Musculoskeletal rehabilitation posture motion analysis system based on visual recognition

Abstract

The present application relates to the field of intelligent medical rehabilitation and computer vision technology, in particular to a musculoskeletal rehabilitation posture motion analysis system based on visual recognition; comprising: a visual analysis management center, a motion feature extraction unit, a dynamics inversion unit, a compensation coupling analysis unit, a time sequence conduction unit and a rehabilitation evaluation decision unit; the geometric compliance is distinguished through explicit kinematic feature analysis; if it is compliant, the dynamics inversion unit is used for visual torque inference consistency analysis; in response to torque dispersion, compensation coupling coefficient and motion chain conduction delay acquisition analysis are carried out respectively; finally, the compensation evaluation index is calculated through multi-dimensional compensation quantitative matching analysis, and the implicit compensation or effective driving signal is generated; the present application constructs a hierarchical filtering mechanism, which makes the implicit problem explicit, and fills the gap of internal force mode monitoring in non-contact rehabilitation evaluation.

Description

Technical Field

[0001] This invention relates to the fields of intelligent medical rehabilitation and computer vision technology, specifically a musculoskeletal rehabilitation posture and movement analysis system based on visual recognition. Background Technology

[0002] Existing computer vision and image data processing technologies typically utilize visual sensors to acquire continuous image sequences, extract key points of the human skeleton using image recognition algorithms, and reconstruct limb movement trajectories in three-dimensional space. When analyzing this image data, current technologies generally employ geometric feature-based comparison methods, which determine the compliance of the pose by calculating the Euclidean distance or angular difference between the bone node to be detected and the standard template node in three-dimensional space.

[0003] This image analysis method based on explicit geometry has significant technical limitations. First, existing visual image analysis algorithms struggle to extract implicit dynamic features from pure geometric data. Due to the lack of mechanical sensor data, the visual system cannot directly acquire joint torques and muscle activation states in image sequences, making it impossible to identify image frames with seemingly standard geometric trajectories but actually extremely low dynamic efficiency or abnormal joint forces.

[0004] When processing complex human motion image sequences, traditional algorithms often focus on the analysis of the main motion vectors of the target joints, while neglecting the calculation of the interrelationship of micro-motion features in non-target related regions. This makes it difficult for image analysis systems to capture the subtle linkage signals of non-related parts, resulting in the inability to effectively remove false compliant data generated by compensatory movements when reconstructing the kinematic chain.

[0005] Existing image processing workflows typically perform computational analysis of the same depth on all captured video frames, lacking a pre-filtering mechanism based on geometric compliance. When faced with high frame rate video streams, this approach leads to a significant waste of computational resources on invalid image frames with obvious geometric deviations, reducing the system's real-time response capability for image data processing. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a musculoskeletal rehabilitation posture and movement analysis system based on visual recognition. Specifically, the technical solution of this invention includes:

[0007] Visual analysis and management center, motion feature extraction unit, dynamics inversion unit, compensatory coupling analysis unit, temporal transmission unit, and rehabilitation assessment and decision-making unit;

[0008] The visual analysis management center is configured to retrieve the visual image sequence of rehabilitation training and send the visual image sequence to the motion feature extraction unit for explicit kinematic feature analysis to obtain standard posture frames and posture frames to be analyzed. The geometric compliance of the posture frames to be analyzed is judged. If the judgment result is geometric anomaly, a posture correction command is directly generated. If the judgment result is geometric compliance, a geometric compliance signal is generated and the dynamic inversion unit is triggered.

[0009] The dynamic inversion unit is configured to perform visual torque inference consistency analysis on the kinematic parameters of the acquired attitude frames to be analyzed in response to the geometric compliance signal, and obtain a torque consistency signal or a torque discrete signal; if a torque consistency signal is obtained, it is directly transmitted to the rehabilitation assessment decision unit to generate an effective driving signal.

[0010] The compensatory coupling analysis unit and the timing transmission unit are configured to respond to the torque discrete signal, respectively perform compensatory coupling coefficient acquisition analysis on the attitude frame to be analyzed to obtain non-associated micro-motion features, and perform kinematic chain transmission delay acquisition analysis on the muscle group activation features of the attitude frame to be analyzed to obtain start-up timing features.

[0011] The rehabilitation assessment decision unit is configured to perform multidimensional compensation quantitative matching analysis on the obtained unrelated micro-motion characteristics and initiation timing characteristics, calculate the compensation assessment index, and compare the compensation assessment index with the preset compensation index threshold: if the compensation assessment index is greater than or equal to the preset compensation index threshold, a latent compensation signal is generated; if the compensation assessment index is less than the preset compensation index threshold, an effective driving signal is generated.

[0012] Preferably, the explicit kinematic feature analysis is configured to perform the following steps:

[0013] The motion capture period of the visual image sequence is used as the time threshold. Each skeletal key point in the visual image sequence is extracted as node data. The trajectory coordinates of each node data within the time threshold in three-dimensional space are obtained as spatial displacement information.

[0014] Geometric angle calculations are performed on the spatial displacement information to obtain the spatial deviation angle;

[0015] If the calculated joint angle is greater than the preset physiological angle threshold, a geometric abnormality signal is generated;

[0016] If the calculated joint angle is less than or equal to the preset physiological angle threshold, a geometric compliance signal is generated, and the visual image sequence corresponding to the geometric compliance signal is set as the posture frame to be analyzed.

