Human body posture abnormity assessment method based on body measurement all-in-one machine

Abstract

The invention provides a human body posture abnormity assessment method based on a body measurement all-in-one machine, and the method comprises the steps: obtaining the video data, collected by multiple cameras, of the bending and foot touching actions of a subject, carrying out the preliminary positioning of a spine, a lumbar vertebra and thoracic vertebra transition region, and key points of legs, and extracting the coordinates of the key points; leg joint angle change is calculated according to the leg key point coordinates, if the angle change amplitude is lower than a preset threshold value, it is judged that leg flexibility is insufficient, a key point tracking strategy is adjusted, and the tracking density of a leg area is increased; according to the posture abnormity score, generating a posture abnormity report of the bending and foot touching actions of the subject, and marking specific positions of spine curvature abnormity, stiffness increase of a transition area between the lumbar vertebra and the thoracic vertebra and insufficient leg flexibility; and dynamically updating a key point tracking strategy and an attitude anomaly judgment threshold according to the action normativity score, and optimizing the accuracy and real-time performance of subsequent action evaluation.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and particularly to a method for evaluating abnormal human postures based on a physical fitness integrated machine. Background Art

[0002] In recent years, the field of human posture evaluation has received much attention due to its crucial role in health monitoring, sports rehabilitation, and intelligent interaction. By accurately analyzing the dynamic postures of the human body through a physical fitness integrated machine, not only can the individual's motor ability and health status be revealed, but it also has an irreplaceable value in preventing injuries and optimizing training. Key points play a role in accurately positioning and tracking key body parts such as the spine, lumbar-thoracic transition area, and legs in this solution. Dynamically adjusting key point tracking is crucial for adapting to posture changes, uneven spinal curvatures, and compensation patterns during movements, which are important abnormal indicators that the system can capture but static tracking would miss. However, the limitations of existing methods are gradually emerging. Most solutions rely on static key point detection or single action templates, making it difficult to adapt to complex posture changes and individual differences in dynamic tests, resulting in insufficient accuracy and universality of the evaluation results. Especially when it comes to multi-joint coordinated movements, traditional technologies often ignore the continuity of joint angle change trajectories and the dynamic compensation mechanisms between different body parts, which leads to significant deviations in the recognition of abnormal postures. The core technical challenges faced by current research focus on how to capture and analyze key factors such as the continuity of joint angle changes, non-uniform spinal curvature distributions, and flexibility compensation patterns in dynamic actions. Since these factors have not been fully addressed, existing systems often cannot accurately reflect problems such as uneven spinal bending curvatures, increased stiffness in the lumbar-thoracic transition region, and leg flexibility compensation when tracking complex movements such as bending down to touch the feet. This not only increases the difficulty of key point tracking but also, due to the lack of an adaptive abnormal determination threshold, leads to misjudgments or missed judgments of posture deviations, thereby affecting the reliability of the overall evaluation. Therefore, how to dynamically adjust the key point tracking strategy to adapt to changes in uneven spinal curvatures and joint stiffness during tests such as bending down to touch the feet, while optimizing the abnormal posture determination threshold to identify leg flexibility compensation patterns, has become a key issue in improving the accuracy of body posture abnormal evaluation. Solving this problem requires finding a balance in dynamic data collection and real-time analysis to ensure that technical strategies can accurately address the challenges brought by individual differences and action complexities. Summary of the Invention

[0003] The present invention provides a method for evaluating abnormal human postures based on a physical fitness integrated machine, mainly including:

[0004] Obtaining video data of the subject's bending down to touch the feet action collected by multiple cameras, initially positioning key points on the spine, lumbar-thoracic transition area, and legs, and extracting the coordinates of the key points;

[0005] Calculate the range of motion of the lumbar-thoracic transition region based on the key point coordinates of the lumbar-thoracic transition region. If the range of motion is lower than the preset threshold, it is determined that the stiffness has increased, and the posture abnormality determination threshold is adjusted.

[0006] Calculate the change in leg joint angles based on the key point coordinates of the legs. If the amplitude of the angle change is lower than the 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 region.

[0007] Input the determination results of the spinal curvature, the stiffness of the lumbar-thoracic transition region, and the leg flexibility into the posture abnormality assessment model to calculate the posture abnormality score. If the score exceeds the preset threshold, it is determined that the posture is abnormal.

[0008] Generate a posture abnormality report for the subject's forward bending to touch the feet action based on the posture abnormality score, and mark the specific positions of abnormal spinal curvature, increased stiffness in the lumbar-thoracic transition region, and insufficient leg flexibility.

[0009] Obtain the three-dimensional coordinate information of the skeletal key points of the subject, extract the target joint angle data from the real-time posture data, compare the target joint angle data with the preset normal angle range, where the normal angle range includes angle thresholds for different action stages, calculate the deviation of the target joint angle data from the normal angle range, determine the degree of non-standardness of the subject's action execution, and generate an action standardness score.

[0010] Dynamically update the key point tracking strategy and the posture abnormality determination threshold according to the action standardness score to optimize the accuracy and real-time performance of subsequent action assessments.

[0011] Further, obtain the video data of the subject's bending down to touch the feet action collected by multiple cameras, preliminarily locate the key points of the spine, the transitional region between the lumbar and thoracic vertebrae, and the legs, and extract the coordinates of the key points, including: align the frames of the bending-down-to-touch-the-feet videos collected by multiple cameras through timestamps, perform time calibration on each video data according to the video frame synchronization signal, and obtain the pre-arranged red and green marker points on the trunk of the subject in the video as the benchmark points of the action sequence. For the spine region of the subject in each video image, use a deep convolutional neural network to identify the transitional region between the lumbar and thoracic vertebrae and obtain the image patches of the spine region. Locate the preset skeletal key points in the image patches of the spine region, and the skeletal key points include the seventh cervical vertebra point, the twelfth thoracic vertebra point, the fifth lumbar vertebra point, and the upper edge point of the sacrum. Perform multi-view fusion on the located skeletal key points in each video according to the camera calibration parameters, and use the triangulation method to calculate the three-dimensional spatial coordinates of the skeletal key points. Track the changes in the spine curve during the whole process of the subject's bending down to touch the feet according to the sequence of the three-dimensional spatial coordinates of the skeletal key points, and use a Kalman filter to smooth the motion trajectories of the skeletal key points. The state variables of the Kalman filter are the three-dimensional coordinates of the skeletal key points, and the observation variables are the projected coordinates of the detected skeletal key points in the image. Based on the smoothed three-dimensional spatial coordinates of the skeletal key points, calculate the angle between the line connecting adjacent skeletal key points and the vertical direction to obtain the spine bending angle data. According to the three-dimensional spatial coordinates of the skeletal key points and the spine bending angle data, use a three-dimensional reconstruction method to construct a spine structure model of the subject. Further, calculate the movement amplitude of the transitional region between the lumbar and thoracic vertebrae through the coordinates of the key points in the transitional region between the lumbar and thoracic vertebrae. If the movement amplitude is lower than the preset threshold, it is determined that the stiffness has increased, and adjust the posture abnormality determination threshold, including: according to the sequence of the three-dimensional coordinates of the key points in the transitional region between the lumbar and thoracic vertebrae, use the Euler angle decomposition method to calculate the angle between the line connecting adjacent key points and the reference coordinate system, and obtain the movement amplitude values in the sagittal plane, coronal plane, and transverse plane of this region. Perform time series analysis on the movement amplitude values, use a sliding time window with a length of 1 second to calculate the movement speed curve, and perform noise reduction processing on the movement speed curve through wavelet transform to obtain the smoothed movement speed sequence. For the smoothed movement speed sequence, calculate the mean and standard deviation of the movement speed, and set the 75% quantile of the movement speed mean as the preset speed threshold. According to the preset speed threshold, judge the movement stiffness of the transitional region between the lumbar and thoracic vertebrae. If the duration of the movement speed being lower than the preset speed threshold exceeds 2 seconds, it is determined that the movement stiffness has increased. For the time period with increased movement stiffness, use a support vector machine to analyze the movement trajectories of the key points to obtain the compensatory movement feature vectors. According to the compensatory movement feature vectors, use a hidden Markov chain to identify abnormal movement patterns, and adjust the posture abnormality determination threshold to the 60% quantile of the movement speed mean.

