Rehabilitation motion quantitative evaluation system and method fusing posture metric and feedback learning

Abstract

The application discloses a rehabilitation action quantitative evaluation system and method fusing posture measurement and feedback learning, comprising: collecting real-time human posture data of a user when performing a preset rehabilitation action, and extracting a three-dimensional space coordinate sequence of key skeleton points; constructing a multi-level posture measurement index system, and calculating multi-dimensional posture quantitative features; performing dynamic time warping matching on the multi-dimensional posture quantitative features and corresponding standard rehabilitation action templates; constructing a quality evaluation model based on feedback learning, comprehensively evaluating the execution quality of the rehabilitation action through a supervised learning mechanism, and generating a quantitative evaluation score; identifying specific deviation types and deviation degrees in action execution, and generating targeted rehabilitation feedback guidance information, which not only provides accurate deviation identification and individualized guidance, but also adaptively adjusts rehabilitation difficulty, significantly reduces rehabilitation cost, and improves the accessibility and intelligent level of rehabilitation.

Description

Technical Field

[0001] This invention relates to the field of quantitative assessment technology for rehabilitation movements, and specifically to a quantitative assessment system and method for rehabilitation movements that integrates posture measurement and feedback learning. Background Technology

[0002] Quantitative assessment of rehabilitation movements is an important research direction in rehabilitation medicine and intelligent human-computer interaction, and it is of great significance for improving the quality and efficiency of patients' home rehabilitation training. Traditional rehabilitation assessment mainly relies on the visual observation and experience judgment of rehabilitation physicians, which has problems such as strong subjectivity, inconsistent assessment standards, and difficulty in large-scale promotion. With the development of depth cameras and human pose estimation technology, vision-based automatic rehabilitation assessment methods have gradually become a research hotspot. However, existing technologies still have the following shortcomings: (1) Existing posture measurement methods are single-dimensional and difficult to fully characterize the performance quality of rehabilitation movements. Most current automatic rehabilitation assessment systems only focus on single-dimensional measurement indicators such as local joint angles or movement trajectories. For example, they only calculate the difference between the joint angle and the standard value, ignoring the morphological similarity of the global movement trajectory and the dynamic changes in body stability during the movement. The performance quality of rehabilitation movements depends not only on whether the joint angle is in place, but also on the comprehensive influence of multiple factors such as the smoothness and symmetry of the movement trajectory and trunk balance ability. Single-dimensional measurement methods are prone to one-sided assessment results and cannot accurately identify potential problems such as compensatory movements or body imbalance, thus affecting the clinical credibility of rehabilitation assessment. (2) When different individuals perform the same rehabilitation movement, there are natural differences in the rhythm, speed and duration of the movement. Traditional frame-by-frame comparison methods cannot handle nonlinear and scalable changes on the time axis. Although the dynamic time warping algorithm has been initially introduced into the field of movement recognition, existing applications are mostly at the level of frame-level feature matching. They have failed to build an adaptive alignment framework that integrates multi-level posture features, which makes the assessment of movement morphology similarity susceptible to interference from differences in execution speed, resulting in insufficient robustness and accuracy of the assessment results. Summary of the Invention

[0003] The purpose of this invention is to provide a quantitative assessment system and method for rehabilitation movements that integrates posture measurement and feedback learning, in order to solve the technical problems of existing posture measurement methods having a single dimension, making it difficult to comprehensively represent the execution quality of rehabilitation movements, and lacking an effective alignment and similarity measurement mechanism for spatiotemporal morphological differences.

[0004] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A first aspect of the present invention provides a method for quantitative assessment of rehabilitation movements that integrates posture measurement and feedback learning, comprising the following steps: Real-time human posture data is collected when the user performs preset rehabilitation movements, and the real-time human posture data is preprocessed to extract the three-dimensional spatial coordinate sequence of key skeletal points. Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a multi-level posture measurement index system integrating local joint angles, global motion trajectory and body stability is constructed, and multi-dimensional posture quantification features are calculated. The multi-dimensional posture quantification features are dynamically time-warped and matched with the corresponding standard rehabilitation movement templates to calculate a movement consistency score based on morphological similarity. A quality assessment model based on feedback learning is constructed. The multi-dimensional posture quantitative features and the action consistency score are used as inputs. The execution quality of rehabilitation actions is comprehensively evaluated through a supervised learning mechanism to generate a quantitative assessment score. Based on the quantitative assessment score, the specific types and degrees of deviations in the execution of the movement are identified, and targeted rehabilitation feedback guidance information is generated.

[0005] As a preferred embodiment of the present invention, real-time human posture data of the user performing preset rehabilitation movements is collected, and the real-time human posture data is preprocessed to extract the three-dimensional spatial coordinate sequence of key skeletal points, including: The system uses a pose estimation model based on a depth camera to capture video frame sequences in real time as the user performs preset rehabilitation actions. Preprocessing operations such as denoising, timestamp alignment, and background subtraction are performed on the acquired video frame sequence to generate standardized human motion image data; A human skeletal key point detection algorithm is used to locate and track a preset set of human skeletal key points from the standardized human motion image data. The set of key points includes at least the shoulder, elbow, wrist, hip, knee, and ankle joints, as well as the center point of the torso. Extract the coordinates of each key point in each frame of the image in three-dimensional space to construct a time-continuous three-dimensional spatial coordinate sequence, expressed as:

[0006] in, For frame index, Total number of frames For key point indexing, The total number of key points. For the first The first frame The coordinates of the key points in three-dimensional space.

[0007] As a preferred embodiment of the present invention, based on the three-dimensional spatial coordinate sequence of the key skeletal points, a multi-level posture measurement index system integrating local joint angles, global motion trajectory, and body stability is constructed, including: Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a local joint angle metric is constructed. This local joint angle metric calculates the spatial angle formed by three adjacent key points for each joint, and its expression is as follows:

[0008] in, For the first The first frame The angle of each joint and These are the vectors of the two bone segments that make up the joint; Construct a local angle feature vector for all joint angles in time series. ; Construct a global motion trajectory measurement index, which includes: calculating the motion trajectory curves of preset representative key points in three-dimensional space; The motion trajectory curve is parameterized, and the trajectory length, trajectory curvature, rate of change of direction, and spatial offset between the trajectory and the standard template are extracted to form a global trajectory feature vector. ; A body stability metric is constructed, comprising the standard deviation of displacement at the center point, the height and distance differences between bilaterally symmetrical key points, and the range of movement of the pressure center within the support surface. These body stability metrics are then fused to form a body stability feature vector. ; The local angle feature vector Global trajectory feature vector and body stability feature vector Temporal alignment and normalization are performed, and the results are fused to form a multi-dimensional pose quantization feature set. .