[0017] Preferably, the consistency analysis process for visual moment inference is as follows:

[0018] The visual kinematic parameters of the target joint in the posture frame to be analyzed are obtained. The visual kinematic parameters include joint angle values, angular velocity values, angular acceleration values, distal limb configuration vectors, and human inertial scalars. Based on the preset inverse dynamics mapping model, the theoretical joint torque values ​​are calculated using the visual kinematic parameters.

[0019] At the same time, the standard torque threshold range in the preset benchmark biomechanical model is retrieved, and the theoretical joint torque value is compared and analyzed with the standard torque threshold range;

[0020] If the theoretical joint torque value is within the standard torque threshold range, a torque consistency signal is obtained, and an effective drive signal is directly generated based on this torque consistency signal; if the theoretical joint torque value exceeds the standard torque threshold range, it is determined to be an abnormal force application mode, and a torque discrete signal is directly generated.

[0021] Preferably, the process for obtaining and analyzing the compensatory coupling coefficient is as follows:

[0022] The main motion vector of the target joint in the posture frame to be analyzed is obtained, and the micro motion vector of the non-target associated joint in the posture frame to be analyzed is also obtained. The non-target associated joint is a joint that has a biomechanical connection with the target joint but should remain stationary in the standard movement.

[0023] The main motion vector and the micro-motion vector are converted into a scalar sequence of instantaneous velocity amplitude. The correlation between the main motion velocity scalar sequence and the micro-motion velocity scalar sequence is calculated. The correlation between the main motion velocity scalar sequence and the micro-motion velocity scalar sequence is set as the compensatory coupling coefficient, and the compensatory coupling coefficient is set as the non-correlated micro-motion feature.

[0024] Preferably, the process for obtaining and analyzing kinetic chain conduction delay is as follows:

[0025] The visual activation moments of the proximal core muscle group region in the posture frame to be analyzed are obtained, and the response moments of displacement of the distal limbs in the posture frame to be analyzed are also obtained.

[0026] The value obtained by subtracting the visual activation time from the response time is set as the kinematic chain propagation delay duration, and the kinematic chain propagation delay duration is set as the startup timing feature.

[0027] Preferably, the multidimensional compensation quantitative matching analysis is configured to perform the following steps:

[0028] Obtain the compensatory coupling coefficient and the kinematic chain propagation delay;

[0029] The compensatory coupling coefficient is compared with the preset coupling threshold, and the kinematic chain transmission delay time is compared with the preset delay threshold.

[0030] Filter out the compensatory coupling coefficients whose values ​​are greater than or equal to the corresponding preset coupling thresholds, and the motion chain propagation delay durations whose values ​​are greater than or equal to the corresponding preset delay thresholds;

[0031] The selected compensatory coupling coefficient and the kinematic chain transmission delay time are weighted and calculated to obtain the compensatory evaluation index.

[0032] The compensation assessment index is compared with the preset compensation index threshold: if the compensation assessment index is greater than or equal to the preset compensation index threshold, it is determined that there is implicit compensation and an implicit compensation signal is generated; if the compensation assessment index is less than the preset compensation index threshold, it is determined that there is an effective driving force and an effective driving signal is generated.

[0033] Preferably, the process for obtaining the compensation assessment index is as follows:

[0034] The selected compensatory coupling coefficients are normalized to obtain coupling strength values, and the selected kinematic chain propagation delay times are normalized to obtain timing anomaly values.

[0035] The coupling component is obtained by multiplying the coupling strength value by the preset coupling weight factor.

[0036] The time series components are obtained by multiplying the time series outliers by the preset time series weighting factors.

[0037] Calculate the sum of the coupling component and the time-series component, and set this sum as the compensation evaluation index.

[0038] Preferably, the inverse dynamical mapping model is constructed in the following way:

[0039] Multiple sets of sample data containing synchronized visual image data and force table mechanical data are collected. Spatiotemporal features in the visual image data are extracted as input, and joint torques in the force table mechanical data are extracted as output. A nonlinear mapping relationship from visual kinematic parameters to theoretical joint torque values ​​is established through deep learning network training.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. This invention constructs a hierarchical filtering mechanism from explicit geometric form to implicit mechanical essence, significantly improving the real-time response speed and computing efficiency of rehabilitation video stream processing. Addressing the resource waste caused by the equal calculation of all action frames in existing technologies, this solution prioritizes geometric compliance judgment based on spatial deviation angles, triggering computationally intensive dynamic inversion analysis only when the patient's movement trajectory meets basic standards. This geometry-first-mechanical-second strategy effectively avoids meaningless complex mechanical calculations for obviously erroneous movements such as insufficient amplitude or severe distortion, ensuring the system's smooth operation at high frame rates.

[0042] 2. This invention utilizes deep learning dynamics inversion technology to overcome the limitations of traditional visual assessment, which can only observe position, and achieves non-contact perspective on muscle exertion status. By establishing an inverse dynamics mapping model that includes distal limb configuration vectors and human inertial parameters, the system can accurately infer theoretical joint torques from visual motion parameters in the absence of mechanical sensors. This enables the system to identify high-tension pathological movement patterns that appear to have standard geometric postures but are actually extremely inefficient in terms of mechanics or have abnormal joint stress, filling the gap in monitoring internal force exertion patterns in non-contact rehabilitation assessment.

