Scoliosis rehabilitation training posture image monitoring method and system

Abstract

The present application relates to the field of image processing and rehabilitation medical auxiliary technology, disclose a scoliosis rehabilitation training posture image monitoring method and system, the method is through RGB camera acquisition patient rehabilitation training video stream, execute lightweight human detection and multi-joint skeleton key point estimation to construct two-dimensional posture skeleton sequence;The key point sub-sequence of the spine is parameterized curve fitting and extracts the lateral deviation and the rotation angle feature, synchronous calculation double shoulder height difference, double hip inclination angle and body symmetry degree such as body state index, fusion generates body state comprehensive evaluation vector;The real-time posture sequence is matched with the standard rehabilitation action template with weighted dynamic time warping, joint body state evaluation vector identifies action amplitude deficiency, abnormal rhythm and compensation posture deviation;Based on the deviation identification result dynamic adjustment front module evaluation threshold to form a closed loop feedback;Finally generate the visual rehabilitation log containing action score and bias heat map and upload for therapist remote review.

Description

Technical Field

[0001] This invention relates to the fields of image processing and rehabilitation medicine auxiliary technology, and in particular to a method and system for monitoring posture images during scoliosis rehabilitation training. Background Technology

[0002] Scoliosis is a skeletal disorder characterized by three-dimensional spinal deformity, affecting approximately 2% to 4% of adolescents worldwide. For mild to moderate cases with a Cobb angle between 10° and 40°, conservative treatments, such as Schroth 3D corrective exercises and the SEAS scientific exercise method, are widely recommended as first-line interventions. Clinical studies have shown that properly performed Schroth corrective training can reduce the Cobb angle by an average of approximately 7° and improve trunk rotation by an average of approximately 2.5° within a 10-week treatment cycle. However, the effectiveness of rehabilitation training is highly dependent on the patient's ability to maintain correct posture and perform movements correctly throughout the training process. In current clinical practice, patients mainly receive face-to-face guidance from therapists in rehabilitation institutions. However, due to factors such as uneven distribution of medical resources and the cost of medical care, many patients cannot obtain frequent professional guidance and are forced to complete their daily rehabilitation training at home. In home training settings lacking professional supervision, patients often find it difficult to accurately judge whether their movements meet the requirements of the treatment plan. Common problems include insufficient training range, improper rhythm control, and incorrect postures where other muscle groups compensate for the movement of the target muscle group. These deviations not only reduce the rehabilitation effect, but may even exacerbate the progression of scoliosis in severe cases.

[0003] In recent years, computer vision-based human pose estimation technology has made significant progress. Open-source frameworks such as OpenPose and MediaPipe can extract key points of the human skeleton in real time using a monocular RGB camera. Meanwhile, motion assessment technology based on dynamic time warping algorithms has shown promising applications in fitness guidance and physical therapy assistance. However, these general technologies face three challenges when directly applied to scoliosis rehabilitation training monitoring: First, general pose estimation models do not provide detailed modeling of the spinal region, making it difficult to accurately capture the morphological changes in the curve from the cervical to the sacral vertebrae; second, standard dynamic time warping algorithms assign equal weight to all key points, failing to highlight the assessment priority of the trunk region in scoliosis rehabilitation training; and third, existing motion assessment systems lack a fusion mechanism to jointly analyze spinal morphological indicators with motion sequence matching results.

[0004] Among existing human posture monitoring technologies, Chinese patent application CN120339276A discloses a health assessment method and system based on a 3D human posture image model. This method acquires human annular scan data using a 360° annular scanning device, generates a 3D point cloud using phase-shifted fringe projection technology, and identifies skeletal contour features through orthogonal projection onto sagittal, coronal, and transverse planes to establish a dynamic marking system of anatomical landmarks. Finally, it constructs a 3D human posture image model and generates a personalized posture correction report. While this method achieves good technical results in terms of the accuracy of posture health assessment, it relies on specialized 360° annular scanning devices and structured light projection equipment, which are costly and have strict requirements for the scanning environment, making it difficult to deploy in patients' home settings. Furthermore, this method focuses on the assessment and analysis of static postures and lacks the ability to monitor the execution sequence of movements during dynamic rehabilitation training in real time. It cannot provide timely feedback and correction on the amplitude, rhythm, and compensatory behaviors of patients' training movements, thus making it difficult to directly apply to the quality monitoring of home rehabilitation training for scoliosis patients.

[0005] In summary, existing technologies for monitoring scoliosis rehabilitation training have three main shortcomings: First, there is a lack of lightweight technologies that can achieve refined curve analysis of the spinal region and quantitative assessment of postural symmetry based on ordinary consumer-grade cameras. Existing spinal assessment solutions either rely on professional X-ray imaging equipment or expensive 3D scanning devices, making them difficult to use routinely in home settings. Second, there is a lack of a comprehensive assessment mechanism that integrates scoliosis-specific postural indicators with the spatiotemporal matching of rehabilitation movement sequences. Existing movement assessment systems only focus on the matching degree of limb movement trajectories while ignoring the impact of spinal morphological changes on training quality, resulting in an inability to effectively identify deviation patterns unique to scoliosis rehabilitation training, such as compensatory behaviors. Third, there is a lack of an end-to-end closed-loop training guidance scheme that goes from deviation identification to real-time correction to adaptive parameter adjustment. Existing systems mostly use fixed assessment standards and cannot dynamically adjust assessment sensitivity and training difficulty according to the patient's progress during training, making it difficult to achieve personalized, progressive rehabilitation guidance. Therefore, there is an urgent need to develop a technical solution that can achieve intelligent monitoring and adaptive guidance throughout the entire process of scoliosis rehabilitation training using ordinary cameras. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, this invention provides a method and system for monitoring posture images during scoliosis rehabilitation training.

[0007] This invention employs the following approach: a method for monitoring posture images during scoliosis rehabilitation training, which includes the following steps:

[0008] Step S1, Human Skeletal Pose Estimation and Two-Dimensional Skeletal Sequence Construction: A video stream of the rehabilitation training process of a scoliosis patient is acquired by an RGB camera. Lightweight human target detection is performed on each frame of the video stream to obtain the human bounding box region. Real-time estimation of multi-joint skeletal key points is performed within the human bounding box region to obtain a two-dimensional coordinate set including key nodes of the spine, scapula, pelvis, and limbs. The two-dimensional coordinate set of multiple consecutive frames is arranged in a time sequence to construct a two-dimensional pose skeleton sequence.

