Stroke patient balance function intelligent evaluation system based on posture recognition

Abstract

The present application relates to the technical field of medical rehabilitation evaluation, and discloses a stroke patient balance function intelligent evaluation system based on posture recognition, comprising a data acquisition device and a computing terminal. The system assigns an asymmetric pathological compensation monitoring weight to the affected key points according to clinical prior information; generates a binary validity mask by fusing multi-dimensional image data and combining visual confidence; uses a Kalman filter model to predict coordinates when the limbs are blocked, and outputs corrected three-dimensional skeletal coordinates; calculates a dynamic barycentric trajectory based on the coordinates to obtain a basic physical balance score; extracts spatial compensation parameters such as trunk lateral deviation angle and pelvic height difference to calculate a comprehensive dynamic penalty coefficient; and finally, the penalty coefficient is used to deduct the numerical value of the basic physical balance score to output a real balance evaluation score. The present application eliminates the illusion that the patient maintains balance by relying on abnormal compensation movements, and objectively reflects the neural motor control level of the test object.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation assessment technology, specifically to an intelligent assessment system for balance function in stroke patients based on posture recognition. Background Technology

[0002] Stroke patients with hemiplegia often experience balance dysfunction, and reliable balance assessment is the foundation for developing rehabilitation treatment plans. Existing balance assessment methods mostly employ computer vision-based posture recognition and center of gravity trajectory tracking systems.

[0003] Current visual pose recognition solutions typically assign equal computational weights to all skeletal nodes throughout the test subject's body for tracking. Stroke patients exhibit a pathological characteristic of bilateral neuromotor control asymmetry, and the equal-weight tracking method fails to consider this clinical feature, making it difficult to accurately capture abnormal movements on the affected side. Furthermore, during dynamic balance tests, patients' limbs often occlude each other. Existing recognition systems, when faced with decreased visual confidence due to limb occlusion, often experience coordinate data jumps or tracking loss, leading to interruptions in skeletal movement trajectories and affecting the reliability of subsequent data calculations.

[0004] Furthermore, conventional visual center-of-gravity tracking systems primarily rely on calculating the swing path or offset distance of the body's overall center of gravity to output a balance score. This calculation method focuses on assessing whether the patient can maintain a stable center of gravity. However, in reality, stroke patients often use pathological compensatory movements such as trunk tilting, pelvic elevation, or upper limb flexor spasm to maintain their center of gravity in order to prevent falls. Existing assessment algorithms fail to quantify and incorporate these abnormal compensatory movements from a spatial geometric perspective, and cannot distinguish between simple center-of-gravity maintenance and the actual level of neuromotor control. This results in the output assessment score being mixed with the compensatory balance performance established by the patient using compensatory movements, failing to objectively reflect the patient's recovery status. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent assessment system for balance function in stroke patients based on posture recognition. This system solves the problem that existing posture recognition assessment systems fail to quantify and eliminate abnormal compensatory movements when stroke patients maintain their center of gravity, resulting in assessment scores that cannot objectively reflect the patient's true neuromotor control level.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: This invention provides an intelligent assessment system for balance function in stroke patients based on posture recognition, including a data acquisition device and a computing terminal.

[0007] The data acquisition device is used to acquire continuous video stream data of the test object during the evaluation period and transmit it to the computing terminal. The continuous video stream data includes image frames arranged in time sequence, as well as color image streams and depth data streams. The computing terminal includes a constraint initialization module, a data processing module, a basic evaluation module, a compensation extraction module, and a calibration output module.

[0008] The constraint initialization module receives prior clinical information from the test subject, including the test posture pattern and affected-side labeled variables. Based on the affected-side labeled variables, the system determines the affected and healthy sides, classifying anatomical and skeletal key points on the affected side as pathological monitoring nodes and assigning asymmetric pathological compensation monitoring weight variables. Simultaneously, anatomical and skeletal key points on the healthy side are classified as regular tracking nodes and assigned regular posture extraction weight variables. Stroke patients with hemiplegia typically exhibit asymmetric neurological control deficits. By assigning asymmetric weights, the system alters the conventional approach of treating the entire body with equal weight in posture recognition, shifting computational resources and feature extraction priority towards the affected side, thereby increasing the accuracy of capturing pathological compensation movements.

[0009] The data processing module fuses the two-dimensional pixel coordinates of the skeletal keypoints set on the color image with the depth distance values ​​in the depth data stream, mapping them to physical space coordinates in an absolute three-dimensional coordinate system. It also obtains the peak value of the probability heatmap as the visual confidence level and combines them to generate a four-dimensional data tensor. The system compares the visual confidence level in the four-dimensional data tensor with a preset global recognition confidence threshold: when the visual confidence level is greater than or equal to the global recognition confidence threshold, the binary validity mask is assigned the value one; when the visual confidence level is less than the threshold, it is assigned the value zero.

[0010] Simultaneously, the data processing module establishes independent temporal prediction models for each skeletal keypoint using a Kalman filter model. When the binarization validity mask is set to a value of one, the three-dimensional coordinate parameters in the four-dimensional data tensor are used as the observation inputs of the Kalman filter model for weighted calculation and internal covariance updates, outputting the corrected three-dimensional skeletal coordinate data; When the binarization validity mask is zero, the system blocks the observation input path. The Kalman filter model uses the state transition matrix to perform pure state prediction based on the motion velocity variables extracted from previous unoccluded image frames. The predicted 3D spatial coordinates are then output as the corrected 3D skeletal coordinate data. This mechanism establishes a confidence-based spatial occlusion determination logic. When limb occlusion causes a decrease in confidence, it blocks erroneous coordinate inputs and maintains the continuity of the skeletal trajectory by relying on historical motion states.

[0011] The basic assessment module determines the set of body segments to be included in the calculation based on the test posture pattern. Using the 3D coordinates of proximal and distal skeletal key points in the corrected 3D skeletal coordinate data, and combining this with the human segment centroid position constant, linear interpolation is performed to calculate the local centroid 3D coordinates. A weighted summation algorithm is then used to obtain the spatial coordinates of the overall centroid, generating a dynamic 3D centroid trajectory sequence. The system calculates and sums the Euclidean distance of the overall centroid's spatial coordinates in 3D physical space between two adjacent image frames to obtain the total length of the centroid swing path.

[0012] The system obtains the average total length of the center of gravity swing path from healthy individuals of the same age as a reference value. When the total length of the center of gravity swing path is less than or equal to the reference value, the basic physical balance score is assigned to a preset full score. When it is greater than the reference value, the system performs a difference calculation, using an exponential decay function combined with a preset decay constant to subtract the full score from the calculated difference, thus obtaining the basic physical balance score. Using an exponential decay function to handle the center of gravity shift difference conforms to the nonlinear law of the human body's physical balance loss process.

[0013] The compensation extraction module combines the corrected 3D skeletal coordinate data with the pathological compensation monitoring weight variables to calculate the spatial geometric linkage relationship, obtaining the trunk lateral deviation angle, the coronal plane height difference of the pelvis, and the spatial angle of the affected elbow joint. The specific calculation process is as follows: extract the coordinates of the left and right acromions and anterior superior iliac spines, and calculate the arithmetic mean to obtain the 3D spatial coordinates of the shoulder center and the pelvic center; construct a trunk spatial vector with the pelvic center as the starting point and the shoulder center as the ending point, and calculate the angle between this vector and the Y-axis direction vector of the absolute 3D coordinate system to obtain the trunk lateral deviation angle.

[0014] The difference between the Y-axis coordinates of the affected and healthy anterior superior iliac spines is calculated to obtain the coronal plane height difference of the pelvis; a difference less than zero is assigned a value of zero. Starting from the coordinates of the affected elbow joint, upper arm and forearm spatial vectors are constructed, ending at the coordinates of the affected acromion and wrist joint, respectively. The spatial angle of the affected elbow joint is obtained by calculating the dot product of the two vectors. This process quantifies, from a spatial geometric perspective, the typical compensatory movement pattern of trunk tilting, pelvic elevation, and upper limb flexor spasticity associated with maintaining balance in stroke patients.

[0015] The system utilizes a piecewise mapping function to transform the aforementioned spatial angles and height differences into trunk lateral deviation compensation indices, pelvic elevation compensation indices, and upper limb flexor spasticity compensation indices. It then calculates the arithmetic mean over the assessment period to obtain global trunk, pelvic, and upper limb compensation coefficients. The system generates an enable mask based on the test posture pattern and uses this mask to dynamically filter preset basic weight parameters. It then performs a weighted summation of each global compensation coefficient according to the enable mask and basic weight parameters, and divides the result by the sum of the actual effective weights under the current test mode to obtain a comprehensive dynamic penalty coefficient. Using this masking mechanism, the system can adaptively address differences in compensation sites under different assessment actions.

[0016] The calibration output module multiplies the comprehensive dynamic penalty coefficient with a preset maximum penalty ratio factor to obtain the actual deduction ratio. The calibration discount coefficient is then obtained by subtracting the actual deduction ratio from the numerical value. The system multiplies the baseline physical balance score with the calibration discount coefficient to calculate the true balance assessment score, and performs rounding and truncation to limit it to a closed interval of zero to one hundred. Finally, the true balance assessment score is compared with a preset obstacle threshold to output the clinical balance grading result. Various global compensation coefficients are compared with label trigger thresholds to generate specific compensation labels, and the results are integrated to construct and output a comprehensive balance assessment report. This architecture employs a multiplicative approach, using the severity of compensation to apply numerical discounts to the center-of-gravity physical score. This effectively eliminates compensatory balance performance established by patients through pathological abnormal movements, objectively reflecting their neuromotor control level.

[0017] This invention provides an intelligent assessment system for balance function in stroke patients based on posture recognition. It has the following beneficial effects: 1. This invention receives prior clinical information through a constraint initialization module and assigns asymmetric pathological compensation monitoring weight variables to key skeletal points on the affected and healthy sides. This design changes the conventional approach of assigning equal weights to all nodes in the body during posture recognition. Considering the asymmetric nature of bilateral neural control in stroke hemiplegic patients, it increases the system's monitoring weight for pathological compensation movements on the hemiplegic side, thereby improving the accuracy of capturing abnormal movements on the affected side.

[0018] 2. This invention utilizes a data processing module to fuse two-dimensional pixel coordinates and depth data to generate a four-dimensional data tensor, and generates a binary validity mask based on visual confidence to control the observation input of the Kalman filter model. When a patient's limb is occluded, causing the recognition confidence to fall below a threshold, the system uses this mask to block erroneous three-dimensional coordinate input and uses the motion velocity extracted from preceding normal image frames for coordinate prediction. This processing method reduces coordinate jumps caused by limb occlusion and maintains the continuity of the skeletal motion trajectory.