[0043] 3. This invention employs a compensatory coupling analysis method based on interrelation, solving the technical problem that traditional qualitative observation is unable to quantify subtle linkage compensation. By performing statistical correlation calculations at the velocity level between the main motion vector of the target joint and the micro-motion vector of the non-target associated joints, the system can accurately capture micro-motion features of non-associated parts that are difficult to detect with the naked eye, such as shoulder shrugs and pelvic rotations. This method makes implicit synergistic movements explicit, effectively preventing misjudgments in rehabilitation assessments caused by patients forcibly completing movements using incorrect muscle groups.

[0044] 4. This invention introduces kinetic chain conduction delay analysis and a multidimensional compensatory quantitative decision-making mechanism, enabling in-depth assessment of neural control ability and core stability. By quantifying the time difference between proximal core muscle activation and distal limb response, the system can keenly distinguish between active controlled movement and swinging compensatory movement, and identify deep-seated problems such as delayed activation of core muscles. At the same time, the comprehensive compensatory assessment index generated by combining spatial coupling strength and temporal anomalies can flexibly adapt to different rehabilitation focuses in orthopedics and neurology, providing precise quantitative compensatory indicators for clinical practice. Attached Figure Description

[0045] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0046] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0048] Example 1:

[0049] Please see Figure 1This embodiment provides a musculoskeletal rehabilitation posture and motion analysis system based on visual recognition. The system includes a visual analysis management center, a motion feature extraction unit, a dynamics inversion unit, a compensatory coupling analysis unit, a temporal transmission unit, and a rehabilitation assessment and decision-making unit. The visual analysis management center is configured to retrieve visual image sequences from rehabilitation training and send these sequences to the motion feature extraction unit for explicit kinematic feature analysis, obtaining standard posture frames and posture frames to be analyzed. The geometric compliance of the posture frames to be analyzed is then judged. In response to a geometric anomaly, the system directly generates a posture correction command; in response to a geometric compliance result, the system generates a geometric compliance signal and triggers the dynamics inversion unit.

[0050] The dynamic inversion unit is configured to perform visual torque inference consistency analysis on the kinematic parameters of the acquired attitude frames to be analyzed in response to the geometric compliance signal, and obtain a torque consistency signal or a torque discrete signal; if a torque consistency signal is obtained, it is directly transmitted to the rehabilitation assessment decision unit to generate an effective driving signal.

[0051] The compensatory coupling analysis unit and the timing transmission unit are configured to respond to the torque discrete signal, respectively perform compensatory coupling coefficient acquisition analysis on the attitude frame to be analyzed to obtain non-associated micro-motion features, and perform kinematic chain transmission delay acquisition analysis on the muscle group activation features of the attitude frame to be analyzed to obtain start-up timing features.

[0052] The rehabilitation assessment decision unit is configured to perform multidimensional compensation quantitative matching analysis on the obtained unrelated micro-motion characteristics and initiation timing characteristics, calculate the compensation assessment index, and compare the compensation assessment index with the preset compensation index threshold: in response to the compensation assessment index being greater than or equal to the preset compensation index threshold, a latent compensation signal is generated; in response to the compensation assessment index being less than the preset compensation index threshold, an effective driving signal is generated.

[0053] To address the technical pain point of existing technologies that only focus on whether the geometric position of the limbs meets the standard, while ignoring whether the muscle force pattern behind the movement is correct, this system constructs a progressive analysis architecture from explicit geometric form to implicit mechanical essence; the visual analysis management center, as the system's central control point and data flow hub, is not only responsible for retrieving the visual image sequence of rehabilitation training, i.e., RGB-D depth image stream, but also undertakes the responsibility of primary filtering;

[0054] Through explicit kinematic feature analysis, the system not only solves the patient's skeletal data from the image stream to construct the posture frame to be analyzed, but also dynamically maps and generates the corresponding standard posture frame from the standard rehabilitation action library by identifying the category and phase of the current action, thus realizing the process of obtaining the standard posture frame; in view of the functional limitation of mapping and generation, this embodiment specifically adopts the Dynamic Time Warping (DTW) algorithm to achieve spatiotemporal alignment.

[0055] The system calculates the Euclidean distance matrix between the pose frame sequence to be analyzed and the template sequences in the standard action library, and searches for a regularized path with the minimum cumulative distance to determine the current pose. The specific frame in the standard library corresponding to the frame to be analyzed at a given time is used as the standard attitude frame for that time.

[0056] This method effectively solves the problem of timing misalignment with the standard library caused by the varying speed of patients' movements, ensuring the accuracy of the benchmark for subsequent comparisons. Using the standard posture frame as a geometric reference benchmark, the spatial Euclidean distance of each skeletal node in the posture frame to be analyzed relative to the corresponding node in the standard posture frame is calculated to quantify the differences and perform geometric compliance judgment. If the patient cannot even perform basic movements or limb positions, it is judged as a geometric abnormality, and the system outputs posture correction instructions such as raising the target arm. At this time, no computing power is required for subsequent advanced mechanical analysis.