[0012] Further, the pixel coordinates of the key points of the lumbar and thoracic vertebrae are detected using an image recognition algorithm. The angle formed by connecting the key points of the lumbar and thoracic vertebrae is defined as the range of motion of this area. A reference threshold for the range of motion of the lumbar-thoracic transition area when a normal adult bends down to touch the feet is set. If the detected range of motion is less than the reference threshold, it is determined that the stiffness of the lumbar-thoracic transition area increases. According to the degree of increased stiffness, the angle threshold for abnormal posture determination is adjusted upward, including: using a deep convolutional neural network to segment the collected spine image to obtain a spine area image, and performing noise reduction processing on the spine area image through median filtering to obtain a noise-reduced spine image. For the noise-reduced spine image, the Hough transform is used to identify the boundary of the lumbar-thoracic transition area, and the image block of this area is extracted. The vertebral edge point coordinates from the tenth thoracic vertebra to the second lumbar vertebra in this area are obtained through a bone feature point detector. According to the vertebral edge point coordinates, the least squares method is used to fit the position of the vertebral center point to obtain the sequence of vertebral center points in the lumbar-thoracic transition area. For the sequence of vertebral center points, the triangulation method is used to calculate the angle between the line connecting adjacent vertebral center points and the vertical direction to obtain the range of motion value of the lumbar-thoracic transition area. According to the range of motion value of the lumbar-thoracic transition area, it is compared with the pre-stored range of motion reference value of a normal adult. If the current area range of motion value is less than the reference value, a support vector regressor is used to calculate the stiffness quantization value. For the stiffness quantization value, a piecewise function is used to adjust the abnormal posture determination threshold, increasing the threshold by 10% when the stiffness quantization value is in the first interval, 20% when in the second interval, and 30% when in the third interval.

[0013] Further, calculate the change in leg joint angles based on the coordinates of leg key points. If the amplitude of 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 in the leg area, including: using a deep residual network to segment the leg area in the video image, extracting the initial coordinate sequences of the hip joint point, knee joint point, and ankle joint point from the segmented image area, and smoothing the coordinate sequences using a Gaussian filter with a window size of 5×5 pixels. According to the smoothed coordinate sequences, use the vector cross product method to calculate the angle change curve between the two vectors from the hip joint point to the knee joint point and from the knee joint point to the ankle joint point, and perform denoising on the angle change curve using a three-layer discrete wavelet transform. For the denoised angle change curve, extract the difference between the maximum angle value and the minimum angle value as the amplitude of the joint angle change, and compare the amplitude of the angle change with a preset joint range of motion threshold. If the amplitude of the angle change is lower than the preset threshold, mark this frame as a flexibility-insufficient frame. According to the flexibility-insufficient frames, calculate the time interval during which flexibility insufficiency continuously occurs, perform piecewise fitting on the joint angle change curve within this interval, and obtain the curve slope sequence. For the curve slope sequence, use a support vector machine to evaluate the motion process, and the input features of the support vector machine include three dimensions: amplitude of angle change, angular velocity, and acceleration. Based on the evaluation output value of the support vector machine, gradient density tracking point arrays are respectively arranged in the hip joint, knee joint, and ankle joint areas, and the interval between adjacent tracking points is the smallest at the joint center and increases towards the edge. Use the pyramid optical flow method to perform inter-frame tracking on the tracking point arrays and update the spatial position coordinates of each tracking point in real time.

[0014] Furthermore, input the determination results of spinal curvature, stiffness in the transition region between the lumbar and thoracic vertebrae, and leg flexibility into the postural abnormality assessment model to calculate the postural abnormality score. If the score exceeds the preset threshold, it is determined as postural abnormality, including: According to the spinal key point coordinate sequence, use cubic Bezier curves to fit the curves of adjacent four key points, and obtain the curvature value sequence of each spinal segment through the curvature radius calculation formula. Use the maximum-minimum normalization method to map the curvature value sequence to the interval from zero to one. For the stiffness data in the transition region between the lumbar and thoracic vertebrae, use a support vector machine to classify the motion characteristics of this region. The characteristics include three dimensions: joint movement angle, movement speed, and acceleration, and obtain the quantified stiffness value. According to the quantified stiffness value, use a piecewise linear mapping function for numerical conversion. When the quantified stiffness value is less than the first threshold, it is mapped to the interval from zero to three; when it is greater than the first threshold and less than the second threshold, it is mapped to the interval from four to seven; when it is greater than the second threshold, it is mapped to the interval from eight to ten. For the sequence of leg joint angle changes, use the random forest algorithm to extract motion characteristics. The characteristics include three dimensions: joint movement range, movement duration, and speed change rate, and calculate the leg flexibility value through the weighted average method. According to the leg flexibility value, use a piecewise linear mapping function for numerical conversion. When the flexibility value is less than the third threshold, it is mapped to the interval from zero to three; when it is greater than the third threshold and less than the fourth threshold, it is mapped to the interval from four to seven; when it is greater than the fourth threshold, it is mapped to the interval from eight to ten. Construct a four-layer perceptron network. The input layer includes the normalized curvature value, stiffness mapping value, and flexibility mapping value. The hidden layers include sixteen, eight, and four nodes respectively, and the output layer is a single-node postural abnormality score value. According to the postural abnormality score value, compare it with the preset abnormal determination reference value, and output a postural abnormality mark when the score value exceeds the reference value.

[0015] Further, according to the posture anomaly score, a posture anomaly report for the subject's action of bending down to touch the feet is generated, marking the specific positions of abnormal spinal curvature, increased stiffness in the transition region between the lumbar and thoracic vertebrae, and insufficient leg flexibility, including: According to the posture anomaly score, a deep neural network is used to locate the abnormal spinal curvature region, and curvature features are extracted through a five-layer convolutional structure to obtain continuous abnormal segments on the spinal curve where the curvature value exceeds a preset threshold. For the continuous abnormal segments, a morphological processing method is used for boundary extraction, a red boundary line is marked at the edge of the abnormal region, and an arrow is drawn pointing to the center position of the abnormal region. For the transition region between the lumbar and thoracic vertebrae, a density clustering algorithm is used to spatially group the regions with increased stiffness, the clustering radius is set to the distance between adjacent vertebral bodies, and the minimum density threshold is set to three abnormal points. According to the density clustering result, the centroid coordinates of each clustering region are calculated, and the contour of the region with increased stiffness is drawn with a yellow mark, marking the area size of the region. For the region with insufficient leg flexibility, a region growing algorithm is used to analyze the restricted joints, with the joint movement angle as the growth condition, and the connectivity threshold is set to the difference between adjacent pixels. According to the region growing result, three-level color annotation is performed on regions with different degrees of insufficient flexibility, with mildly restricted regions marked green, moderately restricted regions marked yellow, and severely restricted regions marked red. A natural language generator is used to construct the anomaly report, and the report content includes the spatial coordinates of the abnormal region, the abnormal degree classification, the numbers of affected vertebral bodies, and the joint movement restriction angle values.

