A human posture anomaly evaluation method based on a body measurement integrated machine

Abstract

This application provides a method for assessing abnormal human posture based on a body composition analyzer, comprising: acquiring video data of a subject's bending-over-to-touch-the-foot movement captured by multiple cameras; preliminarily locating key points in the spine, the transition area between the lumbar and thoracic vertebrae, and the legs; extracting the coordinates of the key points; calculating the change in leg joint angles based on the coordinates of the leg key points; if the change in angle is lower than a preset threshold, it is determined that the leg flexibility is insufficient, and the key point tracking strategy is adjusted to increase the tracking density in the leg area; generating a postural abnormality report of the subject's bending-over-to-touch-the-foot movement based on the postural abnormality score, marking the specific locations of abnormal spinal curvature, increased stiffness in the transition area between the lumbar and thoracic vertebrae, and insufficient leg flexibility; and dynamically updating the key point tracking strategy and the postural abnormality judgment threshold based on the movement standardization score to optimize the accuracy and real-time performance of subsequent movement assessments.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for assessing abnormal human posture based on an integrated body measurement machine. Background Technology

[0002] Human posture assessment has garnered significant attention in recent years due to its crucial role in health monitoring, sports rehabilitation, and intelligent interaction. Precise analysis of dynamic posture using integrated body composition analyzers not only reveals individual athletic ability and health status but also holds irreplaceable value for injury prevention and training optimization. In this approach, key points play a vital role in accurately locating and tracking key body parts such as the spine, lumbothoracic transition zone, and legs. Dynamically adjusting key point tracking is crucial for adapting to postural changes, uneven spinal curvature, and compensatory patterns during movement; this allows the system to capture important abnormal indicators missed by static tracking. However, the limitations of existing methods are becoming increasingly apparent. Most solutions rely on static key point detection or single action templates, making it difficult to adapt to complex postural changes and individual differences in dynamic testing, resulting in insufficient accuracy and universality of assessment results. Especially when multi-joint coordinated movements are involved, traditional techniques often neglect the continuity of joint angle change trajectories and the dynamic compensation mechanisms between different body parts, leading to significant biases in the identification of abnormal postures. Current research faces core technical challenges focused on capturing and analyzing key factors such as the continuity of joint angle changes, the non-uniformity of spinal curvature distribution, and flexibility compensation patterns during dynamic movements. Because these factors remain unresolved, existing systems often fail to accurately reflect uneven spinal curvature, increased stiffness in the transition zone between the lumbar and thoracic vertebrae, and leg flexibility compensation when tracking complex movements such as bending over and touching the toes. This not only increases the difficulty of keypoint tracking but also leads to misjudgments or omissions of posture deviations due to the lack of adaptive anomaly detection thresholds, thus affecting the reliability of the overall assessment. Therefore, how to dynamically adjust the keypoint tracking strategy to adapt to changes in uneven spinal curvature and joint stiffness during test movements such as bending over and touching the toes, while simultaneously optimizing the posture anomaly detection threshold to identify leg flexibility compensation patterns, has become a key issue in improving the accuracy of body posture anomaly assessment. Solving this problem requires finding a balance between dynamic data acquisition and real-time analysis to ensure that the technical strategy can accurately address the challenges posed by individual differences and the complexity of movements. Summary of the Invention

[0003] This invention provides a method for assessing abnormal human posture based on a body composition analyzer, mainly including:

[0004] Acquire video data of subjects bending over and touching their feet collected by multiple cameras, perform preliminary localization of key points in the spine, the transition area between the lumbar and thoracic vertebrae and the legs, and extract the coordinates of the key points;

[0005] By using the coordinates of key points in the transition area between the lumbar and thoracic vertebrae, the range of motion in the transition area between the lumbar and thoracic vertebrae is calculated. If the range of motion is lower than a preset threshold, it is determined that the stiffness has increased and the abnormal posture judgment threshold is adjusted.

[0006] The change in leg joint angle is calculated based on the coordinates of key points on the leg. If the change in angle is less than a preset threshold, it is determined that the leg flexibility is insufficient. The key point tracking strategy is then adjusted to increase the tracking density in the leg area.

[0007] The results of the assessment of spinal curvature, stiffness in the transition area between the lumbar and thoracic vertebrae, and leg flexibility are input into the posture abnormality assessment model to calculate the posture abnormality score. If the score exceeds the preset threshold, it is judged as a posture abnormality.

[0008] Based on the abnormal posture score, a report on the abnormal posture of the subject's bending over and touching his feet is generated, marking the specific locations of abnormal spinal curvature, increased stiffness in the transition area between the lumbar and thoracic vertebrae, and insufficient leg flexibility.

[0009] The three-dimensional coordinate information of the subject's skeletal key points is obtained, the target joint angle data is extracted from the real-time posture data, the target joint angle data is compared with the preset normal angle range, the normal angle range includes angle thresholds for different movement stages, the deviation of the target joint angle data from the normal angle range is calculated, the degree of non-standardization of the subject's movement is determined, and a movement standardization score is generated.

[0010] Based on the motion standardization score, the key point tracking strategy and posture anomaly detection threshold are dynamically updated to optimize the accuracy and real-time performance of subsequent motion evaluation.

[0011] Furthermore, video data of the subject's bending-over-to-feet movement captured by multiple cameras is acquired. Preliminary localization of key points in the spine, the transition area between the lumbar and thoracic vertebrae, and the legs is performed, and key point coordinates are extracted. This includes: aligning frames of the bending-over-to-feet movement videos captured by multiple cameras using timestamps; time-calibrating each video data stream based on video frame synchronization signals; and extracting pre-placed red and green markers on the subject's torso as reference points for the movement sequence. For the subject's spine region in each video image, a deep convolutional neural network is used to identify the transition area between the lumbar and thoracic vertebrae, obtaining image blocks of the spine region. Pre-defined skeletal key points are located within these image blocks, including the seventh cervical vertebra, the twelfth thoracic vertebra, the fifth lumbar vertebra, and the upper edge of the sacrum. Multi-view fusion of the skeletal key points located in each video stream is performed based on camera calibration parameters, and the three-dimensional spatial coordinates of the skeletal key points are calculated using triangulation. The changes in the spinal curve of the subject during the entire bending and toe-touching process are tracked based on the three-dimensional spatial coordinate sequence of the skeletal key points. A Kalman filter is used to smooth the motion trajectory of the skeletal key points. The state variable of the Kalman filter is the three-dimensional coordinate of the skeletal key points, and the observation variable is the projected coordinate of the skeletal key points detected in the image. Based on the smoothed three-dimensional spatial coordinate of the skeletal key points, the angle between the line connecting adjacent skeletal key points and the vertical direction is calculated to obtain the spinal curvature angle data. Based on the three-dimensional spatial coordinate of the skeletal key points and the spinal curvature angle data, a three-dimensional reconstruction method is used to construct a spinal structure model of the subject. Furthermore, the motion amplitude of the lumbar and thoracic vertebral transition area is calculated using the key point coordinates of the lumbar and thoracic vertebral transition area. If the motion amplitude is lower than a preset threshold, it is determined to be an increase in stiffness, and the abnormal posture judgment threshold is adjusted. This includes: calculating the angle between the line connecting adjacent key points and the reference coordinate system using the Euler angle decomposition method based on the three-dimensional coordinate sequence of the key points in the lumbar and thoracic transition area, and obtaining the motion amplitude values ​​of this area in the sagittal, coronal, and transverse planes. Time series analysis is performed on the motion amplitude values. A sliding time window of 1 second is used to calculate the motion velocity curve. Wavelet transform is used to denoise the motion velocity curve, resulting in a smoothed motion velocity sequence. For the smoothed motion velocity sequence, the mean and standard deviation of the motion velocity are calculated, and the 75th percentile of the mean motion velocity is set as a preset velocity threshold. Based on the preset velocity threshold, motion stiffness in the transition region between the lumbar and thoracic vertebrae is assessed. If the motion velocity is below the preset velocity threshold for more than 2 seconds, it is considered an increase in motion stiffness. For the time period of increased motion stiffness, a support vector machine is used to analyze the motion trajectory of key points to obtain compensatory motion feature vectors. Based on the compensatory motion feature vectors, a hidden Markov chain is used to identify abnormal motion patterns, and the abnormal posture judgment threshold is adjusted to the 60th percentile of the mean motion velocity.