[0057] Subsequent in-depth analysis is triggered only when the patient's movements are determined to be geometrically compliant. The dynamic inversion unit, as one of the core innovations of this system, performs a mapping between vision and mechanics. That is, without wearing mechanical sensors, it infers joint torques from visual motion parameters to determine whether there is torque dispersion, i.e., the movement conforms to geometric standards but has extremely low mechanical efficiency or abnormal force. The compensatory coupling analysis unit is used to capture the micro-movement characteristics of non-target joints, such as the shoulder shrugging during elbow flexion exercises, to solve spatial compensation problems.

[0058] The temporal transmission unit analyzes the initiation sequence of the kinetic chain, determines whether there is delayed activation of the core muscle groups, and addresses temporal compensation issues. The rehabilitation assessment decision unit, as the final judgment module, generates quantitative indicators based on multidimensional data to distinguish between effective drivers and implicit compensations. The specific analysis process of the temporal transmission unit is as follows: the system first extracts the moment when the proximal core muscle group region begins to generate displacement exceeding a preset rest threshold from the posture frame to be analyzed, and sets this as the visual activation moment. Simultaneously, the response time of effective displacement of the distal limb is acquired. Then the conduction delay of the kinematic chain was calculated. And set this duration as the startup timing characteristic; if If the movement falls within the preset physiological normal range, it is considered coordinated movement. If the value is close to 0, negative, or exceeds the upper limit of the normal physiological range, it is judged as an abnormal motor control strategy; among them, the proximal core muscle group area may be the key point of the abdomen or back, and the distal limb may be the key point of the wrist or ankle.

[0059] This embodiment constructs a hierarchical filtering mechanism, prioritizing geometry over mechanics, avoiding meaningless and complex mechanical calculations for obviously erroneous movements. This significantly improves the system's real-time response speed when processing high frame rate rehabilitation video streams. Simultaneously, this approach makes latent problems explicit, identifying hidden rehabilitation risks such as postures that appear standard but actually involve incorrect muscle activation, like common common movement patterns in stroke patients. This fills the gap in monitoring intrinsic force patterns in non-contact rehabilitation assessment.

[0060] Example 2:

[0061] This embodiment details the specific execution logic of explicit kinematic feature analysis. The analysis process is configured to perform the following steps: acquiring motion capture time periods of visual image sequences as time thresholds, extracting key skeletal points from the visual image sequences as node data, and obtaining the trajectory coordinates of each node data within the time threshold in three-dimensional space as spatial displacement information; performing geometric angle calculations on the spatial displacement information to obtain spatial deviation angles; generating geometric anomaly signals in response to calculated joint angles being greater than preset physiological angle thresholds; generating geometric compliance signals in response to calculated joint angles being less than or equal to preset physiological angle thresholds, and setting the visual image sequence corresponding to the geometric compliance signal as the posture frame to be analyzed.

[0062] System settings for motion capture time period As a time threshold, a pose estimation algorithm is used to extract skeletal keypoints from the visual image sequence to obtain node data. In this embodiment, a specific physical definition of the joint angle is defined; that is, the joint angle here specifically refers to the spatial deviation angle between the frame to be analyzed and the standard frame; it should be clarified that the joint angle in this embodiment is physically defined as the angle of spatial deviation between the limb vector of the pose frame to be analyzed and the corresponding limb vector of the standard pose frame. The calculation formula is as follows:

[0063]

[0064] in, For the patient's current limb vector, The limb vectors for standard movements in phase; Represents the magnitude of the vector; To prevent the preset minimum value where the denominator is zero, for example It is configured to be added to the denominator, that is, to add the minimum value after the denominator value is calculated to avoid division by zero error;

[0065] Accordingly, the preset physiological angle threshold is specified here as the maximum permissible deviation threshold. For example, 30 degrees; based on this definition, the discrimination logic fully conforms to the technical intention of hierarchical filtering: when the calculated joint angle, i.e. the deviation angle, is greater than the preset physiological angle threshold, it indicates that the patient's movement is seriously distorted or the amplitude is seriously insufficient. For example, the standard movement requires raising the hand 90 degrees, but the patient only raises it 10 degrees. The vector angle between the two is extremely large. At this time, a geometric abnormality signal is generated and directly enters the correction process to avoid invalid calculation.

[0066] If the calculated joint angle is less than or equal to the preset physiological angle threshold, it indicates that the patient's movement trajectory is basically consistent, a geometric compliance signal is generated, and the visual image sequence corresponding to the geometric compliance signal is set as the posture frame to be analyzed and enters the subsequent dynamic analysis.

[0067] This embodiment corrects the contradiction in the original logic regarding compliance judgment by redefining the calculation benchmark of geometric angles, ensuring that the system can effectively eliminate invalid actions, including those with insufficient amplitude, while also guaranteeing the geometric regularity of subsequent dynamic inversion inputs.

[0068] Example 3:

[0069] This embodiment details the operation mechanism of the dynamic inversion unit and the construction method of the inverse dynamic mapping model. The inverse dynamic mapping model is constructed in the following way: multiple sets of sample data containing synchronized visual image data and force table mechanical data are collected, the spatiotemporal features in the visual image data are extracted as input, the joint torques in the force table mechanical data are extracted as output, and a nonlinear mapping relationship from visual kinematic parameters to theoretical joint torque values ​​is established through deep learning network training.