[0019] 3. This invention introduces a comprehensive dynamic penalty coefficient to calibrate the baseline physical balance score during the assessment phase. The system quantifies the degree of abnormal movements the patient makes while maintaining balance by calculating spatial compensation parameters from multiple locations, such as the trunk lateral deviation angle, pelvic height difference, and the angle of the affected elbow joint, and deducts from the baseline physical balance score accordingly. This calculation logic eliminates the illusion that the patient relies on compensatory movements to maintain balance, ensuring that the output true balance assessment score objectively reflects the test subject's neuromotor control level. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a comparison chart of the clinical consistency of balanced assessment scores under different algorithm models of this invention; Figure 4 This is a distribution map of the spatial compensation characteristics of multiple body segments under different clinical grades according to the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see the appendix Figure 1 This invention provides an intelligent assessment system for balance function in stroke patients based on posture recognition, comprising: a data acquisition device and a computing terminal.

[0023] The data acquisition equipment employs a vision image acquisition device with a depth sensor. It acquires the 3D calibration parameters of the test environment and continuous video stream data of the test object. The data acquisition equipment is communicatively connected to a computing terminal, transmitting the continuous video stream data to the terminal.

[0024] The computing terminal is equipped with multiple data processing modules. These modules include a constraint initialization module, a data processing module, a basic evaluation module, a compensation extraction module, and a calibration output module.

[0025] Please see the appendix Figure 2 This invention provides an intelligent assessment method for balance function in stroke patients based on posture recognition, comprising the following steps: S10, The constraint initialization module receives the clinical prior information of the test object. The clinical prior information includes the affected side marker variable and the test position pattern. The constraint initialization module assigns asymmetric pathological compensation monitoring weight variables to the corresponding anatomical skeletal key points of the test object in the calculation system according to the affected side marker variable. S20, the data processing module receives continuous video stream data, processes the continuous video stream data frame by frame to extract the set of skeletal key points, and generates a four-dimensional data tensor containing three-dimensional spatial coordinates and visual confidence. The data processing module generates a binary validity mask for each skeletal key point according to a preset global recognition confidence threshold. For skeletal key points that are determined to be invalid by the binary validity mask, the data processing module uses a temporal prediction model to perform spatial coordinate interpolation smoothing and outputs the corrected three-dimensional skeletal coordinate data. S30, the basic assessment module receives the corrected three-dimensional skeletal coordinate data, calculates the spatial coordinates of the overall center of gravity of the test object according to the preset multi-body segment dynamic parameters, calculates the spatial swing characteristic value of the center of gravity within the assessment period and calculates the basic physical balance score without compensatory action intervention. S40, the compensation extraction module receives the corrected three-dimensional skeletal coordinate data, validity mask and pathological monitoring weight variables. The compensation extraction module calculates the spatial geometric linkage relationship of the skeletal key points that meet the validity conditions based on the binary validity mask, and obtains the multi-body segment spatial compensation feature parameters. The multi-body segment spatial compensation feature parameters include the trunk lateral deviation angle, the coronal plane height difference of the pelvis and the spatial angle of the affected elbow joint. The compensation extraction module calculates the comprehensive dynamic penalty coefficient. S50, the calibration output module obtains the basic physical balance score and the comprehensive dynamic penalty coefficient. The calibration output module uses the comprehensive dynamic penalty coefficient to numerically correct the basic physical balance score to obtain the true balance assessment score. The calibration output module converts the final balance function score into the corresponding clinical balance grading result and outputs an assessment report containing the scale grading and abnormal posture feature records.

[0026] The constraint initialization module executes the following specific workflow for receiving clinical prior information of the test object and allocating pathological monitoring weight variables.

[0027] Obtain the clinical prior information of the test subjects, which includes the affected side labeled variables and the test position pattern.

[0028] The methods for obtaining prior clinical information of test subjects include receiving manually input data through a graphical user interface, or reading the test subjects' medical history data files from a medical information system through a data communication interface and performing structured parsing. Structured parsing refers to extracting fields related to limb involvement from electronic medical records. In stroke patients, unilateral cerebral hemisphere lesions can lead to contralateral motor function impairment. The visual algorithm evaluation system obtains the affected side to modify the default setting of conventional posture recognition algorithms for left-right lateral movement symmetry. The affected side marker variable reflects the degree of motor function impairment in the limbs contralateral to the lesion in the stroke test subject. The affected side marker variable is defined as... , Affected side labeled variables In Refers to the left side, used to indicate that the left limb of the test subject is the hemiplegic side with impaired motor function; Affected side labeled variable In The term "right side" indicates that the test subject's right limb is the hemiplegic side with impaired motor function. The test position modes are set according to rehabilitation medicine assessment scenarios, and the specific features include a sitting test mode, a bipedal standing test mode, and a walking test mode. The test position mode serves as a filter condition constraining the range of the skeletal key point set. In the sitting test mode, the algorithm ignores the spatial displacement calculation of the knee and ankle joints; in the bipedal standing and walking test modes, the knee and ankle joints are included in the skeletal key point set.

[0029] For the design of view components in graphical user interfaces and the establishment of protocols for data communication interfaces in medical information systems, those skilled in the art can use conventional software programming tools and network transmission technologies. Their data interaction mechanisms are well-known technologies in the field and will not be elaborated here.

[0030] Based on the labeled variables of the affected side, asymmetric pathological compensation monitoring weight variables are assigned to the corresponding anatomical skeletal key points of the test subjects in the calculation system.

[0031] The constraint initialization module establishes a memory mapping relationship between the set of anatomical skeletal key points and the types of weight variables within the computing terminal. The set of anatomical skeletal key points includes the core joints symmetrical on both sides of the human body and anatomical landmarks on the body surface. When the affected side's marker variable... After specifying the target side, the constraint initialization module divides the anatomical and skeletal key points located on the target side into pathological monitoring nodes, and divides the anatomical and skeletal key points located on the opposite side into regular tracking nodes.

[0032] For regular tracking nodes, the constraint initialization module assigns regular attitude extraction weight variables. For pathological monitoring nodes, the constraint initialization module assigns pathological compensation monitoring weight variables. In the computer memory data structure, the regular attitude extraction weight variable has a value of 0, and the pathological compensation monitoring weight variable has a value of 1. The pathological compensation monitoring weight variable serves as a trigger flag for the compensation extraction module to perform abnormal geometric relationship calculations, instructing the computing terminal to perform abnormal compensation displacement calculations or abnormal compensation angle calculations in the pathological monitoring node region. The regular attitude extraction weight variable indicates that the node coordinates are only used for the overall centroid trajectory calculation of the basic assessment module and do not participate in the penalty calculation of local compensation actions.

[0033] The system employs an asymmetric variable labeling pattern to differentiate the processing priority of bilateral limbs in the algorithm calculation. Variables are labeled with the affected side. For example, the system defines the right acromion, right elbow joint, right wrist joint, and right anterior superior iliac spine of the test subject as pathological monitoring nodes and assigns them pathological compensation monitoring weight variables. At the same time, it assigns conventional posture extraction weight variables to the corresponding skeletal key points on the left side. The asymmetric variable allocation logic transforms the physical pathological characteristics of the test subject into constraints in the algorithm space.

[0034] By integrating the affected side marker variables, the test position pattern, and the node mapping relationship including the weight variable for monitoring asymmetric pathological compensation, a test object environment configuration matrix is ​​generated.

[0035] The constraint initialization module writes the test object environment configuration matrix into the shared runtime memory area of ​​the computing terminal. The test object environment configuration matrix is ​​used as a constant parameter for cross-process calls by the data processing module, the basic evaluation module, and the compensation extraction module during runtime.

[0036] The data processing module performs the following specific workflows for 3D calibration of the test environment and extraction of key skeletal points of the test object.

[0037] Establish an absolute three-dimensional coordinate system for the test environment and calibrate the camera parameters.

[0038] Conventional monocular cameras can only acquire two-dimensional planar images, and algorithms used to calculate the forward and backward displacement of the test subject are prone to physical measurement errors. Stroke patients often exhibit forward and backward trunk swaying when maintaining balance, making accurate acquisition of depth dimension physical displacement essential for balance function assessment. The data acquisition equipment employs a visual image acquisition device with a depth sensor. Subsidiary features of visual image acquisition devices with depth sensors include infrared structured light cameras, time-of-flight cameras, or binocular stereo cameras.

[0039] The data processing module acquires depth point cloud data and color image data from the physical testing environment. It extracts a pre-defined static reference object from the physical testing environment as the origin of the spatial coordinate system. The lower-level features of the static reference object include the corner points of the calibration mat or the physical markers for securing the wheelchair. The data processing module sets the optical axis direction of the data acquisition equipment as the primary observation direction and the vertical line direction in the physical testing environment as the absolute three-dimensional coordinate system. The axis is set perpendicular to the main observation direction and... The direction of the axial plane is The axis is defined as the direction parallel to the main observation direction. The axes are used to establish an absolute three-dimensional coordinate system based on the physical environment. For the matrix operations of camera intrinsic and extrinsic parameter calibration and coordinate system transformation, those skilled in the art can use conventional computer vision calibration algorithms. The calculation process is well-known in the field and will not be elaborated upon here.

[0040] Receive continuous video stream data and extract the set of skeletal key points.

[0041] The data processing module receives continuous video stream data. This continuous video stream data includes a color image stream and a depth data stream. To ensure data consistency, the data processing module uses hardware timestamps to perform frame-time matching between the color image stream and the depth data stream, and resamples and projects the depth data stream onto the coordinate system of the color image stream using the camera's factory calibration matrix, achieving pixel-level spatial registration. The data processing module processes the continuous video stream data frame by frame, extracts the total number of frames, and defines the total number of frames as... For the first The frame image data processing module uses a human pose estimation algorithm to process the color images in the color image stream and identify the human topology of the test subject. The frame number is... The lower-level features of human pose estimation algorithms include bottom-up or top-down human topology detection networks based on convolutional neural networks.

[0042] The data processing module extracts the core anatomical nodes of the test object through a human pose estimation algorithm and generates a set of skeletal key points.