[0012] Furthermore, image recognition algorithms are used to detect the pixel coordinates of key points in the lumbar and thoracic vertebrae. The angle formed by connecting these key points is defined as the range of motion in this region. A reference threshold for the range of motion in the lumbar-thoracic transition region when a normal adult bends over to touch their foot is set. If the detected range of motion is less than the reference threshold, it is determined that the stiffness in the lumbar-thoracic transition region has increased. Based on the degree of stiffness increase, the angle threshold for judging abnormal posture is increased. This includes: using a deep convolutional neural network to segment the acquired spinal images to obtain spinal region images; performing noise reduction processing on the spinal region images using median filtering to obtain denoised spinal images; using Hough transform to identify the boundary of the lumbar-thoracic transition region in the denoised spinal images; extracting image blocks of this region; and obtaining the coordinates of the vertebral body edge points from the tenth thoracic vertebra to the second lumbar vertebra in this region using a skeletal feature point detector; and using the least squares method to fit the vertebral body center point position based on the vertebral body edge point coordinates to obtain the vertebral body center point sequence of the lumbar-thoracic transition region. For the vertebral body center point sequence, triangulation is used to calculate the angle between the line connecting adjacent vertebral body center points and the vertical direction, obtaining the motion amplitude value of the lumbar-thoracic transition region. This motion amplitude value is then compared with pre-stored baseline motion amplitude values ​​for normal adults. If the current region's motion amplitude value is less than the baseline value, a stiffness quantification value is calculated using a support vector regression. For this stiffness quantification value, a piecewise function is used to adjust the posture abnormality judgment threshold: the threshold is increased by 10% when the stiffness quantification value is in the first interval, by 20% when it is in the second interval, and by 30% when it is in the third interval.

[0013] Furthermore, the leg joint angle change is calculated based on the coordinates of key leg points. If the angle change is lower than a preset threshold, it is determined to be due to insufficient leg flexibility. The key point tracking strategy is then adjusted to increase the tracking density in the leg region. This includes: segmenting the leg region in the video image using a deep residual network; extracting the initial coordinate sequences of the hip, knee, and ankle joints from the segmented image region; and smoothing the coordinate sequences using a Gaussian filter with a window size of 5×5 pixels. Based on the smoothed coordinate sequences, the angle change curve between the two vectors from the hip joint to the knee joint and from the knee joint to the ankle joint is calculated using the vector cross product method. The angle change curve is then denoised using a three-layer discrete wavelet transform. For the denoised angle change curve, the difference between the maximum and minimum angle values ​​is extracted as the joint angle change amplitude. This amplitude is compared with a preset joint mobility threshold. If the angle change amplitude is lower than the preset threshold, the frame is marked as a frame with insufficient flexibility. Based on the frames indicating insufficient flexibility, the time intervals in which consecutive flexibility deficiencies occur are calculated. The joint angle change curves within these intervals are then piecewise fitted to obtain a curve slope sequence. For this curve slope sequence, a support vector machine (SVM) is used to evaluate the motion process. The input features of the SVM include three dimensions: angle change amplitude, angular velocity, and acceleration. Based on the SVM's evaluation output, gradient density tracking point arrays are deployed in the hip, knee, and ankle joint regions, with the spacing between adjacent tracking points being minimum at the joint center and increasing towards the edges. Pyramid optical flow is used to perform inter-frame tracking of the tracking point arrays, updating the spatial coordinates of each tracking point in real time.

[0014] Furthermore, the results of the assessment of spinal curvature, stiffness in the transition region between the lumbar and thoracic vertebrae, and leg flexibility are input into the posture abnormality assessment model to calculate a posture abnormality score. If the score exceeds a preset threshold, it is judged as a posture abnormality. This includes: based on the coordinate sequence of key points of the spine, using cubic Bézier curves to fit curves to four adjacent key points, obtaining the curvature value sequence of each segment of the spine through the radius of curvature calculation formula, and mapping the curvature value sequence to the zero-to-one interval using the maximum-minimum value normalization method. For the stiffness data of the transition region between the lumbar and thoracic vertebrae, a support vector machine is used to classify the motion characteristics of this region. The features include three dimensions: joint range of motion, motion speed, and acceleration, to obtain a quantitative value of stiffness. Based on the quantitative value of stiffness, a piecewise linear mapping function is used for numerical transformation. When the quantitative value of stiffness is less than a first threshold, it is mapped to the zero-to-three interval; when it is greater than the first threshold but less than a second threshold, it is mapped to the four-to-seven interval; and when it is greater than the second threshold, it is mapped to the eight-to-ten interval. For the sequence of leg joint angle changes, a random forest algorithm is used to extract motion features, including three dimensions: joint range of motion, motion duration, and rate of change of velocity. A weighted average method is used to calculate the leg flexibility value. Based on the leg flexibility value, a piecewise linear mapping function is used for numerical transformation: flexibility values ​​less than a third threshold are mapped to the 0-3 interval; values ​​greater than the third threshold but less than a fourth threshold are mapped to the 4-7 interval; and values ​​greater than the fourth threshold are mapped to the 8-10 interval. A four-layer perceptron network is constructed. The input layer contains normalized curvature values, stiffness mapping values, and flexibility mapping values; the hidden layers contain sixteen, eight, and four nodes respectively; and the output layer is a single-node posture anomaly score. The posture anomaly score is compared with a preset anomaly judgment benchmark value; when the score exceeds the benchmark value, a posture anomaly marker is output.

[0015] Furthermore, based on the posture abnormality score, a posture abnormality report is generated for the subject's bending-over-to-the-foot movement, marking the specific locations of spinal curvature abnormalities, increased stiffness in the lumbar-thoracic vertebral transition region, and insufficient leg flexibility. This includes: locating the spinal curvature abnormality region using a deep neural network based on the posture abnormality score; extracting curvature features through a five-layer convolutional structure; and identifying continuous abnormal segments with curvature values ​​exceeding a preset threshold from the spinal curve. For these continuous abnormal segments, morphological processing methods are used for boundary extraction, marking red boundary lines at the edges of the abnormal regions and drawing arrows pointing to the center of the abnormal region. For the lumbar-thoracic vertebral transition region, a density clustering algorithm is used to spatially group the areas of increased stiffness, with the cluster radius set to the distance between adjacent vertebrae and the minimum density threshold set to three abnormal points. Based on the density clustering results, the centroid coordinates of each cluster region are calculated, and the outline of the increased stiffness region is drawn using yellow markers, with the region size labeled. For areas of insufficient leg flexibility, a region growing algorithm is used to analyze restricted joints, using the joint range of motion as the growth condition and the connectivity threshold set to the difference between adjacent pixels. Based on the growth results of the regions, areas with different degrees of insufficient flexibility were marked with a three-level color scheme: slightly restricted areas were marked in green, moderately restricted areas in yellow, and severely restricted areas in red. An anomaly report was constructed using a natural language generator. The report included the spatial coordinates of the abnormal area, the level of abnormality, the number of the affected vertebra, and the angle of joint movement restriction.

[0016] Furthermore, the abnormal key features are correlated with the detection results of three aspects: spinal curvature, stiffness in the lumbar-thoracic transition region, and leg flexibility. The coordinates of key points in the spinal curvature intervals, lumbar-thoracic transition regions, and leg joints that are determined to be abnormal are extracted. Based on these coordinates, the specific locations of abnormal spinal curvature, stiffness in the lumbar-thoracic transition region, and insufficient leg flexibility in the human posture image are marked. This includes: extracting spinal curvature features from the human posture image using a five-layer deep convolutional neural network; reducing the dimensionality of the feature map using a max-pooling layer; and marking the coordinate points of abnormal regions according to a preset curvature abnormality threshold. Morphological processing is performed on the sequence of abnormal region coordinate points, and a dilation operator is used for region connectivity analysis to obtain a complete set of boundary points for the abnormal spinal curvature regions. For the stiffness data in the lumbar-thoracic transition region, a region growing algorithm is used to segment the stiff region, using the vertebral body center point as the seed point and the difference in motion angles as the growth condition to obtain the contour of the stiff region. Based on the stiff region contour, a support vector machine is used to classify the region features, obtaining the boundaries of stiff regions at three levels: mild, moderate, and severe. For leg joint motion data, the k-nearest neighbor algorithm is used to calculate the spatial distribution of points restricting joint movement. The restricted area is then determined using a region expansion method, with the expansion radius set according to the joint range of motion. Abnormal regions are marked on the human posture image: areas with abnormal spinal curvature are filled with red semi-transparent material, areas of stiffness in the lumbar and thoracic vertebrae are marked with yellow outlines, and areas with insufficient leg flexibility are marked with blue dashed lines. These abnormal regions are quantitatively described by calculating their area, perimeter, and center of gravity, and the quantitative index values ​​of these regions are marked on the image.

