Neck rehabilitation training action guiding and evaluating method and system based on visual feedback

Abstract

The invention discloses a neck rehabilitation training action guiding and evaluating method and system based on visual feedback, and belongs to the technical field of crossing of computer vision and rehabilitation medical treatment. According to the method, a face image is collected through a front camera of intelligent equipment, head and neck key points are extracted based on a multi-scale convolutional neural network, and a three-dimensional posture is estimated; a virtual guide path is adaptively generated and dynamic tracking is performed, a multi-dimensional quality evaluation model is established to quantitatively analyze the joint motion range, the speed smoothness and the compensation action, real-time feedback is provided according to an evaluation result, the training difficulty is dynamically adjusted, a closed-loop optimization mechanism is formed, and a detailed rehabilitation training report is generated. Accurate neck rehabilitation training can be achieved only through common intelligent equipment, the exercise quality quantitative evaluation and compensatory action automatic detection capabilities are achieved, personalized self-adaptive training is achieved through closed-loop feedback, and an effective solution is provided for remote rehabilitation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer vision and rehabilitation medicine, and relates to a neck rehabilitation training action guiding and evaluating method and system based on visual feedback, in particular to a neck rehabilitation training technology using an intelligent device camera to capture head and neck movements in real time, quantifying training effects through a multi-dimensional quality evaluation model, and forming a closed-loop feedback optimization mechanism. The technology is mainly applied to remote rehabilitation training, motor function evaluation and rehabilitation progress monitoring of patients with cervical diseases. BACKGROUND

[0002] Cervical disease is a common chronic disease in modern society. Factors such as long-term poor posture and excessive use of electronic devices have led to an increasing incidence of cervical disease year by year. Neck rehabilitation training is an important means of treating cervical disease. Through standardized exercise training, neck muscle strength can be improved, joint mobility can be increased, and pain symptoms can be relieved. However, traditional rehabilitation training relies on on-site guidance by a rehabilitation therapist, and there are problems such as insufficient rehabilitation resources, difficulty in training supervision, and subjective effect evaluation.

[0003] In recent years, virtual reality technology and computer vision technology have been introduced into the field of rehabilitation training, providing a new solution for rehabilitation training. Existing rehabilitation training systems mostly use dedicated depth cameras or inertial measurement units to capture human motion, but these devices are high in cost and inconvenient to use, limiting their application in home scenarios. Rehabilitation training solutions based on ordinary intelligent device cameras have gradually attracted attention, but existing technologies mainly have the following shortcomings:

[0004] First, existing technologies mostly use algorithms such as OpenPose combined with Kinect depth cameras for full-body posture estimation. For example, Chinese patent application publication CN109003301A discloses a human posture estimation method and rehabilitation training system based on OpenPose and Kinect. The system fuses two-dimensional human joint points obtained by the OpenPose algorithm with depth information from Kinect to obtain three-dimensional human joint points, which are used to control the actions of a virtual agent. However, this solution requires a dedicated Kinect device, which is inconvenient to use in a home environment, and is designed for full-body rehabilitation training without optimization for the special needs of neck rehabilitation, lacking the ability to accurately capture small movements of the head and neck.

[0005] Second, the existing technology mainly focuses on whether the action is completed, lacking quantitative evaluation of the quality of movement. Rehabilitation training not only requires completing the prescribed action, but more importantly, ensuring the quality of action, including whether the joint range of motion meets the standard, whether the movement speed is smooth, whether there is compensatory action, etc. Compensatory action refers to the patient completing the target action by mobilizing other muscle groups or changing the movement pattern when the movement function is limited. Although the action is completed on the surface, it may lead to the solidification of incorrect movement patterns, and even cause new orthopedic problems. The existing technology lacks effective detection means for compensatory action.

[0006] Third, the existing technology lacks a closed-loop feedback mechanism and cannot dynamically adjust the training difficulty according to the training effect. Rehabilitation training is a gradual process that requires gradually increasing the training intensity according to the patient's rehabilitation progress. The existing system mostly uses fixed training programs and cannot achieve personalized adaptive training.

[0007] Fourth, the rehabilitation evaluation report of the existing technology is relatively simple, mostly recording basic data such as joint angles, lacking in-depth analysis of the quality of movement, and difficult to provide valuable rehabilitation decision basis for the rehabilitation therapist.

[0008] Therefore, it is urgent to develop a neck rehabilitation training system based on ordinary intelligent devices, focusing on neck rehabilitation, having the ability to quantitatively evaluate the quality of movement, being able to automatically detect compensatory action, and realizing closed-loop feedback optimization. SUMMARY

[0009] The purpose of the present application is to solve the above-mentioned deficiencies in the prior art, and to provide a neck rehabilitation training action guidance and evaluation method and system based on visual feedback. The method and system can realize neck rehabilitation training only with the front camera of an ordinary intelligent device, accurately extract head and neck key points through a multi-scale convolutional neural network, adaptively generate a virtual guide path, establish a multi-dimensional movement quality evaluation model to quantitatively analyze joint range of motion, speed smoothness and compensatory action, form a closed-loop feedback optimization mechanism to dynamically adjust the training difficulty, and generate a detailed rehabilitation training report to provide decision basis for the rehabilitation therapist.

[0010] The present application is realized by the following technical solutions:

[0011] The method comprises the steps of acquisition, path generation, tracking evaluation, feedback optimization and report generation. The acquisition step acquires a user's face image sequence in real time through the front camera of a smart device, extracts two-dimensional image coordinates of head and neck key points based on a multi-scale convolutional neural network, estimates three-dimensional Euler angles of head posture by combining time sequence information, and generates head and neck motion trajectory data. The path generation step adaptively generates a virtual guide path according to the type of training motion preset by a rehabilitation therapist and the baseline parameters of the user's current cervical spine activity, and the virtual guide path includes a space-time coordinate sequence and a speed constraint parameter. The target path is presented in the form of a dynamic cursor on the screen. The tracking evaluation step calculates the spatial deviation and temporal deviation between the user's head cursor and the virtual guide path in real time, quantitatively analyzes the joint activity completion rate, motion speed smoothness and compensatory action intensity based on a multi-dimensional quality evaluation model, and generates a motion quality score. The feedback optimization step determines the feedback type according to the motion quality score, provides positive incentive feedback when the score is higher than a first threshold, provides corrective feedback and reduces the path difficulty coefficient when the score is lower than a second threshold, and provides real-time warning feedback when compensatory action is detected, forming a closed-loop regulation mechanism. The report generation step statistically analyzes the trajectory deviation data, quality score sequence and compensatory action occurrence frequency of a single training, and generates a rehabilitation training report containing a joint activity curve, a speed distribution histogram and a compensatory action timestamp.