[0016] Further, the abnormal key features are corresponded to the detection results in three aspects: spinal curvature, stiffness of the lumbar-thoracic transition region, and leg flexibility. The key point coordinates of the abnormal spinal curvature interval, lumbar-thoracic transition region, and leg joints determined as abnormal are extracted. According to the abnormal key point coordinates, the specific positions of abnormal spinal curvature, stiffness of the lumbar-thoracic transition region, and insufficient leg flexibility in the human body posture image are marked, including: extracting spinal curvature features from the human body posture image by using a five-layer deep convolutional neural network, reducing the dimension of the feature map through a max pooling layer, and marking the coordinate points of the abnormal region 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 regional connectivity analysis to obtain a complete set of boundary points of the abnormal spinal curvature region. For the stiffness data of the lumbar-thoracic transition region, a region growing algorithm is used to segment the stiff region. The center point of the vertebral body is used as the seed point, and the motion angle difference is used as the growth condition to obtain the contour of the stiff region. According to the contour of the stiff region, a support vector machine is used to classify the regional features to obtain the boundary of the stiff region at three levels: mild, moderate, and severe. For the leg joint movement data, the k-nearest neighbor algorithm is used to calculate the spatial distribution of the joint movement restricted points, and the range of the restricted region is obtained through a region expansion method. The expansion radius is set according to the joint range of motion. Abnormal region annotations are drawn on the human body posture image. The abnormal spinal curvature region is filled with red semi-transparency, the lumbar-thoracic stiff region is marked with a yellow contour line, and the insufficient leg flexibility region is marked with a blue dotted line. Quantitative description of the abnormal region is performed, and the area, perimeter, and centroid position of the abnormal region are calculated, and the quantitative index values of the abnormal region are marked on the image.

[0017] Further, obtain the three-dimensional coordinate information of the skeletal key points of the subject, extract the target joint angle data from the real-time pose data, compare the target joint angle data with a preset normal angle range, where the normal angle range includes angle thresholds for different action stages, calculate the deviation degree of the target joint angle data from the normal angle range, and determine the non-standard degree of the subject's action execution to generate an action standardization score, including: Based on the image sequence collected by multiple cameras, use a deep convolutional neural network to locate the human skeletal key points, and calculate the three-dimensional spatial coordinates of the skeletal key points through the triangulation method, where the baseline distance of the triangulation is the distance between adjacent cameras. For the three-dimensional spatial coordinate sequence, perform smoothing processing using a three-layer discrete wavelet transform to eliminate the coordinate jitter caused by image noise and obtain the smoothed motion trajectory of the skeletal key points. For the motion trajectory of the skeletal key points, use the dynamic time warping algorithm to calculate the time segmentation points of the action sequence, and set the window length of the algorithm to one-fourth of the standard action period. According to the time segmentation points, divide the action sequence into four standard stages: preparation, acceleration, stability, and deceleration, and calculate the included angle between the connecting lines of the skeletal key points within each stage. For the included angle of the connecting lines, compare it with a preset normal angle range, where the normal angle range is the first threshold interval in the preparation stage, the second threshold interval in the acceleration stage, the third threshold interval in the stability stage, and the fourth threshold interval in the deceleration stage. According to the comparison result, use a support vector machine to calculate the deviation degree of the joint angle in each stage, and the feature vector of the deviation degree includes three dimensions: angle difference, angular velocity difference, and acceleration difference. Perform weighted processing on the deviation degree, where the weight in the preparation stage is the first weight, the weight in the acceleration stage is the second weight, the weight in the stability stage is the third weight, and the weight in the deceleration stage is the fourth weight, and calculate the overall action standardization score.

[0018] Further, according to the action standardization score, the key point tracking strategy and the posture abnormality determination threshold are dynamically updated to optimize the accuracy and real-time performance of subsequent action evaluation, including: According to the action standardization score, a deep convolutional neural network is used to adaptively update the key point tracking strategy. The neural network includes a five-layer convolutional structure. By extracting features from the action sequence with a score lower than the first threshold, the key point displacement deviation value is obtained. For the key point displacement deviation value, cluster analysis is performed on the key point spatial distribution, and a density clustering algorithm is used to identify the joint areas with larger deviation values to obtain the spatial range of the area to be optimized. According to the area to be optimized, a gradient boosting tree is used to dynamically adjust the abnormality determination threshold. The input features of the gradient boosting tree include three dimensions: joint angle, movement speed, and acceleration. For different action stages, the abnormality determination threshold is updated in segments. The first reference value is used in the preparation stage, the second reference value is used in the acceleration stage, the third reference value is used in the stable stage, and the fourth reference value is used in the deceleration stage. According to the updated threshold, the real-time action data is segmented using a fixed-length sliding time window, and the length of the time window is set to one-fourth of the standard action cycle. For the segmented data, the key point layout density is increased in the abnormal area, and the density increase amount is inversely proportional to the score value of this area. The coordinates of the newly added key points are generated through an interpolation algorithm. A Kalman filter is used to predict the trajectory of the key point coordinates. The state vector of the filter includes the three-dimensional coordinates and speed of the key point, and the prediction result is corrected through measurement update.

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

[0020] The present invention discloses a method for evaluating human body posture abnormality based on a physical fitness testing integrated machine. The method obtains the video data of the subject's bending down to touch the feet action, locates and extracts the coordinates of the key points of the spine, the transition area between the lumbar vertebra and the thoracic vertebra, and the legs. According to these key point coordinates, the present invention calculates the movement amplitude of the transition area between the lumbar vertebra and the thoracic vertebra and the change of the leg joint angle to determine the stiffness and flexibility. Combining the spinal curvature, the present invention inputs it into the posture abnormality evaluation model to calculate the posture abnormality score. Based on the evaluation result, a detailed posture abnormality report is generated. In addition, the present invention also uses the three-dimensional coordinate information of the bone key points to compare the target joint angle with the preset normal range in real time to evaluate the action standardization. By dynamically updating the key point tracking strategy and the determination threshold, the present invention continuously optimizes the accuracy and real-time performance of the evaluation, providing effective technical support for sports medicine and rehabilitation training. Description of the Drawings

[0021] Figure 1 It is a flowchart of a method for evaluating human body posture abnormality based on a physical fitness testing integrated machine of the present invention.

[0022] Figure 2 Schematic diagram of a method for evaluating abnormal human postures based on a body measurement integrated machine according to the present invention. Specific implementation manners

[0023] In order to enable those skilled in the art of this technology to better understand the technical solutions in this specification, the following will clearly and completely describe the technical solutions in the embodiments of this specification with reference to the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the scope of protection of this specification.

[0024] As Figure 1 -2, a method for evaluating abnormal human postures based on a body measurement integrated machine in this embodiment may specifically include:

[0025] S101. If the evaluation function of the body measurement integrated machine is activated, obtain the video data of the subject's bending down to touch the feet action collected by multiple cameras, preliminarily locate and extract the coordinates of the key points of the spine, the transition area between the lumbar vertebra and the thoracic vertebra, and the legs, and at the same time construct the three-dimensional space coordinates of the bone key points based on the time stamp and the multi-view fusion technology and calculate the spine bending angle data.

[0026] S1011. Synchronously collect video data of the subject's bending down to touch the feet action through multiple cameras, use the time stamp to perform frame alignment and time calibration on the video data of each channel, extract the preset marking points on the subject's trunk as the reference data of the action sequence. The marking points include a red mark placed on the seventh cervical vertebra and a green mark placed on the upper edge of the sacrum. Use a deep convolutional neural network to identify the spine area and output the bone key point coordinates including the seventh cervical vertebra point, the twelfth thoracic vertebra point, the fifth lumbar vertebra point, and the upper edge point of the sacrum.

[0027] S1012. After obtaining the bone key point coordinates, fuse the multi-view video data according to the camera calibration parameters, use the triangulation method to calculate the three-dimensional space coordinates of the bone key points, and use the Kalman filter to smooth the motion trajectory of the key points, calculate the angle between the connection line of adjacent key points and the vertical direction to generate the spine bending angle data, and then construct the spine structure model of the subject through the three-dimensional reconstruction technology.

[0028] The multi-camera system uses high-speed cameras to synchronously collect video data, with a resolution of 1280×1024 and a frame rate of 120 frames per second to ensure high temporal accuracy of action capture. The cameras are arranged at 45 degrees in front of the subject, 45 degrees behind the subject, and 90 degrees on both sides, and the shooting distance is 3 meters. The diameter of the marking point is 15 mm, the red mark is placed on the spinous process of the seventh cervical vertebra, and the green mark is placed on the upper edge of the sacrum for easy reference positioning.