[0017] Furthermore, the three-dimensional coordinate information of the subject's skeletal key points is obtained, target joint angle data is extracted from real-time posture data, and the target joint angle data is compared with a preset normal angle range, which includes angle thresholds for different movement stages. The deviation of the target joint angle data from the normal angle range is calculated to determine the degree of non-standardization of the subject's movement execution and generate a movement standardization score. This includes: locating human skeletal key points using a deep convolutional neural network based on image sequences acquired by multiple cameras; calculating the three-dimensional spatial coordinates of the skeletal key points using triangulation, where the baseline distance of the triangulation is the distance between adjacent cameras; smoothing the three-dimensional spatial coordinate sequence using a three-layer discrete wavelet transform to eliminate coordinate jitter caused by image noise and obtaining the smoothed skeletal key point motion trajectory; calculating the time segment points of the movement sequence using a dynamic time warping algorithm for the skeletal key point motion trajectory, with the window length of the algorithm set to one-quarter of the standard movement cycle; dividing the movement sequence into four standard stages—preparation, acceleration, stabilization, and deceleration—based on the time segment points, and calculating the angle between the lines connecting the skeletal key points in each stage. The angle between the connecting lines is compared with a preset normal angle range. This normal angle range is defined as a first threshold range for the preparation phase, a second threshold range for the acceleration phase, a third threshold range for the stabilization phase, and a fourth threshold range for the deceleration phase. Based on the comparison results, a support vector machine is used to calculate the deviation of the joint angles at each stage. The feature vector of the deviation includes three dimensions: angle difference, angular velocity difference, and acceleration difference. The deviation is then weighted, with the preparation phase receiving the first weight, the acceleration phase the second weight, the stabilization phase the third weight, and the deceleration phase the fourth weight, to calculate an overall motion standardization score.

[0018] Furthermore, based on the motion standardization score, the keypoint tracking strategy and posture anomaly judgment threshold are dynamically updated to optimize the accuracy and real-time performance of subsequent motion evaluation. This includes: Based on the motion standardization score, a deep convolutional neural network is used to adaptively update the keypoint tracking strategy. The neural network contains a five-layer convolutional structure. By extracting features from motion sequences with scores below a first threshold, keypoint displacement deviation values ​​are obtained. For these keypoint displacement deviation values, cluster analysis is performed on the spatial distribution of keypoints. A density clustering algorithm is used to identify joint regions with large deviation values, obtaining the spatial range of the region to be optimized. Based on the region to be optimized, a gradient boosting tree is used to dynamically adjust the anomaly judgment threshold. The input features of the gradient boosting tree include three dimensions: joint angle, motion speed, and acceleration. For different motion stages, the anomaly judgment threshold is updated in segments: a first baseline value is used in the preparation stage, a second baseline value in the acceleration stage, a third baseline value in the stabilization stage, and a fourth baseline value in the deceleration stage. Based on the updated threshold, a fixed-length sliding time window is used to segment the real-time motion data. The length of the time window is set to one-quarter of the standard motion cycle. For the segmented data, the density of keypoints is increased in abnormal regions, with the increase in density inversely proportional to the score value of that region. The coordinates of the newly added keypoints are generated using an interpolation algorithm. A Kalman filter is then used to predict the trajectory of these keypoint coordinates. The state vector of the filter contains the three-dimensional coordinates and velocity of the keypoints. The prediction results are corrected through measurement updates.

[0019] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0020] This invention discloses a method for assessing abnormal human posture based on a body composition analyzer. The method acquires video data of a subject's bending-over-to-the-foot movement, and locates and extracts the coordinates of key points in the spine, the transition area between the lumbar and thoracic vertebrae, and the legs. Based on these key point coordinates, the invention calculates the range of motion in the lumbar-thoracic transition area and changes in leg joint angles to determine stiffness and flexibility. Combined with spinal curvature, the invention inputs the data into a posture abnormality assessment model to calculate a posture abnormality score. Based on the assessment results, a detailed posture abnormality report is generated. Furthermore, this invention utilizes the three-dimensional coordinate information of skeletal key points to compare target joint angles with preset normal ranges in real time, assessing the standardization of movements. By dynamically updating the key point tracking strategy and judgment thresholds, this invention continuously optimizes the accuracy and real-time performance of the assessment, providing effective technical support for sports medicine and rehabilitation training. Attached Figure Description

[0021] Figure 1 This is a flowchart of a human posture abnormality assessment method based on an integrated body measurement machine according to the present invention.

[0022] Figure 2 This is a schematic diagram of a human posture abnormality assessment method based on an integrated body measurement machine according to the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0024] like Figure 1 -2, This embodiment of a method for assessing abnormal human posture based on a body composition analyzer may specifically include:

[0025] S101. If the assessment function of the integrated physical fitness test machine is activated, it acquires video data of the subject's bending and foot-touching movements collected by multiple cameras, performs preliminary localization and extracts coordinates of key points in the spine, lumbar and thoracic vertebrae transition area and legs, and constructs three-dimensional spatial coordinates of key skeletal points based on timestamp and multi-view fusion technology and calculates spinal curvature angle data.

[0026] S1011. Simultaneously capture video of the subject's bending-over-to-feet movement using multiple cameras. Use timestamps to perform frame alignment and time calibration on the video data from each channel. Extract the preset marker points on the subject's torso as the baseline data for the action sequence. The marker points include a red marker placed at the seventh cervical vertebra and a green marker at the upper edge of the sacrum. Use a deep convolutional neural network to identify the spinal region and output the coordinates of key skeletal points including the seventh cervical vertebra, the twelfth thoracic vertebra, the fifth lumbar vertebra, and the upper edge of the sacrum.

[0027] S1012. After obtaining the coordinates of the key points of the skeleton, the multi-view video data is fused according to the camera calibration parameters. The three-dimensional spatial coordinates of the key points of the skeleton are calculated by triangulation. The motion trajectory of the key points is smoothed by Kalman filter. The angle between the line connecting adjacent key points and the vertical direction is calculated to generate spinal curvature angle data. Then, the spinal structure model of the subject is constructed by three-dimensional reconstruction technology.

[0028] The multi-camera system uses high-speed cameras to synchronously acquire video data with a resolution of 1280×1024 and a frame rate of 120 frames per second, ensuring high temporal accuracy of motion capture. The cameras are positioned 45 degrees in front of, 45 degrees behind, and 90 degrees to the sides of the subject, with a shooting distance of 3 meters. The markers are 15 millimeters in diameter; red markers are placed at the spinous process of the seventh cervical vertebra, and green markers are placed at the upper edge of the sacrum for easy reference positioning.

[0029] In the keypoint recognition stage, a deep convolutional neural network takes the original image as input and outputs the location of the spinal region through 5 convolutional layers and 3 fully connected layers, then accurately locates the skeletal keypoints in the image patch. Multi-view fusion is based on camera intrinsic and extrinsic parameter matrices and is calibrated using a 7×9 checkerboard calibration board with a grid side length of 30 mm, ensuring that at least two cameras simultaneously observe each keypoint to calculate its 3D coordinates.

[0030] In trajectory smoothing, the state variables of the Kalman filter include the three-dimensional position and velocity of key points. The measurement noise covariance is set to 5 mm, and the process noise covariance is set to 2 mm / frame. If a sudden change in the velocity of a key point is detected, the process noise covariance is dynamically increased to improve the response speed. Spinal curvature angle data reflects the degree of curvature in each segment. When standing normally, the thoracic kyphosis angle is 20 to 40 degrees, and the lumbar lordosis angle is 40 to 60 degrees. When bending over to touch the toes, the lumbar lordosis gradually decreases, while the thoracic kyphosis increases.