[0012] The application also provides a neck rehabilitation training motion guidance and evaluation system based on visual feedback, which comprises a multi-scale head and neck feature extraction and posture estimation module, an adaptive guide path generation and dynamic tracking module, a multi-dimensional motion quality evaluation and compensation detection module, and a closed-loop feedback optimization and rehabilitation report generation module. The multi-scale head and neck feature extraction and posture estimation module acquires a user's face image sequence in real time through the front camera of a smart device, extracts two-dimensional image coordinates of head and neck key points based on a multi-scale convolutional neural network, estimates three-dimensional Euler angles of head posture by combining time sequence information, and generates head and neck motion trajectory data. The adaptive guide path generation and dynamic tracking module adaptively generates a virtual guide path according to the type of training motion preset by a rehabilitation therapist and the baseline parameters of the user's current cervical spine activity, and the virtual guide path includes a space-time coordinate sequence and a speed constraint parameter. The target path is presented in the form of a dynamic cursor on the screen, and the spatial deviation and temporal deviation between the user's head cursor and the virtual guide path are calculated in real time. The multi-dimensional motion quality evaluation and compensation detection module quantitatively analyzes the joint activity completion rate, motion speed smoothness and compensatory action intensity based on a multi-dimensional quality evaluation model, and generates a motion quality score. The closed-loop feedback optimization and rehabilitation report generation module determines the feedback type according to the motion quality score, provides positive incentive feedback, corrective feedback or real-time warning feedback, dynamically adjusts the path difficulty coefficient according to the score result, and statistically analyzes the training data to generate a rehabilitation training report.

[0013] The beneficial effects of the present application are:

[0014] First, the present application only needs the front camera of an ordinary intelligent device to realize neck rehabilitation training, without the need for a dedicated depth camera or an inertial measurement unit, greatly reducing the cost of the device and the threshold for use, and facilitating the promotion and application in a home scenario. The multi-scale convolutional neural network accurately extracts the head and neck key points, realizing high-precision capture of the slight movement of the head and neck.

[0015] Second, the present application establishes a multi-dimensional motion quality evaluation model, not only evaluating whether the action is completed, but also quantitatively analyzing the joint range of motion completion rate, motion speed smoothness and compensation action intensity, providing a scientific quality evaluation system for rehabilitation training. The innovative compensation action detection algorithm can automatically identify the compensation behaviors such as shoulder lifting and trunk compensation, timely correct the wrong movement mode, and avoid the solidification of compensation actions.

[0016] Third, the present application forms a closed-loop feedback optimization mechanism, automatically adjusts the path difficulty coefficient according to the action quality score, and realizes personalized adaptive training. When the user training effect is good, the difficulty is automatically increased, and when the user training effect is not good, the difficulty is automatically reduced, ensuring that the training intensity matches the user's ability, and improving the safety and effectiveness of rehabilitation training.

[0017] Fourth, the rehabilitation training report generated by the present application contains detailed information such as joint range of motion curve graph, speed distribution histogram and compensation action timestamp, and performs in-depth analysis on the motion quality, providing valuable rehabilitation decision basis for rehabilitation therapists, facilitating remote supervision and rehabilitation progress tracking.

[0018] Fifth, the present application forms a synergistic effect through the deep coupling of the four core modules. The output of the multi-scale head and neck feature extraction and posture estimation module directly drives the path tracking calculation of the adaptive guide path generation and dynamic tracking module, the tracking deviation is input in real time to the multi-dimensional motion quality evaluation and compensation detection module for quality analysis, the evaluation result is adjusted by the closed-loop feedback optimization and rehabilitation report generation module, forming a complete closed loop of forward transmission, performance evaluation, reverse feedback and parameter adjustment, realizing 1+1>2 synergistic effect, and having significant innovation compared with the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is the overall architecture schematic diagram of the system of the present application.

[0020] Figure 2 is the flowchart of the method of the present application.

[0021] Figure 3 is the multi-dimensional motion quality evaluation flowchart. DETAILED DESCRIPTION

[0022] Reference is made to the accompanying drawings Figures 1-3 and the embodiments below to further explain the present application.

[0023] As shown in Figure 1 and Figure 2 , the present application provides a neck rehabilitation training action guidance and evaluation method and system based on visual feedback. The system includes a multi-scale head and neck feature extraction and posture estimation module 1, an adaptive guidance path generation and dynamic tracking module 2, a multi-dimensional motion quality evaluation and compensation detection module 3, and a closed-loop feedback optimization and rehabilitation report generation module 4. Each module is coupled through parameters to form a complete closed-loop system, realizing intelligent rehabilitation training from motion capture, path guidance, quality evaluation to feedback optimization.

[0024] The multi-scale head and neck feature extraction and posture estimation module 1 is the data input source of the system, responsible for real-time acquisition of user face image sequences from the front camera of the intelligent device, and accurate extraction of head and neck key points and posture information through a deep learning model.

[0025] This module uses an improved lightweight convolutional neural network architecture, with the backbone network based on MobileNetV3 design optimized for the computational resource limitations of mobile devices. The network contains an initial convolutional layer, multiple inverted residual bottleneck modules, and a feature pyramid structure. The initial convolutional layer uses a 3x3 convolutional kernel with a stride of 2 to downsample the input RGB image from a resolution of 640x480 to 320x240, extracting shallow texture features. The inverted residual bottleneck module uses an expansion-depth separable convolution-compression structure, expanding the channel number through a 1x1 convolution, then performing a 3x3 depth separable convolution, and finally compressing the channel number through a 1x1 convolution, significantly reducing the computational load while maintaining the feature expression capability.

[0026] The core innovation of the present application is the introduction of a multi-scale receptive field fusion mechanism, using parallel hollow convolution branches with rates of 2, 4, and 8 in the feature extraction stage to capture context information at different scales. The branch with a rate of 2 is suitable for capturing local detail features such as eyebrow, nose tip, and other small-range key points; the branch with a rate of 4 is suitable for capturing medium-range facial structure features; and the branch with a rate of 8 is suitable for capturing global context information such as the overall contour of the head. The features of the three branches are adaptively weighted and fused through a channel attention mechanism. The channel attention module first performs global average pooling and global maximum pooling on the feature map, then learns the channel weights through a shared multi-layer perceptron, and finally applies the weights to the original feature map to strengthen the feature expression of important channels. This multi-scale fusion mechanism significantly improves the detection accuracy of head and neck key points, especially in complex scenarios such as face occlusion or light changes.