[0009] As a preferred embodiment of the present invention, a multi-dimensional attitude quantization feature set is fused together. Specifically: With the global trajectory feature vector Using the time axis as a reference, linear interpolation is employed to interpolate the local angle feature vector. and body stability feature vector Time sampling points are uniformly interpolated to the same level as On the same timestamp sequence, obtain time-aligned feature vectors. , and ,in ; The eigenvalues ​​of each time-aligned eigenvector are mapped to the interval [0, 1] and then normalized. A multi-level fusion weight matrix is ​​constructed, wherein differentiated fusion weight coefficients are assigned to the local angle feature vector, global trajectory feature vector, and body stability feature vector for different rehabilitation movement types or different rehabilitation stages. , and ,satisfy ; The normalized feature vector , and The multi-dimensional pose quantization feature set is constructed by weighting and concatenating the features according to the fusion weight matrix. Its expression is:

[0010] in, For the first The comprehensive feature vector at each time step, This represents the total number of time steps after alignment.

[0011] As a preferred embodiment of the present invention, the multi-dimensional posture quantification features are dynamically time-warped and matched with the corresponding standard rehabilitation movement templates to calculate a movement consistency score based on morphological similarity, including: A pre-built standard rehabilitation movement template library is constructed, which contains standard posture quantification feature sequences for each type of rehabilitation movement. , where M is the time length of the standard action sequence, and the standard template is obtained through multiple examples demonstrated by rehabilitation experts; Obtain the measured posture quantization feature sequence of the user performing rehabilitation movements. ,in The duration of the user action sequence; The measured feature sequence is calculated using the dynamic time warping algorithm. With the standard template feature sequence The minimum cumulative distance between them specifically includes: Construct an M×N cumulative distance matrix D, where the matrix element D(i,j) represents the minimum cumulative matching distance between the i-th point in the standard sequence and the j-th point in the measured sequence; Define a local distance metric function That is, the squared Euclidean distance between two eigenvectors, where This represents the standard posture quantization feature sequence for the i-th type of rehabilitation movement. Represents the measured posture quantization feature sequence of the j-th performed rehabilitation action; The cumulative distance is calculated using a recursive formula, the expression of which is:

[0012] The boundary conditions are: , , ; Extract the final cumulative distance value D(M, N) from the cumulative distance matrix D to characterize the degree of overall morphological difference between the measured action and the standard action; The final cumulative distance value D(M, N) is converted into a motion consistency score based on morphological similarity. Its expression is:

[0013] in, The reference normalization factor is the cumulative distance obtained by dynamically warping and matching the standard action template with itself. The action consistency score is... The value range is [0%, 100%], and the higher the score, the more consistent the action pattern is with the standard template.

[0014] As a preferred embodiment of the present invention, a quality assessment model based on feedback learning is constructed. Using the multi-dimensional posture quantification features and the movement consistency score as input, a supervised learning mechanism is used to comprehensively evaluate the execution quality of rehabilitation movements, generating a quantitative assessment score, including: A quality assessment model based on supervised learning is constructed using a deep neural network, which integrates input feature vectors; The input feature vector is a multi-dimensional pose quantization feature set. The extracted global statistical features include: mean, variance, peak value, and range of motion of each joint angle; length, mean curvature, and mean rate of change of direction of the global trajectory. Obtain the action consistency score and its segmented scoring sequence on the time axis ,in The number of segments into which a sequence of actions is divided according to the phases of motion; Construct a labeled training dataset, wherein each sample in the training dataset contains an input feature vector of a single rehabilitation action. and their corresponding expert-annotated quality scores The expert-marked quality score is comprehensively evaluated by rehabilitation experts based on multi-dimensional evaluation criteria, which include movement accuracy, fluency, stability, and completion. The supervised learning model is trained using the training dataset, and the optimization objective is to minimize the loss function between the predicted score and the expert-annotated score. The expression for the loss function is:

[0015] in, The total number of training samples, The quantitative evaluation score for model prediction. Experts are assigned scores; cross-validation is used during training to prevent overfitting, and an early stopping strategy is employed to determine the optimal model parameters; The input feature vector X of the user to be evaluated is fed into the trained quality assessment model, and the model outputs a comprehensive quantitative assessment score. This score represents the overall quality of the user's rehabilitation movements; the higher the score, the better the quality of the movements. Output the contribution weight of each input feature in the quality assessment model to the final assessment score, and identify key posture measurement indicators that affect the quality of rehabilitation movements.

[0016] As a preferred embodiment of the present invention, based on the quantitative assessment score, specific types and degrees of deviation in the execution of the movement are identified, and targeted rehabilitation feedback guidance information is generated, including: Construct a deviation identification rule base, which contains multiple preset deviation types for each type of rehabilitation movement, and each deviation type corresponds to an abnormal pattern of one or more posture measurement indicators; The multi-dimensional pose quantization feature set The deviation of each posture measurement index is calculated by comparing it frame by frame with the standard rehabilitation movement template, including: joint angle deviation, trajectory space deviation and stability deviation. Based on the magnitude and duration of the deviation, and combined with the deviation identification rule base, the deviation type and severity level of the current action are determined. The severity level is divided into three levels: mild, moderate, and severe, based on the ratio range of the deviation amount relative to a preset threshold. Based on the quantitative evaluation score Based on the identified types and degrees of deviation, targeted rehabilitation feedback guidance information is retrieved from a pre-set feedback information database or dynamically generated. The rehabilitation feedback guidance information is generated using a hierarchical feedback strategy. The generated rehabilitation feedback guidance information is displayed to the user through a terminal device. At the same time, the user's acceptance of the feedback and the improvement effect of subsequent actions are recorded to form a closed-loop feedback optimization mechanism.

[0017] As a preferred embodiment of the present invention, a hierarchical feedback strategy is used to generate the rehabilitation feedback guidance information, wherein the hierarchical feedback strategy includes: Real-time feedback layer: Detects significant deviations in real time during action execution and triggers alarms. Phased feedback layer: After a single action or a group of actions is completed, a comprehensive evaluation report is generated, which includes deviation analysis, improvement suggestions, and demonstration of the action. Long-term feedback layer: Based on historical data from multiple rehabilitation training sessions, it generates a rehabilitation progress trend chart and dynamically adjusts the difficulty coefficient and target threshold of rehabilitation movements.

[0018] A second aspect of the present invention provides a rehabilitation movement quantitative assessment system integrating posture measurement and feedback learning, for implementing the rehabilitation movement quantitative assessment method integrating posture measurement and feedback learning as described in any one of claims 1-8, the system comprising, Data acquisition and preprocessing module: used to acquire real-time human posture data when the user performs preset rehabilitation movements, and preprocess the real-time human posture data to extract the three-dimensional spatial coordinate sequence of key skeletal points; Multi-level posture measurement module: connected to the data acquisition and preprocessing module, used to construct a multi-level posture measurement index system that integrates local joint angles, global motion trajectory and body stability based on the three-dimensional spatial coordinate sequence of the key bone points, and calculate multi-dimensional posture quantification features. Dynamic time warping matching module: connected to the multi-level posture measurement module, used to perform dynamic time warping matching of the multi-dimensional posture quantification features with the corresponding standard rehabilitation movement template, and calculate the movement consistency score based on morphological similarity. Feedback learning evaluation module: It is connected to the multi-level posture measurement module and the dynamic time warping matching module respectively, and is used to construct a quality evaluation model based on feedback learning. It takes the multi-dimensional posture quantification features and the action consistency score as input, and comprehensively evaluates the execution quality of rehabilitation actions through a supervised learning mechanism to generate a quantitative evaluation score. Deviation identification and feedback generation module: connected to the feedback learning and assessment module, used to identify the specific type and degree of deviation in the execution of the action based on the quantitative assessment score, and generate targeted rehabilitation feedback guidance information.

