifying and superimposing each level.

[0074] In some implementations, the Euler video magnification algorithm can be used to amplify respiratory motion: First, a multi-scale Gaussian pyramid is constructed for the input video frames. Then, the difference results between adjacent frames are frequency-domain filtered to extract motion components within the frequency range of interest (e.g., 0.1Hz-0.5Hz, corresponding to a respiratory rate of 6-30 breaths / minute). Finally, the amplified motion components are superimposed back onto the original frames to reconstruct the video. In optional implementations, the motion magnification factor can range from 50 to 200 times. The specific value can be adjusted according to the scale of the newborn's chest region in the image: a smaller magnification factor (e.g., 50-80 times) is used when the chest region occupies a large area of ​​the image, and a larger magnification factor (e.g., 150-200 times) is used when the chest region occupies a small area of ​​the image, to ensure that the amplified respiratory motion amplitude is easy to observe without producing obvious artifacts.

[0075] Frequency domain analysis was performed on the magnified respiratory motion feature sequence to extract respiratory frequency features.

[0076] In some implementations, a Fast Fourier Transform (FFT) can be performed on the time-enlarged pixel values ​​of the chest region to obtain the spectrum, and then the peak frequency can be extracted from the spectrum as the respiratory frequency estimate. In optional implementations, when the respiratory frequency estimate exhibits multi-peak phenomena (such as the simultaneous presence of respiratory frequency and heart rate frequency components), bandpass filtering preprocessing or spectral peak clustering methods can be used to separate the respiratory frequency components; when there are no significant peaks in the spectrum (such as peak power being less than twice the global average power), it can be determined that the respiratory motion signal in the current video frame is too weak to reliably extract the respiratory frequency.

[0077] The respiratory rate characteristics are compared with a normal respiratory rate range model selected based on age and birth weight parameters. When the respiratory rate is lower than the first respiratory threshold or higher than the second respiratory threshold, a respiratory abnormality warning is generated.

[0078] In some implementations, the normal range of respiratory rate can be set as follows: 25-45 breaths / minute for infants with normal weight aged 0-1 month, 30-50 breaths / minute for infants aged 1-3 months, 25-45 breaths / minute for infants aged 3-6 months, and 20-35 breaths / minute for infants aged 6-12 months.

[0079] It should be noted that when the age parameter falls exactly at the boundary of an interval (e.g., 1 month or 3 months old), the system adopts a higher safety priority judgment rule: that is, if the respiratory rate ranges of two adjacent two age intervals overlap, the interval with the lower upper limit or higher lower limit is preferred to ensure sensitivity to respiratory abnormalities. For example, the overlapping interval between 0-1 month and 1-3 month is 30-45 breaths / minute. For an infant who is exactly 1 month old, if their respiratory rate is 30-45 breaths / minute, the system will still judge it as normal to reduce false alarms; if the respiratory rate exceeds this overlapping interval, it will be judged as abnormal according to the corresponding interval. In an optional implementation, when the birth weight parameter is low birth weight, the upper limit of the normal respiratory rate range can be appropriately increased by 5-10 breaths / minute to reflect the physiological characteristic that low birth weight infants usually have a compensatory faster heart rate; when the birth weight parameter is macrosomia, the lower limit of the normal respiratory rate range can be appropriately decreased by 3-5 breaths / minute. The first respiratory threshold can be set to 80% of the lower limit of the normal range, and the second respiratory threshold can be set to 120% of the upper limit of the normal range to achieve a balance between sensitivity and specificity.

[0080] By introducing a respiratory motion analysis step, this embodiment further assesses respiratory function status based on the detected prone position. Prone position is a significant risk factor for Sudden Infant Death Syndrome (SIDS), and prolonged prone positioning accompanied by respiratory dysfunction requires timely clinical attention. Traditional posture monitoring methods focus only on posture type while neglecting respiratory function. This embodiment extracts subtle chest rise and fall movements from the video using motion magnification technology and combines this with frequency domain analysis to obtain an estimated respiratory rate. When the detected respiratory rate deviates from the normal range selected based on age and birth weight, the system generates a respiratory abnormality warning, providing additional clinical reference information for healthcare professionals. The normal range model for respiratory rate considers the impact of age and weight on respiratory function development, making abnormality assessment more in line with the requirements of personalized medicine.

[0081] In one embodiment, an anxiety movement detection step is also included, specifically comprising the following sub-steps: Extract the temporal posture sequence of the newborn's continuous whole-body movements from the real-time video stream.