[0043] Define the set of skeletal keypoints as , Skeletal key point set In Refers to the left acromion, a set of key skeletal points. In Refers to the right acromion, a set of key skeletal points. In Refers to the left anterior superior iliac spine, a set of key skeletal points. In Refers to the right anterior superior iliac spine, a set of key skeletal points. In Refers to the left elbow joint, a set of key skeletal points. In Refers to the right elbow joint, a set of key skeletal points. In Refers to the left wrist joint, a set of key skeletal points. In Refers to the right wrist joint, a set of key skeletal points. In Refers to the left ankle joint, a set of key skeletal points. In Refers to the right ankle joint.

[0044] By fusing depth information and image information, a four-dimensional data tensor is generated.

[0045] For each extracted skeletal keypoint, the data processing module extracts its two-dimensional pixel coordinates on the color image. Based on these coordinates, the module reads the depth distance value from the corresponding depth data stream. Combining this with the camera's intrinsic parameter matrix, the module maps the two-dimensional pixel coordinates and depth distance value to physical space coordinates in an absolute three-dimensional coordinate system. In the implementation of this three-dimensional physical mapping, the system uses the camera's focal length parameters and the image's principal point coordinates, employing similar triangle geometry to backproject the pixel positions of the color image plane into three-dimensional space. This projection is then combined with the depth distance value to obtain the three-dimensional coordinates in the camera coordinate system, which are subsequently transformed to the absolute three-dimensional coordinate system using an extrinsic parameter matrix. Simultaneously, the data processing module obtains the peak value of the probability heatmap output by the human pose estimation algorithm, using this peak value as the visual confidence level of the skeletal keypoint.

[0046] Define the skeletal keypoints as numbered Among them, the key points of the skeleton are numbered. The system will... The skeletal keypoints in the frame image are numbered as follows: The data structure output of the skeletal key points is a four-dimensional data tensor. : ; In the formula, Indicates the first The skeletal keypoints in the frame image are numbered as follows: The four-dimensional data tensor corresponding to the skeletal key points, four-dimensional data tensor In The key points of the skeleton are numbered as follows The key points of the skeleton in the first The absolute three-dimensional coordinate system in the frame image Axis coordinates, four-dimensional data tensor In The key points of the skeleton are numbered as follows The key points of the skeleton in the first The absolute three-dimensional coordinate system in the frame image Axis coordinates, four-dimensional data tensor In The key points of the skeleton are numbered as follows The key points of the skeleton in the first The absolute three-dimensional coordinate system in the frame image Axis coordinates, four-dimensional data tensor In The key points of the skeleton are numbered as follows The key points of the skeleton in the first Visual confidence level corresponding to the frame image, visual confidence level Visual confidence reflects the probability that the human pose estimation algorithm determines that the pixel location corresponds to an anatomical node. The value range is [0,1], and the visual confidence score is... The closer the value is to 1, the higher the probability that the algorithm has accurately determined the coordinate point; visual confidence. The closer the value is to 0, the higher the probability that the anatomical node is occluded or the image is blurred, leading to identification errors. The data processing module will then process the total number of frames. The four-dimensional data tensor of each frame is stored in the system cache sequence as the data source for subsequent calculations.

[0047] After extracting the four-dimensional data tensor, the data processing module performs a dynamic validity mask generation operation. Stroke patients often wear loose hospital gowns during balance function tests, and the hemiplegic limbs are prone to spasticity and torsion due to abnormal muscle tone. Limb torsion and clothing occlusion can cause local anatomical nodes to be invisible in the visual image. When anatomical nodes are invisible, the three-dimensional spatial coordinates output by the human pose estimation algorithm will drift irregularly, and the visual confidence level will decrease accordingly. To prevent drifting coordinates from interfering with subsequent compensation action calculations, the data processing module uses visual confidence level to filter the spatial data.

[0048] The data processing module obtains the preset global recognition confidence threshold. The global recognition confidence threshold is defined as follows: Global identification confidence threshold The lower-level features include empirical constants derived from offline testing based on large-scale pose datasets, or adaptive baseline values ​​obtained through dynamic calibration of the test object while it maintains its initial resting pose. In the specific implementation of dynamic calibration, the data acquisition device collects data on the test object while it maintains its resting pose before the test begins. The initial video image data of the frame. The data processing module extracts the key points of each skeleton in this... The visual confidence score within each frame is calculated and its arithmetic mean is determined. The data processing module multiplies this arithmetic mean by a preset reduction factor to obtain an adaptive baseline value, which is then used as the global recognition confidence threshold. The reduction factor ranges from [0.7, 0.9]. Global identification confidence threshold. The value range is set to [0.5, 0.75].

[0049] The data processing module iterates through the four-dimensional data tensors in the system's cache sequence. For frame number... The image data processing module reads the skeletal keypoint numbers as follows: The four-dimensional data tensor corresponding to the skeletal key points The visual confidence level is then compared with the global recognition confidence threshold. Numerical comparison calculations were performed, and the key points of the skeleton were numbered according to the comparison results. The skeletal key points generate corresponding binary state variables.

[0050] The data processing module generates a binary validity mask using the following mathematical logic: ; In the formula, The key points of the skeleton are numbered as follows The key points of the skeleton in the first The corresponding binarized validity mask in the frame image, The key points of the skeleton are numbered as follows The key points of the skeleton in the first The visual confidence level corresponding to the frame image This represents the global recognition confidence threshold.

[0051] The data processing module completes the status labeling of all skeletal key points throughout the entire evaluation cycle based on the above logic. At that time, the data processing module determined the key point number of the skeleton to be... The skeletal keypoints are not occluded and their coordinates have not shifted, so the values ​​are assigned. .when At that time, the data processing module determined the key point number of the skeleton to be... The skeletal keypoints are occluded or have low coordinate confidence, so assign values ​​accordingly. Binarization validity mask This serves as the underlying logic switch for the basic assessment module and the compensation extraction module to determine the legality of data. The data processing module uses a binarized validity mask. With the corresponding four-dimensional data tensor Establish a memory mapping relationship and input it together into the subsequent coordinate interpolation and smoothing process.

[0052] After generating a binary validity mask, the data processing module performs temporal prediction and correction on occluded or misidentified skeletal key points based on the mask determination results. The limb movements of stroke patients during balance tests exhibit physical continuity; the spatial positions of human skeletal nodes do not jump instantaneously over long distances between adjacent video frames. When the visual algorithm outputs 3D data with coordinate drift due to clothing occlusion or limb twisting, mathematical derivation using motion state parameters from previous time periods can calculate and recover the reasonable physical spatial coordinates of skeletal nodes during the period of data loss.

[0053] The data processing module establishes independent temporal prediction models for each skeletal keypoint. The underlying features of these models include Kalman filter models, particle filter models, or bidirectional exponential smoothing models. In this implementation, we take the Kalman filter model for spatial coordinate interpolation and smoothing as an example. The data processing module defines a state vector within the Kalman filter model. This state vector contains six physical quantities: three spatial position variables and three motion velocity variables for the corresponding skeletal keypoint in the absolute three-dimensional coordinate system.

[0054] The data processing module reads the system cache sequence frame by frame. The data processing module reads the skeletal keypoint numbers as follows: The key points of the skeleton in the first Binarization validity mask corresponding to the frame image Binarization validity mask The value can be 1 or 0. The data processing module uses the binarized validity mask. The logical values ​​trigger different filter state machine branches.

[0055] When the binary validity mask When the value of is 1, the data processing module determines that the spatial observation data of the current frame has not drifted. The data processing module reads the first... The skeletal keypoints in the frame image are numbered as follows: The four-dimensional data tensor corresponding to the skeletal key points The data processing module will process the four-dimensional data tensor. The 3D coordinate parameters are used as the observation inputs to the Kalman filter model. The Kalman filter model combines the prediction state of the previous frame with the observation inputs of the current frame, performs weighted calculations, updates the internal covariance, and outputs the corrected 3D skeleton coordinate data.

[0056] When the binary validity mask When the value is 0, the data processing module determines that the spatial observation data of the current frame has undergone physical drift due to occlusion. The data processing module blocks the four-dimensional data tensor. The observation input path for the Kalman filter model. The Kalman filter model uses a state transition matrix to perform pure state prediction based on motion velocity variables extracted from preceding, unoccluded image frames. The predicted 3D spatial coordinates are directly output as corrected 3D skeletal coordinate data. By isolating drift observations in the underlying logic, the system prevents abnormal visual drift coordinate data from contaminating the internal state of the temporal prediction model, maintaining the physical continuity of the human posture topology.

[0057] For the parameter settings and iterative update equations of the state transition matrix, observation matrix, and covariance matrix in the Kalman filter model, those skilled in the art can perform conventional configurations based on the uniform motion model or the uniformly accelerated motion model. The matrix derivation and calculation process are well-known techniques in this field and will not be elaborated here.

[0058] After processing by the time-series prediction model, the data processing module integrates the interpolated and smoothed data of each skeletal keypoint and outputs a unified set of corrected coordinates. The corrected 3D skeletal coordinate data is defined as follows: Its mathematical structure is as follows: ; In the formula, Indicates the frame number. Indicates the key point number of the skeleton. The key points of the skeleton are numbered as follows The key points of the skeleton in the first Corrected 3D skeleton coordinate data in the frame image. The key points of the skeleton are numbered as follows The key points of the skeleton in the first The corrected absolute 3D coordinate system in the frame image axis coordinate values, The key points of the skeleton are numbered as follows The key points of the skeleton in the first The corrected absolute 3D coordinate system in the frame image axis coordinate values, The key points of the skeleton are numbered as follows The key points of the skeleton in the first The corrected absolute 3D coordinate system in the frame image Axis coordinate values.

[0059] The data processing module will count the total number of frames. The corrected 3D skeletal coordinate data of all key points of the skeleton is transmitted to the basic assessment module and the compensation extraction module, serving as the basic data source for subsequent multi-body segment dynamics calculations and spatial geometric linkage calculations.

[0060] The basic assessment module receives corrected 3D skeletal coordinate data. A stroke patient's ability to maintain posture is directly reflected in the spatial displacement of the body's overall center of gravity. The body's balance in physical space depends on the projected position of the overall center of gravity relative to the supporting surface. By simplifying the human body into multiple rigid segments connected by joints, the overall center of gravity can be considered as a spatial point representing the weighted average of the center of mass positions of each local segment according to its mass proportion. Based on multi-rigid-body kinematics theory, the basic assessment module deconstructs the test subject's human body model into multiple rigid segments and calculates the spatial coordinates of the test subject's overall center of gravity.