[0031] A spinal model is constructed through 3D reconstruction based on key point coordinates and anatomical features. The vertebral body height is set to 20 mm and the intervertebral disc thickness to 8 mm. Spline interpolation is used to generate a smooth spinal curve, visually demonstrating posture changes during movement and providing a data foundation for subsequent anomaly assessment. This embodiment of the invention does not impose excessive limitations on specific parameter settings, which can be adjusted by technicians according to actual needs.

[0032] The above steps enable dynamic data acquisition and key point analysis of the bending-over-touching-the-foot movement, laying the foundation for subsequent posture anomaly assessment, ensuring that the assessment process can adapt to the complexity of the movement and individual differences, and improving the accuracy and reliability of the results.

[0033] S102. Calculate the range of motion and assess stiffness in the transition area between the lumbar and thoracic vertebrae by using the coordinates of key points in the transition area. If the range of motion or speed is lower than a preset threshold, it is determined that the stiffness has increased and the abnormal posture judgment threshold is adjusted. At the same time, image recognition technology is used to further analyze the range of motion and optimize the threshold adjustment strategy.

[0034] S1021. Based on the three-dimensional coordinate sequence of key points in the transition region between the lumbar and thoracic vertebrae, the Euler angle decomposition method is used to calculate the angle between the line connecting adjacent key points and the reference coordinate system, obtaining the motion amplitude values ​​in the sagittal, coronal, and transverse planes. Then, a motion velocity curve is calculated using a 1-second sliding time window. Wavelet transform is used to denoise the curve, resulting in a smoothed motion velocity sequence, and its mean and standard deviation are calculated. If the motion velocity is below a preset threshold for more than 2 seconds, it is judged as increased stiffness. Support vector machine and hidden Markov chain analysis are used to analyze abnormal motion patterns and adjust the judgment threshold. To achieve this process, the acquisition system acquires marker point data from the tenth thoracic vertebra to the second lumbar vertebra at a frequency of 100 Hz. Three marker points are placed on each vertebra, located at the spinous process and the left and right transverse processes, respectively, with a marker point diameter of 10 mm. Euler angle decomposition decomposes three-dimensional spatial motion into three orthogonal plane angles: the sagittal plane reflects flexion and extension, the coronal plane captures lateral bending, and the transverse plane analyzes rotation. In normal individuals, the sagittal flexion angle ranges from 40 to 60 degrees, the coronal lateral bending from 15 to 25 degrees, and the transverse rotation from 10 to 20 degrees. Wavelet transform uses Daubechies basis functions, removing high-frequency noise through a four-level decomposition while preserving the low-frequency trend of motion velocity. The signal-to-noise ratio is improved by approximately 8 dB after noise reduction. The mean motion velocity is set to the 75th percentile as a threshold, and the standard deviation reflects velocity fluctuations. In healthy individuals, the coefficient of variation is typically below 25%, increasing to over 35% in a stiff state. Support vector machine extracts velocity, acceleration, and trajectory curvature as feature vectors. The training sample includes 500 sets of normal and abnormal motion data, achieving a classification accuracy of 92%. A hidden Markov chain is used to construct a 5-state model, including motion initiation, acceleration, constant speed, deceleration, and termination. The normal state transition probability is regularly distributed, while the dwell time is prolonged in abnormal states. Based on this, the threshold is adjusted to the 60th percentile of the mean velocity to improve the detection sensitivity for slight stiffness.

[0035] S1022. In keypoint motion analysis, if increased stiffness is detected, the transmitting end uses a deep convolutional neural network to segment the spinal image and extract features from the transition region between the lumbar and thoracic vertebrae. After median filtering for noise reduction, Hough transform is used to locate the region boundary and calculate the motion amplitude. Support vector regressors are then used to quantify stiffness and adjust the judgment threshold segment by segment. Specifically, image acquisition uses a 2048×2048 pixel resolution DR device. The deep convolutional neural network contains 5 convolutional layers with 3×3 pixel kernels, combined with a max pooling layer, outputting a binary image of the spinal region. Median filtering uses a 7×7 pixel window to remove edge noise while preserving the vertebral contour, achieving 95% edge fidelity. Hough transform detects vertebral boundary lines with a 1-degree angular resolution and a 2-pixel distance resolution. The accumulator threshold is 20% of the image height, accurately locating the region from the tenth thoracic vertebra to the second lumbar vertebra. The skeletal feature point detector extracts 12 edge control points for each vertebra, fits the vertebral center point using the least squares method, sets the residual threshold to 3 pixels, and iterates 50 times to achieve sub-pixel accuracy. The triangulation method calculates the angle between adjacent center points and the vertical direction. When bending over to touch the feet normally, the range of the lumbar-chest transition area is about 45 degrees. If it is lower than this value, the support vector regressor calculates stiffness using the radial basis function (parameter 0.1), which is divided into three intervals: 80% to 90%, 70% to 80%, and below 70%. The thresholds are increased by 10%, 20%, and 30% respectively. After adjustment, the threshold range increases from 85% to 110.5% to adapt to different degrees of stiffness.

[0036] Motion characteristic analysis of the transition region between the lumbar and thoracic vertebrae is crucial for spinal function assessment. The aforementioned methods detect stiffness by measuring both amplitude and velocity of motion. To ensure accuracy, the system integrates Euler angle decomposition and image segmentation techniques to capture multidimensional motion information in the sagittal, coronal, and transverse planes, and optimizes data quality through wavelet transform and Hough transform. Support vector machines and hidden Markov chains further identify compensatory motion patterns, such as abnormal trajectories when the radius of curvature decreases from 500 mm to 300 mm, significantly improving the ability to discriminate irregular movements. Compared to traditional static detection, this invention's dynamic threshold adjustment improves sensitivity to mild stiffness, achieving a 95% accuracy rate in clinical trials on 50 patients with low back pain.

[0037] S1023. In practical applications, if the motion velocity sequence shows a sustained low speed or amplitude below the baseline, the system automatically triggers a threshold adjustment strategy. For example, the threshold for judging abnormal posture is adjusted from the 75th percentile of the mean motion velocity to the 60th percentile. Simultaneously, combined with image analysis results, the angle threshold is adjusted upwards in segments to ensure that the assessment results reflect the true degree of stiffness. To verify the effectiveness, tests were conducted on 100 subjects, including healthy and patient groups. The sensitivity and specificity reached 85% and 90%, respectively, indicating the reliability of this method in dynamic posture assessment.

[0038] This invention constructs a complete motion function assessment system through multi-level analysis, encompassing kinematic feature extraction, statistical analysis, and pattern recognition. It not only achieves objective quantification of stiffness but also optimizes adaptability to complex movements through adaptive thresholding, providing reliable data support for subsequent posture anomaly assessment. This invention does not impose excessive limitations on specific algorithm parameters, which can be adjusted by technical personnel according to actual scenarios. For example, the decomposition level of wavelet transform or the kernel function type of support vector machine can be adjusted to suit the motion characteristics of different populations.

[0039] S103. Calculate joint angle changes based on leg keypoint coordinates to assess flexibility. If the change is below a preset threshold, it is determined to be insufficient flexibility. The tracking strategy is adjusted to increase the tracking density in the leg region to improve analysis accuracy. First, a deep residual network is used to segment the leg region from the video image, extracting the coordinate sequences of the hip, knee, and ankle joints. Then, the angle change curve is calculated using the vector cross product method and denoised. Finally, a support vector machine is used to analyze motion features and optimize the tracking point layout.