[0082] In some implementations, a temporal pose sequence can be constructed based on the body key point positions of consecutive frames (e.g., 15-30 frames, corresponding to 1-2 second video segments). The temporal pose sequence can be represented as a multivariate time series of key point coordinates changing over time.

[0083] In an optional implementation, when the newborn is asleep, the temporal posture sequence shows relatively small variations; when the newborn is awake and active, the temporal posture sequence contains richer information on movement patterns. To accurately assess the presence of restless movements, video clips of the newborn in an awake state can be selected for analysis. Specifically, the newborn's arousal state can be determined by analyzing the variance or spectral characteristics of key point movements.

[0084] The presence of uneasy motion is detected based on the temporal posture sequence, and an uneasy motion quality score is output.

[0085] In some implementations, a pre-trained motion pattern recognition model (such as one based on a temporal convolutional network or a long short-term memory network) can be used to analyze the motion characteristics of a temporal pose sequence and output the probability of the presence of uneasy motion or a quality score.

[0086] In an optional implementation, the restlessness movement quality score can be quantified as a score from 0 to 100, where 0 indicates the complete absence of restlessness movement characteristics (such as maintaining a single posture continuously or only exhibiting slight tremors), and 100 indicates the presence of typical restlessness movement characteristics (such as complex, coordinated movements of various parts of the body). Typical characteristics of restlessness movement include: the complexity and variability of the movements, coordinated movements of both sides of the limbs, and the degree of involvement of various parts of the body. When the restlessness movement quality score is in the middle range (such as 40-60 points), it indicates that the current movement pattern is atypical, and it is necessary to extend the observation time or increase the number of sampling segments to improve the reliability of the judgment.

[0087] Based on the birth weight parameters, the scoring threshold for judging restlessness and abnormal movement is adaptively adjusted, with the tolerance for the judgment of restlessness and abnormal movement in very low birth weight infants being relaxed, and the judgment threshold for restlessness and abnormal movement in macrosomic infants being lowered.

[0088] In some implementations, the scoring threshold for judging restlessness and abnormal movement is adaptively adjusted based on the birth weight parameter. Specifically, the tolerance for judging restlessness and abnormal movement in very low birth weight infants (below 1500g) is lowered by 10% to 20% compared to the judgment threshold for normal birth weight infants (e.g., 60 points) (i.e., the threshold is reduced to 48 to 54 points); the judgment threshold for judging restlessness and abnormal movement in macrosomic infants (above 4000g) is lowered by 5% to 10% (i.e., the threshold is reduced to 54 to 57 points).

[0089] The reasons are as follows: In very low birth weight infants, neurological development may be delayed, and the typicality and coordination of restless movements may be below normal developmental levels; therefore, a more lenient threshold is needed to avoid over-prediction. On the other hand, macrosomic infants may have birth injuries or neurodevelopmental risks, and lowering the threshold can increase the sensitivity of detecting potential abnormalities. When the birth weight parameter is within the normal range, a standard threshold is used (i.e., a quality score below 60 is considered a lack of restless movements).

[0090] When the restlessness movement quality score is lower than the adjusted judgment threshold and continues to exceed the preset time threshold, a neurodevelopmental screening warning is generated.

[0091] In some implementations, the preset time threshold can be set as the duration of absent restless movements for 5-10 minutes consecutively to exclude interference from transient periods of stillness (such as blinking or attention shifting). In an optional implementation, when the daily percentage of restless movement time is below normal for more than 3 consecutive days, in addition to generating a real-time alert, parents can be advised to schedule a follow-up visit to a neurodevelopmental specialist clinic. The neurodevelopmental screening alert may include information such as the time-series change curve of the restless movement quality score, the specific value of the adjusted judgment threshold, and the suggested clinical assessment time window.

[0092] By introducing a restlessness movement detection step, this embodiment combines posture monitoring with neurodevelopmental assessment, providing a technical means for early screening of the neonatal nervous system. Restlessness movements are an important clinical indicator for assessing the developmental status of an infant's nervous system. Typical restlessness movements manifest as complex, coordinated movements of various parts of the body; their absence or abnormality may indicate neurodevelopmental disorders such as cerebral palsy or developmental delay. Traditional clinical assessment relies on the subjective observation of professional physicians. This embodiment achieves objective quantitative scoring of restlessness movements through temporal posture sequence analysis. When the restlessness movement quality score consistently falls below a threshold adjusted for birth weight, the system generates a neurodevelopmental screening warning, providing a time window for early intervention. By introducing birth weight parameters to adaptively adjust the threshold, this embodiment can consider the differences in neurodevelopmental timing between low-birth-weight and macrosomic infants, reducing misjudgments caused by physiological characteristics and improving the clinical reliability of the warning system.