[0061] The basic assessment module acquires preset multi-segment dynamic parameters. These parameters include the relative mass percentages of various body segments. The sub-features of these relative mass percentages include constant values ​​derived from anthropometry standards, specifically the mass percentages of the trunk, upper arm, forearm, and lower limb segments. The basic assessment module reads the test object environment configuration matrix generated by the constraint initialization module and extracts the test posture pattern. Based on the test posture pattern, the basic assessment module determines the set of segments to be included in the calculation. In the seated test mode, the basic assessment module masks the lower limb skeletal data, and the set of segments included in the calculation includes the trunk and upper limb segments; in the bipedal standing and walking test modes, the set of segments included in the calculation includes the trunk, upper limb segments, and lower limb segments. For test modes that mask some body segments, the basic assessment module normalizes the relative mass percentages in the set of segments to be calculated. The basic assessment module calculates the sum of the original relative mass percentages of the set of segments involved in the calculation, divides the original relative mass percentage of each local segment by this sum, and obtains the normalized relative mass percentage, ensuring that the sum of the normalized relative mass percentages of all local segments involved in the calculation is equal to 1.

[0062] The basic assessment module calculates the local centroid spatial position of each local segment in the set of segments involved in the calculation. For a local segment composed of proximal and distal skeletal keypoints, the basic assessment module extracts the proximal skeletal keypoint at the [missing information - likely a specific location or point]. Corrected 3D skeleton coordinate data in frame images, and extraction of distal bone key points in the 1st frame. The corrected 3D skeletal coordinate data in the frame image. The basic evaluation module performs linear interpolation calculations on the 3D coordinates of proximal and distal skeletal keypoints based on the human segment centroid position constant, obtaining the local centroid 3D coordinates of the local segment in the current frame. The human segment centroid position constant represents the proportion of the physical distance between the local centroid and the line connecting the proximal and distal skeletal keypoints. The basic evaluation module uses this physical distance proportion parameter to determine the exact spatial location of the local centroid through point-to-point difference calculations of 3D spatial vectors. For the specific values ​​of the relative mass proportions of each segment in the multi-segment dynamic parameters and the selection of the local centroid interpolation constant, those skilled in the art can refer to standard anthropometry databases for value configuration; the parameter settings are well-known techniques in the field and will not be elaborated upon here.

[0063] The total number of local segments included in the set of segments participating in the calculation is defined as follows: Define the local segment numbering as... The numbering of local body segments Define the number as The normalized relative mass percentage of the local body segments is The basic evaluation module calculates the spatial coordinates of the overall centroid of the test object using a weighted summation algorithm. ; In the formula, Indicates the frame number. Indicates the test object is in the first... Spatial coordinates of the overall centroid in the frame image. It is a system containing a corresponding absolute three-dimensional coordinate system. axis coordinate values axis coordinate values ​​and A three-dimensional spatial vector of axis coordinates. This represents the total number of local segments contained in the set of segments participating in the calculation. Indicates the number of the local body segment. Indicates the number is The relative mass percentage of local body segments after normalization. Indicates the number is The local segment in the first The corresponding local centroid 3D coordinates in the frame image. It is a system containing a corresponding absolute three-dimensional coordinate system. axis coordinate values axis coordinate values ​​and A three-dimensional spatial vector of axis coordinates.

[0064] The basic evaluation module traverses the total number of frames. For each image frame within the evaluation period, the spatial coordinates of the overall center of gravity are calculated frame by frame to generate a dynamic three-dimensional center of gravity trajectory sequence. This dynamic three-dimensional center of gravity trajectory sequence records the physical spatial swaying state of the test object during the process of maintaining balance. This sequence serves as the direct data source for subsequent calculations of the basic physical balance score.

[0065] After generating a dynamic three-dimensional center of gravity trajectory sequence, the basic assessment module extracts the spatial sway characteristic values ​​of the center of gravity within the assessment period and calculates the basic physical balance score without compensatory movement intervention. The process of maintaining balance in the human body is a dynamic process of keeping the body's center of gravity within the support plane through the regulation of the neuromuscular system. Stroke patients, due to impaired motor control pathways, have a reduced ability to maintain center of gravity stability, resulting in continuous displacement of the overall center of gravity's spatial coordinates in three-dimensional physical space. The spatial sway characteristic values ​​of the center of gravity quantify the test subject's ability to control posture at the physical trajectory level.

[0066] The basic evaluation module extracts multi-dimensional quantitative indicators reflecting attitude stability from the dynamic 3D center-of-gravity trajectory sequence. The sub-features of the center-of-gravity spatial oscillation feature value include the total length of the center-of-gravity oscillation path, the area of ​​the center-of-gravity oscillation envelope, and the root mean square of the center-of-gravity velocity. In the specific implementation, calculating the total length of the center-of-gravity oscillation path is taken as an example. The basic evaluation module reads the spatial coordinates of the overall center-of-gravity in each frame within the total number of frames. The spatial coordinates of the overall center-of-gravity include the corresponding absolute 3D coordinate system... axis coordinate values axis coordinate values ​​and Axis coordinate values. The basic evaluation module calculates the Euclidean distance in three-dimensional physical space between the spatial coordinates of the overall center of gravity between two adjacent image frames, and sums up the Euclidean distances of all adjacent image frames to obtain the total length of the center of gravity swing path within the evaluation period.

[0067] The specific mathematical logic for calculating the total length of the center of gravity swing path in the basic assessment module is as follows: ; In the formula, This indicates the total length of the center of gravity swing path within the evaluation period. Indicates the total number of frames. Indicates the frame number. Indicates the first The spatial coordinates of the overall centroid in the frame image correspond to the absolute three-dimensional coordinate system. axis coordinate values, Indicates the first The spatial coordinates of the overall centroid in the frame image correspond to the absolute three-dimensional coordinate system. axis coordinate values, Indicates the first The spatial coordinates of the overall centroid in the frame image correspond to the absolute three-dimensional coordinate system. axis coordinate values, Indicates the first The spatial coordinates of the overall centroid in the frame image correspond to the absolute three-dimensional coordinate system. axis coordinate values, Indicates the first The spatial coordinates of the overall centroid in the frame image correspond to the absolute three-dimensional coordinate system. axis coordinate values, Indicates the first The spatial coordinates of the overall centroid in the frame image correspond to the absolute three-dimensional coordinate system. Axis coordinate values.

[0068] After obtaining the total length of the center of gravity swing path, the basic assessment module maps this total length to a basic physical balance score. The basic assessment module uses a nonlinear decay model for score mapping. The module presets a value of one hundred as the ideal benchmark for a perfect score. The module obtains the statistical mean of the total length of the center of gravity swing path from healthy individuals of the same age as a reference benchmark. The module compares the total length of the center of gravity swing path within the assessment period with the reference benchmark. When the total length of the center of gravity swing path within the assessment period is less than or equal to the reference benchmark, the module determines that the test subject's balance ability is within the normal range and directly assigns the basic physical balance score as the benchmark for a perfect score. When the total length of the center of gravity swing path within the assessment period is greater than the reference benchmark, the module calculates the difference between the total length of the center of gravity swing path within the assessment period and the reference benchmark. Using an exponential decay function combined with a preset decay constant, the module subtracts the benchmark value from the calculated difference to obtain the basic physical balance score.

[0069] The logic for calculating the fundamental physical equilibrium fraction using the exponential decay function in the basic assessment module is implemented through the following core formula: ; In the formula, Represents the basic physical equilibrium fraction. This indicates the total length of the center of gravity swing path during the evaluation period. Indicates the reference base value. This represents the preset attenuation constant. The descent gradient of the fraction as the pendulum length increases is determined by the decay constant. The value range is [0.01, 0.05], and the attenuation constant is... The values ​​were determined by collecting actual center of gravity swing path data from healthy individuals and patients with varying degrees of balance disorders, and then using nonlinear least squares method to fit the parameters of the attenuation model.

[0070] To prevent abnormal calculation results from exceeding limits, the basic assessment module performs boundary truncation on the basic physical equilibrium scores, limiting the highest score to 100 and the lowest score to zero. The acquisition of the statistical mean of tests from healthy individuals of the same age and the construction of the norm reference database can be achieved using conventional medical statistical methods by those skilled in the art. The data distribution patterns and statistical analysis processes are well-known techniques in this field and will not be elaborated upon here.

[0071] The basic assessment module transmits the calculated basic physical balance score to the calibration output module. This basic physical balance score reflects the test subject's ability to maintain posture through overall center of mass shift. At this stage, it does not eliminate scoring interference caused by abnormal compensatory movements such as asymmetric support using the healthy side limbs. This parameter will serve as the basis for subsequent compensation penalty algorithms.

[0072] The compensation extraction module receives corrected three-dimensional skeletal coordinate data transmitted from the basic assessment module or data processing module. When stroke patients maintain sitting or standing balance, due to decreased strength in the core muscles on the affected side, they are unable to maintain a relatively neutral position between the pelvis and ribcage. Influenced by gravity and the pull of the muscles on the healthy side, the trunk often tilts towards the healthy side. This lateral trunk tilt alters the spatial distribution projection of the body's physical center of gravity, representing a typical abnormal movement strategy that masks the true balance control deficiency in the hemiplegic limb. The compensation extraction module quantifies the degree of lateral trunk tilt compensation based on the geometric linkages of key points in the human skeleton in three-dimensional space.

[0073] The compensatory extraction module extracts the left acromion at the first The corrected 3D skeletal coordinate data in the frame image, and the right acromion in the [missing information] . The corrected 3D skeleton coordinate data in the frame image. The compensation extraction module performs an arithmetic mean calculation on these two coordinate points to obtain the first... The three-dimensional spatial coordinates of the shoulder center in the frame image. The compensation extraction module extracts the left anterior superior iliac spine at the [missing information - likely a specific point or location]. The corrected 3D skeletal coordinate data in the frame image, and the right anterior superior iliac spine in the first... The corrected 3D skeleton coordinate data in the frame image. The compensation extraction module performs an arithmetic mean calculation on these two coordinate points to obtain the first... The three-dimensional spatial coordinates of the pelvic center in the frame image.

[0074] The compensation withdrawal module is based on the first Using the three-dimensional spatial coordinates of the pelvic center in the frame image as the starting point of the vector, and with the first... Using the three-dimensional spatial coordinates of the shoulder center in the frame image as the endpoint of the vector, construct the first... The torso spatial vector in the frame image. The compensation extraction module obtains the absolute three-dimensional coordinate system. Axis direction vector. Since an absolute three-dimensional coordinate system was already established during the initial environmental calibration... The axis is parallel to the vertical line direction in the physical testing environment, in an absolute three-dimensional coordinate system. The axial direction vector is the reference point that represents the vertical downward force of gravity.