[0027] The head and neck key point detection adopts a heat map regression method. For an input image, the network outputs a heat map proportional to the input size. The value of each position in the heat map represents the probability that the position is a key point. The present application defines 15 head and neck key points, including the tip of the nose, the inner corner of the left eye, the outer corner of the left eye, the inner corner of the right eye, the outer corner of the right eye, the left corner of the mouth, the right corner of the mouth, the chin, the left earlobe, the right earlobe, the top of the head, the back of the neck, the left shoulder, the right shoulder, and the suprasternal notch. By finding local maxima on the heat map and using a sub-pixel accuracy interpolation method, the two-dimensional image coordinates of each key point are obtained , wherein represents the key point index.

[0028] The head pose estimation is based on the mapping relationship from 2D key points to 3D pose, and the PnP algorithm is used to solve the three-dimensional Euler angles of the head. First, a standard 3D head model is established, which defines the standard three-dimensional coordinates of each key point in the local coordinate system of the head. Then, by using the correspondence between the detected 2D key point coordinates and the 3D model coordinates, the rotation matrix from the camera coordinate system to the head coordinate system is estimated by solving the perspective projection equation and the translation vector . The rotation matrix can be further decomposed into three Euler angles: the pitch angle , which describes the forward and backward inclination of the head, the yaw angle , which describes the left and right turning of the head, and the roll angle , which describes the left and right inclination of the head.

[0029] To improve the stability of pose estimation, the present application introduces a time series smoothing filtering mechanism. For the Euler angles estimated in the th frame, an exponential moving average method is used for smoothing processing. The smoothed angles are obtained by weighted calculation of historical information and current observation. The weight coefficient of the exponential moving average is set to 0.7, which achieves a good balance between response speed and smoothing effect. After smoothing, the jitter of pose estimation is significantly reduced, and the motion trajectory is more continuous and smooth.

[0030] The output of this module includes the two-dimensional image coordinates of the 15 key points, the three Euler angles, and the confidence score of the key points. These data are output in real time at a frequency of 30 frames per second as input for the subsequent modules. The key point confidence score is used to judge the detection quality. When the confidence is lower than 0.6, the system will prompt the user to adjust the camera position or lighting conditions.

[0031] The adaptive guide path generation and dynamic tracking module 2 is the core guide module of the system, which is responsible for generating personalized virtual guide paths according to the user's rehabilitation needs and current ability, and tracking the deviation between the user's head cursor and the path in real time.

[0032] The module first conducts an initial assessment of the user, measuring baseline parameters of the user's cervical spine mobility. The initial assessment includes four basic movements: flexion (head forward low), extension (head backward tilt), rotation (head horizontal rotation), and lateral flexion (head tilt to shoulder). The user completes these four movements in turn under the guidance of the system, and the system records the maximum angle value of each movement. The flexion-extension mobility is defined as the absolute value of the sum of the maximum flexion angle and the maximum extension angle, with a normal value of about 130°. The left-right rotation mobility is defined as the absolute value of the sum of the maximum left rotation angle and the maximum right rotation angle, with a normal value of about 160°. The left-right lateral flexion mobility is defined as the absolute value of the sum of the maximum left lateral flexion angle and the maximum right lateral flexion angle, with a normal value of about 90°.

[0033] According to the measured mobility data, the system calculates the comprehensive mobility percentage , which reflects the gap between the user's current cervical spine function and the normal level. The comprehensive mobility percentage is obtained by dividing the measured flexion-extension, left-right rotation, and left-right lateral flexion mobilities by the corresponding normal values, and then taking the weighted average of the three, with weights of 0.4, 0.4, and 0.2, respectively, because flexion-extension and left-right rotation are used more frequently in daily life. The comprehensive mobility percentage has a value range of 0 to 1, and the closer the value is to 1, the better the cervical spine function.

[0034] The generation of the virtual guide path uses a parametric design method, and the shape, length, and speed requirements of the path are all adjusted adaptively according to the user's comprehensive mobility percentage and rehabilitation stage. The rehabilitation stage is divided into three stages: early, middle, and late. The early stage corresponds to the first 4 weeks of rehabilitation, the middle stage corresponds to 4-8 weeks of rehabilitation, and the late stage corresponds to 8 weeks of rehabilitation or more. Different rehabilitation stages correspond to different difficulty coefficients, with a difficulty coefficient of 0.5 for the early stage, 0.75 for the middle stage, and 1.0 for the late stage.

[0035] The maximum deviation angle of the path is determined according to the comprehensive mobility percentage and the difficulty coefficient, and the calculation formula is , where is the difficulty coefficient of the current rehabilitation stage, is the standard maximum deviation angle for this type of training movement. For example, for flexion-extension training, is set to 60°; for left-right rotation training, is set to 80°; and for left-right lateral flexion training, is set to 45°.

[0036] The speed constraints of the path include the minimum speed and the maximum speed , the speed range is also adaptively adjusted according to the user's ability. The minimum speed is set to 10° / s to ensure that the movement is not too slow to cause muscle fatigue; the maximum speed is determined according to the rehabilitation stage, which is 30° / s at the initial stage, 45° / s at the middle stage, and 60° / s at the later stage. The speed constraint is achieved by limiting the time interval between path points According to the angular difference between adjacent path points and the speed constraint.

[0037] The virtual guide path generates a smooth space-time trajectory curve using a cubic spline interpolation algorithm. Cubic spline interpolation ensures that the path is continuous and differentiable at key control points, avoiding sudden changes in speed and acceleration. For forward flexion and extension training, the path starts from the neutral position, first flexes forward to the maximum angle, then returns to the neutral position, then extends backward to the maximum angle, and finally returns to the neutral position, forming a complete movement cycle. Several control points are evenly distributed on the path, and a dense sequence of path points is generated between the control points through cubic spline interpolation, with a path point interval of 0.1s to ensure smooth guidance.

[0038] The virtual guide path is presented on the screen in the form of a dynamic cursor, with the target path displayed as a yellow moving circle and the real-time cursor corresponding to the user's head position displayed as a red circle. The user needs to control the red cursor to follow the yellow target path through head movement. This visual feedback form is intuitive and easy to understand, converting abstract angle control into a specific cursor following task, improving the fun and participation of the training.