[0093] In one embodiment, the method for training the pose estimation model is also included, specifically comprising the following sub-steps: During the training phase, a newborn posture dataset with annotations of body key points and joint angles was obtained.

[0094] In some implementations, the newborn posture dataset can be manually labeled by professional annotators based on video frames, marking the positions of key body points (such as shoulder joints, elbow joints, hip joints, knee joints, etc.), and the corresponding joint angle values ​​can be simultaneously labeled using joint angle measurement tools (such as protractors or medical image analysis software).

[0095] In optional implementations, to ensure annotation quality, keypoint location annotations can be processed by averaging or majority voting after independent annotation by multiple people. Samples with significant disagreements in joint angle annotations are removed or re-annotated. The dataset size can be set to at least 1000 video clips, each clip containing sample distributions of different ages (0-12 months), different birth weight categories (normal weight, low birth weight, macrosomia), and different posture types (supine, prone, lateral) to cover the main scene variations in practical applications.

[0096] A multi-task loss function is constructed, comprising a keypoint localization loss term and a joint angle alignment loss term. The joint angle alignment loss term is calculated based on the deviation between the joint angles derived from the keypoints output by the model and the actual joint angles. In some implementations, the keypoint localization loss term can use mean squared error loss. ; in, Let i be the location of the i-th keypoint predicted by the model. This represents the actual location of the annotation. The joint angle alignment loss term can be calculated using the angle mean square error loss: ; Or L1 angle loss: ; in, The joint angles are calculated based on predicted key points. The actual joint angles are labeled. In an optional implementation, when multiple joint angles need to be measured in a single video frame (e.g., simultaneously measuring the adductor angle and dorsiflexion angle), the weighted average or maximum value of the loss terms for each joint angle can be taken as the final value. To balance the optimization contributions of different joints.

[0097] The joint angle alignment loss term and the keypoint localization loss term are weighted and combined to form the total loss function for model optimization. In some implementations, the total loss function can be expressed as: ; in, and To balance the weights of the two loss terms, a hyperparameter can be set. In an optional implementation, when the accuracy of joint angle measurement is the primary optimization goal for clinical applications, this hyperparameter can be set... > (like =0.3, =0.7); when the accuracy of key point positioning and the accuracy of angle measurement are equally important, it can be set to 0.7. = When adaptive weight adjustment is required during the training phase, an uncertainty-weighted method can be used, dynamically adjusting the weights based on the variance of the two loss terms during training. and The value of allows the model to automatically balance the optimization difficulty of different tasks.

[0098] The pose estimation model is trained end-to-end based on the total loss function, so that the model can simultaneously optimize the key point localization accuracy and joint angle measurement accuracy during the training process.

[0099] In some implementations, stochastic gradient descent (SGD) or the Adam optimizer can be used to iteratively update the model parameters. The learning rate can be set to 0.001-0.01, and the number of training epochs can be set to 50-200. The specific parameters can be tuned based on the performance on the validation set.

[0100] In optional implementations, to improve the model's generalization ability, data augmentation strategies can be introduced during training, such as random cropping, brightness / contrast adjustment, horizontal flipping (with simultaneous adjustment of keypoint annotations), and random noise injection. When the joint angle measurement error on the validation set no longer decreases for several consecutive rounds (e.g., 5-10 rounds), an early stopping mechanism can be triggered to prevent overfitting.

[0101] By constructing a multi-task loss function and jointly optimizing keypoint localization and joint angle measurement, this embodiment solves the problem of misalignment between the training objectives of posture estimation models and clinical application needs in the background art. Traditional posture estimation models optimize keypoint localization accuracy (such as PCK, OKS), but the limb length of newborns is much smaller than that of adults. The same pixel-level localization error will be amplified when calculating joint angles, leading to the failure of clinical grading.