[0040] S1031. A deep residual network is used to segment the leg region of the input high-definition video image, obtaining the initial coordinate sequences of the hip, knee, and ankle joints. After smoothing by Gaussian filtering, the vector cross product method is used to calculate the joint angle change curve. Then, a three-layer discrete wavelet transform is applied for noise reduction to extract the angle change amplitude. If it is lower than a preset threshold, it is marked as a frame with insufficient flexibility. The video input resolution is 1920×1080 pixels. The deep residual network contains a 50-layer convolutional structure. The segmentation time per frame is about 40 milliseconds. The accuracy of the segmented leg region reaches 98%, and the image block size of each joint region is 200×200 pixels. The Gaussian filtering uses a 5×5 pixel window with a standard deviation of 1.2, effectively reducing the coordinate jitter from ±3 pixels to ±0.5 pixels, improving the stability of subsequent calculations. The vector cross product method calculates the angle between the two vectors from the hip joint to the knee joint and from the knee joint to the ankle joint, avoiding the computational complexity of inverse trigonometric functions, and generating a continuous angle change curve. Wavelet transform was performed using the db4 basis function, followed by a three-level decomposition. After removing high-frequency noise, the correlation coefficient between the curve and the original data reached 0.95. The angular change amplitude was defined as the difference between the maximum and minimum angles. The hip joint range of motion threshold was set at 100 degrees and the knee joint at 120 degrees. If five consecutive frames were below the threshold, the segment was marked as a range of insufficient flexibility. This design is based on the range of motion of the hip joint from 0 to 120 degrees and the knee joint from 0 to 135 degrees in normal individuals, aiming to capture subtle flexibility abnormalities.

[0041] For frames with insufficient flexibility, the motion process was further analyzed. The slope sequence of the angle change curve was obtained through piecewise fitting, and a support vector machine (SVM) was used to evaluate the motion characteristics. Then, a gradient density tracking point array was deployed in the joint region, and inter-frame tracking was achieved using pyramid optical flow to optimize the tracking density. For the denoised angle change curve, the system performed piecewise linear fitting in units of 25 frames, with 50% overlap between adjacent intervals. The slope sequence of each segment was calculated to reflect the rate of angle change. The SVM used the angle change amplitude, angular velocity, and acceleration as input features. The angular velocity threshold was set to 90 degrees / second, and the acceleration threshold to 180 degrees / second². The model was trained based on data from 500 subjects, including 300 normal subjects and 200 subjects with limited mobility, achieving a classification accuracy of 92%. Based on the evaluation results, tracking point arrays were deployed in the hip, knee, and ankle joint regions. The tracking point interval at the joint center was 5 pixels, increasing to 15 pixels towards the edge. For example, 16 points were placed at the center of the knee joint, 32 points in the transition area, and 48 points in the edge area, totaling 96 points. The pyramid optical flow method constructs a three-layer image pyramid with a search window of 21×21 pixels and an iteration limit of 20 times, ensuring that the tracking success rate under high-speed motion is improved from 85% to 95%. This strategy enhances adaptability to complex movements by dynamically adjusting the tracking density, especially providing more refined data support when leg flexibility is insufficient.

[0042] S1032. The assessment of leg joint range of motion is crucial for human motion analysis. The aforementioned method achieves a complete process from image segmentation to motion feature extraction, ensuring accurate identification of flexibility deficiencies. The system not only focuses on the amplitude of angular changes but also incorporates multi-dimensional feature analysis of angular velocity and acceleration, avoiding the limitations of single indicators. For example, in rapid leg bending movements, if the angular velocity is below 90 degrees / second and the angular amplitude is insufficient, the system prioritizes increasing the density of tracking points in the knee joint area, updating spatial coordinates in real time, and capturing compensatory motion details. In clinical trials, the method was validated on 100 subjects, including 50 normal individuals and 50 patients with limited mobility. It achieved a sensitivity of 87% and a specificity of 89% for mild flexibility deficiencies and a recognition rate of 96% for moderate to severe deficiencies, demonstrating its high efficiency in dynamic assessment.

[0043] To further improve tracking accuracy, the system dynamically adjusts its strategy based on time intervals with insufficient flexibility. For example, when continuous low-amplitude movements are detected, the tracking points in the joint center region are automatically densified while maintaining moderate sparseness in the edge regions to balance computational efficiency and data quality. This adaptive adjustment not only improves the ability to capture details of leg movements but also provides more reliable basic data for subsequent posture anomaly assessment. Technicians can adjust the filtering parameters or the number of wavelet decomposition layers to optimize the results according to actual needs.

[0044] S104. Input the analysis results of spinal curvature, stiffness in the transition area between the lumbar and thoracic vertebrae, and leg flexibility into the posture abnormality assessment model. Generate a posture abnormality score through multi-level feature extraction and calculation. If the score exceeds a preset threshold, it is judged as a posture abnormality and a corresponding label is output. First, a curvature sequence is fitted based on the coordinates of the spinal key points and normalized. Then, motion features are extracted for the transition area between the lumbar and thoracic vertebrae and the leg area, and quantized values ​​are mapped. Finally, a four-layer perceptron network is used to fuse features and calculate the final score.

[0045] S1041. Based on the coordinate sequence of key spinal points, the spinal curvature is fitted using cubic Bézier curves, and normalized values ​​are calculated. Simultaneously, support vector machines and random forest algorithms are used to quantify the stiffness of the transition region between the lumbar and thoracic vertebrae and the flexibility of the legs, generating mapping values ​​for subsequent evaluation. Twenty-five marker points are placed on the key spinal points, with each four adjacent points forming a fitting unit. The cubic Bézier curve is fitted with a control point spacing of 20 mm, keeping the error within 0.5 mm. The curvature value for each segment is calculated using the radius of curvature formula; for example, the normal value for the thoracic segment is 0.15 to 0.20, and for the lumbar segment, it is 0.10 to 0.15. The curvature values ​​are then normalized to the 0-1 range using maximum and minimum value normalization. For the transition region between the lumbar and thoracic vertebrae, motion features including joint range of motion, velocity, and acceleration were collected at a sampling frequency of 60 Hz. Support vector machines were used for classification with radial basis function (RBF) kernels. The training data contained 1000 samples, with feature thresholds such as velocity 90 degrees / second and acceleration 180 degrees / second². Stiffness quantification values ​​were generated, with normal individuals ranging from 0.2 to 0.4, mild stiffness from 0.4 to 0.6, and severe stiffness exceeding 0.6. Stiffness was converted to a 0-10 rating using piecewise linear mapping: values ​​less than 0.4 were mapped to 0-3, 0.4-0.6 to 4-7, and greater than 0.6 to 8-10, with mapping slopes of 7.5, 15, and 10, respectively, to highlight the distinguishability of moderate stiffness. For the leg region, the random forest algorithm extracts features from the joint angle change sequence, including range of motion, duration, and rate of change of velocity. The algorithm constructs 100 decision trees with a depth of 8. The baseline range for the hip joint is 180 degrees, and for the knee joint it is 150 degrees. The normal duration is 2 to 4 seconds, the threshold for rate of change of velocity is 30%, and the flexibility value is normally distributed between 0.7 and 0.9. Similarly, it is converted to the 0 to 10 range through segmented mapping. Values ​​less than the third threshold of 0.6 are mapped to 0 to 3, values ​​between 0.6 and 0.8 are mapped to 4 to 7, and values ​​greater than 0.8 are mapped to 8 to 10.

[0046] A four-layer perceptron network was constructed to integrate the aforementioned features and calculate a posture anomaly score. The input layer receives normalized curvature, stiffness, and flexibility mapping values, which are processed by three hidden nodes to output a score, which is then compared with a preset threshold to determine anomalies. The network structure is pyramidal: the first hidden layer (16 nodes) handles initial feature extraction, the second layer (8 nodes) performs feature combination, and the third layer (4 nodes) completes fusion. The output layer is a single-node score. ReLU is used as the activation function to avoid the vanishing gradient problem. The training data includes 1000 samples: 600 normal and 400 abnormal, achieving a test accuracy of 93%. The scoring uses a percentage system, with a baseline of 75 points. Scores below 75 are considered normal, 75-85 are mildly abnormal, and scores above 85 are significantly abnormal. Compared to traditional single-index assessments, multi-dimensional feature fusion more comprehensively reflects motor function. For example, patients with mild stiffness may only show abnormalities in velocity features, while those with significant abnormalities will deviate from the normal range in both curvature and flexibility. This design improves the system's diagnostic resolution.

[0047] S1042. The importance of multidimensional feature analysis in posture assessment lies in its comprehensive capture of complex movement states. It quantifies movement patterns through spinal curvature and reflects the degree of functional limitation by combining stiffness and flexibility, providing objective evidence for clinical practice. In clinical validation, 100 subjects were tested, including 50 normal individuals and 50 patients with abnormalities. The results showed that the sensitivity for identifying mild abnormalities was 88%, the specificity was 92%, and the accuracy for identifying significant abnormalities reached 96%. For example, a patient with lumbar stiffness had a curvature value as low as 0.08 and a stiffness mapping value of 8, scoring over 85 points and correctly marked as abnormal, while normal individuals typically scored between 60 and 70 points.