[0039] Dynamic tracking calculates the deviation between the user's head cursor and the virtual guide path in real time. The spatial deviation is defined as the Euclidean distance between the cursor coordinates corresponding to the user's current head position and the target path point coordinates . The time deviation is defined as the difference between the actual time and the expected time for the user to reach a certain path point . The spatial deviation reflects the accuracy of the user's action, and the time deviation reflects the user's speed control ability.

[0040] This module outputs the real-time calculated spatial deviation and time deviation sequence to the multi-dimensional motion quality evaluation and compensation detection module 3 as an important input for quality evaluation. At the same time, this module receives path difficulty adjustment instructions from the closed-loop feedback optimization and rehabilitation report generation module 4, dynamically adjusts the path parameters according to the user's training performance, and forms a closed-loop feedback mechanism.

[0041] The multi-dimensional motion quality evaluation and compensation detection module 3 is the intelligent analysis core of the system, responsible for multi-dimensional quantitative evaluation of the user's motion quality, including joint range of motion evaluation, speed smoothness evaluation, and compensation action detection.

[0042] Joint range of motion assessment quantifies the degree to which the user's actual range of cervical spine movement matches the preset target range. For a single training session, the system records the maximum angle value achieved by the user throughout the entire movement. and minimum angle value Actual activity level Calculated as the difference between the absolute values ​​of the two. Preset target activity level. Determined by the maximum offset angle of the virtual guide path. Joint range of motion completion rate. Defined as the ratio of actual activity level to target activity level, the mathematical expression is:

[0043] ,

[0044] in, The percentage of joint range of motion completion. The actual activity level completed by the user (unit: degree). The preset target activity level (unit: degrees). When When the joint range of motion is considered to be within acceptable limits; when At that time, the range of motion of the joint was considered to be basically up to standard; when If the joint range of motion is not up to standard, the training difficulty should be reduced or additional guidance should be provided.

[0045] Speed ​​smoothness assessment quantifies the uniformity and fluidity of movement speed, a crucial indicator for judging movement quality. Ideal rehabilitation training movements should maintain a constant or near-constant speed, avoiding muscle tension and joint impact caused by sudden increases or decreases in speed. This invention proposes an adaptive smoothness index to comprehensively evaluate speed quality.

[0046] First, calculate the motion velocity sequence. For the ... Frame, head pose angle is The time interval between adjacent frames is s (corresponding to a sampling frequency of 30 frames per second), instantaneous angular velocity The calculation is the angle difference between adjacent frames divided by the time interval, i.e. For a complete training session, an angular velocity sequence can be obtained. ,in This represents the total number of training frames.

[0047] Secondly, the normalized variance of the velocity is calculated to assess the uniformity of velocity variation. (Velocity variance) The variance indicates the degree to which speed deviates from the average value; a smaller variance indicates a more uniform speed. To eliminate the influence of differences in speed benchmarks between different training movement types and different users, this invention uses normalized variance, defined as the ratio of the speed standard deviation to the speed mean, i.e., the coefficient of variation. :

[0048] ,

[0049] where, is the speed coefficient of variation (unitless), is the speed standard deviation (unit: degree / s), is the speed mean (unit: degree / s), is the instantaneous angular velocity of the i-th frame (unit: degree / s), is the total number of frames. The smaller the speed coefficient of variation, the more uniform the speed, and the smoother the movement.

[0050] Again, the normalized kurtosis of acceleration is calculated to evaluate the degree of speed mutation. The acceleration sequence is obtained by differencing the speed sequence, i.e. The kurtosis of acceleration is an index in statistics to describe the shape of the distribution. The larger the kurtosis value, the thicker the tail of the distribution, i.e. there are more extreme values, corresponding to the violent fluctuations of speed. The normalized kurtosis is defined as:

[0051] ,

[0052] where, is the normalized acceleration kurtosis (unitless), is the acceleration kurtosis, is the reference kurtosis value (set to 5.0, based on statistical analysis of normal rehabilitation training data), is the acceleration of the i-th frame (unit: degree / s²), is the acceleration mean (unit: degree / s²), is the acceleration standard deviation (unit: degree / s²), is the total number of frames. The smaller the normalized kurtosis, the more the acceleration distribution approaches the normal distribution, and the more stable the speed change. Finally, the adaptive smoothness index

[0053] is calculated by combining the speed coefficient of variation and the normalized acceleration kurtosis:

[0054] ,

[0055] where, is the adaptive smoothness index (unitless, value range is 0 to positive infinity), is the speed coefficient of variation (unitless), is the normalized acceleration kurtosis (unitless).​is the normalized acceleration kurtosis (dimensionless). The smoothness index is in the form of harmonic mean, considering both the velocity uniformity and the acceleration stationarity. When both are small, the smoothness index is large, indicating good quality of motion. For ease of understanding and presentation, the smoothness index is mapped to a percentage score between 0 and 100, with the mapping formula wherein is the velocity smoothness score (ranging from 0 to 100). When , the velocity smoothness is considered excellent; when , the velocity smoothness is considered good; when , the velocity smoothness is considered poor, and the velocity control training needs to be strengthened.

[0056] Compensation action detection is an important innovation of the present application, aiming to automatically identify whether the user has used other parts of the body to compensate when completing the neck training, so as to avoid the solidification of the wrong movement pattern. Common compensation actions for neck rehabilitation training include shoulder lifting compensation and trunk compensation.

[0057] Shoulder lifting compensation refers to the user's shoulder involuntarily lifting up when completing the neck action, using the strength of the shoulder muscles to assist in completing the head movement. This compensation action will reduce the training effect of the neck muscles, and may cause shoulder and neck muscle imbalance in the long term. The present application identifies shoulder lifting compensation by monitoring the vertical displacement of the shoulder key points. Before the training starts, the system records the initial vertical coordinates of the user's shoulder key points , which correspond to the longitudinal coordinates of the left and right shoulders, respectively. During the training process, the vertical displacement of the shoulder key points relative to the initial position is calculated in real time and . When the vertical displacement of either shoulder exceeds the compensation threshold , it is determined that shoulder lifting compensation has occurred. The compensation threshold is set according to the image resolution and camera field of view. For a standard resolution of 640x480, the compensation threshold is set to 15 pixels, corresponding to an actual displacement of about 3 cm.