[0102] This embodiment introduces a joint angle alignment loss term into the loss function, enabling the model to perceive angle measurement errors during training and use them as one of the optimization signals in backpropagation. Through joint optimization, the model can learn more feature representations that are more favorable for joint angle measurement, thereby significantly improving angle measurement accuracy while maintaining keypoint localization accuracy. Experiments show that the model trained with multi-task loss reduces the error in joint angle measurement tasks by 30%-50% compared to the model that only optimizes keypoint localization, effectively narrowing the gap between measurement accuracy and clinical grading intervals.

[0103] In one embodiment, the method further includes temporal multi-frame fusion optimization based on angle confidence intervals, specifically comprising the following sub-steps: Obtain multiple measured angle values ​​of the same target joint and their corresponding angle confidence intervals from multiple consecutive frames of images.

[0104] In some implementations, a time window length of 5-15 frames (i.e., a 0.2-1 second video segment) can be set, and the measured angle values ​​and confidence intervals of all frames can be extracted within this time window. In optional implementations, when there are significant differences in the measured angle values ​​within the time window (e.g., the difference between the maximum and minimum values ​​exceeds 15°), outlier detection can be performed on the data within the time window to remove outlier measurement points caused by occlusion or keypoint detection failure; when the width of the angle confidence interval of a frame exceeds a preset wide threshold (e.g., 20°), the weight contribution of the data of that frame in the subsequent fusion process can be reduced or the frame can be directly excluded.

[0105] The confidence interval widths of each frame are compared; the narrower the confidence interval, the more reliable the angle measurement in that frame.

[0106] In some implementations, the reliability index R can be defined as 1 / ,in Let be the standard deviation of the uncertainty of the measured angle in the i-th frame, and let the confidence interval width be . Proportional; when When the value is smaller, the reliability index R is larger, indicating that the measurement results of that frame are more reliable.

[0107] In optional implementations, in addition to assessing reliability based on the confidence interval width, other reliability indicators can be combined for comprehensive judgment, such as the visibility score of keypoint detection, the geometric consistency between keypoints (e.g., limb length should remain relatively stable between adjacent frames), and the degree of deviation from the measurement value of the previous frame. When multiple indicators give conflicting reliability assessments, a weighted voting mechanism can be used to determine the final reliability ranking.

[0108] The multiple measured angle values ​​are weighted and fused, where the weight of each frame's measured angle is determined by an inverse proportional function to the width of its confidence interval. The weight coefficient is calculated using the following formula: ; in, This represents the standard deviation (in degrees) of the uncertainty of the measured angle in the i-th frame, and its value is positively correlated with the width of the confidence interval. The physical meaning of this weighting method is: the more reliable the measurement (…), the higher the confidence interval width. Smaller frames receive higher fusion weights, and the weights are related to... It is inversely proportional to the square of the result, thus effectively suppressing the contamination of the fusion result by low-quality frames.

[0109] In some implementations, when When =1°, the weight of this frame is 1; when When the angle is 2°, the weight of this frame is 1 / 4; when... When the angle is 5°, the weight of this frame is 1 / 25. This shows that frames with narrower confidence intervals have a significantly increased weight in the fusion result, inversely proportional to the square of the sum of its parts, meaning that the fused measurement angle is mainly contributed by high-confidence frames.

[0110] In alternative implementations, other forms of weighting functions can also be used, such as weight calculation based on information entropy: = ; in Using a preset reference standard deviation, this function form exhibits a relatively mild weight amplification effect on high-confidence frames. When the number of frames involved in fusion is small (e.g., less than 3 frames), equal-weighted averaging can be used as an alternative fusion scheme.

[0111] The fused measurement angle value is output, and the width of the fused angle confidence interval is smaller than the weighted average of the confidence interval widths of each frame, thereby achieving multi-frame temporal fusion noise reduction.

[0112] In some implementations, the fused measurement angle is: ; The angular uncertainty after fusion can be calculated as follows: ; The corresponding confidence interval after fusion is [ - , + ],in =k· , where k is a constant corresponding to the confidence level (e.g., k=2 corresponds to approximately a 95% confidence level, meaning the probability that the true angle value falls within the confidence interval is 95%. This confidence interval provides crucial reliability information for subsequent probability determination models). Due to the properties of the harmonic mean, Strictly less than Σ · (Weighted arithmetic mean), therefore the width of the fused confidence interval must be smaller than the weighted average of the confidence intervals of each frame.