[0070] To address the nonlinear challenge of inferring internal forces from visual appearances, this embodiment constructs a deep neural network model; data acquisition is performed to obtain a synchronous dataset. To ensure the reproducibility of the model and the correctness of its principles, this embodiment corrects the defect in the original scheme where missing input parameters prevented the unique mapping of the true torque. According to the principles of rigid body dynamics, the driving torque of a joint depends not only on its local motion state but also on the decisive influence of the distal limb configuration, i.e., the distribution of rotational inertia. For example, during shoulder flexion and extension, whether the elbow joint is straightened or bent will significantly change the rotational inertia of the entire arm around the shoulder joint. Therefore, this embodiment adds a distal limb configuration vector to the input layer.

[0071] Specifically, the network input layer has 1 node The number of remotely associated joints, carrying the current moment. eigenvectors ;

[0072] in, This corresponds to the angle, angular velocity, and angular acceleration of the target joint at a single frame moment; This refers to the set of angles of the distal associated joints, such as when analyzing the shoulder joint. Including the angles of the elbow and wrist joints, used to describe the real-time mass distribution and changes in rotational inertia of the system; For composite human inertial scalar , used to introduce individualized quality dimensions;

[0073] formula Characteristic quantities constructed based on the principle of rigid body rotational inertia Although it is not an analytical solution for the moment of inertia in rigid body dynamics, in deep learning networks, this feature can effectively characterize the nonlinear influence of distal limb mass distribution on joint torque, and a stable mapping relationship can be established through training convergence.

[0074] in, The unit of human body mass is kg. Height is measured in meters (m). This is the body shape inertia correction factor; this factor The method of obtaining the BMI is as follows: the system directly calculates the BMI value using the input mass M and height H, that is... ;when At that time, the inertia correction factor Values ;when When the value exceeds this range, it is determined according to the preset body shape-inertia lookup table. The specific values ​​and corresponding relationships are as follows: when <18.5, The value is 0.92; when 24 < At 28:00, The value is 1.08; when >28 o'clock, The value is set to 1.15 to correct the nonlinear effect of limb mass distribution on rotational inertia under different body sizes;

[0075] By introducing necessary parameters to describe the dynamic inertial distribution of the system and clearly defined human inertial parameters The model successfully constructed a fully bijective mathematical relationship from visual kinematic parameters to joint torques, solving the problem that the model could not converge due to the lack of state variables. The network contains three hidden layers with the number of nodes designed to be 64, 128 and 64 respectively. The ReLU activation function is used between layers to fit nonlinear features, and the output layer has one node to output the predicted torque.

[0076] The consistency analysis process for visual torque inference is as follows: Visual kinematic parameters of the target joint in the posture frame to be analyzed are obtained. These parameters include joint angle, angular velocity, and angular acceleration values, distal limb configuration vector, and human inertial scalar. Based on a preset inverse dynamics mapping model, the theoretical joint torque value is calculated using the visual kinematic parameters. Simultaneously, the standard torque threshold range in a preset benchmark biomechanical model is retrieved, and the theoretical joint torque value is compared with the standard torque threshold range. If the theoretical joint torque value is within the standard torque threshold range, a torque consistency signal and a torque discrete signal are obtained. In the actual operation phase, the system calculates the angular velocity of the target joint based on the posture frame to be analyzed. and angular acceleration The calculation formula is as follows:

[0077]

[0078] It should be noted that, since the visually acquired data is a discrete-time series, the above differential operations are performed numerically using the second-order central difference method, i.e.:

[0079]

[0080] in, The video frame sampling interval;

[0081] When inputting the inverse dynamics model, in addition to the aforementioned visual kinematic parameters, the system will simultaneously extract the distal limb configuration vector. and human body parameters This allows for the construction of a complete input vector; the trained inverse dynamics mapping model is then utilized. The theoretical joint torque value was calculated. The calculation formula is as follows:

[0082]

[0083] The system retrieves the corresponding standard torque threshold range from the benchmark biomechanical model. It should be noted that the baseline biomechanical model is a parameterized statistical model constructed based on a large amount of pre-collected biomechanical data of healthy individuals under standard rehabilitation movements, combined with height, weight, and BMI index. If the theoretical joint torque value is within the standard torque threshold range, it indicates that the force application pattern is normal, and an effective driving signal is directly generated based on this consistent torque signal. If the theoretical joint torque value exceeds the standard torque threshold range, it indicates that the patient may have used excessive force to maintain a seemingly simple movement, generating a discrete torque signal.

[0084] This embodiment utilizes deep learning inversion technology to overcome the limitations of traditional vision, which can only observe position, enabling the system to see through the state of muscle exertion. By establishing a nonlinear mapping between visual kinematic parameters and joint torques, it can accurately identify pathological movement patterns with high tension and low efficiency without contact.

[0085] Example 4:

[0086] This embodiment details the process of obtaining and analyzing the compensatory coupling coefficient, which is as follows: The main motion vector of the target joint in the attitude frame to be analyzed is obtained, and the micro-motion vectors of the non-target associated joints in the attitude frame to be analyzed are also obtained. The non-target associated joints are joints that have a biomechanical connection with the target joint but should remain stationary in standard movements. The main motion vector and the micro-motion vector are converted into a scalar sequence of instantaneous velocity amplitude. The cross-correlation degree of the main motion velocity scalar sequence and the micro-motion velocity scalar sequence is calculated. The cross-correlation degree of the main motion velocity scalar sequence and the micro-motion velocity scalar sequence is set as the compensatory coupling coefficient, and the compensatory coupling coefficient is set as the non-associated micro-motion feature.