[0029] In the key point recognition stage, the deep convolutional neural network takes the original image as input, and outputs the position of the spine region through 5 convolutional layers and 3 fully connected layers. Subsequently, the bone key points are accurately located in the image patches. The multi-view fusion is based on the camera's internal and external parameter matrices, calibrated by a 7×9 checkerboard calibration board with a grid side length of 30 mm, ensuring that at least two cameras simultaneously observe each key point to calculate the three-dimensional coordinates.

[0030] In the trajectory smoothing process, the state variables of the Kalman filter include the three-dimensional position and velocity of the 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 key point velocity is detected, the process noise covariance is dynamically increased to improve the response speed. The spine curvature angle data reflects the curvature degree of 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 down to touch the feet, the lumbar lordosis gradually decreases, and the thoracic kyphosis increases.

[0031] Through 3D reconstruction, a spine model is constructed based on the key point coordinates and anatomical features. The vertebral body height is set to 20 mm, and the intervertebral disc thickness is 8 mm. Spline interpolation is used to generate a smooth spine curve, visually showing the posture changes during the movement, providing a data basis for subsequent abnormal assessment. The embodiments of the present invention do not make excessive limitations on the specific parameter settings, which can be adjusted by technicians according to actual needs.

[0032] Through the above steps, the dynamic data collection and key point analysis of the action of bending down to touch the feet are realized, laying a foundation for subsequent posture abnormal assessment, ensuring that the assessment process can adapt to the action complexity and individual differences, and improving the accuracy and reliability of the results.

[0033] S102. Calculate the movement amplitude of this region and evaluate the stiffness through the key point coordinates in the transition region between the lumbar and thoracic vertebrae. If it is detected that the movement amplitude or speed is lower than the preset threshold, it is determined that the stiffness has increased and the posture abnormal determination threshold is adjusted. At the same time, image recognition technology is used to further analyze the movement amplitude and optimize the threshold adjustment strategy.

[0034] S1021. Based on the three-dimensional coordinate sequence of key points in the lumbar-thoracic transition region, the Euler angle decomposition method is used to calculate the angle between the line connecting adjacent key points and the reference coordinate system, and the motion amplitude values in the sagittal plane, coronal plane, and transverse plane are obtained. Subsequently, a sliding time window with a length of 1 second is used to calculate the motion speed curve. Wavelet transform is applied to denoise the curve to obtain a smoothed motion speed sequence, and its mean and standard deviation are calculated. If the motion speed is lower than the preset threshold for more than 2 seconds continuously, it is determined that the stiffness has increased. The support vector machine and hidden Markov chain are combined to analyze abnormal motion patterns and adjust the decision threshold. To achieve this process, the acquisition system obtains the marker point data from the tenth thoracic vertebra to the second lumbar vertebra at a frequency of 100 Hz. Three marker points are arranged on each vertebra, located at the spinous process and the left and right transverse processes respectively, and the diameter of the marker point is 10 mm. The Euler angle decomposition decomposes the three-dimensional space motion into three orthogonal plane angles. The sagittal plane reflects flexion and extension, the coronal plane captures lateral curvature, and the transverse plane analyzes rotation. The sagittal plane flexion angle range for the normal population is 40 to 60 degrees, the coronal plane lateral curvature is 15 to 25 degrees, and the transverse plane rotation is 10 to 20 degrees. The Daubechies basis function is selected for wavelet transform, and high-frequency noise is removed through 4-level decomposition, retaining the low-frequency trend of the motion speed. The signal-to-noise ratio of the denoised signal is increased by about 8 dB. The mean value of the motion speed is set as the 75th percentile as the threshold. The standard deviation reflects the speed fluctuation. The coefficient of variation of the healthy population is usually lower than 25%, and it increases to more than 35% in the stiff state. The support vector machine extracts speed, acceleration, and trajectory curvature as feature vectors. The training samples include 500 groups of normal and abnormal motion data, and the classification accuracy rate reaches 92%. The hidden Markov chain constructs a 5-state model, including motion start, acceleration, constant speed, deceleration, and termination. The normal state transition probability shows a regular distribution, and the residence time is prolonged in the abnormal state. Based on this, the threshold is adjusted to the 60th percentile of the speed mean value to improve the detection sensitivity to mild stiffness.

[0035] S1022. In the key-point motion analysis, if an increase in stiffness is detected, the transmitting end uses a deep convolutional neural network to segment the spinal image and extract the features of the lumbar-thoracic transition region. After noise reduction by median filtering, the Hough transform is used to locate the region boundary and calculate the motion amplitude. Combining with a support vector regressor, the stiffness is quantified and the decision threshold is adjusted in segments. Specifically, a DR device with a pixel resolution of 2048×2048 is used for image acquisition. The deep convolutional neural network contains 5 convolutional layers with a convolutional kernel of 3×3 pixels, paired with a max-pooling layer, and outputs a binary image of the spinal region. The median filter uses a 7×7 pixel window to remove edge noise while retaining the vertebral contour, with an edge fidelity of 95%. The Hough transform detects the vertebral boundary line with an angular resolution of 1 degree and a distance resolution of 2 pixels, and the accumulator threshold is 20% of the image height, accurately locating the region from the tenth thoracic vertebra to the second lumbar vertebra. The bone feature point detector extracts 12 edge control points on each vertebra, fits the center point of the vertebra by the least squares method, sets the residual threshold to 3 pixels, and iterates 50 times to reach sub-pixel accuracy. The triangulation method calculates the angle between adjacent center points and the vertical direction. When bending down normally to touch the feet, the amplitude of the lumbothoracic transition region is about 45 degrees. If it is lower than this value, the support vector regressor calculates the stiffness with a radial basis kernel 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 stiffness levels.

[0036] The analysis of the motion characteristics of the lumbar-thoracic transition region is crucial for spinal function evaluation. The stiffness detection is realized from two aspects of motion amplitude and speed through the above method. To ensure accuracy, the system comprehensively combines Euler angle decomposition and image segmentation techniques to capture multi-dimensional motion information in the sagittal plane, coronal plane, and transverse plane, and optimizes the data quality through wavelet transform and Hough transform. Support vector machines and hidden Markov chains are further used to identify compensatory motion patterns, such as abnormal trajectories when the radius of curvature drops from 500 mm to 300 mm, significantly improving the discrimination ability for irregular motions. Compared with traditional static detection, the method of dynamically adjusting the threshold in the present invention improves the sensitivity to slight stiffness, and the recognition accuracy for 50 patients with low back pain in clinical tests reaches 95%.

[0037] S1023. In practical applications, if the motion speed sequence shows continuous low speed or the amplitude is lower than the benchmark, the system automatically triggers a threshold adjustment strategy. For example, the posture abnormality decision threshold is adjusted from the 75th percentile of the motion speed mean to the 60th percentile. At the same time, combined with the image analysis results, the angle threshold is increased in segments to ensure that the evaluation result reflects the true degree of stiffness. To verify the effect, 100 subjects are tested, including a healthy group and a patient group, and the sensitivity and specificity reach 85% and 90% respectively, indicating the reliability of this method in dynamic posture evaluation.

[0038] A complete motion function evaluation system is constructed through multi-level analysis, covering kinematic feature extraction, statistical analysis, and pattern recognition. It not only objectively quantifies stiffness but also optimizes the adaptability to complex movements through adaptive thresholds, providing reliable data support for subsequent abnormal posture evaluation. The present invention does not overly limit specific algorithm parameters, which can be adjusted by technicians according to the actual scenario, such as the decomposition level of wavelet transform or the kernel function type of support vector machine, to adapt to the motion characteristics of different populations.

[0039] S103. Calculate the joint angle change based on the coordinates of the key points of the leg to evaluate flexibility. If the change amplitude is lower than the preset threshold, it is determined that the flexibility is insufficient, and the tracking density of the leg area is increased by adjusting the tracking strategy to improve the analysis accuracy. First, use the deep residual network to segment the leg area from the video image, extract the coordinate sequences of the hip joint, knee joint, and ankle joint, then calculate the angle change curve through the vector cross product method and perform denoising processing, and analyze the motion characteristics and optimize the tracking point layout in combination with the support vector machine.