[0009] Step S2, Joint Analysis of Spinal Curve Fitting and Postural Symmetry: Extract the spinal key point subsequence between the cervical and sacral nodes from the two-dimensional posture skeleton sequence, perform parametric curve fitting on the spinal key point sequence to generate a spinal midline fitting curve, extract lateral offset features and rotation angle features based on the spinal midline fitting curve, simultaneously calculate the shoulder height difference index, hip tilt angle index, and trunk left and right symmetry index, and fuse the features and indices to generate a comprehensive postural evaluation vector.

[0010] Step S3, Rehabilitation Movement Spatiotemporal Sequence Matching and Training Deviation Identification Step: The real-time posture frame sequence in the two-dimensional posture skeleton sequence is dynamically time-warped and matched with the preset standard rehabilitation movement template. The movement similarity score is calculated based on the matching path, and the deviations of insufficient movement amplitude, abnormal rhythm, and compensatory posture are identified in conjunction with the posture comprehensive evaluation vector.

[0011] Step S4, Real-time Correction Feedback and Adaptive Training Guidance: Generate instant correction prompts based on training deviations, and dynamically adjust the judgment threshold of body symmetry analysis in step S2 and the sensitivity parameters of dynamic time warping matching in step S3.

[0012] This invention also provides a posture image monitoring system for scoliosis rehabilitation training, used to execute the above-mentioned method. The system includes: a skeletal posture estimation module, used to acquire video streams through an RGB camera and perform human target detection and multi-joint skeletal key point estimation to construct a two-dimensional posture skeleton sequence; a spinal curve and symmetry joint analysis module, used to extract spinal key point sub-sequences from the skeleton sequence and perform cubic B-spline curve fitting, extract lateral offset and rotation angle features, and simultaneously calculate postural balance indicators such as shoulder height difference, hip tilt angle, and trunk left and right symmetry, and fuse them to generate a comprehensive postural assessment vector; a motion matching and deviation recognition module, used to perform weighted dynamic time warping matching and jointly identify insufficient motion amplitude deviation, abnormal rhythm deviation, and compensatory posture deviation with the comprehensive postural assessment vector; a correction feedback and adaptive adjustment module, used to generate visual and voice correction prompts according to the deviation type, and feed back parameter adjustment signals to the spinal curve and symmetry joint analysis module and the motion matching and deviation recognition module to achieve closed-loop optimization; and a log generation and remote management module, used to summarize the data processed by each module to generate a visualized rehabilitation training log and upload it to a remote server for therapist review. The five modules form a closed-loop topology with feedback loops, enabling the system to continuously optimize assessment accuracy based on the patient's real-time training status.

[0013] This invention offers the following advantages: By jointly analyzing the lateral offset and rotation angle features obtained from fitting the spinal midline curve with quantitative indicators of postural symmetry, and synergistically fusing the results of this joint analysis with the similarity score of spatiotemporal sequence matching of rehabilitation movements, integrated monitoring of postural assessment and rehabilitation movement quality assessment specifically for scoliosis is achieved. Specifically, the lateral offset and rotation angle features output by the spinal curve fitting module reflect the degree of morphological abnormalities in the patient's spine, while the indicators such as shoulder height difference, hip tilt angle, and trunk symmetry output by the postural symmetry analysis module quantify the impact of scoliosis on overall posture from the perspective of whole-body posture balance. This synergistic fusion mechanism allows compensatory behavioral patterns that cannot be captured by individual postural assessments or individual movement matching to be accurately identified in the combined space of both, thus producing a nonlinear gain effect that surpasses the simple superposition of individual modules. Furthermore, the closed-loop feedback mechanism enables the system to adaptively optimize assessment parameters based on the patient's real-time training status, improving monitoring sensitivity while avoiding false alarms, effectively solving the technical challenge of lacking objective monitoring and real-time guidance for home rehabilitation training of scoliosis patients. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the posture image monitoring method for scoliosis rehabilitation training provided in this embodiment of the invention.

[0015] Figure 2This is a schematic diagram of the architecture of the scoliosis rehabilitation training posture image monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to explain this invention and are not intended to limit the scope of protection of this invention.

[0017] Reference Figure 1 As shown, in one embodiment of the present invention, a method for monitoring posture images during scoliosis rehabilitation training is provided. This method includes steps S1 to S5, where there is a data flow dependency between each step, and the output of step S4 is fed back to steps S2 and S3 to form a closed-loop adjustment mechanism. The specific implementation process of each step is described in detail below.

[0018] Step S1: Human skeleton pose estimation and 2D skeleton sequence construction. In a preferred embodiment of the present invention, step S1 first requires configuring the video acquisition hardware environment. Specifically, the system uses a standard RGB camera as the video input source. This camera can be a built-in camera integrated into a laptop, tablet, or smartphone, or a standalone external USB camera. Preferably, the camera's frame rate is not less than 25fps, and the spatial resolution is set to 640×480 pixels to control computational overhead while ensuring the accuracy of skeletal key point detection. In a home use scenario, the camera is preferably installed approximately 2m to 3m directly in front of the patient, with the lens height roughly level with the patient's waist, to ensure complete capture of full-body images of the patient's standing and bending rehabilitation training movements.

[0019] After acquiring the video stream data, the system performs lightweight human target detection processing on each frame of the image. In one embodiment of the present invention, the human target detection adopts a detection network architecture based on single-stage anchor box regression. This network uses a lightweight backbone network of the MobileNet series as a feature extractor, and the detection latency is controlled within 15ms under the condition of an input resolution of 640×480 pixels. During the detection process, the network outputs the confidence score of the candidate human bounding boxes. A confidence threshold of 0.5 is set to filter low-quality detection results. Subsequently, a non-maximum suppression algorithm is used to remove redundancy from overlapping candidate boxes, wherein the cross-union ratio threshold of the non-maximum suppression is set to 0.45. After the above processing, the system extracts the coordinate information of the rectangular bounding box containing the complete human body region from each frame of the image.

[0020] Furthermore, within the extracted human bounding box region, the system employs a top-down multi-joint skeletal keypoint estimation network to accurately locate the positions of each joint in the human body. Preferably, in this embodiment, the number of estimated skeletal keypoints is 17, specifically including: nose tip, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle. Based on this, the system further derives two virtual keypoints—the neck midpoint and the hip midpoint—using a midpoint calculation method for adjacent keypoints. The neck midpoint is the arithmetic mean of the coordinates of the left and right shoulder keypoints, and the hip midpoint is the arithmetic mean of the coordinates of the left and right hip keypoints. These 19 keypoints together constitute a complete set of posture keypoints covering the spinal axis, shoulder girdle, pelvic ring, and limb joints.