[0075] The compensation extraction module calculates the first The torso spatial vector in the frame image and the absolute three-dimensional coordinate system The angle between the axial direction vectors is calculated. The core formula for the torso lateral slip angle in a frame image is as follows: ; In the formula, Indicates the frame number. Indicates the first The torso side profile angle in the frame image. Indicates the first The torso spatial vector in the frame image. Representing an absolute three-dimensional coordinate system Axial direction vector.

[0076] After obtaining the trunk lateral tilt angle for each frame, the compensation extraction module converts it into a trunk lateral tilt compensation index through piecewise mapping operations. The compensation extraction module obtains preset physiological tolerance angles and maximum penalty angles. The physiological tolerance angle is set based on the inherent slight trunk sway amplitude of healthy individuals maintaining resting balance, and its value ranges from [3°, 5°]. The maximum penalty angle represents the trunk tilt threshold when the human body reaches its compensation limit, and its value ranges from [15°, 20°].

[0077] The compensation extraction module uses a piecewise mapping function to calculate the first... The mathematical logic behind the trunk lateral deviation compensation index in frame images is as follows: ; In the formula, Indicates the frame number. Indicates the first Trunk lateral deviation compensation index in frame images Indicates the first The torso side profile angle in the frame image. Indicates the physiological tolerance level. Indicates the angle of maximum penalty.

[0078] When the When the torso lateral tilt angle in the frame image is less than or equal to the physiological tolerance angle, the compensation extraction module determines that the angle is within the normal physiological posture sway range of the human body, and then... The trunk lateral deviation compensation index in the frame image is assigned a value of zero. When the torso lateral slip angle in the frame image is greater than or equal to the maximum penalty angle, the compensation extraction module determines that the torso lateral slip compensation has reached its limit, and then... The trunk lateral deviation compensation index in the frame image is assigned a value of one. When the... When the torso lateral slip angle in the frame image is between the physiological tolerance angle and the maximum penalty angle, the compensation extraction module calculates the first... The difference between the torso lateral slip angle and the physiological tolerance angle in the frame image is used as the numerator, and the difference between the maximum penalty angle and the physiological tolerance angle is used as the denominator. The first value in the interval between zero and one is obtained through ratio calculation. The trunk lateral deviation compensation index in the frame image.

[0079] The compensation extraction module traverses the total number of frames. For all image frames within the evaluation period, the above vector angle calculation and segmented mapping operations are repeated. The compensation extraction module summarizes the trunk lateral deviation compensation index corresponding to all image frames within the evaluation period, calculates its arithmetic mean, and obtains the global trunk compensation coefficient. The formula for calculating the global trunk compensation coefficient is as follows: ; In the formula, This represents the global trunk compensation coefficient. Indicates the total number of frames. Indicates the frame number. Indicates the first The trunk lateral tilt compensation index in the frame image, and the global trunk compensation coefficient, range from [0,1]. The larger the value of the global trunk compensation coefficient, the higher the degree of dependence of the test subject on trunk tilt to maintain balance throughout the balance assessment period. This global trunk compensation coefficient is stored in the system memory as an input parameter for subsequent multidimensional compensation fusion and basic score penalty deduction.

[0080] After acquiring the corrected 3D skeletal coordinate data, the compensation extraction module extracts the compensatory features of pelvic elevation on the affected side of the test subject. Stroke patients often experience difficulty maintaining balance or performing lower limb movements due to decreased muscle strength in the flexor muscles of the hemiplegic lower limb, making it difficult to achieve normal hip and knee flexion. This leads to difficulty in clearing the ground from the affected side during movement or posture shifts. To prevent the affected lower limb from falling or to maintain the body's center of gravity, patients often excessively contract the quadratus lumborum muscle on the hemiplegic side, abnormally elevating the pelvis upwards in compensation. This pelvic elevation is an abnormal coordinated movement strategy that disrupts the horizontal stability of the pelvis in the coronal plane. The compensation extraction module numerically quantifies the degree of pelvic elevation compensation based on the absolute height changes of local pelvic anatomical nodes in physical 3D space.

[0081] The compensation extraction module reads the test object environment configuration matrix generated by the constraint initialization module and extracts the hemiplegic side identifier of the test object. The hemiplegic side identifier records the specific pathological lateral orientation of the test object, including left-sided or right-sided hemiplegia. Based on the hemiplegic side identifier, the compensation extraction module extracts the first... The three-dimensional spatial coordinates of the affected anterior superior iliac spine and the unaffected anterior superior iliac spine in the frame image. Absolute three-dimensional coordinate system. The axis is parallel to the vertical line in the physical testing environment. The compensation extraction module reads the absolute three-dimensional coordinate system corresponding to the anterior superior iliac spine on the affected side. The axis coordinate values, and the absolute three-dimensional coordinate system corresponding to the contralateral anterior superior iliac spine. Axis coordinate values. In setting an absolute three-dimensional coordinate system... With the positive axis direction vertically upward, the compensatory extraction module will extract the anterior superior iliac spine on the affected side. The axial coordinate values ​​and the contralateral anterior superior iliac spine The difference between the axis coordinate values ​​is calculated, that is, using the anterior superior iliac spine of the affected side. The axis coordinate value minus the healthy side anterior superior iliac spine obtain the axis coordinate value, and get the first axis coordinate value. The height difference in the coronal plane of the pelvis in the frame image. When the calculated difference is less than zero, the compensation extraction module determines that no upward compensation has occurred on the affected side, and then... The height difference of the coronal plane of the pelvis in the frame image is assigned a value of zero.

[0082] In obtaining the After determining the coronal height difference of the pelvis in the frame image, the compensation extraction module converts it into a pelvic elevation compensation index through piecewise mapping operations. The compensation extraction module obtains preset physiological tolerance height differences and maximum penalty height differences. The values ​​of the physiological tolerance height difference and maximum penalty height difference are determined through offline statistical analysis of gait and balance databases of clinically healthy individuals and stroke patients. In normal individuals, the pelvis exhibits slight physiological height fluctuations in the coronal plane when maintaining dynamic balance. Based on human physiological structure and posture control characteristics, the physiological tolerance height difference ranges from [1.0, 2.0] cm. The maximum penalty height difference represents the spatial height difference threshold when abnormal compensation reaches its limit, and its value ranges from [5.0, 8.0] cm.

[0083] The compensation extraction module uses a piecewise mapping function to calculate the first... The pelvic elevation compensation index in a frame image is calculated using the following core formula: ; In the formula, Indicates the frame number. Indicates the first The pelvic elevation compensation index in the frame image. Indicates the first The height difference in the coronal plane of the pelvis in the frame image. This indicates a high degree of physiological tolerance. This indicates the maximum penalty height difference.

[0084] When the When the height difference in the coronal plane of the pelvis in the frame image is less than or equal to the physiological tolerance height difference, the compensation extraction module determines that the height difference is within the normal physiological postural swaying range, and then... The pelvic elevation compensation index in the frame image is assigned a value of zero. When the coronal height difference of the pelvis in the frame image is greater than or equal to the maximum penalty height difference, the compensation extraction module determines that the pelvic elevation compensation has reached its limit, and then... The pelvic elevation compensation index in the frame image is assigned a value of one. When the first... When the height difference of the pelvic coronal plane in the frame image is between the physiological tolerance height difference and the maximum penalty height difference, the compensation extraction module calculates the first... The difference between the coronal plane height difference of the pelvis and the physiological tolerance height difference in the frame image is used as the numerator, and the difference between the maximum penalty height difference and the physiological tolerance height difference is used as the denominator. The first value in the interval between zero and one is obtained through ratio calculation. The pelvic elevation compensation index in the frame image.

[0085] The compensation extraction module traverses the total number of frames. For all image frames within the evaluation period, the coordinate reading, height difference comparison, and segmented mapping operations described above are repeated. The compensation extraction module summarizes the pelvic elevation compensation index corresponding to all image frames within the evaluation period. The compensation extraction module obtains the global pelvic compensation coefficient by calculating the arithmetic mean, and its mathematical formula is as follows: ; In the formula, This represents the global pelvic compensation coefficient. Indicates the total number of frames. Indicates the frame number. Indicates the first The pelvic elevation compensation index in the frame image.

[0086] The global pelvic compensation coefficient ranges from [0,1]. A higher global pelvic compensation coefficient indicates a higher frequency and amplitude of abnormal pelvic elevation coordination movements on the affected side during the entire balance assessment period. This global pelvic compensation coefficient is stored in system memory and, together with the previously obtained global trunk compensation coefficient, serves as input data for the subsequent multidimensional compensation fusion engine.

[0087] After acquiring the corrected 3D skeletal coordinate data, the compensation extraction module extracts the compensatory features of flexor spasticity in the affected upper limb of the test subject. Following upper motor neuron damage in stroke patients, as muscle tone abnormally evolves, the affected upper limb typically exhibits the Brunnstrom recovery phase flexor synergistic movement pattern. During body balance testing, influenced by gravitational counterforce and nerve tension, the affected upper limb displays a contraction posture with shoulder adduction, elbow flexion, and forearm pronation. This abnormal flexion of the upper limb causes the local center of mass of the affected upper limb to move closer to the trunk center, altering the body's original natural biomechanical distribution, and is a compensatory manifestation of decreased control ability after central nervous system damage. The spatial flexion angle of the elbow joint reflects the state of flexor muscle tone in the affected upper limb. The compensation extraction module quantifies the degree of compensation for flexor spasticity in the affected upper limb based on spatial vector angle calculations.

[0088] The compensation extraction module reads the test object environment configuration matrix generated by the constraint initialization module and extracts the hemiplegic side identifier of the test object. Based on the hemiplegic side identifier, the compensation extraction module extracts the first... The three-dimensional spatial coordinates of the affected acromion, elbow joint, and wrist joint in the frame image. The compensation extraction module uses the first... Using the three-dimensional spatial coordinates of the affected elbow joint in the frame image as the starting point of the spatial vector, and taking the first... The three-dimensional spatial coordinates of the affected acromion in the frame image are used as the endpoints of the spatial vector to construct the first... The spatial vector of the upper arm in the frame image. The compensation extraction module uses the first... Using the three-dimensional spatial coordinates of the affected elbow joint in the frame image as the starting point of the spatial vector, and taking the first... The three-dimensional spatial coordinates of the affected wrist joint in the frame image are used as the endpoints of the spatial vector to construct the first... Forearm spatial vector in the frame image.