[0048] To enhance the model's applicability, the system allows adjustment of feature thresholds or mapping functions according to actual needs. For example, the second threshold for stiffness can be adjusted from 0.6 to 0.65 to suit the elderly population, or the number of random forest decision trees can be increased to improve the accuracy of flexibility assessment. This flexibility ensures the reliability of the technology in different scenarios, while providing high-quality data support for subsequent anomaly report generation.

[0049] S105. Generate a postural abnormality report of the subject’s bending-over-touching-the-foot movement based on the postural abnormality score. Extract abnormal spinal curvature features and mark specific locations through a deep neural network. At the same time, perform spatial analysis and visualization marking on the stiffness in the transition area between the lumbar and thoracic vertebrae and the areas with insufficient leg flexibility to ensure that the report intuitively reflects the spatial distribution and severity of the abnormalities.

[0050] S1051. A deep neural network driven by posture anomaly scores was used to analyze spinal curvature features. A five-layer convolutional structure was used to locate segments with abnormal curvature, and morphological processing was employed to extract boundaries. Density clustering and region growing algorithms were used to segment abnormal regions in the lumbar-thoracic transition area and the leg region, respectively, and multi-level annotations were performed. The deep neural network was fed a 1280×1024 pixel image and contained three convolutional layers and two fully connected layers. The convolutional kernel size was 3×3 pixels. After max pooling, the feature map was progressively reduced to 80×64 pixels. The training data included 2000 samples, of which 800 were abnormal and 1200 were normal, achieving a localization accuracy of 94%. Taking thoracic kyphosis as an example, the normal angle range is 20 to 40 degrees. If the angle corresponding to the curvature exceeds 45 degrees, it is marked as an abnormal segment. Morphological processing uses an 8-neighborhood 3×3 structuring element, iterating 3 times to connect abnormal points within a 15-pixel distance, forming a boundary point set of 200 to 300 points. The boundary is marked with a solid red line with a line width of 2 pixels, and the arrow points to the center of the abnormality. The length is one-quarter of the region diameter. The stiffness analysis of the transition region between the lumbar and thoracic vertebrae uses a density clustering algorithm. The cluster radius is set to 25 mm, approximately equal to the distance between the two vertebrae, and the minimum density threshold is 3 abnormal points. The clustering results show a high-density abnormal cluster from the 12th thoracic vertebra to the 2nd lumbar vertebra, with an area of ​​approximately 400 square millimeters, marked with a yellow semi-transparent outline, and the centroid coordinates are displayed in white font. The region growing algorithm was used to analyze areas of insufficient leg flexibility. The hip and knee joints were used as seed points. The growth conditions were that the angle difference between adjacent pixels was less than 5 degrees, the normal hip joint flexion was 120 degrees, and the knee joint flexion was 135 degrees. Areas with insufficient leg flexibility were marked as severely restricted, 90 to 105 degrees as moderate, and 105 to 120 degrees as mild. These were marked with red, yellow, and green gradients, respectively.

[0051] To accurately locate and visualize abnormal areas, the annotation strategy was further optimized. The spatial distribution of restricted points in the leg joints was calculated using the k-nearest neighbor algorithm and region expansion methods, combined with support vector machine (SVM) classification of stiffness levels, ensuring the comprehensiveness and readability of the abnormality report. In leg analysis, the k-nearest neighbor algorithm was set to k=5 with a search radius of 30 mm to identify abnormal point clusters around the hip and knee joints. The expansion radius was dynamically adjusted according to the range of motion; for example, it increased to 25 mm when the hip joint was at 90 degrees, generating blue dashed line annotations with a 5-pixel interval and a 1-pixel line width. Stiffness in the transition region between the lumbar and thoracic vertebrae was classified using the vertebral body center point as a seed, with a motion angle difference of less than 5 degrees, covering an area of ​​approximately 400 square millimeters. SVM classification was performed using radial basis function (RBF) classification: mild restricted motion reduced by 20% to 30%, moderate by 30% to 50%, and severe (over 50%), marked with yellow outlines, with the dashed / solid ratio varying according to severity. This multi-layered visualization not only highlights the extent of the abnormality but also uses color to differentiate severity, improving doctors' intuitive judgment of the abnormal location.

[0052] S1052. The abnormality report generation utilizes a natural language generator, integrating quantitative and qualitative descriptions. It provides the spatial coordinates, severity level, and affected body parts of the abnormal area. Simultaneously, annotations are overlaid on the human posture image, visually displaying the specific locations of spinal curvature abnormalities, stiffness, and insufficient flexibility. Quantitative records in the report include, for example, a kyphosis of 47 degrees in the sixth thoracic vertebra (exceeding normal by 7 degrees) and a mobility of 8 degrees in the twelfth thoracic vertebra (below normal by 15 degrees). Indicators such as area, perimeter, and center of gravity position are annotated at the image edge in square millimeters and millimeters, with a font size of 12 pixels. Qualitative descriptions, such as "increased kyphosis in the mid-thoracic spine, affecting the sixth thoracic vertebra, impacting two vertebrae above and below," are generated using a structured template. The underlying image annotation is the original posture image. Spinal curvature abnormalities are filled with red semi-transparent material (40% transparency), lumbar stiffness areas are outlined in yellow, and restricted leg areas are displayed with blue dashed lines. The overlay effect achieved a resolution score of 9.2 in clinical testing.

[0053] The above method enables precise localization and visualization of abnormal features. Clinical validation was conducted on 100 subjects, including 50 patients with spinal abnormalities and 50 normal subjects. The spatial accuracy reached 5 mm, the boundary repeatability error was less than 2 mm, and the reporting consistency reached 92%, significantly improving diagnostic efficiency and providing a reliable basis for treatment plan formulation. Technicians can adjust the clustering radius or growth conditions as needed to optimize the effect.

[0054] S106. The three-dimensional coordinate information of the subject's skeletal key points is obtained through a multi-camera system. The target joint angle is extracted from the real-time posture data and compared with the preset normal range. The deviation is calculated to evaluate the standardization of the action and generate a score. At the same time, deep convolutional neural networks, triangulation and wavelet transform are used to optimize the data quality. The phased analysis is achieved through dynamic time warping and support vector machines.

[0055] S1061. Image sequences were acquired using four high-speed cameras. A deep convolutional neural network was used to locate key skeletal points, and triangulation was used to calculate 3D coordinates. The coordinate sequences were then smoothed using a three-layer discrete wavelet transform to obtain the motion trajectory. A dynamic time warping algorithm was then used to divide the action into stages. Finally, a support vector machine was used to calculate joint angle deviations and generate a weighted, standardized score. The camera resolution was 1920×1080 pixels, the frame rate was 120 frames per second, and the adjacent spacing was 2 meters to ensure overlapping fields of view and improve reconstruction accuracy. The deep convolutional neural network located 15 key points, including the hip, knee, and ankle, with positioning errors controlled at the sub-pixel level. Triangulation was used to calculate spatial coordinates based on calibration parameters, with an accuracy better than 2 millimeters. The three-layer wavelet transform used the db4 basis function, which improved the signal-to-noise ratio by approximately 8 dB after removing high-frequency noise. The smoothed trajectory accurately reflected the continuity of motion. The dynamic time warping algorithm analyzes the rate of displacement change in a 30-frame window (approximately one-quarter of the motion cycle), dividing the bending-over-to-foot movement into four phases: preparation, acceleration, stabilization, and deceleration, accounting for 15%, 35%, 35%, and 15% of the cycle, respectively. The joint angles in each phase are compared to preset ranges. For example, the hip joint angles are 0 to 30 degrees in the preparation phase, 30 to 90 degrees in the acceleration phase, 90 to 120 degrees in the stabilization phase, and 120 to 90 degrees in the deceleration phase; the knee joint angles are 0 to 20 degrees, 20 to 60 degrees, 60 to 90 degrees, and 90 to 60 degrees, respectively. Support Vector Machines (SVMs) use radial basis function kernels to handle deviation features, including angle difference, angular velocity difference, and acceleration difference. The normal ranges are 10 degrees, 20 degrees / second, and 40 degrees / second², respectively. The weights are assigned as follows: preparation 0.2, acceleration 0.3, stability 0.3, and deceleration 0.2. The scoring uses a percentage system, with 90 points or above considered standard, 80 to 90 points considered slightly non-standard, 70 to 80 points considered moderate, and below 70 points considered severely non-standard.