[0021] For each estimated keypoint, the system records its two-dimensional coordinates in the image coordinate system. and the corresponding key point confidence level Preferably, when the confidence level of a certain key point... When the value is below 0.3, the key point is marked as occluded or invisible. In one embodiment of the present invention, for key points that are temporarily lost due to self-occlusion or viewing angle limitations, a linear interpolation method is used to complete the key point based on the effective coordinates of the key point in the preceding and following frames. The condition for completion is that the number of consecutive lost frames does not exceed 5 frames. When the number of consecutive lost frames exceeds this threshold, the system issues an alarm prompt indicating that the posture tracking has failed, guiding the patient to adjust their position or the camera angle.

[0022] Finally, the system arranges the set of two-dimensional coordinates of key points from multiple consecutive frames according to the chronological order of the acquisition timestamps, constructing a structure in the form of... A two-dimensional pose skeleton sequence, in which Indicates the first In the frame image The coordinates and confidence levels of each key point This represents the total number of frames in the video stream. This two-dimensional pose skeleton sequence serves as the core input data for subsequent steps S2 and S3.

[0023] Step S2: Joint analysis of spinal curve fitting and body symmetry. Step S2 receives the two-dimensional posture skeleton sequence output from Step S1 as input and performs refined analysis of the spinal region and quantitative assessment of whole-body posture symmetry. This step is the core innovation of the technical solution of this invention, which organically integrates spinal morphological analysis and trunk symmetry assessment into a unified processing framework, enabling the two types of assessment results to produce synergistic effects in the joint feature space.

[0024] First, the system extracts a sequence of key points for the spine from the constructed two-dimensional pose skeleton sequence. In one specific embodiment of the invention, the sequence of key points for the spine is defined as follows: the cervical vertebrae are approximated by the midpoint of the neck (denoted as ). The sacral vertebra is approximately represented by the midpoint of the hip (denoted as ). At the same time, define an approximate point of the upper thoracic vertebra at the midpoint of the line connecting the left and right shoulders (denoted as ). The approximate point of the lower thoracic vertebra is defined approximately one-third of the trunk length above the midpoint of the line connecting the left and right hips (denoted as ). ), and in and The midpoint between the two points defines the approximate point of the lumbar spine (denoted as ). Therefore, the sequence of spinal keypoints extracted from each frame of the image is as follows: .

[0025] After obtaining the subsequence of key spinal points, the system performs parametric curve fitting to generate a fitted curve of the spinal midline. In a preferred embodiment of the present invention, a cubic B-spline curve fitting method is used. Specifically, five key spinal points are used as reference points, and the coordinates of control points are obtained by solving the linear equations of the B-spline basis functions. The number of control points is set to 5 to 8. During the fitting process, the least squares method is used to minimize the fitting residual. Preferably, the fitting residual threshold is set to no more than 2 pixels. The fitted spinal midline curve can be represented by a parametric equation. ,in For normalized curve parameters, Corresponding to the cervical spine end, Corresponding to the sacral end.

[0026] Based on the obtained spinal midline fitting curve, the system extracts two key features: lateral offset features and rotation angle features. The calculation method for lateral offset is as follows: First, the cervical vertebrae are determined. With sacral nodes The line connecting the two points is used as the spinal reference baseline. Then, the vertical distance from each uniformly sampled point on the spinal midline fitting curve to this reference baseline is calculated, and the maximum value is taken as the lateral offset feature. The calculation formula is as follows:

[0027] ,

[0028] in: The Euclidean distance function from a point to a line, in pixels; The equation for the line connecting the cervical and sacral vertebrae; Here are the curve parameters, and their range is: The sampling interval is 0.01. This lateral offset reflects the maximum degree to which the spinal midline deviates from the midline of the human body. A higher value indicates a more pronounced degree of scoliosis. At a resolution of 640×480 pixels, when the patient is 2.5m away from the camera, 1 pixel corresponds to approximately 0.4cm of actual body distance. It can be further converted into physical distance units for clinical reference.

[0029] Rotation angle features Defined as the angle between the tangent direction of the fitted curve along the spinal axis at the point of maximum lateral offset and the vertical direction. Let the curve parameters at which the lateral offset reaches its maximum value be... Then the formula for calculating the rotation angle feature is:

[0030] ,

[0031] in: and The fitted curves of the spinal midline are respectively in terms of parameters The first derivative values ​​in the horizontal and vertical directions at the specified location, in pixels per parameter unit; The result is the arctangent function. The unit is degrees (°), and the range of values ​​is... . The larger the absolute value, the more severe the rotational tilt of the spine at the point of maximum offset. Preferably, when If the angle exceeds 15°, the system will mark the frame as a significant rotational anomaly.

[0032] While performing spinal curve analysis, the system simultaneously calculates three postural symmetry indicators. The first is the shoulder height difference indicator. It is defined as the absolute value of the difference between the ordinate values ​​of the left shoulder keypoint and the right shoulder keypoint:

[0033] ,

[0034] in: and These are the ordinate values ​​of the key points on the left and right shoulders in the image coordinate system, in pixels. Preferably, when When the difference exceeds 8 pixels (approximately 3.2cm in physical distance), it is considered as asymmetry in shoulder height.

[0035] The second item is the bihip tilt angle index. It is defined as the angle between the line connecting the left and right hip key points and the horizontal reference line of the image:

[0036] ,

[0037] in: and The images show the coordinates of key points on the left and right hips, respectively. The unit is degrees (°), and the range of values ​​is... Preferably, when If the angle exceeds 3°, it is considered an abnormal pelvic tilt.

[0038] The third item is the left-right symmetry index of the torso. It measures the geometric symmetry of the trunk by the ratio of the area enclosed on both sides by the fitted curve of the spinal axis and the reference baseline:

[0039] ,

[0040] in: and The fitted curve of the spinal midline and the reference baseline are respectively. The area to the left and right of the baseline within the enclosed region is calculated using a numerical integration method, with the unit being pixels². It is a dimensionless ratio, and its value range is... ,when A value of 1 indicates perfect symmetry. Preferably, when... A value below 0.75 is considered a significant asymmetry of the trunk.

[0041] Finally, step S2 integrates the above five features and indicators into a comprehensive posture assessment vector. :

[0042] ,

[0043] This vector serves as an important auxiliary input for identifying rehabilitation movement deviations in step S3, enabling the movement matching module to not only assess the spatiotemporal trajectory consistency of limb movements, but also to simultaneously perceive the dynamic changes in spinal morphology and trunk symmetry, thereby achieving the synergistic fusion of the two assessment dimensions.

[0044] Step S3: Rehabilitation movement spatiotemporal sequence matching and training deviation identification step. Step S3 receives the two-dimensional posture skeleton sequence output from step S1 and the body posture comprehensive evaluation vector output from step S2 as dual inputs. By spatiotemporally aligning and matching the posture frame sequence of the patient's real-time rehabilitation training with the standard movement template, and combining body posture indicators, deviation classification and identification are performed.