[0089] The compensation extraction module calculates the first... The upper arm spatial vector in the frame image and the first The dot product of the forearm spatial vectors in the frame image is used to obtain the spatial angle between the two vectors, which serves as a physical variable reflecting the degree of elbow flexion. The calculation of the... The core formula for the spatial angle of the affected elbow joint in a frame image is as follows: ; In the formula, Indicates the frame number. Indicates the first The spatial angle of the affected elbow joint in the frame image. Indicates the first The upper arm spatial vector in the frame image. Indicates the first Forearm spatial vector in the frame image.

[0090] In order to arrive at the first After determining the spatial angle of the affected elbow joint in the frame image, the compensation extraction module transforms it into the first frame through a piecewise mapping operation. The upper limb flexor spasticity compensation index in the frame image. The compensation extraction module obtains the preset physiological tolerance extension angle and maximum penalty flexion angle. For the specific numerical mapping relationship between the physiological tolerance extension angle and the maximum penalty flexion angle, those skilled in the art can refer to the kinematic data distribution of the clinical spasticity assessment criteria for routine configuration; the parameter calibration process is a well-known technique in the field and will not be elaborated here. The physiological tolerance extension angle represents the lower limit of the angle when the affected upper limb is in a relatively relaxed state, and its value ranges from [130°, 150°]. The maximum penalty flexion angle represents the upper limit of the angle when the affected upper limb experiences severe flexor spasticity, and its value ranges from [60°, 90°].

[0091] The compensation extraction module uses a piecewise mapping function to calculate the first... The upper limb flexor spasticity compensation index in frame images, since a smaller elbow joint spatial angle indicates more severe flexion spasticity, is calculated using the following formula: ; In the formula, Indicates the frame number. Indicates the first Compensation index for upper limb flexor spasticity in frame images Indicates the first The spatial angle of the affected elbow joint in the frame image. Indicates the physiological tolerance stretch angle. This indicates the maximum penalty bend angle.

[0092] When the When the spatial angle of the affected elbow joint in the frame image is greater than or equal to the physiological tolerance extension angle, the compensation extraction module determines that the affected upper limb is in a normal relaxed extension state and no flexion spasticity has occurred. The upper limb flexor spasticity compensation index in the frame image is assigned a value of zero. When the... When the spatial angle of the affected elbow joint in the frame image is less than or equal to the maximum penalized flexion angle, the compensation extraction module determines that the flexor spasticity of the affected upper limb has reached its limit, and then... The upper limb flexor spasticity compensation index in the frame image is assigned a value of one. When the first... When the spatial angle of the affected elbow joint in the frame image is between the maximum penalized flexion angle and the physiological tolerance extension angle, the compensation extraction module calculates the physiological tolerance extension angle and the first... The difference in the spatial angle of the affected elbow joint in each frame image is used as the numerator, and the difference between the physiologically tolerable extension angle and the maximum penalized flexion angle is used as the denominator. The first value in the interval between zero and one is obtained through ratio calculation. Compensation index for upper limb flexor spasticity in frame images.

[0093] The compensation extraction module traverses the total number of frames. For all image frames within the evaluation period, the above vector construction, angle calculation, and reverse segmentation mapping operations are repeated. The compensation extraction module summarizes the upper limb flexor spasticity compensation index corresponding to all image frames within the evaluation period. The compensation extraction module obtains the global upper limb compensation coefficient by calculating the arithmetic mean, and its mathematical formula is as follows: ; In the formula, This represents the global upper limb compensation coefficient. Indicates the total number of frames. Indicates the frame number. Indicates the first Compensation index for upper limb flexor spasticity in frame images.

[0094] The global upper limb compensation coefficient ranges from [0,1]. A higher global upper limb compensation coefficient indicates a higher frequency and amplitude of abnormal synergistic contraction of the affected upper limb flexor muscles during the entire balance assessment period. This global upper limb compensation coefficient is stored in the system memory and, together with the previously obtained global trunk compensation coefficient and global pelvic compensation coefficient, serves as the basic feature input for the subsequent multidimensional compensation fusion engine.

[0095] After calculating the compensation coefficients for each dimension, the compensation extraction module performs a comprehensive dynamic penalty coefficient calculation based on the fusion mask state. When stroke patients maintain balance in different positions, the available compensation strategies are influenced by physical constraints. For example, in the seated test mode, the subject's pelvis is fixed by the seat support surface, the ischial tuberosities bear gravity, and the pelvis's freedom of movement in the coronal plane is limited. In this situation, quadratus lumborum muscle contraction cannot cause overall pelvic elevation, and the pelvic elevation compensation mechanism cannot occur. If all dimensions of compensation features are directly fused with fixed weights, the zero values ​​of inapplicable indicators will lower the overall penalty intensity, causing the quantitative results to deviate from the true compensation state. The compensation extraction module introduces a masking mechanism to shield inapplicable compensation features, ensuring the rigor of the evaluation results.

[0096] The compensation extraction module reads the test object environment configuration matrix generated by the constraint initialization module and extracts the test posture mode corresponding to the current test task. Based on the test posture mode, the compensation extraction module generates an enable mask for each compensation feature. The enable mask uses Boolean logic; a value of one indicates that the compensation feature participates in the fusion calculation of the current test mode, while a value of zero indicates that the compensation feature is masked by the system in the current test mode.

[0097] The compensation extraction module configures trunk, pelvic, and upper limb masks. In both bipedal standing and walking test modes, trunk, pelvic, and upper limb compensations all meet the anatomically required conditions for occurrence, and the compensation extraction module assigns a value of one to all three masks. In the seated test mode, the compensation extraction module assigns a value of one to the trunk and upper limb masks, and a value of zero to the pelvic mask.

[0098] The compensation extraction module acquires the preset basic weight parameters for each compensation feature. These basic weight parameters reflect the relative severity of the interference of abnormal coordinated movements of different body parts on the overall balance state. From a biomechanical perspective, trunk lateral deviation directly alters the spatial projection position of the overall center of gravity, thus masking the true balance ability to the greatest extent; pelvic movement is next; and upper limb spasticity, being a local limb center of gravity shift, has a relatively smaller interference level. The basic weight parameters include trunk weight, pelvic weight, and upper limb weight. For the specific values ​​of these basic weight parameters, those skilled in the art can refer to the evaluation standards of rehabilitation medicine scales for routine settings. The numerical allocation rules are well-known in the field and will not be elaborated here. Under normal configuration, the trunk weight ranges from [0.4, 0.6], the pelvic weight ranges from [0.2, 0.4], and the upper limb weight ranges from [0.1, 0.3]. In the preset configuration, the sum of the trunk weight, pelvic weight, and upper limb weight is usually set to one.

[0099] The compensation extraction module dynamically filters and masks the basic weight parameters using an enable mask, and normalizes the weights involved in the calculation. The module then performs a weighted summation of the global trunk compensation coefficient, global pelvic compensation coefficient, and global upper limb compensation coefficient according to their respective enable masks and basic weight parameters to calculate the comprehensive dynamic penalty coefficient. The specific formula for calculating the comprehensive dynamic penalty coefficient is as follows: ; In the formula, This represents the comprehensive dynamic penalty coefficient. Indicates the weight of the torso. Indicates the torso mask. This represents the global trunk compensation coefficient. Indicates pelvic weight. Indicates pelvic mask, This represents the global pelvic compensation coefficient. Indicates upper limb weight. Indicates an upper limb mask. This represents the global upper limb compensation coefficient. The denominator of the formula constitutes the sum of the actual effective weights under the current test mode, which is used to mathematically normalize the compensation terms involved in the calculation.

[0100] To prevent the program from running out of bounds due to a zero denominator in case of abnormal system configuration, the compensation extraction module checks the denominator value before executing the above formula calculation. When the denominator is equal to zero, the compensation extraction module determines that there is no applicable compensation assessment item and directly assigns the comprehensive dynamic penalty coefficient to zero.

[0101] The calculated comprehensive dynamic penalty coefficient ranges from [0,1]. This coefficient quantifies the overall severity of all abnormal synergistic movements of the hemiplegic side invoked by the test subject to maintain body balance under corresponding postural constraints. The comprehensive dynamic penalty coefficient, as the data output of the compensation extraction module, is transmitted to the calibration output module. This coefficient serves as a downward reduction penalty scaling factor, used to numerically correct the baseline data output by the baseline assessment module, thereby eliminating errors in the baseline data related to inaccurate balance ability caused by abnormal compensatory movements.

[0102] The calibration output module receives the baseline physical balance score transmitted from the baseline assessment module and the comprehensive dynamic penalty coefficient transmitted from the compensation extraction module. In clinical motor function assessment, stroke patients often rely on the unaffected limbs for support or activate abnormal synergistic movements on the affected side to prevent falls. This strategy makes the overall center of gravity appear stable in physical space. Although compensatory behavior temporarily maintains the body posture at the level of physical center of gravity distribution, at the neurophysiological level, this abnormal force pattern and synergistic movement hinder the reconstruction of normal motor programs on the affected side. The baseline physical balance score obtained solely from the trajectory of the center of gravity swing cannot reflect the patient's true central nervous system control deficit. The calibration output module introduces a pathological feature penalty mechanism, using the comprehensive dynamic penalty coefficient to reduce the baseline physical balance score, restoring the test subject's true level of motor control.

[0103] The calibration output module obtains the preset maximum penalty scaling factor. The maximum penalty scaling factor determines the maximum deduction limit of the system from the basic physical balance score under maximum compensation. The setting of the maximum penalty scaling factor needs to comprehensively consider the interference weight of different compensatory actions on the score. Its value is obtained by collecting clinical test data of stroke patients and comparing it with the error of expert manual scoring through offline statistical regression. The value range of the maximum penalty scaling factor is [0.2, 0.4].

[0104] The calibration output module multiplies the comprehensive dynamic penalty coefficient and the maximum penalty ratio factor to obtain the actual deduction ratio under the current test state. The calibration output module subtracts the actual deduction ratio from a constant to obtain the calibration discount coefficient. The calibration output module multiplies the basic physical balance score by the calibration discount coefficient to calculate the true balance assessment score. The system sets the range of the basic physical balance score to the real number interval [0, 100].

[0105] The core formula for calculating the true balanced assessment score is as follows: ; In the formula, Indicates the true balanced assessment score. Represents the basic physical equilibrium fraction. This represents the maximum penalty ratio factor. This represents the comprehensive dynamic penalty coefficient.