[0087] System defines the main motion vector Define the displacement vector in space for the target joint, such as the affected shoulder joint; define the micro-motion vector. For non-target related joints, such as the displacement vectors of the torso or contralateral shoulder that should remain stationary within the same time period; before substituting into the following cross-correlation calculation formula, in order to resolve the contradiction between the variable definition and the subsequent parameters in physical meaning, the system pre-performs differentiation processing: based on the sampling time interval. , the three-dimensional displacement vector and Converted to an instantaneous velocity scalar sequence, the specific calculation formula is revised as follows:

[0088]

[0089]

[0090] This step ensures that the physical quantity that reflects the speed of the action is substituted into the formula, thus avoiding the failure of correlation analysis caused by using position data.

[0091] To quantify the degree of correlation between the two, this embodiment introduces a cross-correlation calculation formula, performs normalized cross-correlation calculation on the time series of the two vectors, and obtains the compensatory coupling coefficient. To address the logical trap of the original formula having a zero denominator when the non-target joint is in an ideal static state, this embodiment introduces a minimal regularization term into the denominator. The revised calculation formula is as follows:

[0092]

[0093] in, The calculated compensatory coupling coefficient, physically representing the motion synchronization between the non-target joint and the target joint, has a value range of... Dimensionless unit; To represent the total number of frames in the attitude frame sequence to be analyzed; For frame indexing; It is a scalar of instantaneous velocity amplitude; This is the arithmetic mean of the corresponding sequence; For example, a regularization constant. It is configured to be added to the denominator to prevent the denominator from being zero, i.e. To prevent program crashes caused by time-related events, ensure the logical consistency of the algorithm under ideal recovery conditions;

[0094] This embodiment calculates the statistical correlation between non-related parts and the main moving parts, which can accurately capture subtle linkages that are difficult to detect with the naked eye, such as the slight rotation of the pelvis when raising the leg. The closer the coefficient value is to 1, the more serious the linkage compensation phenomenon is, thereby accurately locating the source of compensation and solving the technical problem that traditional qualitative observation cannot quantify subtle compensatory movements.

[0095] Example 5:

[0096] This embodiment details the process of obtaining and analyzing kinetic chain conduction delay, which is as follows: the visual activation time of the proximal core muscle group region in the posture frame to be analyzed is obtained, and the response time of the distal limb displacement in the posture frame to be analyzed is obtained. The value obtained by subtracting the visual activation time from the response time is set as the kinetic chain conduction delay duration, and the kinetic chain conduction delay duration is set as the start-up timing feature.

[0097] The system extracts key moments and defines visual activation moments. The moment when the proximal core muscle group, such as the abdomen or back, begins to produce displacement exceeding the resting threshold is defined as the response moment. For distal limbs, such as the wrist or ankle, the moment of effective displacement is calculated; the duration of kinetic chain conduction delay is also calculated. The calculation formula is as follows:

[0098]

[0099] in, The calculated delay duration, in physical terms, is the time difference between the transmission of a motor command from the core to the limbs, measured in milliseconds. In normal open-chain movements, the core muscles typically activate before the limbs. And within a specific physiological range; this range is preset based on exercise statistics of healthy individuals, for example, 50ms to 200ms; responding to Abnormalities, such as values ​​close to 0 or even negative, or excessive delays, indicate abnormal motion control strategies.

[0100] This embodiment effectively distinguishes between active controlled movement and swinging compensatory movement by quantifying the initiation time difference between proximal stabilizing muscles and distal motor muscles. The extraction of this feature value provides a quantitative basis for the assessment of neural control ability, especially in the context of stroke rehabilitation, and can keenly capture the pathological compensatory strategies adopted by patients due to insufficient core stability.

[0101] Example 6:

[0102] This embodiment details how the rehabilitation assessment decision unit integrates multidimensional features to generate the final assessment result. The multidimensional compensation quantitative matching analysis is configured to perform the following steps: obtain the compensation coupling coefficient and the kinetic chain conduction delay duration; compare the compensation coupling coefficient with a preset coupling threshold, and compare the kinetic chain conduction delay duration with a preset delay threshold;

[0103] The compensatory coupling coefficients with values ​​greater than or equal to the corresponding preset coupling thresholds, and the kinematic chain propagation delay durations with values ​​greater than or equal to the corresponding preset delay thresholds are selected; the selected compensatory coupling coefficients and kinematic chain propagation delay durations are weighted and calculated to obtain the compensatory evaluation index;

[0104] The compensation assessment index is compared with the preset compensation index threshold: if the compensation assessment index is greater than or equal to the preset compensation index threshold, it is determined that there is implicit compensation and an implicit compensation signal is generated; if the compensation assessment index is less than the preset compensation index threshold, it is determined that there is an effective driving force and an effective driving signal is generated.

[0105] The data obtained in the previous steps are filtered using a threshold, retaining only the data that is significantly abnormal, i.e., filtered out... The compensatory coupling coefficient, and The kinematic chain conduction delay time, here symbol The threshold for defined temporal delay is related to the aforementioned motion capture period. Symbols are distinguished to eliminate ambiguity in physical meaning; to address the issue of missing variables in subsequent weighting formulas due to a single indicator failing to meet the standard during the screening process, this embodiment employs zero-value filling and conditional activation logic.