[0045] Regarding the construction of a standard rehabilitation movement template library, in one embodiment of this invention, the system incorporates standard movement templates from two major mainstream scoliosis rehabilitation programs: one is the Schroth three-dimensional corrective gymnastics movement template set, covering 12 standard movement categories including breath-guided spinal derotation exercises, lateral stretching corrective exercises, and core stability training; the other is the SEAS scientific exercise method movement template set, containing 10 standard movement categories including active self-correction postures, functional integration movements, and daily posture maintenance training. Each standard movement template... Data was collected by professional therapists in a standardized environment and manually reviewed and confirmed. For the first The number of frames per standard motion template, For this action The key point coordinate set of the frame. Preferably, the system automatically recommends a subset of training movements based on the patient's scoliosis type (C-type or S-type) and Cobb angle range (10° to 25° for mild, 25° to 40° for moderate).

[0046] In terms of spatiotemporal sequence matching, this invention employs a weighted dynamic time warping algorithm to align real-time pose frame sequences with standard motion templates. Unlike traditional dynamic time warping algorithms, this invention introduces a keypoint weighting mechanism tailored to scoliosis rehabilitation scenarios. Specifically, let the system-defined keypoint weight vector be... The weights of key points related to the spine (midpoint of the neck, left shoulder, right shoulder, left hip, right hip, midpoint of the hip) are as follows: Set to 2.0, and the weights of the key points of the limbs (left and right elbows, left and right wrists, left and right knees, left and right ankles) are... The weight of the head key points (nose tip, left and right eyes, left and right ears) is set to 0.8. The weight is set to 0.5. That is, in this embodiment, the matching weight of spine-related key points is 2.5 times that of limb key points, a ratio falling within the preferred range of 1.5 to 3 times. When the patient's scoliosis is mild (Cobb angle less than 20°), a lower limit of 1.5 times can be used to avoid oversensitivity in spinal region assessment and misjudgment of normal physiological posture fluctuations. When the patient's scoliosis is severe (Cobb angle greater than 30°), an upper limit of 3 times can be used to further strengthen the bias assessment weight of the spinal region, thereby forming an assessment sensitivity adapted to the severity of scoliosis between the matching contributions of spine-related key points and limb key points. The formula for calculating the weighted inter-frame distance is:

[0047] ,

[0048] in: The first of the real-time attitude sequence Frame key point set; The first standard action template Frame key point set; For the first The weighting coefficients for each key point are dimensionless. This is the Euclidean distance calculation operator, with units in pixels. Before calculating the inter-frame distance, the system first performs translational normalization processing on the key point coordinates of the real-time frame and the template frame with the midpoint of the hip as the origin and scale normalization processing based on the trunk length, thereby eliminating the differences in body shape and standing distance between the patient and the template collector.

[0049] Based on the aforementioned weighted inter-frame distance, the system constructs a cumulative distance matrix using a standard dynamic time warping recursive formula and backtracks to the optimal matching path. Preferably, to control computational complexity and avoid overstretching the matching, the Sakoe-Chiba constraint band width is set to 20% of the template length. After matching is completed, an action similarity score is calculated. Defined as the negative exponential transformation of the normalized average path distance:

[0050] ,

[0051] in: The cumulative distance of the optimal path output by the dynamic time warping algorithm, in pixels; The length of the optimal matching path is the number of matching pairs contained in the path; it is a dimensionless positive integer. The normalized baseline distance is the median of the average inter-frame distance of all action templates in the template library, expressed in pixels. The range of values ​​is The closer the value is to 1, the higher the degree of conformity between the real-time action and the standard template.

[0052] In terms of deviation recognition, the system is based on action similarity scores. and posture comprehensive assessment vector The joint analysis automatically identifies three common types of training biases.

[0053] The first type is insufficient range of motion deviation. After dynamic time warping and matching are completed, the system extracts the angle change range of each target joint in the real-time motion sequence within the motion execution cycle. The range of angle changes of the corresponding joints in the standard template When comparing, If the angle is insufficient, it is determined that the joint angle is not sufficient. Preferably, the joint angle is calculated using the three-point angle method, that is, two vectors are determined by the coordinates of three adjacent key points, and the angle between the two vectors is taken as the bending angle of the joint.

[0054] The second category is rhythmic anomaly deviation. The system analyzes the local slope characteristics of the optimal matching path in dynamic time warping. Under normal rhythm, the matching path should approximate a diagonal line with a slope of 1. When the local slope of multiple consecutive matching pairs on the path... The deviation from the unit slope exceeds the preset rhythm tolerance threshold. The time is judged as an abnormal rhythm, and the specific conditions are as follows: ,in The preferred setting is 0.4. When A value significantly greater than 1 indicates that the patient's actions are performed too quickly, and vice versa.

[0055] The third category is compensatory postural deviation. Identifying this deviation requires the combined use of the comprehensive postural assessment vector output from step S2. During the execution of the target movement, the system monitors the angular offset of non-target joints (i.e., joints where the current movement does not require active motion). When the angular offset of a non-target joint exceeds a preset compensatory tolerance range... (Preferably set to 10°) and simultaneously the trunk symmetry in the overall posture assessment vector. When a sudden drop occurs with an amplitude exceeding 0.1, it is judged as a compensatory postural deviation. This joint judgment mechanism is one of the key innovations of this invention. Its essence is that relying solely on changes in the angle of non-target joints may produce many false alarms, because patients will also have a certain range of natural movements in non-target joints during normal training; however, by combining it with the abrupt change characteristics of trunk symmetry, it is possible to accurately distinguish between normal coordinated movements and true compensatory behaviors, thereby significantly improving the accuracy of deviation identification.

[0056] Step S4: Real-time correction feedback and adaptive training guidance step. Step S4 receives the training deviation identification result output from step S3 as input and performs two core functions: correction information generation and closed-loop parameter adjustment.

[0057] Regarding the generation of correction information, the system automatically generates corresponding real-time correction prompts based on the identified deviation type and severity. In one embodiment of the invention, the correction prompts include two modalities: visual prompts and voice prompts. Visual prompts are presented by overlaying skeletal annotations onto the real-time video screen, highlighting the joint positions with deviations in red, and indicating the correct direction of movement and target position with arrows. Voice prompts generate corresponding natural language instructions based on the deviation type. For example, when an insufficient range of motion deviation is detected, the voice prompt is "Please extend further to the left; the target angle has not yet been reached." When an excessively fast rhythm deviation is detected, the voice prompt is "The movement speed is too fast; please slow down and coordinate with your breathing." When a compensatory posture deviation is detected, the voice prompt is "Please keep your pelvis stable and avoid hip compensation." Preferably, the trigger interval for voice prompts is not less than 3 seconds to avoid interrupting the training rhythm with overly frequent prompts.