[0019] As a preferred embodiment of the present invention, the dynamic time warping matching module includes: A standard template library storage unit is used to pre-store standard posture quantification feature sequences for each type of rehabilitation movement; The cumulative distance calculation unit is used to calculate the minimum cumulative distance between the measured feature sequence and the standard template feature sequence using the dynamic time warping algorithm. A consistency score conversion unit is used to convert the minimum cumulative distance into an action consistency score based on morphological similarity.

[0020] Compared with the prior art, the present invention has the following advantages: This invention utilizes a multi-level posture measurement index system, integrating three dimensions: local joint angles, global movement trajectory, and body stability. This achieves a comprehensive and refined representation of rehabilitation movement quality, overcoming the limitations of single-dimensional assessment. A dynamic time warping algorithm is introduced to eliminate individual differences in movement speed, enabling precise measurement of morphological similarity between measured movements and standard templates, significantly improving the robustness of the assessment. A supervised learning assessment model is constructed based on deep neural networks, transferring expert knowledge into objective, reproducible, and interpretable quantitative scores. Employing a three-tiered feedback strategy (immediate, phased, and long-term) combined with a closed-loop optimization mechanism, it not only provides accurate deviation identification and personalized guidance but also adaptively adjusts rehabilitation difficulty, forming a continuous evolutionary capability of "assessment-feedback-improvement-optimization." This significantly reduces rehabilitation costs and enhances the accessibility and intelligence of home-based rehabilitation. Attached Figure Description

[0021] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0022] Figure 1 The flowchart illustrates the quantitative assessment method for rehabilitation movements that integrates posture measurement and feedback learning, as provided in this embodiment of the invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0024] like Figure 1 As shown, this invention provides a method for quantitative assessment of rehabilitation movements that integrates posture measurement and feedback learning, including the following steps: Real-time human posture data is collected when the user performs preset rehabilitation movements, and the real-time human posture data is preprocessed to extract the three-dimensional spatial coordinate sequence of key skeletal points. Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a multi-level posture measurement index system integrating local joint angles, global motion trajectory and body stability is constructed, and multi-dimensional posture quantification features are calculated. In this embodiment, a multi-level posture measurement index system is constructed by integrating three dimensions of measurement indicators: local joint angle, global movement trajectory, and body stability. This system can simultaneously capture the local detail precision, overall morphological similarity, and dynamic balance ability of rehabilitation movements, thereby more comprehensively and accurately representing the overall quality of the user's performance of rehabilitation movements. This multi-level fusion mechanism effectively solves the shortcomings of existing technologies, such as single evaluation dimensions, easy neglect of compensatory movements and body imbalances, and significantly improves the clinical credibility and completeness of the evaluation results.

[0025] The multi-dimensional posture quantification features are dynamically time-warped and matched with the corresponding standard rehabilitation movement templates to calculate a movement consistency score based on morphological similarity. In this embodiment, a dynamic time warping algorithm is used to nonlinearly align and match the user's measured feature sequence with the standard rehabilitation movement template. This effectively eliminates the interference of natural differences in execution speed, movement rhythm, and duration among different individuals on similarity measurement. By constructing a cumulative distance matrix and converting it into a consistency score based on morphological similarity, the overall deformation degree between the user's movement trajectory and the standard template can be objectively evaluated. This avoids misjudgments caused by time axis misalignment in traditional frame-by-frame comparison methods, significantly improving the robustness and accuracy of movement consistency assessment. It is suitable for user groups of different rehabilitation stages, ages, and motor abilities.

[0026] A quality assessment model based on feedback learning is constructed. The multi-dimensional posture quantitative features and the action consistency score are used as inputs. The execution quality of rehabilitation actions is comprehensively evaluated through a supervised learning mechanism to generate a quantitative assessment score. Based on the quantitative assessment score, the specific types and degrees of deviations in the execution of the movement are identified, and targeted rehabilitation feedback guidance information is generated.

[0027] In this embodiment, a quality assessment model based on supervised learning is constructed. Using multi-dimensional posture quantification features and movement consistency scores as inputs, a deep neural network is used to comprehensively evaluate the execution quality of rehabilitation movements. The model is trained using multi-dimensional evaluation criteria labeled by rehabilitation experts. It can learn from expert evaluation experience and output objective and quantitative evaluation scores. Compared with existing methods that rely entirely on manual rules or simple threshold judgments, the supervised learning mechanism of this invention effectively overcomes the defects of strong subjectivity and inconsistent evaluation standards in traditional rehabilitation assessment, and realizes the standardization, repeatability and quantification of evaluation results.

[0028] Real-time human posture data is collected when the user performs preset rehabilitation movements, and the real-time human posture data is preprocessed to extract the three-dimensional spatial coordinate sequence of key skeletal points, including: The system uses a pose estimation model based on a depth camera to capture video frame sequences in real time as the user performs preset rehabilitation actions. Preprocessing operations such as denoising, timestamp alignment, and background subtraction are performed on the acquired video frame sequence to generate standardized human motion image data; In this embodiment, the acquired video frame sequence is subjected to preprocessing operations such as denoising, timestamp alignment, and background subtraction in sequence. This effectively eliminates the interference of ambient lighting changes, sensor noise, and complex backgrounds on subsequent key point detection. Among them, the timestamp alignment operation ensures the synchronization of multi-source data in the time dimension, while the background subtraction operation reduces redundant information in non-human areas, making the image data input to the skeletal key point detection algorithm cleaner and more standardized, thereby significantly improving the accuracy and stability of key point detection.

[0029] A human skeletal key point detection algorithm is used to locate and track a preset set of human skeletal key points from the standardized human motion image data. The set of key points includes at least the shoulder, elbow, wrist, hip, knee, and ankle joints, as well as the center point of the torso. In this embodiment, the human skeleton key point detection algorithm is specifically as follows: The preprocessed human motion image data and the preset key point index set are used as inputs to the human skeleton key point detection algorithm; The data processing procedure is as follows: Perform human region detection on the input image and locate the bounding boxes of one or more human bodies in the image; If there are multiple users, the target user is selected according to preset rules, such as the closest to the image center, the largest area, or a pre-specified user ID; Output the bounding box coordinates of the target user's human body, and crop out the human body region image as input for subsequent processing; The cropped human body region image is input into a pre-trained 2D pose estimation network; An encoder-decoder architecture is adopted, and image features are extracted through a convolutional neural network to output a confidence heatmap for each key point. Peak detection is performed on the heatmap of each key point to locate the two-dimensional pixel coordinates and the corresponding confidence score. If there is occlusion or self-occlusion that causes the confidence score to be lower than the preset threshold, the key point is marked as "invisible". The detected key points are associated according to the topology of the human skeleton to form a complete skeleton model. The coordinates of the key points are transformed from the camera coordinate system to the human coordinate system, and a standardized reference coordinate system is defined with the center of gravity of the human body as the origin. Spatial normalization is performed on the converted 3D coordinates by scaling the human height or torso length to eliminate numerical changes caused by differences in body shape among different users. Output the normalized three-dimensional spatial coordinates.