[0106] Specifically, if a certain characteristic, for example If the value is less than the preset threshold and is filtered by the filtering logic, the system will not remove it from the calculation process, but will mark its corresponding normalized value as 0 and participate in subsequent calculations. The specific process of obtaining the compensation evaluation index is as follows: the selected compensation coupling coefficient is normalized to obtain the coupling strength value, and the selected kinematic chain transmission delay time is normalized to obtain the timing anomaly value.

[0107] The coupling component is obtained by multiplying the coupling strength value by a preset coupling weight factor; the time-series component is obtained by multiplying the time-series outlier by a preset time-series weight factor; the sum of the coupling component and the time-series component is calculated and set as the compensation assessment index; in order to fuse features of different dimensions, this embodiment uses a weighted calculation method to generate the compensation assessment index. Perform normalization; if the feature value fails the screening, set it directly. If the feature value passes the screening, its calculation formula is:

[0108]

[0109]

[0110] in, The coupling strength value is derived from the normalization of the compensatory coupling coefficient. Its physical meaning is the severity of spatial compensation, and it is dimensionless. This indicates taking the smaller value between the calculated value and 1 to prevent timing delays. Much larger hour, An excessively large value leads to an imbalance in weights. It is a time series outlier, and it is specified that when hour, It originates from the normalization of the delay duration, and its physical meaning is the degree of abnormality in time control; it is dimensionless. The normalized boundary constant for the pre-defined compensatory coupling coefficient is dimensionless. The normalized boundary constant for the pre-defined kinematic chain conduction delay time is denoted as , where Corresponding to the lower limit of physiological normal delay, Weighted summation is performed based on the upper limit of pathological delay:

[0111]

[0112] in, This is a preset coupling weight factor, for example, 0.6, which physically represents the degree to which an action emphasizes spatial stability. This is a preset temporal weighting factor, for example, 0.4. Its physical meaning is the requirement of the action on the timing of neural control, and it satisfies... , The final calculated compensation assessment index; the system executes the decision output, which will... Compared with the preset compensation index threshold Perform a comparison; if If implicit compensation is detected, the system generates an implicit compensation signal, such as triggering a red alert and marking the compensation site; if If the signal is deemed valid, a valid drive signal is generated.

[0113] This embodiment uses a multi-dimensional fusion algorithm to unify spatial compensatory motion in the spatial dimension and temporal anomalies in the temporal dimension into a single quantitative indicator, avoiding the one-sidedness of evaluation by a single indicator; at the same time, it introduces an adjustable weighting factor. and The algorithm can flexibly adapt to different types of rehabilitation training, such as orthopedic rehabilitation which focuses on spatial angles and neurological rehabilitation which focuses on temporal control, thereby achieving personalized and high-precision compensatory assessment in complex clinical rehabilitation scenarios.

[0114] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A musculoskeletal rehabilitation posture and movement analysis system based on visual recognition, characterized in that, It includes a visual analysis and management center, a motion feature extraction unit, a dynamics inversion unit, a compensatory coupling analysis unit, a temporal transmission unit, and a rehabilitation assessment and decision-making unit; The visual analysis management center is configured to retrieve visual image sequences from rehabilitation training and send the visual image sequences to the motion feature extraction unit for explicit kinematic feature analysis to obtain standard posture frames and posture frames to be analyzed. The geometric compliance of the posture frames to be analyzed is then judged. If the judgment result is geometric anomaly, a posture correction command is directly generated. If the judgment result is geometric compliance, a geometric compliance signal is generated and the dynamic inversion unit is triggered. The dynamic inversion unit is configured to perform visual torque inference consistency analysis on the kinematic parameters of the acquired attitude frame to be analyzed in response to the geometric compliance signal, so as to obtain a torque consistency signal or a torque discrete signal. If a torque-consistent signal is obtained, it is directly transmitted to the rehabilitation assessment decision unit to generate an effective driving signal; The compensatory coupling analysis unit and the timing transmission unit are configured to respond to the torque discrete signal, respectively perform compensatory coupling coefficient acquisition analysis on the attitude frame to be analyzed to obtain non-associated micro-motion features, and perform kinematic chain transmission delay acquisition analysis on the muscle group activation features of the attitude frame to be analyzed to obtain start-up timing features. The rehabilitation assessment decision unit is configured to perform multidimensional compensation quantitative matching analysis on the obtained unrelated micro-motion characteristics and initiation timing characteristics, calculate the compensation assessment index, and compare the compensation assessment index with a preset compensation index threshold: if the compensation assessment index is greater than or equal to the preset compensation index threshold, a latent compensation signal is generated; if the compensation assessment index is less than the preset compensation index threshold, an effective driving signal is generated. The consistency analysis process for visual moment inference is as follows: The visual kinematic parameters of the target joint in the posture frame to be analyzed are obtained. The visual kinematic parameters include joint angle values, angular velocity values, angular acceleration values, distal limb configuration vectors, and human inertial scalars. Based on the preset inverse dynamics mapping model, the theoretical joint torque values ​​are calculated using the visual kinematic parameters. At the same time, the standard torque threshold range in the preset benchmark biomechanical model is retrieved, and the theoretical joint torque value is compared and analyzed with the standard torque threshold range; If the theoretical joint torque value is within the standard torque threshold range, a torque consistency signal is obtained, and an effective drive signal is directly generated based on this torque consistency signal. If the theoretical joint torque value exceeds the standard torque threshold range, a torque discrete signal is obtained. The process for obtaining and analyzing the compensatory coupling coefficient is as follows: The main motion vector of the target joint in the posture frame to be analyzed is obtained, and the micro motion vector of the non-target associated joint in the posture frame to be analyzed is also obtained. The non-target associated joint is a joint that has a biomechanical connection with the target joint but should remain stationary in the standard movement. The main motion vector and the micro motion vector are converted into a scalar sequence of instantaneous velocity amplitude. The cross-correlation degree of the main motion velocity scalar sequence and the micro motion velocity scalar sequence is calculated. The cross-correlation degree of the main motion velocity scalar sequence and the micro motion velocity scalar sequence is set as the compensatory coupling coefficient, and the compensatory coupling coefficient is set as the non-correlated micro motion feature. The inverse dynamic mapping model is constructed in the following way: Multiple sets of sample data containing synchronized visual image data and force table mechanical data are collected. Spatiotemporal features in the visual image data are extracted as input, and joint torques in the force table mechanical data are extracted as output. A nonlinear mapping relationship from visual kinematic parameters to theoretical joint torque values ​​is established through deep learning network training. The network input layer has 1 node The number of remotely associated joints, carrying the current moment. eigenvectors ; in, This corresponds to the angle, angular velocity, and angular acceleration of the target joint at a single frame moment; The set of angles for the distal associated joints; For composite human inertial scalar ,in For human body quality, For human height, This is the body shape inertia correction factor; The body shape inertia correction coefficient pass Calculated and valued according to segmentation rules: when hour, Values ;when <18.5, The value is 0.92; when 24 < BMI BMI The value is 1.15; Using a trained inverse dynamics mapping model The theoretical joint torque value was calculated. The calculation formula is as follows: ； Based on sampling time interval , the three-dimensional displacement vector and Converted to an instantaneous velocity scalar sequence, the specific calculation formula is revised as follows: ； The specific formula for calculating the compensatory coupling coefficient is as follows: ； in, The calculated compensatory coupling coefficient, physically representing the motion synchronization between the non-target joint and the target joint, has a value range of... Dimensionless unit; To represent the total number of frames in the attitude frame sequence to be analyzed; For frame indexing; It is a scalar of instantaneous velocity amplitude; This is the arithmetic mean of the corresponding sequence; is the regularization constant.