[0058] Regarding closed-loop parameter adjustment, the output signal of step S4 is fed back to steps S2 and S3 to dynamically adjust the evaluation parameters of the front-end module. This is the core mechanism for constructing an end-to-end closed-loop monitoring system in this invention. The specific adjustment strategy includes the following two levels.

[0059] Firstly, adaptive adjustment of the threshold for determining body symmetry analysis in step S2. When the system detects continuous... Frame (preferred) Within a training duration of approximately 5 seconds, the lateral offset features in the overall posture evaluation vector. It shows a continuous downward trend and the cumulative decline exceeds the preset convergence threshold. (Preferred) When the threshold for abnormal shoulder height difference is reduced to 6 pixels (pixels), the system determines that the patient's posture is improving. At this point, it automatically reduces the threshold for abnormal shoulder tilt angle from the initial 8 pixels to 6 pixels and the threshold for abnormal pelvic tilt angle from the initial 3° to 2°. This adjustment allows the system to gradually increase its assessment sensitivity as the patient's posture continues to improve, avoiding the oversight of subtle postural deviations in later stages due to fixed thresholds.

[0060] Secondly, adaptive adjustment of the dynamic time warping matching sensitivity parameter in step S3. When the system detects continuous... One action cycle (preferred) Internal action similarity score Continuously improving and in the most recent cycle When the value exceeds 0.85, the system determines that the patient's motor execution ability is improving. At this point, it automatically increases the threshold for insufficient amplitude judgment from the initial 0.7 to 0.8, and adjusts the rhythm tolerance threshold. The compensation tolerance range was reduced from the initial 0.4 to 0.3. The angle was reduced from an initial 10° to 7°. This adjustment increases the training difficulty while ensuring continuous improvement in training quality, achieving personalized, progressive training guidance.

[0061] Preferably, the above adaptive adjustment process includes a safety rollback mechanism. When the adjusted threshold causes the deviation detection rate to suddenly exceed 80% in a certain assessment, the system automatically rolls back the relevant threshold to the value before adjustment and records the rollback event in the training log for subsequent analysis and evaluation by therapists. It is worth noting that the above closed-loop feedback mechanism is not a simple linear scaling of the threshold, but rather an intelligent decision-making process based on the statistical characteristics of the training data. In one embodiment of the invention, the system internally maintains a sliding window buffer to store the most recent... One action cycle (preferred) The system uses evaluation data from a buffer to determine whether a patient's training status has steadily improved, based on the variance and trend slope of the data within that buffer. Parameter tightening is triggered only when the variance is below a preset threshold and the trend slope is positive; when the variance suddenly increases, the system either maintains the current parameters or appropriately loosens them to accommodate potential fluctuations in the patient's status. This statistical confidence-based adjustment strategy effectively avoids over-adjustment of parameters due to isolated, occasional performance.

[0062] Furthermore, in a preferred embodiment of the present invention, the standard rehabilitation movement template in step S3 is not fixed; the system supports therapists uploading customized movement templates via a remote management interface. When recording standard movements, therapists use the same RGB camera and skeletal keypoint estimation process as the patient to capture their own movement sequences, which are then injected into the template library on the patient's end after quality review. This mechanism allows therapists to customize more personalized training programs based on the specific scoliosis characteristics and rehabilitation stages of different patients, further enhancing the clinical applicability of the system.

[0063] Step S5: Rehabilitation training log generation and remote collaborative management step. Step S5 receives the processing data from each step from S1 to S4, performs comprehensive data aggregation and visualization of a complete rehabilitation training process, and generates a structured rehabilitation training log.

[0064] Regarding the assessment of movement completion, the system calculates a comprehensive completion score for each rehabilitation movement performed during this training session. In one embodiment of the present invention, the score comprehensively considers performance across three dimensions: movement similarity, body balance, and rhythm control. The specific calculation formula is as follows:

[0065] ,

[0066] in: The action similarity score calculated in step S3 has a value range of [value range missing]. ; The normalized mapping function for the comprehensive posture assessment vector maps the five-dimensional posture indicators to... A single score value within the interval; The rhythm control score is defined as the proportion of matched pairs whose local slope in the matching path is within the normal range, with a value range of [value missing]. ; , , The weighting coefficients are preferably 0.4, 0.35, and 0.25, with the sum of the three being 1. The range of values ​​is Preferably, when It was judged to be good at the time. It was judged as qualified at that time. It was determined that improvements were needed.

[0067] In generating posture deviation heatmaps, the system uses the human skeletal outline as a base map and maps the cumulative deviation values ​​of each skeletal key point during training into color gradients for visualization. Specifically, for each skeletal key point... The system calculates the average deviation distance of the target relative to the corresponding key points of the standard action template throughout the entire training process. The deviation value is then converted into a chromatographic value from blue (cool tone, indicating small deviation) to red (warm tone, indicating large deviation) using a linear mapping function, generating a heat map that is overlaid on the skeletal diagram. This heat map allows patients and therapists to visually identify body parts where deviations are concentrated during training.

[0068] Regarding the generation of improvement suggestions, the system automatically generates textual improvement suggestions based on the deviation identification results and scoring data. Preferably, the improvement suggestions are arranged according to priority, with the actions with the lowest scores and the body parts with the most serious deviations placed at the top of the suggestion list, and specific corrective instructions are provided in conjunction with the standard execution guidelines corresponding to the action.

[0069] Furthermore, in one embodiment of the present invention, the rehabilitation training log also includes a personalized training plan automatically recommended based on the patient's scoliosis type and Cobb angle range. The system dynamically updates the recommended types and intensity levels of training movements based on historical training data and postural improvement trends recorded in the rehabilitation training log. Preferably, when training data for two consecutive weeks shows that the average completion score for a certain type of movement is consistently higher than 0.85, the system automatically suggests increasing the difficulty level of that movement by one level or increasing the number of training sets.

[0070] After the training log is generated, the system encapsulates it in a structured data format (preferably JSON) and uploads it to a remote server via an encrypted network communication channel, allowing therapists to view it through a dedicated client. Therapists can add review comments and personalized guidance to the log, and the comments are automatically pushed to the client for display the next time the patient starts training.

[0071] In summary, the scoliosis rehabilitation training posture image monitoring method provided by this invention provides basic data for subsequent analysis through skeletal posture estimation in step S1; extracts scoliosis-specific assessment features through spinal curve fitting and symmetry joint analysis in step S2; achieves objective quantitative assessment of rehabilitation training quality through motion matching and deviation identification in step S3; enables the system to adaptively optimize assessment parameters through a closed-loop feedback mechanism in step S4; and achieves persistent recording and remote collaboration of training data through log generation and remote management in step S5. There is a deep data flow dependency between the five steps, and the feedback adjustment from step S4 to steps S2 and S3 constitutes a complete closed-loop control structure, enabling the entire system to continuously optimize its monitoring accuracy and training guidance effect during continuous use.