[0030] Extract the coordinates of each key point in each frame of the image in three-dimensional space to construct a time-continuous three-dimensional spatial coordinate sequence, expressed as:

[0031] in, For frame index, Total number of frames For key point indexing, The total number of key points. For the first The first frame The coordinates of the key points in three-dimensional space.

[0032] In this embodiment, the coordinate positions of each key point in each frame in three-dimensional space are extracted to construct a time-continuous three-dimensional spatial coordinate sequence, realizing a complete digital expression of human rehabilitation movements in time and space dimensions. Compared with the method that only uses two-dimensional image coordinates, the three-dimensional coordinate information retains the motion information of the depth dimension z-axis, which can accurately represent the user's movement characteristics in the forward and backward directions, such as leaning forward or backward. At the same time, the time-continuous sequence expression provides structured data input for subsequent dynamic time warping matching, multi-level feature extraction and closed-loop feedback learning, which is the data foundation of the entire rehabilitation assessment method.

[0033] Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a multi-level posture measurement index system integrating local joint angles, global motion trajectory, and body stability is constructed, including: Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a local joint angle metric is constructed. This local joint angle metric calculates the spatial angle formed by three adjacent key points for each joint, and its expression is as follows:

[0034] in, For the first The first frame The angle of each joint and These are the vectors of the two bone segments that make up the joint; Construct a local angle feature vector for all joint angles in time series. ; In this embodiment, a local joint angle metric is constructed by calculating the spatial angle formed by three adjacent key points. The inverse cosine function is then used to accurately solve for the joint angle values ​​in three-dimensional space. The three-dimensional spatial angle calculation fully considers depth dimension information, accurately reflecting the range of motion of the joint in real three-dimensional space. This is particularly suitable for evaluating rehabilitation movements involving anterior-posterior movements, such as shoulder flexion and hip extension. Simultaneously, a local angle feature vector is constructed for all joint angles according to a time series. It fully preserves the dynamic changes of each joint during movement.

[0035] Construct a global motion trajectory measurement index, which includes: calculating the motion trajectory curves of preset representative key points in three-dimensional space; The motion trajectory curve is parameterized, and the trajectory length, trajectory curvature, rate of change of direction, and spatial offset between the trajectory and the standard template are extracted to form a global trajectory feature vector. ; In this embodiment, preset representative key points are used, such as dynamic selection based on the type of action (e.g., wrist for upper limb actions, ankle for lower limb actions), to construct a global action trajectory curve. The trajectory is then parameterized, and multi-dimensional features such as trajectory length, trajectory curvature, rate of change of direction, and spatial offset between the trajectory and the standard template are extracted. The trajectory length reflects the amplitude of the action, the trajectory curvature and rate of change of direction reflect the smoothness and fluency of the action, and the spatial offset directly quantifies the morphological difference between the user's actual trajectory and the standard template. This allows for a comprehensive characterization of the similarity and abnormal patterns of the action trajectory from a geometric morphological perspective.

[0036] A body stability metric is constructed, comprising the standard deviation of displacement at the center point, the height and distance differences between bilaterally symmetrical key points, and the range of movement of the pressure center within the support surface. These body stability metrics are then fused to form a body stability feature vector. ; The local angle feature vector Global trajectory feature vector and body stability feature vector Temporal alignment and normalization are performed, and the results are fused to form a multi-dimensional pose quantization feature set. .

[0037] In this embodiment, the local angle feature vector, global trajectory feature vector, and body stability feature vector are aligned and normalized in the time dimension to form a unified multi-dimensional pose quantization feature set. The time alignment operation solves the inconsistency problem of the three types of features in terms of original sampling frequency and timestamp. The normalization process eliminates the numerical imbalance problem caused by the difference in scale and magnitude between different features. This fusion framework provides a feature input with a unified format and consistent scale for subsequent dynamic time warping matching and supervised learning evaluation, ensuring the numerical stability and model convergence of the entire evaluation process.

[0038] The fusion constitutes a multi-dimensional pose quantization feature set Specifically: With the global trajectory feature vector Using the time axis as a reference, linear interpolation is employed to interpolate the local angle feature vector. and body stability feature vector Time sampling points are uniformly interpolated to the same level as On the same timestamp sequence, obtain time-aligned feature vectors. , and ,in ; In this embodiment, the time axis of the global trajectory feature vector is used as the benchmark. The local angle feature vector and the body stability feature vector are uniformly interpolated to the same timestamp sequence using a linear interpolation method. The advantage of this strategy is that the global trajectory feature usually has the highest temporal resolution and completely records the entire process information from the start of the action to the end. Using it as the benchmark can preserve the temporal integrity of the action to the maximum extent. In contrast, if the stability feature with a lower sampling rate is used as the benchmark, the temporal details of the trajectory and angle information will be lost. The linear interpolation method is simple to calculate and has good continuity, and can achieve accurate alignment of multi-source features without introducing obvious distortion.

[0039] The eigenvalues ​​of each time-aligned eigenvector are mapped to the interval [0, 1] and then normalized. A multi-level fusion weight matrix is ​​constructed, wherein differentiated fusion weight coefficients are assigned to the local angle feature vector, global trajectory feature vector, and body stability feature vector for different rehabilitation movement types or different rehabilitation stages. , and ,satisfy ; In this embodiment, for rehabilitation movements primarily involving fine motor skills of the upper limbs, such as finger grasping and wrist flexion and extension, the accuracy of local joint angles is of paramount importance and can be assigned a higher weight. For rehabilitation exercises that primarily involve lower limb walking or gait training, global trajectory and body stability are of greater importance and are assigned higher weight. and For balance training exercises, such as standing on one leg and sitting-standing transfers, the weights for body stability are... It should be significantly higher than other dimensions.

[0040] The normalized feature vector , and The multi-dimensional pose quantization feature set is constructed by weighting and concatenating the features according to the fusion weight matrix. Its expression is:

[0041] in, For the first The comprehensive feature vector at each time step, This represents the total number of time steps after alignment.

[0042] In this embodiment, a differentiated weight allocation mechanism is adopted, which can be flexibly configured according to different clinical scenarios and individual needs, significantly improving the clinical adaptability and personalization of the assessment method. The weight coefficients can be preset by expert experience or dynamically optimized through subsequent feedback learning modules to achieve personalized rehabilitation assessments for each individual.

[0043] The multi-dimensional posture quantification features are dynamically time-warped and matched with the corresponding standard rehabilitation movement templates to calculate a movement consistency score based on morphological similarity, including: A pre-built standard rehabilitation movement template library is constructed, which contains standard posture quantification feature sequences for each type of rehabilitation movement. , where M is the time length of the standard action sequence, and the standard template is obtained through multiple examples demonstrated by rehabilitation experts; Obtain the measured posture quantization feature sequence of the user performing rehabilitation movements. ,in The duration of the user action sequence; The measured feature sequence is calculated using the dynamic time warping algorithm. With the standard template feature sequence The minimum cumulative distance between them specifically includes: Construct an M×N cumulative distance matrix D, where the matrix element D(i,j) represents the minimum cumulative matching distance between the i-th point in the standard sequence and the j-th point in the measured sequence; In this embodiment, the dynamic time warping algorithm is used to calculate the minimum cumulative distance between the measured feature sequence and the standard template feature sequence. This effectively solves the problem of nonlinear misalignment of the time axis caused by differences in movement speed, rhythm and duration when different users perform the same rehabilitation movement. Through nonlinear time warping, one point in one sequence is allowed to match multiple points in another sequence, thereby finding the optimal alignment path and ensuring that the evaluation of movement morphology similarity is not affected by differences in execution speed.