[0058] Trunk compensation refers to the user's trunk following the head movement when completing the neck action, rather than only the head and neck moving relative to the trunk. This compensation action also reduces the training effect and may cause fatigue of the waist and back muscles. The present application identifies trunk compensation by analyzing the coordination of head and neck movement and trunk movement. Ideally, the trunk should remain relatively stationary during neck training, and the head and neck should move independently relative to the trunk. When the trunk movement and head movement show a high degree of correlation, it indicates that the trunk is assisting the head movement, i.e., trunk compensation has occurred.

[0059] Specifically, the trunk posture angle is defined as the angle of the suprasternal notch key point relative to the vertical direction , and the head posture angle is the pitch component of the head Euler angle For one complete training, two time series and are obtained.The Pearson correlation coefficient of the two is calculated :

[0060] ,

[0061] wherein, is the Pearson correlation coefficient of head and trunk movement (ranging from -1 to 1), is the head pitch angle of the i-th frame (in degrees), is the trunk posture angle of the i-th frame (in degrees), is the mean of the head pitch angle (in degrees), is the mean of the trunk posture angle (in degrees), is the total number of frames. When the correlation coefficient exceeds the synergy threshold , it is determined that trunk compensation occurs. The synergy threshold is set to 0.7, which is determined based on statistical analysis of a large amount of normal and abnormal training data, and can effectively distinguish between normal head and neck independent movement and abnormal trunk assisted movement. Based on the evaluation results of the above three dimensions, the system calculates the comprehensive motion quality score :

[0062]

[0063] ,

[0064] wherein, is the comprehensive motion quality score (ranging from 0 to 100), is the joint range of motion completion rate (expressed in percentage, ranging from 0 to 100), is the speed smoothness score (ranging from 0 to 100), is the compensation action penalty term (set to 20 when compensation action is detected, otherwise 0), is the weight coefficient, set to 0.4, 0.4, 1.0 respectively. The weight design ensures that the joint range of motion and the speed smoothness are equally important, and the compensation action directly deducts a fixed score to emphasize its negative impact. The value range of the comprehensive score is 0-100, the higher the score, the better the motion quality. When , the rating is excellent; when , the rating is good; when , the rating is qualified; when , the rating is unqualified and needs to be improved.

[0065] ​​The module outputs the real-time motion quality score to the closed-loop feedback optimization and rehabilitation report generation module 4, triggering the corresponding feedback mechanism and difficulty adjustment strategy.

[0066] The closed-loop feedback optimization and rehabilitation report generation module 4 is the decision output module of the system, responsible for providing real-time feedback based on multi-dimensional motion quality assessment results, dynamically adjusting training difficulty, and generating detailed rehabilitation training reports.

[0067] The real-time feedback mechanism provides three types of feedback according to different levels of motion quality scores: positive incentive feedback, corrective feedback, and real-time warning feedback. When the motion quality score is higher than the first threshold (set to 85), the system provides positive incentive feedback, displaying encouraging text such as "standard motion, continue to maintain!" on the screen, and enhancing visual feedback through green highlighting, enhancing the user's training confidence and enthusiasm. When the motion quality score is lower than the second threshold (set to 60), the system provides corrective feedback, pointing out current problems such as "insufficient joint range of motion, please try to increase the movement amplitude" or "speed is not smooth, please move at a constant speed", while attracting the user's attention through orange warning color. When compensatory motion is detected, the system immediately provides real-time warning feedback, displaying "shoulder elevation detected, please relax your shoulders" or "trunk compensation detected, please keep your trunk stable" in red flashing text on the screen, accompanied by a prompt sound to remind the user to correct the motion in a timely manner.

[0068] The closed-loop difficulty adjustment mechanism dynamically adjusts the difficulty coefficient of the virtual guide path according to the user's historical training performance, achieving personalized adaptive training. The system maintains a sliding window to record the average quality score of the last several training times . When the average quality score of consecutive training (usually set to 3-5 times) is higher than the excellent threshold (set to 85), the system determines that the current training difficulty is too low and automatically increases the path difficulty coefficient. Difficulty increase is achieved by increasing the maximum offset angle of the path or increasing the movement speed requirement, with an adjustment range of 5%-15%. For example, if the current path has a maximum offset angle of 45°, a 10% increase will result in a maximum offset angle of 49.5°. Difficulty increase adopts a gradual strategy to avoid sudden difficulty jumps that may cause frustration to the user.

[0069] Correspondingly, when the average quality score of consecutive training (usually set to 3-5 times) is lower than the passing thresholdWhen the difficulty coefficient is higher than the threshold (set to 60), the system determines that the current training difficulty of the user is too high, and automatically reduces the path difficulty coefficient. Difficulty reduction is achieved by reducing the maximum deviation angle of the path or reducing the movement speed requirement, and the adjustment amplitude is also 5%-15%. For example, if the upper limit of the movement speed of the current path is 40° / s, it will be reduced by 10% to 36° / s. The difficulty reduction mechanism ensures that the training intensity always matches the user's ability, avoiding the risk of excessive fatigue or injury.

[0070] The adjustment instruction of the difficulty coefficient is transmitted to the adaptive guide path generation and dynamic tracking module 2 through a closed-loop feedback channel. After receiving the adjustment instruction, the module 2 automatically applies the new difficulty parameter in the next training and regenerates the virtual guide path. In this way, from quality assessment to difficulty adjustment, to path regeneration, a complete closed-loop feedback link is formed, realizing the self-optimization and adaptability of the system.

[0071] The rehabilitation training report generation function provides detailed training data analysis for the therapist and the user. The report content includes four main parts:

[0072] First, the joint range of motion graph. This graph shows the trend of the angle of the cervical spine in each direction over time during training. The horizontal axis is time (unit: seconds), and the vertical axis is angle (unit: degrees). For flexion and extension training, the graph shows the change in pitch angle; for left and right rotation training, the graph shows the change in yaw angle; for left and right lateral flexion training, the graph shows the change in roll angle. The maximum and minimum values are marked on the curve, making it easy to see the range of motion achieved by the user. At the same time, the reference line of the target range of motion is superimposed on the graph, and the comparison between the actual curve of the user and the reference line intuitively reflects the completion.