[0113] In optional implementations, when the measured angle within the time window exhibits a trend (e.g., the angle value continuously increases or decreases), a weighted moving average or exponentially weighted moving average method can be used to better track the trend of angle changes rather than simple smoothing. When a rapid change in the measured angle is detected that exceeds the range of normal physiological activity (e.g., a change of more than 30° within 1 second), a measurement data quality alarm can be triggered, indicating possible sensor malfunction or severe external interference.

[0114] By introducing a temporal multi-frame fusion optimization step, this embodiment effectively utilizes the temporal redundancy information between consecutive video frames to improve the reliability and stability of angle measurements. Single-frame measurements are inevitably affected by factors such as the detection uncertainty of the pose estimation model, video noise, and slight occlusion, leading to fluctuations in the measurement results. By weighted fusion of angle measurements from multiple frames within a time window, this embodiment can amplify the contribution of high-confidence frames and compress the contribution of low-confidence frames, thereby obtaining more stable measurement results than any single frame. The compression effect of the confidence interval width after fusion means an overall reduction in measurement uncertainty, which is of great significance for clinical grading: when the confidence interval narrows, the probability of measurement data crossing the clinical grading threshold decreases, and the uncertainty of the judgment result is also reduced, thus significantly reducing the risk of false positives or missed reports. The temporal fusion denoising technology effectively improves the system's measurement performance without increasing hardware costs, demonstrating the significant benefits of software algorithm optimization.

[0115] 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. A real-time early warning method for abnormal neonatal postures using video analysis, characterized in that, include: Acquire real-time video streams of the newborn's sleep area; The pose of the real-time video stream is estimated frame by frame to obtain the position of the newborn's body key points, and the measurement angle of the target joint is calculated based on the position of the body key points. Obtain a measure of the uncertainty in the detection of the body key point positions by the pose estimation model; An error propagation model is established from the uncertainty of key point location to the uncertainty of joint angle measurement. Based on the error propagation model and the detection uncertainty metric, the angle confidence interval or angle probability distribution of the measured angle is calculated. Based on the measured angle and the angle confidence interval, a probability judgment model is used to compare with the preset clinical grading angle threshold, calculate the conditional probability of the current posture relative to each clinical grading level, and obtain the abnormal posture risk level. When the risk level of the abnormal posture exceeds the preset warning threshold, a graded warning message is generated.

2. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 1, characterized in that: The establishment of the error propagation model from the uncertainty of key point location to the uncertainty of joint angle measurement includes: Using the detection uncertainty of each key body point as input, the partial derivative of the position of each key point is calculated using the geometric calculation formula of the joint angle to obtain the sensitivity coefficient of the angle error to the position error of each key point. The formula for calculating the absolute value of the sensitivity coefficient is as follows: ； Where θ is the joint angle. Let be the coordinates of the i-th key point. Let i be the position offset of the i-th key point. The Euclidean distance connecting two key points; Based on the sensitivity coefficient and the detection uncertainty of each key point, the total uncertainty of the target joint angle is calculated according to the error propagation theory. The confidence interval for the measured angle is output as follows: ； in, The total uncertainty is a function of the total uncertainty, calculated by coupling the detection error and sensitivity coefficient of each key point.

3. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 2, characterized in that: It also includes an adaptive distortion correction step: Obtain reference features of known geometric constraints of the scene in the current video frame, the reference features including the rectangular border of the crib or the edge line of the mattress; Based on the projection deformation of the reference feature in the image plane, the shooting posture parameters of the camera relative to the central axis of the newborn's body are inferred. The shooting posture parameters include pitch angle, yaw angle and roll angle. Based on the shooting posture parameters, the perspective distortion mapping relationship between the image plane and the physical space under the current shooting perspective is calculated, and the inverse perspective projection transformation from image coordinates to physical space coordinates is constructed. The coordinates of the key body points are reprojected and corrected using the inverse perspective projection transformation, and the measured angle is recalculated based on the corrected key point coordinates. The recalculated measurement angle is used as the input to the error propagation model.

4. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 3, characterized in that: It also includes a coupled processing based on the error propagation model and the adaptive correction of shooting distortion: Obtain the perspective distortion error influence coefficient of each key body point under the shooting posture parameters; The perspective distortion error influence coefficient and the key point position detection uncertainty are jointly quantified to obtain a comprehensive angle uncertainty measure under the composite error source. The angle confidence interval is updated based on the comprehensive angle uncertainty measure.

5. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 1, characterized in that: The step of comparing the measured angle and the angle confidence interval with a preset clinical grading angle threshold using a probability determination model includes: Obtain the newborn's age in months and birth weight parameters, wherein the birth weight parameters include interval markers for three subcategories: normal weight, low birth weight, and macrosomia. Based on the age parameter and the birth weight parameter, retrieve the corresponding clinical grading angle threshold subset from the database, where different age ranges and different birth weight subcategories correspond to different grading angle thresholds; Establish a threshold adjustment function for clinical grading based on differences in age and birth weight; The threshold adjustment function formula is as follows: ； in, The basic angle threshold; Weight correction factor; Age-corrected factor; Birth weight category; Age in months.

6. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 5, characterized in that: The generated tiered early warning information includes: When the abnormal posture risk level is higher than the first risk threshold but lower than the second risk threshold, and the angle confidence interval crosses the clinical grading angle threshold, it is determined that the current state is at the monitoring boundary, and a prompting message is sent to the user terminal. The prompting message includes the current measurement angle, the angle confidence interval, and suggestions for improving the shooting conditions. When the risk level of the abnormal posture is higher than the second risk threshold, and the angle confidence interval is completely within the clinical abnormality grading interval, it is determined to be a high-confidence abnormal state, and an emergency warning message is sent to the user terminal. The first risk threshold and the second risk threshold are dynamically adjusted based on the birth weight and age parameters, wherein the risk threshold for low birth weight infants is lowered, the risk threshold for macrosomic infants is maintained at the normal baseline level, and the risk threshold for normal weight infants is raised.

7. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 1, characterized in that: It also includes respiratory motion analysis steps: When the current posture is detected to be a prone posture and the upper limit of the confidence interval of the measured angle exceeds the preset prone angle threshold, the spatiotemporal feature sequence of the infant's chest region is extracted from the real-time video stream; The spatiotemporal feature sequence is subjected to motion amplification processing, which includes constructing a Gaussian pyramid, identifying weakly moving pixels based on features at each level, constructing a Laplacian pyramid, and amplifying and superimposing each level. Frequency domain analysis was performed on the magnified respiratory motion feature sequence to extract respiratory frequency features; The respiratory rate characteristics are compared with a normal respiratory rate range model selected based on age and birth weight parameters. When the respiratory rate is lower than the first respiratory threshold or higher than the second respiratory threshold, a respiratory abnormality warning is generated.

8. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 5, characterized in that: It also includes a restlessness movement detection step: Extract the temporal posture sequence of the newborn's continuous whole-body movements from the real-time video stream; Based on the temporal posture sequence, the existence of uneasy motion is detected, and an uneasy motion quality score is output. Based on the birth weight parameters, the scoring threshold for judging restlessness and abnormal movement is adaptively adjusted, with the tolerance for judging restlessness and abnormal movement in very low birth weight infants being relaxed, and the threshold for judging restlessness and abnormal movement in macrosomic infants being lowered. When the restlessness movement quality score is lower than the adjusted judgment threshold and continues to exceed the preset time threshold, a neurodevelopmental screening warning is generated.

9. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 5, characterized in that: It also includes training methods for pose estimation models: During the training phase, a newborn posture dataset with annotations of body key points and joint angles was obtained. Construct a multi-task loss function, which includes a key point localization loss term and a joint angle alignment loss term. The calculation of the joint angle alignment loss term is based on the deviation between the joint angles derived from the key points output by the model and the actual joint angles. The joint angle alignment loss term and the key point localization loss term are weighted and combined to form the total loss function for model optimization. The pose estimation model is trained end-to-end based on the total loss function, so that the model can simultaneously optimize the key point localization accuracy and joint angle measurement accuracy during the training process.

10. The real-time early warning method for abnormal neonatal postures using video analysis according to claim 5, characterized in that: It also includes temporal multi-frame fusion optimization based on angle confidence intervals: Obtain multiple measured angle values ​​of the same target joint and their corresponding angle confidence intervals in multiple consecutive frames of images; The widths of the confidence intervals for each frame are compared; the narrower the confidence interval, the more reliable the angle measurement for that frame. The multiple measured angle values ​​are weighted and fused, where the weight of each frame's measured angle is determined by an inverse proportional function of its confidence interval width. The weight coefficient is calculated using the following formula: ； in, For the first Standard deviation of the uncertainty of the frame measurement angle; The fused measurement angle value is output, and the width of the fused angle confidence interval is smaller than the weighted average of the confidence interval widths of each frame, thereby achieving multi-frame temporal fusion noise reduction.

Images