[0044] Define a local distance metric function That is, the squared Euclidean distance between two eigenvectors, where This represents the standard posture quantization feature sequence for the i-th type of rehabilitation movement. Represents the measured posture quantization feature sequence of the j-th performed rehabilitation action; The cumulative distance is calculated using a recursive formula, the expression of which is:

[0045] The boundary conditions are: , , ; In this embodiment, an M×N cumulative distance matrix D is constructed, and dynamic programming is used to recursively calculate the cumulative distance. For any regular path from the starting point (1,1) to the ending point (M,N), each step can only move to the right, down, or down-right, ensuring the temporal order constraint of the matching process, i.e., time reversal is not allowed, ensuring that the matching result conforms to the natural temporal order of action execution. By solving the global optimal path, the matching scheme that minimizes the cumulative distance can be found among all possible alignment methods, thereby obtaining the most reasonable morphological similarity measure between the measured action and the standard template.

[0046] Extract the final cumulative distance value D(M, N) from the cumulative distance matrix D to characterize the degree of overall morphological difference between the measured action and the standard action; The final cumulative distance value D(M, N) is converted into a motion consistency score based on morphological similarity. Its expression is:

[0047] in, The reference normalization factor is the cumulative distance obtained by dynamically warping and matching the standard action template with itself. The action consistency score is... The value range is [0%, 100%], and the higher the score, the more consistent the action pattern is with the standard template.

[0048] In this embodiment, the action consistency evaluation As a measure of overall morphological similarity between the measured action and the standard template, it provides important contextual information for deviation identification. A higher value indicates better overall quality of user actions, allowing the deviation detection module to focus on subtle local deviations; when... When the value is low, it indicates a significant overall problem. Feedback should prioritize prompting the user to adjust the rhythm or amplitude of their movements. This "coarse-to-fine" tiered evaluation strategy makes feedback guidance more efficient and targeted.

[0049] A quality assessment model based on feedback learning is constructed. Using the multi-dimensional posture quantification features and the movement consistency score as input, a supervised learning mechanism is used to comprehensively evaluate the execution quality of rehabilitation movements, generating a quantitative assessment score, including: A quality assessment model based on supervised learning is constructed using a deep neural network, which integrates input feature vectors; The input feature vector is a multi-dimensional pose quantization feature set. The extracted global statistical features include: mean, variance, peak value, and range of motion of each joint angle; length, mean curvature, and mean rate of change of direction of the global trajectory. In this embodiment, a quality assessment model based on supervised learning is constructed using a deep neural network to quantize multi-dimensional pose feature sets. The extracted global statistical features include the mean, variance, peak value, and range of motion of each joint angle, as well as the length, mean curvature, and mean rate of change of direction of the global trajectory, and are correlated with the motion consistency score. The deep neural network integrates its segmented scoring sequences as input. It has a strong nonlinear fitting ability and can automatically learn the complex mapping relationship between low-level features, such as the mean of joint angles, trajectory curvature and high-level evaluation scores.

[0050] Obtain the action consistency score and its segmented scoring sequence on the time axis ,in The number of segments into which a sequence of actions is divided according to the phases of motion; In this embodiment, not only the overall value of the action consistency score is used. Furthermore, the model incorporates a segmented scoring sequence along the timeline. Many quality issues in rehabilitation movements are not evenly distributed throughout the entire movement but are concentrated in specific phases. For example, shoulder abduction movements may perform well at the initial stage but exhibit scapular compensation at maximum abduction; sit-to-stand transfer movements may lack stability at the standing stage but have normal posture after standing. The introduction of the segmented scoring sequence allows the model to perceive quality differences at different stages, thereby reflecting the importance of stage performance in the final quantitative assessment score, improving the refinement and clinical relevance of the assessment.

[0051] Construct a labeled training dataset, wherein each sample in the training dataset contains an input feature vector of a single rehabilitation action. and their corresponding expert-annotated quality scores The expert-marked quality score is comprehensively evaluated by rehabilitation experts based on multi-dimensional evaluation criteria, which include movement accuracy, fluency, stability, and completion. In this embodiment, each sample in the constructed labeled training dataset contains an input feature vector of one rehabilitation action. and their corresponding expert-annotated quality scores The expert annotation is based on multi-dimensional evaluation criteria, such as the accuracy, fluency, stability and completion of movements, to make a comprehensive assessment. Through supervised learning, the model can learn the assessment experience and judgment standards of rehabilitation experts, transforming subjective and implicit expert knowledge into an objective and reproducible quantitative assessment model. This effectively overcomes the individualized assessment bias in traditional rehabilitation assessments, achieves consistency of assessment standards among different users and at different time points, and significantly improves the credibility and repeatability of assessment results.

[0052] The supervised learning model is trained using the training dataset, and the optimization objective is to minimize the loss function between the predicted score and the expert-annotated score. The expression for the loss function is:

[0053] in, The total number of training samples, The quantitative evaluation score for model prediction. Experts are assigned scores; cross-validation is used during training to prevent overfitting, and an early stopping strategy is employed to determine the optimal model parameters; The input feature vector X of the user to be evaluated is fed into the trained quality assessment model, and the model outputs a comprehensive quantitative assessment score. This score represents the overall quality of the user's rehabilitation movements; the higher the score, the better the quality of the movements. Output the contribution weight of each input feature in the quality assessment model to the final assessment score, and identify key posture measurement indicators that affect the quality of rehabilitation movements.

[0054] In this embodiment, a quality assessment model based on supervised learning is constructed using a deep neural network, specifically as follows: The global statistical feature vector extracted from the multi-dimensional pose quantization feature set includes local joint angle statistical features, global trajectory statistical features, body stability statistical features, and motion consistency score features, and is fused with the input feature vector. As input to deep neural network architecture; The data processing procedure is as follows: Missing values ​​in the input feature vector are filled with the mean or median, and all continuous features are Z-score standardized to have zero mean and unit variance. The quality assessment model is based on a fully connected deep neural network (DNN). Input layer: The number of nodes is equal to the input feature dimension D, and it receives the preprocessed input feature vector; Hidden layers: Employ a multi-layer fully connected structure, with each layer containing 64-256 neurons, and ReLU activation function is used; Batch normalization is added after each hidden layer to accelerate convergence and stabilize training. Dropout is added after each hidden layer with a dropout rate of 0.2-0.5 to prevent overfitting. Output layer: single neuron, the activation function uses Sigmoid to map the output to the interval [0, 1], and then multiplies it by 100 to obtain the quantitative evaluation score [0, 100].