[0073] Second, the speed distribution histogram. This histogram statistics the proportion of movement time in different speed intervals, with the horizontal axis representing the speed interval (unit: degrees / second) and the vertical axis representing the time proportion (unit: percentage). The histogram divides the speed into several intervals, such as 0-10° / s, 10-20° / s, 20-30° / s, etc., and counts the number of frames in each interval, then divides by the total number of frames to get the proportion. The ideal speed distribution should be concentrated in the target speed range, showing a single peak distribution. If the speed distribution is too dispersed or has multiple peaks, it indicates that the user's speed control ability is weak and needs to strengthen the uniform motion training. The average speed and speed standard deviation are marked on the histogram as a supplement to quantitative evaluation.

[0074] Third, the compensatory action timestamp list. This list records in detail the time, type and duration of each compensatory action. The time is expressed in seconds after the start of training, the type includes shoulder lifting compensation and trunk compensation, and the duration is the time span from the appearance to the disappearance of the compensatory action. The list is arranged in chronological order, which facilitates the therapist to locate the specific link where the problem occurs and analyze the causes of compensatory action, such as whether it occurs when moving to the maximum amplitude or when the speed increases, etc. The list counts the total number of compensations and the proportion of total compensation time, which is used as a quantitative indicator of the severity of compensatory action. When the proportion of compensatory time exceeds 20%, it indicates that the user has a significant compensatory tendency and needs to strengthen the establishment of correct action patterns.

[0075] Fourth, the comprehensive score trend chart. This trend chart shows the quality score changes of multiple training, the horizontal axis is the number of training, and the vertical axis is the comprehensive score (0-100). The comprehensive score of each training is plotted on the trend chart and connected with a broken line to show the rising or falling trend of the score. At the same time, the qualified line (score 60), good line (score 70) and excellent line (score 85) are superimposed on the chart to facilitate intuitive judgment of the user's rehabilitation progress. Ideally, as the rehabilitation training progresses, the score should show a gradual upward trend, reflecting the improvement of the user's cervical function. If the score continues to decline or stagnates for a long time, it indicates that the rehabilitation effect is not good, and the training program needs to be adjusted or professional guidance is needed. The trend chart also calculates the linear regression slope of the score to quantify the improvement rate of the score, and the larger the slope, the faster the rehabilitation progress.

[0076] The rehabilitation training report is exported in PDF format for easy saving and sharing. Users can send the report to the therapist for remote evaluation, and the therapist can provide personalized rehabilitation suggestions and training plan adjustment schemes based on the report content. The generation of the report is automated and does not require human intervention. The system automatically organizes data, draws charts and generates reports after each training, greatly improving the efficiency of rehabilitation management.

[0077] In summary, the present application forms a complete cervical rehabilitation training closed-loop system through the deep coupling of the four core modules. The output of the multi-scale head and neck feature extraction and posture estimation module 1 (head and neck key point coordinates and Euler angles) is directly used as the input of the adaptive guide path generation and dynamic tracking module 2, driving the real-time update of the virtual cursor and the deviation calculation; the spatial deviation and time deviation sequence calculated by module 2 is transmitted to the multi-dimensional motion quality assessment and compensation detection module 3 in real time as the key parameter of quality assessment; the action quality score generated by module 3 triggers the feedback mechanism and difficulty adjustment decision of the closed-loop feedback optimization and rehabilitation report generation module 4; the difficulty adjustment instruction of module 4 acts on module 2 in reverse, changing the parameter setting of the virtual guide path, forming a complete closed-loop feedback link.

[0078] The synergistic relationship between the deeply coupled modules achieves the superposition and enhancement of multiple technical effects. First, the multi-scale feature extraction of module 1 improves the pose estimation accuracy, providing high-quality input data for subsequent modules, which is the basis for the performance of the entire system. Second, the adaptive path generation of module 2 ensures that the training difficulty matches the user's ability, avoiding poor training results caused by excessive ease or difficulty, and the dynamic tracking calculation converts abstract angle control into intuitive cursor following, improving the operability of training. Third, the multi-dimensional quality evaluation of module 3 not only judges whether the action is completed, but also quantitatively analyzes the motion quality, especially the innovative compensatory movement detection algorithm can timely detect incorrect movement patterns and avoid compensatory movement solidification, which is the core advantage of the present invention over existing technologies. Finally, the closed-loop feedback mechanism of module 4 dynamically adjusts the training parameters based on the quality evaluation results, achieving self-optimization of the system, and the detailed rehabilitation report provides data support for rehabilitation decision-making, bridging the gap caused by the lack of supervision in remote rehabilitation.

[0079] The synergistic effect of the four modules produces a nonlinear synergistic effect of 1+1>2. Taking compensatory movement detection as an example, if only module 3 has a compensatory detection algorithm without real-time warning feedback, the user cannot timely realize the error and correct it; if only real-time feedback is available without closed-loop difficulty adjustment, the user may frequently appear compensatory due to excessive difficulty, making it difficult to establish a correct pattern; only the combination of precise detection, timely feedback, and adaptive adjustment can effectively suppress compensatory movements and promote the formation of correct movement patterns. This multi-module synergistic, closed-loop optimized system design is the core embodiment of the present invention, which is in sharp contrast to the linear, open-loop design of existing technologies, with a technical similarity of less than 20%, and has significant innovation and practicality.

[0080] A cervical spondylosis patient is taken as an example to perform forward and backward rehabilitation training, and the implementation process of the present invention is described in detail.

[0081] The patient uses an intelligent tablet computer equipped with the system of the present invention for rehabilitation training. The patient places the tablet computer on a desktop support about 50 cm away from the face to ensure that the front camera can clearly capture the face. After the system is started, the user registration and initial evaluation are performed first. The patient inputs basic information (name, age, gender, diagnosis information, etc.), and then completes the range of motion test of four basic actions under the guidance of the system.

[0082] For the flexion-extension test, the system instructs the patient to first lower their head to the maximum angle and hold for 2 seconds. The system records the maximum flexion angle as 45° (normal is approximately 50°). Then, the patient is instructed to raise their head to the maximum angle and hold for 2 seconds. The system records the maximum extension angle as 40° (normal is approximately 60°). The calculated range of motion for flexion-extension is 85°, representing 77% of the normal value (110°). Similarly, the measured range of motion for left and right rotation is 120° (75% of the normal value of 160°), and the range of motion for left and right lateral flexion is 60° (67% of the normal value of 90°). The combined percentage of range of motion is calculated as follows: This indicates that the patient's current cervical spine function is approximately 74% of the normal level.