[0040] S1031. Use the deep residual network to segment the leg area of the input high-definition video image, obtain the initial coordinate sequences of the hip joint point, knee joint point, and ankle joint point, and after smoothing through Gaussian filtering, calculate the joint angle change curve using the vector cross product method, and then extract the angle change amplitude after denoising by three-layer discrete wavelet transform. If it is lower than the preset threshold, it is marked as a flexibility-insufficient frame. The video input resolution is 1920×1080 pixels. The deep residual network contains 50 convolutional structures, and the segmentation time for each frame is about 40 milliseconds. The accuracy of the segmented leg area reaches 98%. The image block size of each joint area is 200×200 pixels. The Gaussian filter 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 and 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 the inverse trigonometric function and generating a continuous angle change curve. The wavelet transform selects the db4 basis function and performs three-layer decomposition. After removing the high-frequency noise, the correlation coefficient between the curve and the original data reaches 0.95. The angle change amplitude is defined as the difference between the maximum and minimum angles. The hip joint mobility threshold is set at 100 degrees, and the knee joint is 120 degrees. If it is lower than the threshold for 5 consecutive frames, this section is marked as a flexibility-insufficient interval. This design is based on the normal range of motion of the hip joint from 0 to 120 degrees and the knee joint from 0 to 135 degrees for the normal population, aiming to capture subtle flexibility abnormalities.

[0041] For frames with insufficient flexibility, the motion process is further analyzed. The slope sequence of the angle change curve is obtained by piecewise fitting, and the support vector machine is used to evaluate the motion characteristics. Then, a gradient density tracking point array is arranged in the joint area, and the pyramid optical flow method is used to achieve inter-frame tracking to optimize the tracking density. For the denoised angle change curve, the system performs piecewise linear fitting in units of 25 frames, with a 50% overlap between adjacent intervals, and calculates the slope sequence of each segment to reflect the angle change rate. The support vector machine uses the angle change amplitude, angular velocity, and acceleration as input features. The angular velocity threshold is set at 90 degrees per second, and the acceleration is 180 degrees per second squared. The model is trained based on the data of 500 subjects, including 300 in the normal group and 200 in the activity-limited group, and the classification accuracy reaches 92%. According to the evaluation results, a tracking point array is arranged in the hip joint, knee joint, and ankle joint areas. The interval between the tracking points at the joint center is 5 pixels, increasing to 15 pixels towards the edge. For example, 16 points are arranged at the knee joint center, 32 points in the transition area, and 48 points in the edge area, with a total of 96 points. The pyramid optical flow method constructs a three-layer image pyramid, with a search window of 21×21 pixels and a maximum of 20 iterations, ensuring that the tracking success rate under high-speed motion is increased from 85% to 95%. This strategy enhances the adaptability to complex movements by dynamically adjusting the tracking density, especially providing more detailed data support when the leg flexibility is insufficient.

[0042] S1032. The evaluation of the range of motion of leg joints is of great significance for human motion analysis. Through the above method, a complete process from image segmentation to motion feature extraction is realized to ensure the accurate identification of insufficient flexibility. The system not only focuses on the angle change amplitude but also combines the multi-dimensional feature analysis of angular velocity and acceleration to avoid the limitations of a single index. For example, in the action of quickly bending the leg, if the angular velocity is lower than 90 degrees per second and the angle amplitude is insufficient, the system will preferentially increase the tracking point density in the knee joint area, update the spatial coordinates in real time, and capture the details of compensatory movements. In clinical tests, 100 subjects, including 50 normal people and 50 patients with activity limitations, were verified. The sensitivity of this method to mild insufficient flexibility reached 87%, the specificity reached 89%, and the recognition rate for moderate to severe limitations reached 96%, proving its efficiency in dynamic evaluation.

[0043] To further improve the tracking accuracy, the system dynamically adjusts the strategy according to the time interval of insufficient flexibility. For example, when continuous low-amplitude movements are detected, the tracking points in the joint center area are automatically encrypted, while maintaining an appropriate sparsity in the edge area to balance the calculation efficiency and data quality. This adaptive adjustment not only improves the ability to capture the details of leg movements but also provides more reliable basic data for subsequent abnormal posture evaluation. Technicians can adjust the filtering parameters or the number of wavelet decomposition layers according to actual needs to optimize the effect.

[0044] S104. Input the analysis results of spinal curvature, stiffness of the lumbar-thoracic transition region, and leg flexibility into the postural abnormality assessment model. Generate a postural abnormality score through multi-level feature extraction and calculation. If the score exceeds the preset threshold, it is determined as postural abnormality and the corresponding mark is output. First, fit the curvature sequence based on the spinal key point coordinates and perform normalization processing. Subsequently, extract motion features and map quantization values for the lumbar-thoracic transition region and the leg region respectively. Finally, use a four-layer perceptron network to fuse features and calculate the final score.

[0045] S1041. Fit the spinal curvature using a cubic Bezier curve according to the spinal key point coordinate sequence and calculate the normalized value. At the same time, quantify the stiffness of the lumbar-thoracic transition region and the flexibility of the legs through the support vector machine and random forest algorithms respectively, and generate mapping values for subsequent evaluation. 25 marker points are arranged on the spine. Every four adjacent points form a fitting unit. The cubic Bezier curve fits the curve with a control point spacing of 20 mm, and the error is controlled within 0.5 mm. Calculate the curvature values of each segment using the curvature radius formula. For example, the normal value of the thoracic segment is 0.15 to 0.20, and that of the lumbar segment is 0.10 to 0.15. Then map the curvature values to the 0 to 1 interval through maximum-minimum normalization. For the lumbar-thoracic transition region, the collected motion features include joint movement angles, movement speeds, and accelerations, with a sampling frequency of 60 Hz. The support vector machine classifies with a radial basis kernel function. The training data contains 1000 groups of samples. Feature thresholds such as a speed of 90 degrees / second and an acceleration of 180 degrees / second² are used to generate the stiffness quantization value. The normal population is distributed between 0.2 and 0.4, mild stiffness is between 0.4 and 0.6, and severe stiffness exceeds 0.6. The stiffness is converted to a 0 to 10 score through piecewise linear mapping. Values less than 0.4 are mapped to 0 to 3, values between 0.4 and 0.6 are mapped to 4 to 7, and values greater than 0.6 are mapped to 8 to 10. The mapping slopes are 7.5, 15, and 10 respectively to highlight the discrimination of moderate stiffness. For the leg region, the random forest algorithm extracts features from the joint angle change sequence, including the range of motion, duration, and rate of speed change. The algorithm constructs 100 decision trees with a depth of 8. The reference range of the hip joint is 180 degrees, and that of the knee joint is 150 degrees. The normal duration is 2 to 4 seconds, and the rate of speed change threshold is 30%. The flexibility values are normally distributed between 0.7 and 0.9. Similarly, they are converted to the 0 to 10 interval through piecewise 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] Construct a four-layer perceptron network to integrate the above features and calculate the posture anomaly score. The input layer receives the normalized curvature value, stiffness mapping value, and flexibility mapping value. After being processed by three layers of hidden nodes, the scoring result is output and compared with a preset threshold to determine anomalies. The network structure is designed in a pyramid shape. The first hidden layer with 16 nodes is responsible for preliminary feature extraction, the second layer with 8 nodes for feature combination, the third layer with 4 nodes for fusion, and the output layer is a single-node scoring value. The ReLU activation function is selected to avoid the gradient vanishing problem. The training data contains 1000 samples, 600 normal and 400 abnormal, and the test accuracy reaches 93%. The scoring uses a 100-point system, with a benchmark value of 75 points. Scores less than 75 are normal, scores between 75 and 85 are mildly abnormal, and scores greater than 85 are significantly abnormal. Compared with traditional single-index evaluation, multi-dimensional feature fusion can more comprehensively reflect the motor function. For example, patients with mild stiffness may only be abnormal in speed features, while those with significant abnormalities will deviate from the normal range in both curvature and flexibility. This design improves the diagnostic resolution of the system.