[0072] Reference Figure 2 As shown, in one embodiment of the present invention, a scoliosis rehabilitation training posture image monitoring system is provided. The system is used to perform the scoliosis rehabilitation training posture image monitoring method described in the above method embodiment. It includes five functional modules: skeletal posture estimation module 1, spinal curve and symmetry joint analysis module 2, action matching and deviation recognition module 3, correction feedback and adaptive adjustment module 4, and log generation and remote management module 5.

[0073] The skeletal pose estimation module 1 performs the functions described in step S1 above. This module acquires video stream data of the patient's rehabilitation training process through a video acquisition interface connected to an RGB camera. Internally, it integrates a lightweight human target detection subunit and a multi-joint skeletal keypoint estimation subunit. The human target detection subunit performs single-stage anchor box detection on each input image frame, outputting the coordinates of the human bounding box. The skeletal keypoint estimation subunit performs top-down multi-joint localization within the bounding box region, outputting two-dimensional coordinates and confidence data containing 17 basic keypoints and 2 derived keypoints. Preferably, this module is deployed on an edge computing device with GPU acceleration capabilities, enabling pipelined processing of detection and estimation with a single-frame latency of no more than 30ms. The output of the skeletal pose estimation module 1 is a two-dimensional pose skeleton sequence, which is simultaneously transmitted to the spinal curve and symmetry joint analysis module 2 and the action matching and deviation recognition module 3.

[0074] The spinal curve and symmetry joint analysis module 2 performs the functions described in step S2 above. This module receives the two-dimensional posture skeleton sequence output by the skeletal posture estimation module 1, extracts the spinal keypoint sub-sequence, and performs cubic B-spline curve fitting, lateral offset calculation, rotation angle feature extraction, and calculation of three posture symmetry indices. This module has a configurable threshold parameter register to store judgment parameters such as abnormal thresholds for shoulder height difference, abnormal thresholds for pelvic tilt angle, and abnormal thresholds for trunk symmetry. These threshold parameters are controlled by feedback signals from the correction feedback and adaptive adjustment module 4, enabling dynamic updates during training. The output of this module is a comprehensive posture evaluation vector, which is transmitted to the action matching and deviation recognition module 3.

[0075] The motion matching and deviation recognition module 3 performs the functions described in step S3 above. This module simultaneously receives two inputs: a two-dimensional posture skeleton sequence and a comprehensive posture evaluation vector. Internally, it integrates a standard rehabilitation motion template library storage unit and a weighted dynamic time warping matching engine. The standard rehabilitation motion templates are stored in the standard rehabilitation motion template library. The template library storage unit contains pre-set standard motion template data for two rehabilitation programs: Schroth and SEAS. The matching engine automatically selects a subset of matching templates based on the patient's current scoliosis type and training program. Sensitivity parameters within this module, such as the rhythm tolerance threshold, amplitude judgment ratio, and compensation tolerance range, are also controlled by feedback signals from the correction feedback and adaptive adjustment module 4. The output of this module includes motion similarity scores and training deviation recognition results, which are transmitted to the correction feedback and adaptive adjustment module 4.

[0076] The correction feedback and adaptive adjustment module 4 performs the function of step S4 above. This module receives the deviation recognition result output by the action matching and deviation recognition module 3, and internally integrates a correction information generation subunit and a parameter adaptive adjustment subunit. The correction information generation subunit generates visual overlay annotations and voice prompts based on the deviation type and severity. The parameter adaptive adjustment subunit calculates new threshold parameters based on the improvement trend of the training state, and writes the adjusted parameters into the threshold parameter registers of the spinal curve and symmetry joint analysis module 2 and the action matching and deviation recognition module 3 respectively through the feedback channel, thereby realizing a closed-loop feedback from deviation recognition to parameter optimization. Preferably, this module also has a safety rollback function; when the threshold adjustment causes an abnormal spike in the deviation detection rate, it automatically restores the parameter value to the value before adjustment.

[0077] The log generation and remote management module 5 performs the functions described in step S5 above. This module aggregates all training data from the four modules mentioned above, including keypoint coordinates for each frame, time series of posture assessment vectors, similarity scores for each movement, deviation identification records, and correction feedback logs. Based on this, the module generates a visualized rehabilitation training log containing movement completion scores, posture deviation heatmaps, and improvement suggestions, and uploads the log to a remote server via an encrypted network communication interface. This module also receives review comments and training plan adjustment instructions sent by therapists remotely and pushes them to the patient's client interface when the next training session begins. Preferably, this module has a built-in data compression and incremental synchronization mechanism. For basic data that changes little in multiple consecutive training sessions, the complete version is transmitted only during the first upload; subsequent training sessions only upload the differing data and newly added assessment results, thereby effectively reducing network bandwidth consumption and improving the user experience in weak network environments.

[0078] It is important to note that the data flow between the five modules is not a unidirectional linear structure, but rather forms a closed-loop topology with feedback loops. Specifically, the output of the skeletal posture estimation module 1 flows simultaneously to the spinal curve and symmetry joint analysis module 2 and the motion matching and deviation recognition module 3, forming a bifurcated structure with one input and two parallel processing paths. The output of the spinal curve and symmetry joint analysis module 2, which is the overall posture assessment vector, is further fed into the motion matching and deviation recognition module 3, where it is fused with the skeletal sequence data as a dual input for analysis. The output of the motion matching and deviation recognition module 3 is transmitted to the correction feedback and adaptive adjustment module 4, while the parameter adjustment signal of the correction feedback and adaptive adjustment module 4 is transmitted in reverse to the spinal curve and symmetry joint analysis module 2 and the motion matching and deviation recognition module 3, forming a closed-loop feedback path. The log generation and remote management module 5 acts as a data aggregation terminal, collecting full process data from each module. This system architecture with closed-loop feedback characteristics enables the entire monitoring system to continuously optimize based on the patient's real-time training status and historical improvement trends, unlike traditional open-loop assessment systems.