[0083] Based on the overall activity level percentage and the rehabilitation stage (assuming it is the initial stage, with a difficulty level of...) The system calculates the maximum offset angle during flexion-extension training as follows: The movement speed ranges from 10-30° / s. The system generates a virtual guidance path, starting from the neutral position (0°), then flexing forward to 22° in 2.2s, returning to the neutral position in 2.2s, then extending backward to -22° in 2.2s, and finally returning to the neutral position in 2.2s, for a total duration of 8.8s, forming a complete movement cycle. 88 path points are generated on the path (intervals of 0.1s), and cubic spline interpolation is used to ensure path smoothness.

[0084] The patient sees a yellow target cursor moving downwards from the center of the screen. The patient controls a red real-time cursor to follow the yellow target by tilting their head down. The multi-scale head and neck feature extraction and pose estimation module 1 detects key facial points in real time and calculates the head pitch angle. In the 0-2.2s phase, the patient's head flexes forward from 0° to 18°, approaching but not fully reaching the target of 22°. The adaptive guidance path generation and dynamic tracking module 2 calculates the spatial deviation, with an average deviation of 15 pixels (corresponding to approximately 4°) and a temporal deviation of 0.1s (slightly lagging).

[0085] During the forward flexion process, the multidimensional motion quality assessment and compensation detection module 3 monitored the motion quality in real time. Joint range of motion assessment revealed that the patient's actual forward flexion angle was 18°, with a completion rate of [percentage missing]. The performance did not fully meet the standards. Velocity smoothness assessment calculates the coefficient of variation of the velocity sequence. (Indicates relatively small and uniform velocity changes), normalized acceleration kurtosis (Indicates that the acceleration distribution is close to normal, with no drastic abrupt changes), adaptive smoothness index Corresponding speed smoothness score The rating was good. Compensatory movement detection analysis revealed that at 1.8 seconds, the vertical displacement of the key point on the left shoulder reached 18 pixels, exceeding the compensation threshold of 15 pixels, indicating a shoulder shrugging compensation, lasting 0.5 seconds. The system immediately displayed flashing red text on the screen: "Shoulder shrugging detected, please relax your shoulder," accompanied by an audio prompt. After seeing the warning, the patient consciously relaxed their shoulder, and the shoulder displacement returned to the normal range.

[0086] After completing one movement cycle, the system calculates a comprehensive movement quality score. Joint range of motion completion rate. Speed ​​smoothness score A compensatory action occurred, therefore a compensation penalty was imposed. Overall score The rating is unsatisfactory. Since the score is below the passing threshold of 60, the closed-loop feedback optimization and rehabilitation report generation module 4 provides corrective feedback, showing that "joint mobility and exercise quality need to be improved, and it is recommended to reduce the training difficulty."

[0087] The system recorded a training score of 42. The patient continued with a second training session, and the system maintained a sliding window to record scores. Assuming the scores for the first three training sessions were 42, 48, and 52 respectively, the average score was 47.3, below the acceptable threshold of 60. The closed-loop difficulty adjustment mechanism determined the current difficulty was too high and decided to reduce the path difficulty coefficient. The reduction was set to 10%, and the new maximum offset angle was adjusted to... The maximum speed limit has been adjusted to In the fourth training iteration, the system applied new difficulty parameters to generate easier virtual guidance paths.

[0088] As the number of training sessions increases and the difficulty level is adjusted, the patient's movement quality gradually improves. Assuming the scores for training sessions 4-6 are 55, 62, and 68 respectively, with an average score of 61.7, this meets the passing standard. The scores for training sessions 7-9 are 70, 75, and 78 respectively, with an average score of 74.3, which is rated as good. The system continues to maintain a sliding window; when the average score for three consecutive training sessions exceeds 85, the difficulty will be automatically increased, guiding the patient to a higher level of challenge.

[0089] After the training, the system automatically generates a rehabilitation training report. The joint range of motion curve in the report shows that in the first 9 training sessions, the patient's maximum flexion and extension angles gradually increased from 85° (1st session) to 95° (9th session), approaching the normal value of 110°, reflecting a significant improvement in cervical spine mobility. The velocity distribution histogram shows that in the first training session, the velocity distribution was relatively dispersed, with the 5-15° / s range accounting for 30%, the 15-25° / s range accounting for 40%, and the 25-35° / s range accounting for 30%, indicating weak velocity control. In the 9th training session, the velocity distribution became more concentrated, with the 15-25° / s range increasing to 70%, and the velocity standard deviation decreasing from 8.5° / s to 5.2° / s, indicating a significant improvement in velocity control. The compensatory movement timestamp list recorded two instances of shoulder shrugging compensation during the first training session (at 1.8s and 6.2s, totaling 1.2s, accounting for 13.6%), one instance of compensation during the third training session (at 2.0s, lasting 0.8s, accounting for 9.1%), and no further compensation occurred from the fifth session onwards, indicating that the patient gradually established the correct movement pattern. The comprehensive score trend chart clearly shows the process of the score increasing from 42 to 78 points, with a linear regression slope of 4.5 points / session, indicating good rehabilitation progress.

[0090] After reviewing the report, the rehabilitation therapist affirmed the patient's rehabilitation progress and recommended continuing the current training program. Once the average score stabilizes above 85 points, lateral rotation and lateral flexion exercises can be added to achieve comprehensive cervical spine function recovery. This example fully demonstrates the application process, core functions, and technical effects of the system of the present invention in actual rehabilitation training, proving the system's feasibility and effectiveness.

[0091] The technical solutions of this invention are not limited to the specific embodiments described above. Any improvements and modifications based on the technical concept of this invention are within the scope of protection of this invention. For example, the specific structure of the multi-scale convolutional neural network can adopt other lightweight network architectures, such as EfficientNet and ShuffleNet; the virtual guidance path generation algorithm can adopt other smooth curve algorithms such as Bézier curves and B-spline curves; the detection of compensatory actions can be extended to more types of compensatory behaviors, such as eye movement compensation and facial muscle tension compensation; the feedback form can be expanded to include multimodal feedback methods such as tactile feedback and voice feedback. These changes and extensions are all within the scope of protection of this invention.