[0106] After calculating the true balance assessment score, the calibration output module performs a rounding operation on the score, removing the decimal part to obtain an integer result. To ensure the standardization of the assessment results, the calibration output module performs amplitude truncation on the rounded data. If the calculated value is greater than one hundred, the calibration output module assigns it the value of one hundred; if the calculated value is less than zero, the calibration output module assigns it the value of zero. Through the above amplitude truncation, the range of the true balance assessment score is strictly limited to a closed interval between zero and one hundred. When the comprehensive dynamic penalty coefficient is equal to zero, it indicates that the test subject did not exhibit abnormal synergistic compensatory behavior on the hemiplegic side during the balance maintenance process, and the true balance assessment score is equal to the basic physical balance score. As the comprehensive dynamic penalty coefficient increases, the calibration discount coefficient decreases, and the true balance assessment score decreases relative to the basic physical balance score. This dynamic deduction logic based on the degree of compensation eliminates the interference caused by abnormal posture and asymmetrical force application strategies on the system assessment results.

[0107] After calculating the true balance assessment score, the calibration output module performs a clinical grading mapping operation. Traditional clinical balance assessments typically rely on manual observation. The calibration output module, through preset segmented mapping rules, converts the quantified true balance assessment score into a balance impairment level that aligns with clinical cognitive habits. This numerical-to-grading mapping combines the data distribution characteristics of the physical test with the assessment criteria in the rehabilitation pathway, providing medical personnel with a reference for determining the patient's current neurological recovery stage.

[0108] The calibration output module acquires preset severe impairment thresholds and moderate impairment thresholds. Regarding the specific values ​​of the severe and moderate impairment thresholds, those skilled in the art can refer to the scoring distribution patterns of standardized clinical scales such as the Berg Balanced Scale for routine configuration; the calibration methods are well-known techniques in this field and will not be elaborated upon here. In the preset configuration, the severe impairment threshold ranges from [35, 45], and the moderate impairment threshold ranges from [75, 85].

[0109] The calibration output module compares the actual balance assessment score with a preset threshold and outputs the corresponding clinical balance grading result. The clinical grading mapping logic is implemented through the following piecewise function: ; In the formula, Indicates the clinical balance grading results. Indicates the true balanced assessment score. Indicates the level of severe balance disorder. Indicates the threshold for severe disability. Indicates a moderate level of balance impairment. Indicates the threshold for moderate obstacle. This indicates a mild balance disorder or a normal level.

[0110] When the actual balance assessment score is less than the severe obstacle threshold, the calibration output module outputs a severe balance obstacle level. Test subjects at this level have poor center of gravity control and are at risk of falling; the system simultaneously outputs a fall prevention warning. When the actual balance assessment score is greater than or equal to the severe obstacle threshold but less than the moderate obstacle threshold, the calibration output module outputs a moderate balance obstacle level. When the actual balance assessment score is greater than or equal to the moderate obstacle threshold, the calibration output module outputs a mild balance obstacle or normal level.

[0111] In addition to outputting the overall score and grading, the calibration output module further performs posture deconstruction feature label output. Balance deficits in stroke patients are often accompanied by compensatory behaviors in different anatomical locations. Simply providing a comprehensive assessment score cannot pinpoint specific motor control deficit nodes, hindering the development of targeted rehabilitation prescriptions. The calibration output module utilizes the global compensation coefficients of each single dimension generated by the compensation extraction module to generate specific pathological feature labels. By comparing the global compensation coefficients with the label trigger threshold, the system discretizes continuous quantitative data into concrete medical feature labels.

[0112] The calibration output module acquires a preset tag trigger threshold. This tag trigger threshold is used to determine whether a certain abnormal coordinated movement has reached a level requiring medical intervention. The value of the tag trigger threshold is determined by collecting a large number of compensation coefficient samples from stroke patients and combining them with clinical diagnostic data from rehabilitation physicians, using receiver operating characteristic (ROC) curve analysis. The range of the tag trigger threshold is [0.4, 0.6].

[0113] The calibration output module reads the global trunk compensation coefficient, global pelvic compensation coefficient, and global upper limb compensation coefficient from the system memory. It then compares the global trunk compensation coefficient with the tag trigger threshold. When the global trunk compensation coefficient is greater than or equal to the tag trigger threshold, the calibration output module generates a trunk lateral deviation compensation tag. This tag reflects the asymmetry in core muscle strength on both sides of the test subject when maintaining balance, indicating a lack of trunk control.

[0114] The calibration output module compares the global pelvic compensation coefficient with the tag trigger threshold. When the global pelvic compensation coefficient is greater than or equal to the tag trigger threshold, the calibration output module generates a pelvic elevation compensation tag. The pelvic elevation compensation tag reflects that the affected lower limb of the test subject is in an abnormal flexor synergistic movement phase, with compensatory overcontraction of the quadratus lumborum muscle.

[0115] The calibration output module compares the global upper limb compensation coefficient with the tag trigger threshold. When the global upper limb compensation coefficient is greater than or equal to the tag trigger threshold, the calibration output module generates an upper limb flexor spasticity tag. The upper limb flexor spasticity tag reflects abnormally increased local muscle tone after damage to the central nervous system in the test subject.

[0116] The calibration output module packages the actual balance assessment score, clinical balance grading results, and all triggered posture deconstruction feature labels into a comprehensive balance assessment report. This comprehensive balance assessment report is output to an external display terminal or stored in a medical database, transforming 3D skeletal kinematic data into medical evaluation indicators, thus realizing a closed-loop business process from physical motion capture and compensatory feature quantification to the output of multidimensional clinical indicators.

[0117] Specific application examples: Experimental setup and data acquisition: To verify the feasibility of this assessment system and method, a clinical comparative experiment was conducted in this embodiment. A total of 60 subjects were recruited, including 30 stroke hemiplegic patients at different stages of rehabilitation (experimental group, including mild, moderate, and severe motor dysfunction) and 30 age- and body-matched healthy subjects (control group). Data acquisition was performed using a binocular stereo camera with a depth sensor, acquiring continuous video stream data at a sampling rate of 30 frames per second (FPS). All subjects completed a 60-second static and dynamic balance test in a bipedal standing test mode. As a ground truth, two senior rehabilitation physicians independently assessed all patients using the Berg Balance Scale (BBS), and the BBS scores (out of 56) were proportionally mapped to a 0-100 percentage point range, defined as the expert clinical score.

[0118] Experiment 1: The actual operation test of the time series prediction model for coordinate correction was conducted on the data processing module. This experiment verified the binarization mask. The effect of Kalman filter interpolation smoothing was compared. In the test, the camera lens was randomly occluded for varying durations (simulating real-world occlusion scenarios such as patient gowns or limb twisting). Experimental results showed that when visual confidence... Below the global recognition confidence threshold Furthermore, when the skeletal spatial coordinates output by traditional pose recognition algorithms show a shift (error exceeding 15cm), the data processing module of this invention determines... The input of abnormal observations was blocked.

[0119] Based on state transition prediction using Kalman filtering, the corrected 3D skeleton coordinate data During the occlusion period, the mean absolute error (MAE) between the actual physical trajectory and the actual trajectory remained within 2.1 cm, ensuring the continuity of subsequent centroid trajectory calculations in three-dimensional physical space and avoiding out-of-bounds calculations of the basic physical equilibrium fraction due to coordinate drift.

[0120] Experiment 2: Validation of the accuracy of assessment scores based on compensatory penalties (corresponding appendix) Figure 3 This experiment compared the baseline physical equilibrium fraction without compensatory intervention. The final output of this invention is the true balance assessment score. (and expert clinical scores).

[0121] like Figure 3 As shown, the experiment sorted 30 patients according to the severity of their disease from severe to mild (horizontal axis 1 to 30). Traditional algorithms (relying solely on the total length of the path of the center of gravity swing)... In the attenuation model, some patients with moderate to severe stroke have higher baseline physical balance scores because they have compensatory movements such as lateral trunk tilting dependent on the healthy side for support and pelvic elevation, which reduces the sway of the center of gravity in space. , manifested as Figure 3 The basic physics equilibrium score deviates from the expert score on the left side of the curve (severe range). This invention introduces a comprehensive dynamic penalty coefficient. According to the formula The score is reduced.

[0122] from Figure 3 As can be seen, after the penalty was applied, the true balance assessment score (dark gray square broken line) representing the method of the present invention remained consistent with the data trend of the expert score throughout the entire patient interval.

[0123] Statistical results show: Traditional basic physics equilibrium fraction The Pearson correlation coefficient between the clinical scores and expert assessments was r=0.68; After calibration using the compensatory penalty of this invention, the true balance assessment score is... The Pearson correlation coefficient with expert clinical scores improved to r=0.94.

[0124] This demonstrates that the present invention reduces the scoring bias caused by abnormal coordinated movements, making the assessment results closer to the objective motor control state after central nervous system damage.

[0125] Experiment 3: Distribution verification of spatial compensation features of multi-body segments (corresponding appendix) Figure 4 To verify the rationality of the quantitative indicators of the compensation extraction module, the global trunk compensation coefficient of 30 stroke patients was extracted in the experiment. Global pelvic compensation coefficient and global upper limb compensation coefficient Based on the clinical grading mapping results (severe, moderate, and mild), the compensatory indices of these three dimensions were grouped and statistically averaged. The results are as follows: Figure 4 As shown.

[0126] The experimental results confirmed the correlation between different levels of impairment and compensatory strategies: As the rehabilitation level improved, all three global compensation coefficients showed an overall downward trend. Specifically, the global trunk compensation coefficient (average 0.72) and global pelvic compensation coefficient (average 0.65) of patients with severe impairment were higher than those with mild impairment (0.15 and 0.12, respectively). This is consistent with the clinical pathological pattern that "the more severe the patient's condition, the more they rely on proximal trunk and pelvic compensation."

[0127] also, Figure 4This reveals the dynamic transfer characteristics of compensatory strategies: During the severe impairment phase, the compensatory range of the trunk and pelvis was greater than that of the upper limbs (trunk 0.72, pelvis 0.65, upper limbs 0.58). In the moderate and mild stages of impairment, due to the recovery of proximal core control, trunk and pelvic compensation decreases, but the upper limb compensation coefficients (0.45 and 0.20, respectively) are relatively higher than the pelvic and trunk coefficients during the same period. This data objectively reflects the clinical fact that mild to moderate patients have shifted to primarily using distal upper limb swinging to maintain balance.