[0079] Furthermore, in a preferred embodiment of the present invention, the software portion of the system can be deployed on various hardware platforms. For laptops or desktops with dedicated GPUs, each module can run using a native deep learning inference framework to achieve optimal processing performance. For mobile devices (such as smartphones or tablets) without dedicated GPUs, the skeletal keypoint estimation network can be replaced with a lightweight model variant optimized for mobile devices, which can still meet the basic needs of rehabilitation training monitoring while moderately reducing the accuracy of keypoint localization. Preferably, when the system detects that the computing power of the running platform is insufficient to support a real-time processing rate of 25fps, it automatically enables the inter-frame processing mode, that is, a complete skeletal keypoint estimation and evaluation process is performed every other frame, and the keypoint coordinates are completed through linear interpolation in intermediate frames, reducing the processing load by about 50% while maintaining acceptable monitoring continuity.

[0080] To verify the effectiveness of the technical solution of this invention, the inventors set up the following test environment: a USB camera with a resolution of 640×480 pixels and a frame rate of 30fps was used as the video acquisition device, and the computing platform was a laptop equipped with an Intel Core i5 processor and an NVIDIA GTX 1650 graphics card. The test dataset included Schroth corrective gymnastics training videos recorded by 20 patients with scoliosis whose Cobb angle ranged from 15° to 35° under the guidance of rehabilitation therapists. Each patient recorded 5 standard movements, each movement was repeated 3 times, for a total of 300 movement execution samples. The therapists simultaneously provided manual scores for each movement execution sample as real labels.

[0081] Regarding the accuracy of skeletal keypoint estimation, the top-down pose estimation network used in this invention has an average keypoint localization error of 4.2 pixels on the test set, which is equivalent to a physical distance of about 1.7 cm. The average localization error of the five keypoints in the spine region is 3.8 pixels, which is slightly better than the average level of keypoints throughout the body. This is due to the inherent advantage of the keypoints in the spine region being less obstructed in a frontal view.

[0082] Regarding the accuracy of spinal curve fitting, using the spinal midline manually marked by the therapist as a reference, the average fitting residual of the cubic B-spline fitting method of this invention is 1.6 pixels, and the maximum fitting residual is 3.1 pixels, both meeting the preset requirement of a 2-pixel average residual threshold. The Pearson correlation coefficient between the extracted lateral offset feature and the Cobb angle measured by the therapist based on X-rays reaches 0.82, indicating that the spinal curve fitting results based on a common RGB camera can reflect the severity of clinical scoliosis to a certain extent.

[0083] Regarding the accuracy of motion matching, the weighted dynamic time warping algorithm of this invention improves the matching accuracy from 81.3% to 89.7% compared to the traditional equal-weighted dynamic time warping algorithm, an improvement of 8.4 percentage points. This improvement is mainly attributed to the higher matching weights given to key points in the spine and trunk regions, allowing the algorithm to pay more attention to the consistency of motion trajectories in the spinal region when evaluating rehabilitation movements, which aligns with the clinical assessment focus of scoliosis rehabilitation training.

[0084] In terms of deviation identification, the accuracy rate of compensatory posture deviation identification using the combined postural comprehensive evaluation vector reached 87.2%, while the accuracy rate of the single judgment method relying solely on joint angle changes was only 72.6%, representing an improvement of 14.6 percentage points. This verifies that the synergistic fusion of postural indicators in step S2 and action matching in step S3 can produce a significant nonlinear gain effect. Specific analysis shows that among the 22 compensatory behaviors missed by the single judgment method, 18 were accompanied by a significant drop in trunk symmetry and were successfully captured by the combined judgment method. The accuracy rate for identifying insufficient movement amplitude deviation was 91.5%, the accuracy rate for identifying rhythmic abnormality deviation was 88.3%, and the weighted average accuracy rate for the comprehensive identification of the three types of deviations was 89.0%.

[0085] Regarding closed-loop feedback, after enabling the adaptive parameter adjustment mechanism in step S4, the overall evaluation error rate of the system decreased from the initial 12.5% ​​to 7.3% after 10 training cycles, indicating that the closed-loop feedback mechanism can continuously optimize the monitoring accuracy of the system. In contrast, the fixed parameter scheme without closed-loop feedback only reduced the error rate from 12.5% ​​to 10.8% after the same training cycle, a significantly smaller improvement than the closed-loop scheme.

[0086] In terms of real-time processing performance, the overall processing latency for skeletal keypoint estimation and motion matching on the test platform is 45ms per frame, meeting the real-time requirement of over 25fps. Specifically, skeletal keypoint estimation takes approximately 15ms, spine curve fitting and symmetry analysis take approximately 8ms, dynamic time warping matching takes approximately 18ms, and correction information generation takes approximately 4ms. Even on an Intel Core i5 integrated graphics platform without a dedicated GPU, an effective processing rate of 15fps can still be achieved after enabling the interleaved frame processing mode.

[0087] Regarding user experience, among the 20 patients who participated in the test, 17 reported that the real-time correction prompts provided by the system significantly helped improve the quality of their training movements, with 85% giving a satisfaction rating of "good" or above. Meanwhile, all 5 therapists who participated in the remote review stated that the visualized posture deviation heatmap and quantified movement completion scores significantly improved the efficiency and relevance of remote guidance. The average time for therapists to review a single training log was reduced from approximately 15 minutes using traditional video playback to approximately 5 minutes, increasing remote guidance efficiency by about 3 times. This fully demonstrates the practical value of the rehabilitation training log visualization solution of this invention in clinical remote collaborative management.

[0088] The above test results show that the scoliosis rehabilitation training posture image monitoring method and system provided by the present invention, through the innovative design of spinal curve fitting and symmetry joint analysis, the deviation identification mechanism of weighted dynamic time warping and postural index coordination, and the end-to-end closed-loop feedback adjustment strategy, can achieve accurate monitoring and intelligent guidance of scoliosis patients' home rehabilitation training movements under ordinary consumer-grade camera conditions. It effectively solves the technical problem of lack of objective monitoring and real-time guidance for patients' home training, and provides a practical and feasible technical support solution for the remote rehabilitation management of scoliosis patients.