[0056] Action standardization assessment requires precise capture of three-dimensional motion features and multi-dimensional analysis. The aforementioned methods, from coordinate acquisition to trajectory smoothing, stage division, and deviation calculation, form a complete process, ensuring the objectivity and reliability of the assessment results. Compared to traditional two-dimensional analysis, three-dimensional coordinates combined with wavelet transform effectively eliminate noise interference; for example, coordinate offsets caused by image jitter are reduced from 5 mm to less than 1 mm. Dynamic time warping resolves errors caused by differences in motion speed through time axis alignment, while multi-feature classification using support vector machines improves sensitivity to subtle anomalies. In clinical testing, the sensitivity for identifying minor irregularities in 100 subjects (50 patients and 50 healthy individuals) reached 85%, and the accuracy for moderate to severe irregularities exceeded 95%.

[0057] S1062. To further improve the accuracy of the assessment, the system dynamically adjusts the analysis strategy according to the characteristics of the movement stage. For example, it increases the weight in the acceleration and stabilization stages to highlight the quality of key movements. At the same time, it allows technicians to adjust the threshold range according to the subject group. For example, the hip joint stabilization stage of the elderly can be relaxed to 80 to 110 degrees. This flexibility enhances the applicability of the system in different scenarios and provides a more scientific quantitative basis for rehabilitation training.

[0058] S107. Based on the action standardization score, dynamically adjust the key point tracking strategy and the posture anomaly judgment threshold. Extract the deviation features of low-scoring actions through a deep convolutional neural network and combine density clustering to identify the region to be optimized. Use gradient boosting tree to update the threshold and use Kalman filter to optimize trajectory prediction, thereby improving the accuracy and real-time performance of subsequent evaluations.

[0059] S1071. Based on the action standardization score-triggered adaptive update mechanism, a five-layer deep convolutional neural network is used to analyze action sequences with scores below 80 to extract key point displacement deviation values. Then, a density clustering algorithm is used to identify joint regions with large deviations and determine the optimization range. Gradient boosting trees are then used to dynamically adjust the anomaly detection threshold based on joint angle, velocity, and acceleration features. Finally, a Kalman filter is used to predict and correct the key point trajectory. The deep convolutional neural network contains a 5-layer convolutional structure with 3×3 pixel kernels, combined with a max-pooling layer. It extracts features from 10 consecutive frames of data, calculates key point displacement deviations, and sets a normal deviation of less than 5 mm. If the deviation exceeds this, an update is triggered. Density clustering analyzes spatial distribution with a radius of 30 mm and a density threshold of 5 points. For example, the knee joint abnormality area covers 80 square millimeters around the patella, with a mean deviation of 15 mm, accurately locating the functional impairment position. Gradient boosting trees are used to construct 100 decision trees of depth 5. The input feature weights are joint angle 0.4, velocity 0.3, and acceleration 0.3. The normal range of hip joint is 0 to 120 degrees, with a velocity of 90 degrees / second and an acceleration of 180 degrees / second². The threshold fluctuates by 15% based on the degree of deviation, with staged baseline values ​​such as 20 degrees for knee joint preparation, 45 degrees for acceleration, 90 degrees for stabilization, and 60 degrees for deceleration. The Kalman filter state vector includes three-dimensional coordinates and velocity. The measurement noise covariance is set to 2 mm, the process noise is 5 mm / frame², and the prediction error is less than 3 mm within 20 frames to ensure trajectory continuity.

[0060] Dynamic adjustments to the target area are achieved through a sliding time window and density increments to optimize tracking. The window length is approximately 30 frames, one-quarter of the motion cycle. Keypoint density is increased by a factor of 2 when the score is below 60 and by a factor of 3 when it is below 40. Cubic spline interpolation is used to generate new points to maintain smoothness. This strategy can flexibly increase the density of tracking points according to the severity of the abnormality; for example, in the low-scoring area of ​​the hip joint, the number of points is increased from the original 10 to 30, improving the ability to capture subtle deviations and significantly enhancing real-time performance. In clinical trials, the sensitivity for minor abnormalities reached 92% in 100 subjects (50 patients and 50 healthy individuals), and the accuracy for moderate to severe abnormalities exceeded 97%.

[0061] S1072. The importance of dynamic optimization strategy in human motion analysis lies in its adaptability. Through multi-level algorithm fusion, from deviation extraction to region recognition to threshold adjustment, a closed-loop optimization process is formed, which not only improves the evaluation accuracy, but also provides personalized adaptation for different action stages. For example, the acceleration stage is more prone to deviation due to the large range of motion, so the threshold adjustment range is larger, ensuring the analysis effect of complex actions.

[0062] To further improve the system's applicability, technicians can adjust the clustering radius or filtering parameters based on the characteristics of the subjects. For example, the density threshold can be relaxed to 40 mm for the elderly population to ensure adaptability to low-speed movements. At the same time, the feature weights of the gradient boosting tree can also be optimized according to the type of movement.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing abnormal human posture based on a body composition analyzer, characterized in that, The method includes: acquiring video data of a subject's bending-over-to-feet movement captured by multiple cameras; preliminarily locating key points in the spine, the transition area between the lumbar and thoracic vertebrae, and the legs; extracting the coordinates of these key points; calculating the range of motion in the transition area between the lumbar and thoracic vertebrae using the coordinates of these key points; if the range of motion is lower than a preset threshold, it is determined to be an increase in stiffness, and the abnormal posture threshold is adjusted, including: calculating the angle between the line connecting the key points in the transition area between the lumbar and thoracic vertebrae and the reference coordinate system using Euler angle decomposition, and obtaining the range of motion values ​​in the sagittal, coronal, and transverse planes; calculating the motion velocity curve using a sliding time window based on the motion velocity values; and performing noise reduction processing on the motion velocity curve using wavelet transform to obtain a smoothed curve. Motion speed sequence; calculate the mean and standard deviation of the motion speed for the smoothed motion speed sequence to determine if motion stiffness has increased; for time periods when motion stiffness increases, use a support vector machine to analyze the motion trajectory of key points, identify abnormal motion patterns through a hidden Markov chain, and adjust the abnormal posture judgment threshold; calculate the change in leg joint angle based on the coordinates of leg key points, and if the angle change is lower than a preset threshold, it is determined that the leg flexibility is insufficient, and the key point tracking strategy is adjusted to increase the tracking density of the leg region, including: using a deep residual network to segment the leg region in the video image, and extracting the initial coordinate sequence of the hip joint, knee joint, and ankle joint from the segmented region; based on the initial coordinate sequence, use a support vector machine to analyze the motion trajectory of key points, identify abnormal motion patterns through a hidden Markov chain, and adjust the abnormal posture judgment threshold; calculate the change in leg joint angle based on the coordinates of the leg key points, and if the change in angle is lower than a preset threshold, it is determined that the leg flexibility is insufficient, and the key point tracking strategy is adjusted to increase the tracking density of the leg region, including: using a deep residual network to segment the leg region in the video image, and extracting the initial coordinate sequence of the hip joint, knee joint, and ankle joint; using a support vector machine to analyze the motion trajectory of key points, identify abnormal motion patterns through a hidden Markov chain, and adjust ... The cross-product method is used to calculate the angle change curve between joints, and the angle change curve is subjected to three-level discrete wavelet transform denoising processing. For the denoised angle change curve, the difference between the maximum and minimum angle values ​​is extracted as the angle change amplitude. If the angle change amplitude is lower than a preset threshold, the frame with the angle change amplitude lower than the preset threshold is marked as a frame with insufficient flexibility. Based on the curve slope sequence of the frames with insufficient flexibility, a support vector machine is used to evaluate the motion process, and a gradient density tracking point array is deployed in the joint region. The judgment results of spinal curvature, stiffness in the transition area between the lumbar and thoracic vertebrae, and leg flexibility are input into the posture abnormality assessment model to calculate the posture abnormality score. If the score exceeds a preset threshold, it is judged as a posture abnormality. An abnormal posture score is generated to produce a report on the subject's posture abnormality during the bending-over-to-touch-foot movement, marking the specific locations of abnormal spinal curvature, increased stiffness in the transition area between the lumbar and thoracic vertebrae, and insufficient leg flexibility. Three-dimensional coordinate information of the subject's skeletal key points is obtained, and target joint angle data is extracted from real-time posture data. This target joint angle data is compared with a preset normal angle range, which includes angle thresholds for different movement stages. The deviation between the target joint angle data and the normal angle range is calculated to determine the degree of non-standardization in the subject's movement execution, generating a movement standardization score. Based on the movement standardization score, the key point tracking strategy and posture abnormality judgment threshold are dynamically updated to optimize the accuracy and real-time performance of subsequent movement assessments.