[0055] Based on the quantitative assessment score, the specific types and degrees of deviations in movement execution are identified, and targeted rehabilitation feedback guidance information is generated, including: Construct a deviation identification rule base, which contains multiple preset deviation types for each type of rehabilitation movement, and each deviation type corresponds to an abnormal pattern of one or more posture measurement indicators; The multi-dimensional pose quantization feature set The deviation of each posture measurement index is calculated by comparing it frame by frame with the standard rehabilitation movement template, including: joint angle deviation, trajectory space deviation and stability deviation. Based on the magnitude and duration of the deviation, and combined with the deviation identification rule base, the deviation type and severity level of the current action are determined. The severity level is divided into three levels: mild, moderate, and severe, based on the ratio range of the deviation amount relative to a preset threshold. In this embodiment, the constructed deviation identification rule base presets multiple deviation types for each type of rehabilitation movement. Each deviation type corresponds to an abnormal pattern of one or more posture measurement indicators. The deviation types covered by the rule base include at least: insufficient joint range of motion, joint hyperextension, abnormal movement speed, trunk compensatory displacement, body imbalance, bilateral asymmetry, movement trajectory deviation, movement interruption or tremor, and other common movement quality problems in rehabilitation clinical practice. This systematic rule design enables the present invention to cover the deviation patterns that may occur during the execution of most rehabilitation movements.

[0056] Based on the quantitative evaluation score Based on the identified types and degrees of deviation, targeted rehabilitation feedback guidance information is retrieved from a pre-set feedback information database or dynamically generated. In this embodiment, the multi-dimensional posture quantification feature set is compared frame by frame with the standard rehabilitation movement template. This frame-by-frame, multi-dimensional deviation quantification method enables deviation identification to move beyond the overall judgment level and accurately locate fine-grained information such as "which joint, at what time, and by how much." For example, the system can not only determine "insufficient shoulder joint movement" but also clearly point out that "during the rising phase of the movement in frames 15-25, the angle of the right shoulder joint is 15 degrees smaller than the standard value." This fine-grained positioning provides data support for subsequent accurate feedback.

[0057] The rehabilitation feedback guidance information is generated using a hierarchical feedback strategy. The generated rehabilitation feedback guidance information is displayed to the user through a terminal device. At the same time, the user's acceptance of the feedback and the improvement effect of subsequent actions are recorded to form a closed-loop feedback optimization mechanism.

[0058] In this embodiment, the type and severity level of the deviation are determined based on the magnitude and duration of the deviation, combined with the deviation identification rule base. In clinical rehabilitation practice, short-lived instantaneous deviations, such as slight tremors at the start and end points of a movement, are usually not clinically significant. However, persistent systemic deviations, such as a continuously smaller shoulder joint angle throughout the entire movement, are the issues that require attention and correction. By setting a duration threshold, such as more than 5 consecutive frames or accounting for more than 20% of the total movement duration, random noise and normal fluctuations can be effectively filtered out. Only persistent and significant deviations are identified and reported, avoiding excessive alarms that interfere with user training, while ensuring that deviations that truly require intervention are dealt with in a timely manner.

[0059] The rehabilitation feedback guidance information is generated using a hierarchical feedback strategy, which includes: Real-time feedback layer: Detects significant deviations in real time during action execution and triggers alarms. Phased feedback layer: After a single action or a group of actions is completed, a comprehensive evaluation report is generated, which includes deviation analysis, improvement suggestions, and demonstration of the action. Long-term feedback layer: Based on historical data from multiple rehabilitation training sessions, it generates a rehabilitation progress trend chart and dynamically adjusts the difficulty coefficient and target threshold of rehabilitation movements.

[0060] In this embodiment, rehabilitation feedback guidance information is divided into three layers: immediate feedback, phased feedback, and long-term feedback. This forms a complete time-coverage system, from millisecond-level real-time response to daily / weekly phased summaries and monthly / quarterly long-term tracking. The three layers of feedback each have their own focus and complement each other: immediate feedback ensures movement safety and real-time correction, phased feedback promotes systematic improvement and skill consolidation, and long-term feedback provides macro-trend insights and rehabilitation path optimization. This organic and synergistic architecture design enables users to receive appropriate guidance and support at every time scale of rehabilitation training.

[0061] A rehabilitation movement quantitative assessment system integrating posture measurement and feedback learning, used to implement the rehabilitation movement quantitative assessment method integrating posture measurement and feedback learning as described in any one of claims 1-8, the system comprising, Data acquisition and preprocessing module: used to acquire real-time human posture data when the user performs preset rehabilitation movements, and preprocess the real-time human posture data to extract the three-dimensional spatial coordinate sequence of key skeletal points; In this embodiment, the data acquisition and preprocessing module serves as the system's data entry point. It acquires user action data in real time through a depth camera or a monocular visual pose estimation model, and performs preprocessing operations such as noise reduction, timestamp alignment, and background subtraction, ultimately outputting a standardized three-dimensional spatial coordinate sequence.

[0062] Multi-level posture measurement module: connected to the data acquisition and preprocessing module, used to construct a multi-level posture measurement index system that integrates local joint angles, global motion trajectory and body stability based on the three-dimensional spatial coordinate sequence of the key bone points, and calculate multi-dimensional posture quantification features. In this embodiment, the multi-level posture measurement module receives a three-dimensional spatial coordinate sequence, constructs a multi-level posture measurement index system that integrates local joint angles, global motion trajectory and body stability, and outputs a multi-dimensional posture quantification feature set. The three-level index system decouples motion quality to specific joint, time point and stability dimensions.

[0063] Dynamic time warping matching module: connected to the multi-level posture measurement module, used to perform dynamic time warping matching of the multi-dimensional posture quantification features with the corresponding standard rehabilitation movement template, and calculate the movement consistency score based on morphological similarity. In this embodiment, the dynamic time warping matching module performs dynamic time warping matching between the multi-dimensional posture quantification feature set and the standard rehabilitation movement template, and outputs a movement consistency score based on morphological similarity.

[0064] Feedback learning evaluation module: It is connected to the multi-level posture measurement module and the dynamic time warping matching module respectively, and is used to construct a quality evaluation model based on feedback learning. It takes the multi-dimensional posture quantification features and the action consistency score as input, and comprehensively evaluates the execution quality of rehabilitation actions through a supervised learning mechanism to generate a quantitative evaluation score. In this embodiment, the feedback learning evaluation module takes multi-dimensional posture quantitative features and movement consistency scores as inputs, and uses supervised learning models such as deep neural networks to comprehensively evaluate the quality of rehabilitation movement execution and generate a quantitative evaluation score.

[0065] Deviation identification and feedback generation module: connected to the feedback learning and assessment module, used to identify the specific type and degree of deviation in the execution of the action based on the quantitative assessment score, and generate targeted rehabilitation feedback guidance information.

[0066] In this embodiment, the deviation identification and feedback generation module identifies the specific type and degree of deviation in the execution of the action based on the quantitative evaluation score, and generates targeted rehabilitation feedback guidance information. Through the deviation identification rule base and multi-dimensional deviation quantification, the deviation type and its severity can be accurately located.