[0047] S1042. The importance of multi-dimensional feature analysis in posture assessment lies in its comprehensive capture of complex motion states. It quantifies the motion form through spinal curvature, combines stiffness and flexibility to reflect the degree of functional limitation, providing an objective basis for clinical practice. In clinical verification, 100 subjects were tested, including 50 normal people and 50 abnormal patients. The results showed that the recognition sensitivity for mild abnormalities was 88%, the specificity was 92%, and the accuracy for 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, and the score exceeded 85 points and was correctly marked as abnormal, while the scores of normal people were mostly in the range of 60 to 70 points.

[0048] To enhance the applicability of the model, the system allows adjusting the feature threshold or mapping function according to actual needs. For example, adjusting the second stiffness threshold from 0.6 to 0.65 to adapt to the elderly population, or increasing the number of random forest decision trees to improve the accuracy of flexibility assessment. This flexibility ensures the reliability of the technology in different scenarios and provides high-quality data support for subsequent anomaly report generation.

[0049] S105. Generate a posture anomaly report for the subject's forward bending to touch the feet action according to the posture anomaly score. Extract the abnormal features of the spinal curvature through a deep neural network and mark the specific positions. At the same time, conduct spatial analysis and visual marking on the stiffness in the transition area between the lumbar and thoracic vertebrae and the insufficient flexibility area of the legs to ensure that the report intuitively reflects the spatial distribution and severity of the anomalies.

[0050] S1051. Use the posture anomaly score to drive the deep neural network to analyze the spinal curvature characteristics. Use a five-layer convolution structure to locate the curvature abnormal segment and use morphological processing to extract the boundary. At the same time, use density clustering and region growing algorithms to divide the abnormal area for the transition area between the lumbar and thoracic spine and the leg area and perform multi-level annotation. The deep neural network inputs a 1280×1024 pixel image, contains three convolution layers and two fully connected layers, and the convolution kernel size is 3×3 pixels. After maximum pooling, the feature map is gradually reduced to 80×64 pixels. The training data includes 2000 samples, including 800 abnormal samples and 1200 normal samples. The positioning accuracy rate is 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 8 domains and 3×3 structural elements. Iterate 3 times to connect abnormal points within 15 pixels to form a boundary point set of 200 to 300 points. The boundary is marked with a red solid line with a line width of 2 pixels. The arrow points to the center of the abnormality and the length is one-fourth of the diameter of the region. The stiffness analysis of the transition area between the lumbar and thoracic vertebrae uses a density clustering algorithm. The clustering radius is set to 25 mm, which is approximately equal to the distance between the two vertebrae. The minimum density threshold is 3 abnormal points. The clustering results show that the high-density abnormal cluster from the twelfth thoracic vertebra to the second lumbar vertebra has an area of about 400 square millimeters, marked with a yellow semi-transparent outline, and the centroid coordinates are displayed in white font. The areas of insufficient leg flexibility were analyzed by the regional growing algorithm, with the hip and knee joints as seed points. The growth condition was that the angle difference between adjacent pixels was less than 5 degrees. Normal hip flexion reached 120 degrees and knee flexion reached 135 degrees. Below 90 degrees was marked as severely limited, 90 to 105 degrees as moderate, and 105 to 120 degrees as mild, and were marked with red, yellow, and green gradients, respectively.

[0051] For the precise positioning and visualization of abnormal areas, the annotation strategy is further optimized. The spatial distribution of leg joint restriction points is calculated by the k-nearest neighbor algorithm and the regional expansion method, and the support vector machine is combined to classify the degree of stiffness to ensure the comprehensiveness and readability of the abnormal report. In the leg analysis, the k-nearest neighbor algorithm sets the k value to 5, the search radius is 30 mm, and the abnormal point groups around the hip and knee joints are identified. The expansion radius is dynamically adjusted according to the range of motion, for example, it increases to 25 mm when the hip joint is 90 degrees, and a blue dotted line annotation is generated, with an interval of 5 pixels and a line width of 1 pixel. The stiffness of the transition area between the lumbar and thoracic vertebrae is grown with the center point of the vertebral body as the seed, and the difference in the movement angle is less than 5 degrees. The coverage area is about 400 square millimeters. The support vector machine is classified by the radial basis kernel function. Mild restricted movement is reduced by 20% to 30%, moderate by 30% to 50%, and severe by more than 50%. It is marked with a yellow outline, and the virtual-real ratio varies with the degree. This multi-level visualization not only highlights the range of abnormality, but also distinguishes the severity by color, which improves the doctor's intuitive judgment of abnormal parts.

[0052] S1052. The abnormal report is generated using a natural language generator, integrating quantitative and qualitative descriptions, providing the spatial coordinates of the abnormal area, the degree grading, and information on the affected parts. At the same time, annotations are superimposed on the human posture image to visually display the specific locations of spinal curvature abnormalities, stiffness, and insufficient flexibility. The quantitative part of the report records, for example, that the kyphosis of the sixth thoracic vertebra is 47 degrees, exceeding the normal value by 7 degrees, and the mobility of the twelfth thoracic vertebra is 8 degrees, lower than the normal value by 15 degrees. Indicators such as area, perimeter, and center of gravity position are marked on the image edge in square millimeters and millimeters, with a font size of 12 pixels. The qualitative description, such as "the kyphosis in the middle thoracic region increases, with the apex vertebra being the sixth thoracic vertebra, affecting the two vertebrae above and below", is generated through a structured template. The underlying layer of the image annotation is the original posture image. The spinal curvature abnormality is filled with red semi-transparency, with a transparency of 40%. The stiff lumbar region is marked with a yellow contour line, and the restricted leg region is shown with a blue dashed line. The superimposed effect obtained a recognition score of 9.2 in clinical tests.

[0053] Through the above method, the accurate positioning and visual expression of abnormal features are achieved. In clinical verification, 100 subjects were tested, 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 report consistency reached 92%, significantly improving the diagnostic efficiency and providing a reliable basis for formulating treatment plans. Technicians can adjust the clustering radius or growth conditions according to requirements 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 angles are extracted from the real-time posture data and compared with the preset normal range. The deviation degree is calculated to evaluate the action normality and generate a score. At the same time, a deep convolutional neural network, triangulation, and wavelet transform are used to optimize the data quality, and staged analysis is achieved through dynamic time warping and support vector machines.

[0055] S1061. Four high-speed cameras are used to collect image sequences. A deep convolutional neural network is employed to locate the skeletal key points and the three-dimensional coordinates are calculated through triangulation. Subsequently, a three-layer discrete wavelet transform is performed on the coordinate sequence for smoothing processing to obtain the motion trajectory. The dynamic time warping algorithm is combined to divide the action phases. Finally, the support vector machine is used to calculate the joint angle deviation and generate a normalized score by weighting. The camera resolution is 1920×1080 pixels, the frame rate is 120 frames per second, and the adjacent spacing is 2 meters to ensure field-of-view overlap to improve the reconstruction accuracy. The deep convolutional neural network locates 15 key points, including the hip, knee, ankle, etc., and the positioning error is controlled at the sub-pixel level. The triangulation calculates the spatial coordinates based on the calibration parameters, and the accuracy is better than 2 mm. The three-layer wavelet transform selects the db4 basis function, and the signal-to-noise ratio is increased by about 8 dB after removing the high-frequency noise. The smoothed trajectory can accurately reflect the motion continuity. The dynamic time warping algorithm analyzes the displacement change rate with a 30-frame window (about one-fourth of the action period) to divide the bending-to-touch-foot action into four phases: preparation, acceleration, stability, and deceleration, accounting for 15%, 35%, 35%, and 15% of the period respectively. The joint angles in each phase are compared with the preset ranges. For example, the hip joint is 0 to 30 degrees in the preparation phase, 30 to 90 degrees in the acceleration phase, 90 to 120 degrees in the stability phase, and 120 to 90 degrees in the deceleration phase. The knee joint is 0 to 20 degrees, 20 to 60 degrees, 60 to 90 degrees, and 90 to 60 degrees in turn. The support vector machine processes the deviation features with a radial basis kernel function, including the angle difference, angular velocity difference, and acceleration difference. The normal ranges are 10 degrees, 20 degrees per second, and 40 degrees per second squared respectively. The weight distribution is 0.2 for preparation, 0.3 for acceleration, 0.3 for stability, and 0.2 for deceleration. The score is in a percentile system. A score above 90 is considered normal, 80 to 90 is slightly abnormal, 70 to 80 is moderately abnormal, and below 70 is severely abnormal.