2. The method according to claim 1, characterized in that, The process of acquiring video data of the subject's bending-over-to-feet movement captured by multiple cameras, and preliminarily locating key points in the spine, the transition area between the lumbar and thoracic vertebrae, and the legs, and extracting the coordinates of these key points includes: time-calibrating the video data acquired by multiple cameras using timestamps; obtaining preset marker points on the subject's torso as baseline data for the movement sequence based on the time-calibration results; using a deep convolutional neural network to identify the spinal region based on the baseline data for the movement sequence to obtain the coordinates of the skeletal key points; and performing multi-view fusion based on the coordinates of the skeletal key points and camera calibration parameters, and using triangulation to obtain the three-dimensional spatial coordinates of the skeletal key points.

3. The method according to claim 1, characterized in that, Also includes: Image recognition algorithms are used to detect the pixel coordinates of key points in the lumbar and thoracic vertebrae. The angle formed by connecting these key points is defined as the range of motion in that region. A reference threshold is set for the range of motion in the lumbar-thoracic transition region when a normal adult bends over to touch their foot. If the detected range of motion is less than the reference threshold, the stiffness in the lumbar-thoracic transition region is considered to be increased. Based on the degree of stiffness increase, the angle threshold for judging abnormal posture is adjusted upwards. Specifically, this involves: segmenting the spinal image using a deep convolutional neural network; denoising the segmented spinal region image using median filtering to obtain a denoised spinal image; using the denoised spinal image, Hough transform is used to identify the boundary of the lumbar-thoracic transition region, and the coordinates of the vertebral edge points in this region are obtained using a skeletal feature point detector; for the coordinates of the vertebral edge points, the least squares method is used to fit and obtain the sequence of vertebral center points, and the angle between the line connecting adjacent vertebral center points and the vertical direction is calculated using triangulation; the angle value is compared with a preset benchmark value. If the angle value is less than the benchmark value, a support vector regressor is used to calculate the stiffness quantification value, and the abnormal posture judgment threshold is adjusted using a piecewise function.

4. The method according to claim 1, characterized in that, The process involves inputting the results of spinal curvature, stiffness in the lumbar-thoracic transition region, and leg flexibility assessment into a posture abnormality evaluation model to calculate a posture abnormality score. If the score exceeds a preset threshold, the posture is deemed abnormal. This includes: fitting a cubic Bézier curve to the coordinate sequence of key spinal points; obtaining a curvature value sequence using the radius of curvature calculation formula; mapping the curvature value sequence to the zero-to-one interval using a maximum-minimum normalization method to obtain a normalized curvature value; and using a support vector machine to classify the joint motion angle, movement velocity, and acceleration characteristics of the lumbar-thoracic transition region corresponding to the normalized curvature value to obtain stiffness. The quantified values ​​are transformed using a piecewise linear mapping function to obtain stiffness mapping values. For the leg region corresponding to the stiffness mapping values, a random forest algorithm is used to extract joint range of motion, motion duration, and velocity change rate features from the joint angle change sequence. The leg flexibility values ​​are calculated using a weighted average method and transformed using a piecewise linear mapping function to obtain flexibility mapping values. A four-layer perceptron network is used to input the normalized curvature values, stiffness mapping values, and flexibility mapping values ​​into the input layer. After processing through the hidden layer, the posture anomaly score value is obtained from the output layer. If the score value exceeds the preset anomaly judgment benchmark value, a posture anomaly marker is output.

5. The method according to claim 1, characterized in that, The process of generating a postural abnormality report based on the postural abnormality score, indicating the specific locations of spinal curvature abnormalities, increased stiffness in the lumbar-thoracic spine transition region, and insufficient leg flexibility, includes: obtaining spinal curvature features based on the postural abnormality score; extracting continuous abnormal segments with curvature values ​​exceeding a preset threshold using a five-layer convolutional structure of a deep neural network; obtaining the boundaries of abnormal regions using morphological processing methods for the continuous abnormal segments and marking red boundary lines at the edges of the abnormal regions; spatially grouping the areas of increased stiffness in the lumbar-thoracic spine transition region using a density clustering algorithm and labeling the area of ​​the region based on the density clustering results; and analyzing the areas of insufficient flexibility using a region growing algorithm, classifying them into mildly restricted, moderately restricted, and severely restricted regions based on joint range of motion values.

6. The method according to claim 5, characterized in that, Also includes: The abnormal key features are correlated with the detection results of three aspects: spinal curvature, stiffness in the lumbar-thoracic transition area, and leg flexibility. The coordinates of key points of the spinal curvature range, lumbar-thoracic transition area, and leg joints that are judged to be abnormal are extracted. Based on the coordinates of the abnormal key points, the specific locations of abnormal spinal curvature, stiffness in the lumbar-thoracic transition area, and insufficient leg flexibility in the human posture image are marked. Specifically, this includes: using a deep convolutional neural network to extract spinal curvature feature maps from human posture images, and marking the coordinate points of abnormal areas through a max pooling layer. Morphological processing using the dilation operator is performed based on the sequence of coordinate points in the abnormal region to obtain the boundary point set of the abnormal spinal curvature region. The region growing algorithm is then used for segmentation, and the outline of the rigid region is obtained with the vertebral center point as the seed point. Based on the outline of the rigid region, the k-nearest neighbor algorithm is used to calculate the spatial distribution of the joint movement restriction points, and the boundary of the restricted region is obtained through the region expansion method.

7. The method according to claim 1, characterized in that, The process involves acquiring the three-dimensional coordinate information of the subject's skeletal key points, extracting target joint angle data from real-time posture data, comparing the target joint angle data with a preset normal angle range (which includes angle thresholds for different movement stages), calculating the deviation of the target joint angle data from the normal angle range, determining the degree of non-standardization in the subject's movement execution, and generating a movement standardization score. This includes: using a deep convolutional neural network to locate the human skeletal key points; obtaining a three-dimensional spatial coordinate sequence of the skeletal key points through triangulation between adjacent cameras; smoothing the three-dimensional spatial coordinate sequence using a three-layer discrete wavelet transform to obtain the movement trajectory of the skeletal key points; calculating time segment points for the skeletal key point movement trajectory using a dynamic time warping algorithm to obtain four standard stages: preparation, acceleration, stabilization, and deceleration; and calculating the deviation of the joint angles at each stage using a support vector machine based on the comparison results of the angles between the lines connecting the skeletal key points within the standard stages and the preset normal angle range, and obtaining a movement standardization score through weighted processing.

8. The method according to claim 1, characterized in that, The step of dynamically updating the keypoint tracking strategy and posture anomaly detection threshold based on the action standardization score to optimize the accuracy and real-time performance of subsequent action evaluation includes: obtaining keypoint tracking data based on the action standardization score; using a deep convolutional neural network to extract features from action sequences with scores below a first threshold to obtain keypoint displacement deviation values; performing spatial distribution clustering analysis on the keypoint displacement deviation values; using a density clustering algorithm to identify joint regions with large deviation values ​​to obtain the spatial range of the region to be optimized; using a gradient boosting tree to dynamically adjust the anomaly detection threshold based on the joint angle, motion velocity, and acceleration features of the region to be optimized to obtain an updated threshold; and using a Kalman filter to predict the trajectory of the keypoint coordinates, wherein the state vector of the filter contains the three-dimensional coordinates and velocity of the keypoint, and the corrected prediction result is obtained through measurement and updating.

Images