[0067] In this embodiment, a modular architecture is adopted to decouple the complete process of quantitative assessment of rehabilitation movements into five functionally independent modules with clear interfaces. The modules communicate with each other through standardized data interfaces, such as three-dimensional spatial coordinate sequences, multi-dimensional posture quantitative feature sets, movement consistency scores, and quantitative assessment scores. The internal implementation details of the modules are transparent to the outside world.

[0068] The dynamic time warping matching module includes: A standard template library storage unit is used to pre-store standard posture quantification feature sequences for each type of rehabilitation movement; The cumulative distance calculation unit is used to calculate the minimum cumulative distance between the measured feature sequence and the standard template feature sequence using the dynamic time warping algorithm. A consistency score conversion unit is used to convert the minimum cumulative distance into an action consistency score based on morphological similarity.

[0069] In this embodiment, the standard template library storage unit pre-stores the standard posture quantitative feature sequence for each type of rehabilitation movement, forming a structured template database. The templates, constructed by rehabilitation experts through multiple demonstrations and averaging results, directly incorporate the professional judgment and experience of clinical experts. The templates are authoritative and clinically interpretable, and are particularly suitable for cold start scenarios of new movement types.

[0070] In this embodiment, common features are automatically extracted from a large number of high-quality samples to construct templates through multi-instance learning. This can eliminate individual differences and random fluctuations, and obtain statistically more representative standard templates, which is particularly suitable for template optimization after sufficient sample accumulation.

[0071] This invention utilizes a multi-level posture measurement index system, integrating three dimensions: local joint angles, global movement trajectory, and body stability. This achieves a comprehensive and refined representation of rehabilitation movement quality, overcoming the limitations of single-dimensional assessment. A dynamic time warping algorithm is introduced to eliminate individual differences in movement speed, enabling precise measurement of morphological similarity between measured movements and standard templates, significantly improving the robustness of the assessment. A supervised learning assessment model is constructed based on deep neural networks, transferring expert knowledge into objective, reproducible, and interpretable quantitative scores. Employing a three-tiered feedback strategy (immediate, phased, and long-term) combined with a closed-loop optimization mechanism, it not only provides accurate deviation identification and personalized guidance but also adaptively adjusts rehabilitation difficulty, forming a continuous evolutionary capability of "assessment-feedback-improvement-optimization." This significantly reduces rehabilitation costs and enhances the accessibility and intelligence of home-based rehabilitation.

[0072] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A quantitative assessment method for rehabilitation movements that integrates posture measurement and feedback learning, characterized in that, Includes the following steps: Real-time human posture data is collected when the user performs preset rehabilitation movements, and the real-time human posture data is preprocessed to extract the three-dimensional spatial coordinate sequence of key skeletal points. Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a multi-level posture measurement index system integrating local joint angles, global motion trajectory and body stability is constructed, and multi-dimensional posture quantification features are calculated. The multi-dimensional posture quantification features are dynamically time-warped and matched with the corresponding standard rehabilitation movement templates to calculate a movement consistency score based on morphological similarity. A quality assessment model based on feedback learning is constructed. The multi-dimensional posture quantitative features and the action consistency score are used as inputs. The execution quality of rehabilitation actions is comprehensively evaluated through a supervised learning mechanism to generate a quantitative assessment score. Based on the quantitative assessment score, the specific types and degrees of deviations in the execution of the movement are identified, and targeted rehabilitation feedback guidance information is generated.

2. The method for quantitative assessment of rehabilitation movements integrating posture measurement and feedback learning according to claim 1, characterized in that, Real-time human posture data is collected when the user performs preset rehabilitation movements, and the real-time human posture data is preprocessed to extract the three-dimensional spatial coordinate sequence of key skeletal points, including: The system uses a pose estimation model based on a depth camera to capture video frame sequences in real time as the user performs preset rehabilitation actions. Preprocessing operations such as denoising, timestamp alignment, and background subtraction are performed on the acquired video frame sequence to generate standardized human motion image data; A human skeletal key point detection algorithm is used to locate and track a preset set of human skeletal key points from the standardized human motion image data. The set of key points includes at least the shoulder, elbow, wrist, hip, knee, and ankle joints, as well as the center point of the torso. Extract the coordinates of each key point in each frame of the image in three-dimensional space to construct a time-continuous three-dimensional spatial coordinate sequence, expressed as: in, For frame index, Total number of frames For key point indexing, The total number of key points. For the first The first frame The coordinates of the key points in three-dimensional space.

3. The method for quantitative assessment of rehabilitation movements integrating posture measurement and feedback learning according to claim 2, characterized in that, Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a multi-level posture measurement index system integrating local joint angles, global motion trajectory, and body stability is constructed, including: Based on the three-dimensional spatial coordinate sequence of the key skeletal points, a local joint angle metric is constructed. This local joint angle metric calculates the spatial angle formed by three adjacent key points for each joint, and its expression is as follows: in, For the first The first frame The angle of each joint and These are the vectors of the two bone segments that make up the joint; Construct a local angle feature vector for all joint angles in time series. ; Construct a global motion trajectory measurement index, which includes: calculating the motion trajectory curves of preset representative key points in three-dimensional space; The motion trajectory curve is parameterized, and the trajectory length, trajectory curvature, rate of change of direction, and spatial offset between the trajectory and the standard template are extracted to form a global trajectory feature vector. ; A body stability metric is constructed, comprising the standard deviation of displacement at the center point, the height and distance differences between bilaterally symmetrical key points, and the range of movement of the pressure center within the support surface. These body stability metrics are then fused to form a body stability feature vector. ; The local angle feature vector Global trajectory feature vector and body stability feature vector Temporal alignment and normalization are performed, and the results are fused to form a multi-dimensional pose quantization feature set. .

4. The method for quantitative assessment of rehabilitation movements integrating posture measurement and feedback learning according to claim 3, characterized in that, The fusion constitutes a multi-dimensional pose quantization feature set Specifically: With the global trajectory feature vector Using the time axis as a reference, linear interpolation is employed to interpolate the local angle feature vector. and body stability feature vector Time sampling points are uniformly interpolated to the same level as On the same timestamp sequence, obtain time-aligned feature vectors. , and ,in ; The eigenvalues ​​of each time-aligned eigenvector are mapped to the interval [0, 1] and then normalized. A multi-level fusion weight matrix is ​​constructed, wherein differentiated fusion weight coefficients are assigned to the local angle feature vector, global trajectory feature vector, and body stability feature vector for different rehabilitation movement types or different rehabilitation stages. , and ,satisfy ; The normalized feature vector , and The multi-dimensional pose quantization feature set is constructed by weighting and concatenating the features according to the fusion weight matrix. Its expression is: in, For the first The comprehensive feature vector at each time step, This represents the total number of time steps after alignment.