[0128] This data distribution further supports the data basis for the calibration output module to generate posture deconstruction feature labels such as trunk lateral deviation compensation, pelvic elevation compensation, and upper limb swinging compensation.

Claims

1. A smart assessment system for balance function in stroke patients based on posture recognition, characterized in that, Includes data acquisition equipment and computing terminals; The data acquisition device is used to acquire continuous video stream data of the test object during the evaluation period and transmit it to the computing terminal, wherein the continuous video stream data includes image frames arranged in time sequence; the computing terminal includes: The constraint initialization module is used to receive the clinical prior information of the test subject, which includes the test position pattern and the affected side marker variable. Based on the affected side marker variable, asymmetric pathological compensation monitoring weight variables are assigned to preset anatomical skeletal key points. The data processing module is used to extract a set of skeletal key points from the continuous video stream data, generate a binarized validity mask and a four-dimensional data tensor, and output corrected three-dimensional skeletal coordinate data based on the binarized validity mask. The basic assessment module is used to calculate the basic physical balance score based on the test posture pattern and the corrected three-dimensional skeletal coordinate data. The compensation extraction module is used to calculate a comprehensive dynamic penalty coefficient by combining the corrected three-dimensional skeletal coordinate data with the pathological compensation monitoring weight variables; The calibration output module is used to correct the basic physical balance score using the comprehensive dynamic penalty coefficient, obtain the true balance evaluation score, and output it.

2. The intelligent assessment system for balance function in stroke patients based on posture recognition according to claim 1, characterized in that, When assigning asymmetric pathological compensation monitoring weight variables to preset anatomical skeletal key points, the constraint initialization module is specifically used for: The affected side and the healthy side are determined based on the marked variables of the affected side. The key anatomical and skeletal points located on the affected side are divided into pathological monitoring nodes and the pathological compensation monitoring weight variables are assigned. The anatomical skeletal key points located on the healthy side are divided into regular tracking nodes and assigned regular pose extraction weight variables. 3.The posture recognition based stroke patient balance function intelligent evaluation system according to claim 1, characterized in that, The continuous video stream data includes a color image stream and a depth data stream; when the data processing module generates the binarized validity mask and the four-dimensional data tensor, it is specifically used for: The two-dimensional pixel coordinates of the skeletal keypoint set on the color image contained in the color image stream are fused with the depth distance value read from the depth data stream, and mapped to physical space coordinates in an absolute three-dimensional coordinate system. The peak value of the probability heatmap is obtained as the visual confidence level, and the four-dimensional data tensor is generated by combining them. The visual confidence score in the four-dimensional data tensor is compared with a preset global recognition confidence score threshold. When the visual confidence score is greater than or equal to the global recognition confidence score threshold, the binarization validity mask is assigned a value of 1. When the visual confidence level is less than the global recognition confidence level threshold, the binarization validity mask is assigned a value of 0.

4. The posture recognition based balance function intelligent assessment system for stroke patients according to claim 3, characterized in that, When the data processing module outputs the corrected 3D skeleton coordinate data, it is specifically used for: An independent temporal prediction model is established for each skeletal key point in the set of skeletal key points, and the temporal prediction model adopts the Kalman filter model. When the value of the binarization validity mask is 1, the three-dimensional coordinate parameters in the four-dimensional data tensor are used as the observation input of the Kalman filter model for weighted calculation and internal covariance is updated, and the corrected three-dimensional skeleton coordinate data is output. When the value of the binarization validity mask is 0, the observation input path of the four-dimensional data tensor to the Kalman filter model is blocked. The Kalman filter model uses the state transition matrix to perform pure state prediction based on the motion velocity variables extracted from the preceding image frames that have not been occluded, and outputs the predicted three-dimensional spatial coordinates directly as the corrected three-dimensional skeleton coordinate data. 5.The posture recognition based stroke patient balance function intelligent evaluation system according to claim 1, characterized in that, When the basic assessment module calculates the basic physical equilibrium score, it is specifically used for: Based on the test posture pattern, the set of body segments involved in the calculation is determined. The three-dimensional coordinates of the proximal and distal bone key points corresponding to the set of body segments are extracted from the corrected three-dimensional skeletal coordinate data. The local three-dimensional coordinates of the centroid are calculated by linear interpolation in combination with the human segment centroid position constant. The spatial coordinates of the overall centroid are calculated by weighted summation algorithm, and a dynamic three-dimensional centroid trajectory sequence within the evaluation period is generated. Calculate and sum the Euclidean distances of the spatial coordinates of the overall center of gravity between two adjacent image frames in three-dimensional physical space to obtain the total length of the center of gravity swing path. The average total length of the center of gravity swing path of healthy people of the same age is obtained as a reference benchmark value; when the total length of the center of gravity swing path is less than or equal to the reference benchmark value, the basic physical balance score is directly assigned to the preset full score benchmark. When the total length of the center of gravity swing path is greater than the reference benchmark value, the difference between the total length of the center of gravity swing path and the reference benchmark value is calculated. Using an exponential decay function combined with a preset decay constant, the full score benchmark is numerically reduced based on the calculated difference to obtain the basic physical balance score. 6.The posture recognition based stroke patient balance function intelligent evaluation system according to claim 1, wherein, The compensation extraction module combines the corrected three-dimensional skeletal coordinate data with the pathological compensation monitoring weight variables to calculate the spatial geometric linkage relationship, thereby obtaining multi-segment spatial compensation feature parameters. These multi-segment spatial compensation feature parameters include the trunk lateral deviation angle and the difference in coronal plane height of the pelvis. Specifically, the compensation extraction module is used for: The corresponding three-dimensional spatial coordinates are extracted from the corrected three-dimensional skeletal coordinate data. The coordinates of the extracted left and right acromion are arithmetically averaged to obtain the three-dimensional spatial coordinates of the shoulder center. The coordinates of the extracted left and right anterior superior iliac spines are arithmetically averaged to obtain the three-dimensional spatial coordinates of the pelvic center. The three-dimensional spatial coordinates of the pelvic center are used as the starting point of the vector and the three-dimensional spatial coordinates of the shoulder center are used as the ending point of the vector to construct the trunk spatial vector. The angle between the trunk spatial vector and the Y-axis direction vector of the absolute three-dimensional coordinate system is calculated to obtain the trunk lateral deviation angle. Read the Y-axis coordinate values ​​of the affected and healthy anterior superior iliac spines in the absolute three-dimensional coordinate system. Subtract the Y-axis coordinate value of the healthy anterior superior iliac spine from the Y-axis coordinate value of the affected anterior superior iliac spine to perform a difference calculation to obtain the coronal plane height difference of the pelvis. When the calculated difference is less than zero, the coronal plane height difference of the pelvis is assigned to zero.

7. The stroke patient balance function intelligent evaluation system based on gesture recognition according to claim 6, characterized in that, The multi-segment spatial compensation characteristic parameters also include the spatial angle of the affected elbow joint. Using the corrected three-dimensional skeletal coordinate data, the compensation extraction module constructs an upper arm spatial vector by taking the three-dimensional spatial coordinates corresponding to the affected elbow joint as the starting point of the spatial vector and the three-dimensional spatial coordinates corresponding to the affected acromion as the ending point of the spatial vector. It also constructs a forearm spatial vector by taking the three-dimensional spatial coordinates corresponding to the affected elbow joint as the starting point of the spatial vector and the three-dimensional spatial coordinates corresponding to the affected wrist joint as the ending point of the spatial vector. The spatial angle of the affected elbow joint is obtained by calculating the dot product of the upper arm spatial vector and the forearm spatial vector. Using a piecewise mapping function, the trunk lateral deviation angle, the coronal plane height difference of the pelvis, and the spatial angle of the affected elbow joint are converted into the trunk lateral deviation compensation index, the pelvic elevation compensation index, and the upper limb flexor muscle spasm compensation index, respectively. The trunk lateral deviation compensation index, pelvic elevation compensation index, and upper limb flexor spasticity compensation index corresponding to all image frames within the evaluation period are summarized and their arithmetic averages are calculated to obtain the global trunk compensation coefficient, global pelvic compensation coefficient, and global upper limb compensation coefficient, respectively.

8. The stroke patient balance function intelligent evaluation system based on gesture recognition according to claim 7, characterized in that, When calculating the comprehensive dynamic penalty coefficient, the compensation extraction module is specifically used for: An enable mask containing a trunk mask, a pelvic mask, and an upper limb mask is generated based on the test posture pattern; a preset basic weight parameter containing trunk weight, pelvic weight, and upper limb weight is obtained, and the basic weight parameter is dynamically masked and filtered using the enable mask; The global trunk compensation coefficient, the global pelvic compensation coefficient, and the global upper limb compensation coefficient are weighted and summed according to their respective enable masks and the basic weight parameters. The weighted summation result is divided by the total effective weights in the current test mode to calculate the comprehensive dynamic penalty coefficient. The total effective weights are used to mathematically normalize the compensation items involved in the calculation. When the sum of the actual effective weights is equal to zero, the comprehensive dynamic penalty coefficient is directly assigned the value of zero.

9. The stroke patient balance function intelligent evaluation system based on gesture recognition according to claim 8, characterized in that, When the calibration output module obtains the true balance assessment score, it is specifically used for: The actual deduction ratio is obtained by multiplying the comprehensive dynamic penalty coefficient with the preset maximum penalty ratio factor, and the calibration discount coefficient is obtained by subtracting the actual deduction ratio from the value. The true balance assessment score is calculated by multiplying the basic physical balance score by the calibration discount factor. The true balance assessment score is rounded and truncated to strictly limit its value range to a closed interval of zero to one hundred.

10. The intelligent assessment system for balance function in stroke patients based on posture recognition according to claim 9, characterized in that, When the calibration output module outputs the evaluation report, it is specifically used for: The actual balance assessment score is compared with the preset severe and moderate impairment thresholds, and the corresponding clinical balance classification result is output. The calculated global trunk compensation coefficient, global pelvic compensation coefficient, and global upper limb compensation coefficient are read and compared with preset tag trigger thresholds to generate trunk lateral deviation compensation tags, pelvic elevation compensation tags, and upper limb flexor muscle spasm tags. The real balance assessment score, the clinical balance classification result and the generated trunk laterodeviation compensation label, the pelvis lifting compensation label and the upper limb flexor spasm label are data packed to construct a comprehensive balance assessment report, and the comprehensive balance assessment report is output as the assessment report.

Images