[0089] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for monitoring posture images during scoliosis rehabilitation training, characterized in that, Includes the following steps: Step S1, Human skeleton pose estimation and two-dimensional skeleton sequence construction steps: The video stream of the rehabilitation training process of scoliosis patients is acquired by an RGB camera. Lightweight human target detection is performed on each frame of the video stream to obtain the human bounding box region. Multi-joint skeletal key points are estimated in real time within the human bounding box region to obtain a two-dimensional coordinate set including key nodes of the spine, scapula, pelvis and limbs. The two-dimensional coordinate sets of multiple consecutive frames are arranged in time sequence to construct a two-dimensional pose skeleton sequence. Step S2, Joint Analysis of Spinal Curve Fitting and Postural Symmetry: Extract the spinal key point subsequence between the cervical and sacral nodes from the two-dimensional posture skeleton sequence, perform parametric curve fitting on the spinal key point sequence to generate the spinal midline fitting curve, extract lateral offset features and rotation angle features based on the spinal midline fitting curve, simultaneously calculate the shoulder height difference index, hip tilt angle index and trunk left and right symmetry index, and fuse the lateral offset features, rotation angle features and various postural balance indicators to generate a comprehensive postural evaluation vector; Step S3, Rehabilitation movement spatiotemporal sequence matching and training deviation identification steps: The real-time posture frame sequence in the two-dimensional posture skeleton sequence is dynamically time-warped and matched with the preset standard rehabilitation movement template. The movement similarity score is calculated based on the matching path, and the deviations of insufficient movement amplitude, abnormal rhythm, and compensatory posture are jointly identified based on the comprehensive posture evaluation vector and the movement similarity score. Step S4, Real-time Correction Feedback and Adaptive Training Guidance: Generate instant correction prompts based on the type and severity of training deviations, and dynamically adjust the judgment threshold of body symmetry analysis in step S2 and the sensitivity parameters of dynamic time warping matching in step S3 based on the training state after correction.

2. The method for monitoring posture images during scoliosis rehabilitation training according to claim 1, characterized in that, In step S1, the lightweight human target detection uses a single-stage anchor box detection network, the input image resolution is 640×480 pixels, the detection confidence threshold is set to 0.5, and the non-maximum suppression threshold of the human bounding box is set to 0.45; the real-time estimation of multi-joint skeletal key points uses a top-down pose estimation network, and the number of estimated skeletal key points is not less than 17. The skeletal key points include at least the tip of the nose, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, left and right ankles, and the midpoint of the neck and the midpoint of the hip.

3. The method for monitoring posture images during scoliosis rehabilitation training according to claim 1, characterized in that, In step S2, the parametric curve fitting adopts the cubic B-spline curve fitting method, the number of control points is set to 5 to 8, and the fitting residual threshold is set to no more than 2 pixels; the lateral offset feature is the maximum horizontal offset distance of each sampling point on the spinal axis fitting curve relative to the line connecting the cervical and sacral nodes, and the rotation angle feature is the angle between the tangent of the spinal axis fitting curve at the point of maximum lateral offset and the vertical direction.

4. The method for monitoring posture images during scoliosis rehabilitation training according to claim 1, characterized in that, In step S2, the shoulder height difference index is the absolute value of the difference in the vertical coordinate between the left and right shoulder key points, the hip tilt angle index is the angle between the line connecting the left and right hip key points and the horizontal reference line, and the trunk left-right symmetry index is obtained by calculating the area ratio of the left and right regions of the spine axis fitting curve. The closer the area ratio is to 1, the higher the symmetry.

5. The method for monitoring posture images during scoliosis rehabilitation training according to claim 1, characterized in that, In step S3, the dynamic time warping matching adopts a weighted dynamic time warping algorithm, wherein the matching weight of the spinal-related key points is set to 1.5 to 3 times the matching weight of the limb key points; the standard rehabilitation movement template is stored in the standard rehabilitation movement template library, which contains Schroth three-dimensional corrective gymnastics movement templates and SEAS scientific exercise method movement templates.

6. The method for monitoring posture images during scoliosis rehabilitation training according to claim 1, characterized in that, In step S3, the determination condition for insufficient movement amplitude deviation is: the target joint angle change range is less than 70% of the corresponding joint angle change range in the standard movement template; the determination condition for abnormal rhythm deviation is: the slope of the dynamic time warping path between the real-time movement sequence and the standard movement template deviates from the unit slope by more than a preset rhythm tolerance threshold; the determination condition for compensatory posture deviation is: during the execution of the target movement, the angle deviation of the non-target joint exceeds the preset compensation allowable range.

7. The method for monitoring posture images during scoliosis rehabilitation training according to claim 1, characterized in that, In step S4, the dynamic adjustment includes: when the lateral offset feature in the body posture comprehensive evaluation vector continuously decreases within a preset number of consecutive frames and the decrease exceeds a preset convergence threshold, narrowing the range of the judgment threshold for body posture symmetry analysis in step S2 to improve evaluation sensitivity; when the action similarity score continuously increases for multiple consecutive action cycles, increasing the similarity judgment threshold for dynamic time warping matching in step S3 to increase training difficulty.

8. The method for monitoring posture images during scoliosis rehabilitation training according to claim 7, characterized in that, It also includes step S5, the rehabilitation training log generation and remote collaborative management step: summarizing the processing data of each step from S1 to S4, generating a visualized rehabilitation training log containing the completion score of each movement, a posture deviation heatmap and improvement suggestions, and uploading the rehabilitation training log to the remote server for therapists to review. The posture deviation heatmap uses the outline of the human skeleton as the base map and maps the cumulative deviation value at each key bone point to a color gradient for visualization. The area with the larger the cumulative deviation value is presented in a warm color, and the area with the smaller the cumulative deviation value is presented in a cool color.

9. The method for monitoring posture images during scoliosis rehabilitation training according to claim 8, characterized in that, In step S5, the rehabilitation training log also includes a personalized training plan automatically recommended based on the patient's scoliosis type and Cobb angle range. The personalized training plan is dynamically updated based on the historical training data and posture improvement trends recorded in the rehabilitation training log.

10. A posture image monitoring system for scoliosis rehabilitation training, characterized in that, For performing the posture image monitoring method for scoliosis rehabilitation training as described in claim 9, the system comprises: The skeletal pose estimation module is used to acquire video streams of the rehabilitation training process of scoliosis patients through an RGB camera, perform lightweight human target detection on each frame of the video stream to obtain the human bounding box region, perform real-time estimation of multi-joint skeletal key points within the human bounding box region, and construct a two-dimensional pose skeleton sequence. The spinal curve and symmetry joint analysis module is used to extract the spinal key point subsequence from the two-dimensional posture skeleton sequence and perform parametric curve fitting to generate the spinal midline fitting curve, extract lateral offset features and rotation angle features, and simultaneously calculate the shoulder height difference index, hip tilt angle index and trunk left and right symmetry index, and fuse them to generate a comprehensive posture evaluation vector. The motion matching and deviation recognition module is used to dynamically time-warp match real-time posture frame sequences with standard rehabilitation motion templates, calculate motion similarity scores, and identify training deviations in conjunction with the comprehensive posture assessment vector. The correction feedback and adaptive adjustment module is used to generate real-time correction prompts based on training deviations, and dynamically adjust the judgment threshold of the spinal curve and symmetry joint analysis module and the matching sensitivity parameters of the action matching and deviation recognition module. The log generation and remote management module is used to aggregate the data processed by each module to generate a visual rehabilitation training log that includes a performance score, a heatmap of postural deviations, and improvement suggestions, and then uploads the rehabilitation training log to a remote server.

Images