2. The musculoskeletal rehabilitation posture and movement analysis system based on visual recognition according to claim 1, characterized in that, The explicit kinematic feature analysis is configured to perform the following steps: The motion capture period of the visual image sequence is used as the time threshold. Each skeletal key point in the visual image sequence is extracted as node data. The trajectory coordinates of each node data within the time threshold in three-dimensional space are obtained as spatial displacement information. Geometric angle calculations are performed on the spatial displacement information to obtain the spatial deviation angle; If the calculated joint angle is greater than the preset physiological angle threshold, a geometric abnormality signal is generated; If the calculated joint angle is less than or equal to the preset physiological angle threshold, a geometric compliance signal is generated, and the visual image sequence corresponding to the geometric compliance signal is set as the posture frame to be analyzed.

3. The musculoskeletal rehabilitation posture and movement analysis system based on visual recognition according to claim 1, characterized in that, The process for obtaining and analyzing the kinematic chain conduction delay is as follows: The visual activation moments of the proximal core muscle group region in the posture frame to be analyzed are obtained, and the response moments of displacement of the distal limbs in the posture frame to be analyzed are also obtained. The value obtained by subtracting the visual activation time from the response time is set as the kinematic chain propagation delay duration, and the kinematic chain propagation delay duration is set as the startup timing feature.

4. The musculoskeletal rehabilitation posture and movement analysis system based on visual recognition according to claim 1, characterized in that, The multidimensional compensation quantitative matching analysis is configured to perform the following steps: Obtain the compensatory coupling coefficient and the kinematic chain propagation delay; The compensatory coupling coefficient is compared with the preset coupling threshold, and the kinematic chain transmission delay time is compared with the preset delay threshold. Filter out the compensatory coupling coefficients whose values ​​are greater than or equal to the corresponding preset coupling thresholds, and the motion chain propagation delay durations whose values ​​are greater than or equal to the corresponding preset delay thresholds; The selected compensatory coupling coefficient and the kinematic chain transmission delay time are weighted and calculated to obtain the compensatory evaluation index. The compensation assessment index is compared with the preset compensation index threshold: if the compensation assessment index is greater than or equal to the preset compensation index threshold, it is determined that there is implicit compensation and an implicit compensation signal is generated; if the compensation assessment index is less than the preset compensation index threshold, it is determined that there is an effective driving force and an effective driving signal is generated.

5. The musculoskeletal rehabilitation posture and movement analysis system based on visual recognition according to claim 4, characterized in that, The specific process for obtaining the compensation assessment index is as follows: The selected compensatory coupling coefficients are normalized to obtain coupling strength values, and the selected kinematic chain propagation delay times are normalized to obtain timing anomaly values. The coupling component is obtained by multiplying the coupling strength value by the preset coupling weight factor. The time series components are obtained by multiplying the time series outliers by the preset time series weighting factors. Calculate the sum of the coupling component and the time-series component, and set this sum as the compensation evaluation index.

Images