[0056] The assessment of action normality needs to accurately capture three-dimensional motion features and conduct multi-dimensional analysis. Through the above methods, from coordinate acquisition to trajectory smoothing, then to phase division and deviation calculation, a complete process is formed to ensure the objectivity and reliability of the assessment results. Compared with traditional two-dimensional analysis, the combination of three-dimensional coordinates and wavelet transform effectively eliminates noise interference. For example, the coordinate offset caused by image jitter is reduced from 5 mm to less than 1 mm. The dynamic time warping solves the error caused by the action speed difference through time-axis alignment. The multi-feature classification of the support vector machine improves the sensitivity to subtle abnormalities. In clinical tests, the sensitivity for identifying slight abnormalities in 100 subjects (50 patients and 50 normal individuals) reaches 85%, and the accuracy for moderate and severe abnormalities exceeds 95%.

[0057] S1062. To further improve the evaluation accuracy, the system dynamically adjusts the analysis strategy according to the action stage characteristics. For example, during the acceleration and stable stages, the weight is increased to highlight the execution quality of key actions. At the same time, technicians are allowed to adjust the threshold range according to the subject group. For example, the hip joint stable stage for the elderly population 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. Dynamically adjust the key point tracking strategy and the posture abnormality determination threshold according to the action normality score. Extract the deviation features of low-score actions through a deep convolutional neural network and combine density clustering to identify the area to be optimized. Use the gradient boosting tree to update the threshold and the Kalman filter to optimize the trajectory prediction, so as to improve the accuracy and real-time performance of subsequent evaluations.

[0059] S1071. Trigger the adaptive update mechanism based on the action normality score. Use a five-layer deep convolutional neural network to analyze the action sequence with a score lower than 80 points to extract the key point displacement deviation value. Subsequently, use the density clustering algorithm to identify the joint area with a large deviation and determine the optimization range. Then, use the gradient boosting tree to dynamically adjust the abnormality determination threshold according to the joint angle, speed, and acceleration characteristics. Finally, predict and correct the key point trajectory through the Kalman filter. The deep convolutional neural network contains 5 convolutional structures, the convolutional kernel is 3×3 pixels, and it is paired with a max pooling layer to extract features from 10 consecutive frames of data, calculate the key point displacement deviation, and the normal deviation is less than 5 mm. If it exceeds, an update is triggered. Density clustering analyzes the spatial distribution with a radius of 30 mm and a density threshold of 5 points. For example, the abnormal area of the knee joint covers 80 square millimeters around the patella, and the average deviation reaches 15 mm, accurately locating the position of the dysfunction. The gradient boosting tree constructs 100 decision trees with a depth of 5. The input feature weights are 0.4 for joint angle, 0.3 for speed, and 0.3 for acceleration. The normal range of the hip joint is 0 to 120 degrees, the speed is 90 degrees / second, and the acceleration is 180 degrees / second². The threshold floats by 15% according to the deviation degree, and the phased reference values are such as 20 degrees for knee joint preparation, 45 degrees for acceleration, 90 degrees for stability, and 60 degrees for deceleration. The state vector of the Kalman filter includes three-dimensional coordinates and speed. 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, ensuring the continuity of the trajectory.

[0060] The dynamic adjustment for the area to be optimized realizes optimized tracking by sliding a time window and density increment. The window length is about 30 frames, which is one-fourth of the action period. When the score is lower than 60 points, the key point density increases to 2 times, and when it is lower than 40 points, it increases to 3 times. Cubic spline interpolation is used to generate new points to maintain smoothness. This strategy can flexibly encrypt tracking points according to the severity of anomalies. For example, the low-score area of the hip joint increases from the original 10 points to 30 points, improving the ability to capture subtle deviations and significantly enhancing real-time performance. In clinical tests, the sensitivity to minor anomalies of 100 subjects (50 patients and 50 normal people) reaches 92%, and the accuracy of moderate and severe anomalies exceeds 97%.

[0061] The importance of the dynamic optimization strategy in human motion analysis lies in its adaptability. Through multi-level algorithm fusion, from deviation extraction to area recognition and then 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, during the acceleration stage, due to the large movement amplitude, deviations are more likely to occur, so the threshold adjustment amplitude is larger, ensuring the analysis effect of complex actions.

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

[0063] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for evaluating abnormal human postures based on a physical fitness integrated machine, characterized in that, The method includes: Obtaining video data of the subject's bending-down-to-touch-toes action collected by multiple cameras, preliminarily locating key points in the spine, the transitional region between the lumbar and thoracic vertebrae, and the legs, and extracting the coordinates of the key points; Calculating the movement amplitude of the transitional region between the lumbar and thoracic vertebrae based on the coordinates of the key points in this region. If the movement amplitude is lower than the preset threshold, it is determined that the stiffness has increased, and the posture abnormality determination threshold is adjusted; Calculating the change in leg joint angles based on the coordinates of the key points of the legs. If the amplitude of the angle change is lower than the 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 region; Inputting the determination results of the spinal curvature, the stiffness of the transitional region between the lumbar and thoracic vertebrae, and the leg flexibility into the posture abnormality evaluation model to calculate the posture abnormality score. If the score exceeds the preset threshold, it is determined that the posture is abnormal; Generating a posture abnormality report for the subject's bending-down-to-touch-toes action according to the posture abnormality score, and marking the specific positions of abnormal spinal curvature, increased stiffness in the transitional region between the lumbar and thoracic vertebrae, and insufficient leg flexibility; Obtaining the three-dimensional coordinate information of the skeletal key points of the subject, extracting target joint angle data from the real-time posture data, comparing the target joint angle data with the preset normal angle range, where the normal angle range includes angle thresholds for different action stages, calculating the deviation degree of the target joint angle data from the normal angle range, determining the degree of non-standardness of the subject's action execution, and generating an action standardness score; Dynamically updating the key point tracking strategy and the posture abnormality determination threshold according to the action standardness score to optimize the accuracy and real-time performance of subsequent action evaluation.

2. The method according to claim 1, wherein The obtaining video data of the subject's bending-down-to-touch-toes action collected by multiple cameras, preliminarily locating key points in the spine, the transitional region between the lumbar and thoracic vertebrae, and the legs, and extracting the coordinates of the key points includes: Performing time calibration on the video data collected by multiple cameras through timestamps, and obtaining the preset marker points on the subject's trunk as the reference data of the action sequence according to the time calibration result; For the reference data of the action sequence, using a deep convolutional neural network to identify the spine region to obtain the coordinates of the skeletal key points; Performing multi-view fusion based on the coordinates of the skeletal key points and the camera calibration parameters, and using the triangulation method to obtain the three-dimensional spatial coordinates of the skeletal key points.

3. The method according to claim 1, wherein The calculating the movement amplitude of the transitional region between the lumbar and thoracic vertebrae based on the coordinates of the key points in this region. If the movement amplitude is lower than the preset threshold, it is determined that the stiffness has increased, and the posture abnormality determination threshold is adjusted, includes: Using the Euler angle decomposition method to calculate the angle between the line connecting the key points in the transitional region between the lumbar and thoracic vertebrae and the reference coordinate system, and obtaining the movement amplitude values in the sagittal plane, coronal plane, and transverse plane of the region; According to the movement amplitude values, calculating the movement speed curve through a sliding time window, and performing noise reduction processing on the movement speed curve using wavelet transform to obtain a smoothed movement speed sequence; Calculating the mean and standard deviation of the movement speed for the smoothed movement speed sequence to determine whether the movement stiffness has increased; During the time period when the movement stiffness increases, a support vector machine is used to analyze the movement trajectories of key points, an abnormal movement pattern is identified through a hidden Markov chain, and the posture abnormality determination threshold is adjusted.