Claims

1. A method for guiding and assessing neck rehabilitation training movements based on visual feedback, characterized in that, Includes the following steps: The acquisition process involves real-time acquisition of user facial image sequences using the front-facing camera of a smart device, extraction of two-dimensional image coordinates of key points on the head and neck based on a multi-scale convolutional neural network, estimation of three-dimensional Euler angles of head posture by combining temporal information, and generation of head and neck motion trajectory data. The path generation step adaptively generates a virtual guidance path based on the training movement type preset by the rehabilitation therapist and the user's current cervical spine mobility baseline parameters. The virtual guidance path includes a spatiotemporal coordinate sequence and velocity constraint parameters, and is presented on the screen as a dynamic cursor. The system tracks and evaluates the steps, calculates the spatial and temporal deviations between the user's head cursor and the virtual guidance path in real time, and quantitatively analyzes the joint range of motion completion rate, movement speed smoothness, and compensatory movement intensity based on a multidimensional quality assessment model to generate a movement quality score. The feedback optimization step determines the feedback type based on the action quality score. When the score is higher than a first threshold, positive incentive feedback is provided; when the score is lower than a second threshold, corrective feedback is provided and the path difficulty coefficient is reduced; when a compensatory action is detected, real-time warning feedback is provided, forming a closed-loop adjustment mechanism. The report generation process involves statistically analyzing trajectory deviation data, quality score sequences, and the number of compensatory movements from a single training session to generate a rehabilitation training report that includes a joint range of motion curve, a velocity distribution histogram, and timestamps of compensatory movements.

2. The method according to claim 1, characterized in that, The multi-scale convolutional neural network adopts an improved residual structure, including multiple parallel dilated convolutional branches with dilation rates of 2, 4, and 8 for different branches. It integrates multi-scale receptive field information through a feature fusion module to improve the robustness of head and neck key point detection.

3. The method according to claim 1, characterized in that, The process of adaptively generating virtual navigation paths includes: The overall range of motion percentage is calculated based on the cervical spine flexion-extension, lateral rotation, and lateral flexion ranges of motion measured in the initial assessment. Based on the comprehensive activity percentage and rehabilitation stage parameters, determine the maximum deviation angle and movement speed range of the path; A cubic spline interpolation algorithm is used to generate a smooth spatiotemporal trajectory curve, ensuring that the velocity and acceleration of the path are continuously differentiable.

4. The method according to claim 1, characterized in that, The multidimensional quality assessment model includes three dimensions: joint range of motion assessment, velocity smoothness assessment, and compensatory motion detection. Joint range of motion assessment calculates the completion rate by comparing the user's actual range of angles with the preset target range of angles. Velocity smoothness assessment is based on the time-domain characteristics of motion velocity sequences, calculating the standard deviation of velocity changes and peak acceleration to quantify the smoothness of motion; Compensatory movement detection identifies compensatory behaviors such as trunk compensation and shoulder shrugging by analyzing abnormal movement amplitudes at key points in the shoulder and coordinated head and neck movements.

5. The method according to claim 4, characterized in that, The speed smoothness assessment uses an adaptive smoothness index, which comprehensively quantifies the uniformity and abrupt changes in speed by calculating the normalized variance of motion speed and the normalized kurtosis of acceleration. A higher smoothness index indicates better motion quality.

6. The method according to claim 4, characterized in that, The compensatory action detection employs a multi-joint collaborative analysis method. Real-time monitoring of three-dimensional coordinate changes at key points in the head, neck, and shoulders; Calculate the vertical displacement of key shoulder points relative to their initial positions. When the vertical displacement exceeds a compensation threshold, it is determined to be shoulder shrugging compensation. Calculate the correlation coefficient between head and neck movements and trunk movements. When the correlation coefficient exceeds the coordination threshold, it is determined to be trunk compensation.

7. The method according to claim 1, characterized in that, The closed-loop adjustment mechanism is achieved by dynamically adjusting the path difficulty coefficient: When the average quality score of N consecutive training sessions is higher than the excellent threshold, the path difficulty coefficient is automatically increased, the maximum offset angle of the path is increased, or the movement speed requirement is increased. When the average quality score of M consecutive training sessions is lower than the passing threshold, the path difficulty coefficient is automatically reduced, the maximum offset angle of the path is decreased, or the movement speed requirement is lowered. The values ​​of N and M range from 3 to 5, and the adjustment range of the difficulty coefficient is from 5% to 15%.

8. The method according to claim 1, characterized in that, The rehabilitation training report includes: The joint range of motion curves show the trends of cervical spine flexion-extension, left-right rotation, and lateral flexion angles over time during training. Speed ​​distribution histogram, statistically analyzing the percentage of motion time in each speed range, to assess speed control capability; A list of timestamps for compensatory actions, recording the time, type, and duration of each compensatory action; A comprehensive scoring trend chart compares the changes in quality scores across multiple training sessions, quantifying rehabilitation progress.

9. The method according to claim 1, characterized in that, The training movements include cervical flexion and extension exercises, left and right rotation exercises, left and right lateral flexion exercises, and combined exercise exercises. Different training movement types correspond to different path templates and evaluation parameters.

10. A visual feedback-based neck rehabilitation training movement guidance and assessment system, used to implement the method described in any one of claims 1-9, characterized in that, include: The multi-scale head and neck feature extraction and pose estimation module is used to acquire user facial image sequences in real time through the front-facing camera of smart devices, extract two-dimensional image coordinates of key points of the head and neck based on multi-scale convolutional neural networks, estimate the three-dimensional Euler angles of head pose by combining temporal information, and generate head and neck motion trajectory data; The adaptive guidance path generation and dynamic tracking module, connected to the multi-scale head and neck feature extraction and posture estimation module, is used to adaptively generate a virtual guidance path based on the training movement type preset by the rehabilitation therapist and the user's current cervical spine mobility baseline parameters. The virtual guidance path includes a spatiotemporal coordinate sequence and velocity constraint parameters, and is presented on the screen as a dynamic cursor. The module also calculates the spatial and temporal deviations between the user's head cursor and the virtual guidance path in real time. A multidimensional motion quality assessment and compensation detection module, connected to the adaptive guidance path generation and dynamic tracking module, is used to quantitatively analyze joint range of motion completion rate, motion speed smoothness, and compensatory action intensity based on a multidimensional quality assessment model, and generate a motion quality score. The closed-loop feedback optimization and rehabilitation report generation module is connected to the multidimensional motion quality assessment and compensation detection module, and is also connected to the adaptive guidance path generation and dynamic tracking module. It is used to determine the feedback type based on the motion quality score, provide positive incentive feedback, corrective feedback or real-time warning feedback, dynamically adjust the path difficulty coefficient based on the score results, and statistically analyze the training data to generate a rehabilitation training report.

Images