5. The method for quantitative assessment of rehabilitation movements integrating posture measurement and feedback learning according to claim 4, characterized in that, The multi-dimensional posture quantification features are dynamically time-warped and matched with the corresponding standard rehabilitation movement templates to calculate a movement consistency score based on morphological similarity, including: A pre-built standard rehabilitation movement template library is constructed, which contains standard posture quantification feature sequences for each type of rehabilitation movement. , where M is the time length of the standard action sequence, and the standard template is obtained through multiple examples demonstrated by rehabilitation experts; Obtain the measured posture quantization feature sequence of the user performing rehabilitation movements. ,in The duration of the user action sequence; The measured feature sequence is calculated using the dynamic time warping algorithm. With the standard template feature sequence The minimum cumulative distance between them specifically includes: Construct an M×N cumulative distance matrix D, where the matrix element D(i,j) represents the minimum cumulative matching distance between the i-th point in the standard sequence and the j-th point in the measured sequence; Define a local distance metric function That is, the squared Euclidean distance between two eigenvectors, where This represents the standard posture quantization feature sequence for the i-th type of rehabilitation movement. Represents the measured posture quantization feature sequence of the j-th performed rehabilitation action; The cumulative distance is calculated using a recursive formula, the expression of which is: The boundary conditions are: , , ; Extract the final cumulative distance value D(M, N) from the cumulative distance matrix D to characterize the degree of overall morphological difference between the measured action and the standard action; The final cumulative distance value D(M, N) is converted into a motion consistency score based on morphological similarity. Its expression is: in, The reference normalization factor is the cumulative distance obtained by dynamically warping and matching the standard action template with itself. The action consistency score is... The value range is [0%, 100%], and the higher the score, the more consistent the action pattern is with the standard template.

6. The method for quantitative assessment of rehabilitation movements integrating posture measurement and feedback learning according to claim 5, characterized in that, A quality assessment model based on feedback learning is constructed. Using the multi-dimensional posture quantification features and the movement consistency score as input, a supervised learning mechanism is used to comprehensively evaluate the execution quality of rehabilitation movements, generating a quantitative assessment score, including: A quality assessment model based on supervised learning is constructed using a deep neural network, which integrates input feature vectors; The input feature vector is a multi-dimensional pose quantization feature set. The extracted global statistical features include: mean, variance, peak value, and range of motion of each joint angle; length, mean curvature, and mean rate of change of direction of the global trajectory. Obtain the action consistency score and its segmented scoring sequence on the time axis ,in The number of segments into which a sequence of actions is divided according to the phases of motion; Construct a labeled training dataset, wherein each sample in the training dataset contains an input feature vector of a single rehabilitation action. and their corresponding expert-annotated quality scores The expert-marked quality score is comprehensively evaluated by rehabilitation experts based on multi-dimensional evaluation criteria, which include movement accuracy, fluency, stability, and completion. The supervised learning model is trained using the training dataset, and the optimization objective is to minimize the loss function between the predicted score and the expert-annotated score. The expression for the loss function is: in, The total number of training samples, The quantitative evaluation score for model prediction. Experts are assigned scores; cross-validation is used during training to prevent overfitting, and an early stopping strategy is employed to determine the optimal model parameters; The input feature vector X of the user to be evaluated is fed into the trained quality assessment model, and the model outputs a comprehensive quantitative assessment score. This score represents the overall quality of the user's rehabilitation movements; the higher the score, the better the quality of the movements. Output the contribution weight of each input feature in the quality assessment model to the final assessment score, and identify key posture measurement indicators that affect the quality of rehabilitation movements.

7. The method for quantitative assessment of rehabilitation movements integrating posture measurement and feedback learning according to claim 6, characterized in that, Based on the quantitative assessment score, the specific types and degrees of deviations in movement execution are identified, and targeted rehabilitation feedback guidance information is generated, including: Construct a deviation identification rule base, which contains multiple preset deviation types for each type of rehabilitation movement, and each deviation type corresponds to an abnormal pattern of one or more posture measurement indicators; The multi-dimensional pose quantization feature set The deviation of each posture measurement index is calculated by comparing it frame by frame with the standard rehabilitation movement template, including: joint angle deviation, trajectory space deviation and stability deviation. Based on the magnitude and duration of the deviation, and combined with the deviation identification rule base, the deviation type and severity level of the current action are determined. The severity level is divided into three levels: mild, moderate, and severe, based on the ratio range of the deviation amount relative to a preset threshold. Based on the quantitative evaluation score Based on the identified types and degrees of deviation, targeted rehabilitation feedback guidance information is retrieved from a pre-set feedback information database or dynamically generated. The rehabilitation feedback guidance information is generated using a hierarchical feedback strategy. The generated rehabilitation feedback guidance information is displayed to the user through a terminal device. At the same time, the user's acceptance of the feedback and the improvement effect of subsequent actions are recorded to form a closed-loop feedback optimization mechanism.

8. The method for quantitative assessment of rehabilitation movements integrating posture measurement and feedback learning according to claim 7, characterized in that, The rehabilitation feedback guidance information is generated using a hierarchical feedback strategy, which includes: Real-time feedback layer: Detects significant deviations in real time during action execution and triggers alarms. Phased feedback layer: After a single action or a group of actions is completed, a comprehensive evaluation report is generated, which includes deviation analysis, improvement suggestions, and demonstration of the action. Long-term feedback layer: Based on historical data from multiple rehabilitation training sessions, it generates a rehabilitation progress trend chart and dynamically adjusts the difficulty coefficient and target threshold of rehabilitation movements.

9. A quantitative assessment system for rehabilitation movements that integrates posture measurement and feedback learning, characterized in that, The system is used to implement the rehabilitation movement quantitative assessment method integrating posture measurement and feedback learning as described in any one of claims 1-8, and includes, Data acquisition and preprocessing module: used to acquire real-time human posture data when the user performs preset rehabilitation movements, and preprocess the real-time human posture data to extract the three-dimensional spatial coordinate sequence of key skeletal points; Multi-level posture measurement module: connected to the data acquisition and preprocessing module, used to construct a multi-level posture measurement index system that integrates local joint angles, global motion trajectory and body stability based on the three-dimensional spatial coordinate sequence of the key bone points, and calculate multi-dimensional posture quantification features. Dynamic time warping matching module: connected to the multi-level posture measurement module, used to perform dynamic time warping matching of the multi-dimensional posture quantification features with the corresponding standard rehabilitation movement template, and calculate the movement consistency score based on morphological similarity. Feedback learning evaluation module: It is connected to the multi-level posture measurement module and the dynamic time warping matching module respectively, and is used to construct a quality evaluation model based on feedback learning. It takes the multi-dimensional posture quantification features and the action consistency score as input, and comprehensively evaluates the execution quality of rehabilitation actions through a supervised learning mechanism to generate a quantitative evaluation score. Deviation identification and feedback generation module: connected to the feedback learning and assessment module, used to identify the specific type and degree of deviation in the execution of the action based on the quantitative assessment score, and generate targeted rehabilitation feedback guidance information.

10. The rehabilitation movement quantitative assessment system integrating posture measurement and feedback learning according to claim 9, characterized in that, The dynamic time warping matching module includes: A standard template library storage unit is used to pre-store standard posture quantification feature sequences for each type of rehabilitation movement; The cumulative distance calculation unit is used to calculate the minimum cumulative distance between the measured feature sequence and the standard template feature sequence using the dynamic time warping algorithm. A consistency score conversion unit is used to convert the minimum cumulative distance into an action consistency score based on morphological similarity.

Images