4. The method according to claim 3, wherein It also includes: Using an image recognition algorithm to detect the pixel coordinates of the key points of the lumbar and thoracic vertebrae, defining the angle formed by connecting the key points of the lumbar and thoracic vertebrae as the movement amplitude of this region, setting a reference threshold for the movement amplitude of the lumbar-thoracic transition region when a normal adult bends down to touch the feet. If the detected movement amplitude is less than the reference threshold, it is determined that the stiffness of the lumbar-thoracic transition region increases. According to the degree of increased stiffness, the angle threshold for posture abnormality determination is adjusted upwards, specifically including: Using a deep convolutional neural network to segment the spine image, and performing noise reduction on the segmented spine region image through median filtering to obtain the noise-reduced spine image; According to the noise-reduced spine image, using the Hough transform to identify the boundary of the lumbar-thoracic transition region, and obtaining the coordinates of the vertebral body edge points in this region through a bone feature point detector; For the coordinates of the vertebral body edge points, using the least squares method to fit to obtain the sequence of vertebral body center points, and calculating the angle between the connection line of adjacent vertebral body center points and the vertical direction through triangulation; Comparing the angle value with a preset reference value. If the angle value is less than the reference value, a support vector regressor is used to calculate the stiffness quantization value, and the posture abnormality determination threshold is adjusted through a piecewise function.

5. The method according to claim 1, characterized in that, Calculating the change in leg joint angles based on the coordinates of the leg key points. If the change in angle amplitude 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 region, including: Using a deep residual network to segment the leg region in the video image, and extracting the initial coordinate sequences of the hip joint point, knee joint point, and ankle joint point from the segmented region; According to the initial coordinate sequence, using the vector cross product method to calculate the change curve of the angle between joint points, and performing three-layer discrete wavelet transform denoising processing on the change curve of the angle; For the denoised change curve of the angle, extracting the difference between the maximum angle value and the minimum angle value as the angle change amplitude. If the angle change amplitude is lower than the preset threshold, this frame is marked as a flexibility-insufficient frame; According to the curve slope sequence of the flexibility-insufficient frame, using a support vector machine to evaluate the movement process, and arranging a gradient density tracking point array in the joint region.

6. The method according to claim 1, characterized in that Inputting the determination results of the spine bending curvature, the stiffness of the lumbar-thoracic transition region, and the leg flexibility into the posture abnormality evaluation model to calculate the posture abnormality score. If the score exceeds the preset threshold, it is determined that the posture is abnormal, including: Fitting according to the spine key point coordinate sequence using a cubic Bezier curve, obtaining a sequence of curvature values through the curvature radius calculation formula, and using the maximum-minimum normalization method to map the sequence of curvature values to the interval from zero to one to obtain the normalized curvature value; For the lumbar-thoracic transition region corresponding to the normalized curvature value, using a support vector machine to classify the joint movement angle, movement speed, and acceleration characteristics of this region, obtaining the stiffness quantization value and converting it through a piecewise linear mapping function to obtain the stiffness mapping value; For the leg region corresponding to the stiffness mapping value, the random forest algorithm is used to extract the joint range of motion, movement duration, and speed change rate features from the joint angle change sequence. The leg flexibility value is calculated by the weighted average method and converted into a flexibility mapping value through a piecewise linear mapping function. Using a four-layer perceptron network, the normalized curvature value, the stiffness mapping value, and the flexibility mapping value are input into the input layer. After being processed by the hidden layer, the posture abnormality score value is obtained from the output layer. If the score value exceeds the preset abnormality determination reference value, a posture abnormality mark is output.

7. The method according to claim 1, characterized in that Generating a posture abnormality report for the subject's bending-down-to-touch-the-feet action according to the posture abnormality score, and marking the specific positions of spinal curvature abnormality, increased stiffness in the lumbar-thoracic transition region, and insufficient leg flexibility, including: Obtaining the spinal curvature features according to the posture abnormality score, and extracting the continuous abnormal segments with curvature values exceeding the preset threshold through the five-layer convolutional structure of the deep neural network. Adopting a morphological processing method for the continuous abnormal segments to obtain the boundary of the abnormal region, and marking a red boundary line at the edge of the abnormal region. For the lumbar-thoracic transition region, the density clustering algorithm is used to spatially group the regions with increased stiffness, and the area of the region is marked according to the density clustering result. The region growing algorithm is used to analyze the region with insufficient flexibility, and the mildly restricted region, moderately restricted region, and severely restricted region are divided according to the joint movement angle value.

8. The method according to claim 7, wherein It also includes: Corresponding the abnormal key features with the detection results in three aspects: spinal curvature, stiffness in the lumbar-thoracic transition region, and leg flexibility, extracting the key point coordinates of the spinal curvature interval, lumbar-thoracic transition region, and leg joints determined to be abnormal. According to the abnormal key point coordinates, marking the specific positions of spinal curvature abnormality, stiffness in the lumbar-thoracic transition region, and insufficient leg flexibility in the human body posture image, specifically including: Using a deep convolutional neural network to extract the spinal curvature feature map from the human body posture image, and the feature map marks the abnormal region coordinate points through the max pooling layer. Performing morphological processing on the abnormal region coordinate point sequence using the dilation operator to obtain the boundary point set of the spinal curvature abnormal region, and using the region growing algorithm for segmentation to obtain the contour line of the stiff region with the vertebral body center point as the seed point. Calculating the spatial distribution of the joint movement restricted points according to the contour line of the stiff region using the k-nearest neighbor algorithm, and obtaining the boundary of the restricted region range through the region expansion method.

9. The method according to claim 1, characterized in that, Obtaining the three-dimensional coordinate information of the skeletal key points of the subject, extracting the target joint angle data from the real-time posture data, comparing the target joint angle data with the preset normal angle range, where the normal angle range includes the angle thresholds in different action stages, calculating the deviation degree of the target joint angle data from the normal angle range, and determining the non-standard degree of the subject's action execution to generate an action standardization score, including: Using a deep convolutional neural network to locate the human skeletal key points, and obtaining the three-dimensional spatial coordinate sequence of the skeletal key points through triangulation between adjacent cameras. According to the three-dimensional space coordinate sequence, perform smoothing processing using three-layer discrete wavelet transform to obtain the motion trajectories of the skeletal key points; For the motion trajectories of the skeletal key points, use the dynamic time warping algorithm to calculate the time segmentation points and obtain four standard stages: preparation, acceleration, stability, and deceleration; According to the comparison result between the included angle of the connection lines of the skeletal key points within the standard stage and the preset normal angle range, use a support vector machine to calculate the deviation degree of the joint angles in each stage, and obtain the action normality score through weighted processing.

10. The method according to claim 1, wherein According to the action normality score, dynamically update the key point tracking strategy and the posture anomaly determination threshold to optimize the accuracy and real-time performance of subsequent action evaluation, including: Obtain key point tracking data according to the action normality score, and use a deep convolutional neural network to extract features from the action sequence with a score lower than the first threshold to obtain the key point displacement deviation value; Perform spatial distribution clustering analysis on the key point displacement deviation value, use the density clustering algorithm to identify the joint areas with large deviation values, and obtain the spatial range of the area to be optimized; For the joint angle, motion speed, and acceleration characteristics of the area to be optimized, use the gradient boosting tree to dynamically adjust the anomaly determination threshold to obtain the updated threshold; Use a Kalman filter to predict the trajectory of the key point coordinates. The state vector of the filter includes the three-dimensional coordinates and speed of the key point, and the corrected prediction result is obtained through measurement update.

Images