Intelligent mirror multi-dimensional health assessment system

Abstract

The invention provides an intelligent mirror multi-dimensional health assessment system, and relates to the technical field of health assessment, and the system comprises an acquisition module which is used for carrying out human body skeletonization processing on an acquired continuous three-dimensional image sequence of user motion and position information of user limbs in a space, extracting three-dimensional space coordinate data of key joints of a user body, and transmitting the three-dimensional space coordinate data to a processing module; obtaining a real-time skeleton model of user motion; and the pointing module is used for establishing a first reference vector pointing to the central points of the left and right shoulder joints and a second reference vector pointing to the central points of the left and right hip joints by taking the central point of the trunk of the user as a space base point based on the real-time skeleton model, so as to determine a main reference plane of the movement of the user. According to the method, through extraction of motion skeleton data, construction of a reference and dynamic feature plane, calculation and evaluation of key indexes, generation of motion attitude correction parameters and visual calibration guidance, evaluation and real-time optimization of motion quality are realized, the motion effect is improved, and the damage risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of health assessment technology, and in particular to a smart mirror multi-dimensional health assessment system. Background Technology

[0002] Smart mirrors with health monitoring functions typically capture images of the user's movements through integrated cameras and perform preliminary posture analysis based on these images. However, when analyzing movements involving the swinging of the core body parts, such as bending over or twisting the body, the evaluation dimensions of such systems may be limited.

[0003] Specifically, some systems may rely primarily on changes in the position of joints within a two-dimensional image plane to analyze posture, such as estimating the range of motion by monitoring the pixel coordinates of joints in the image. However, this method is sometimes difficult to accurately quantify the spatial displacement and rotation of limbs in the direction perpendicular to the mirror.

[0004] For example, when a user bends to the side, the tilt angle of their torso and the stability of their pelvis are key to assessing the quality of the movement. However, if observed only from a two-dimensional perspective, it may not be possible to fully capture the actual displacement of the body in three-dimensional space, resulting in incomplete assessment information. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a smart mirror multi-dimensional health assessment system to provide scientific guidance for optimizing exercise posture.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a smart mirror multi-dimensional health assessment system includes:

[0008] The acquisition module is used to perform human skeletonization processing on the acquired continuous three-dimensional image sequence of user movement and the position information of user limbs in space, extract the three-dimensional spatial coordinate data of key joints of the user's body, and obtain a real-time skeleton model of user movement.

[0009] The pointing module is used to establish a first reference vector pointing to the center points of the left and right shoulder joints and a second reference vector pointing to the center points of the left and right hip joints, based on the real-time skeleton model and using the center point of the user's torso as the spatial reference point, so as to determine the main reference plane of the user's movement.

[0010] The construction module is used to select four feature points—the left and right shoulder joints, the left and right hip joints—based on the main reference plane and the real-time skeleton model, in order to construct a dynamic feature plane that represents the limb movement state.

[0011] The calculation module is used to obtain the spatial attitude offset by calculating the angle between the normal vectors of the dynamic feature plane and the main reference plane, and to construct the minimum bounding rectangle based on the spatial positions of the four feature points to obtain the motion stability index.

[0012] The evaluation module is used to conduct a health assessment of the user's movement quality based on spatial posture offset and motion stability indicators, combined with the three-dimensional spatial coordinate data of the user's key joints, through a pre-trained evaluation model, and generate a comprehensive evaluation result including motion posture correction parameters.

[0013] The display module generates visual content, including spatial attitude calibration guidance and motion stability optimization suggestions, based on the comprehensive evaluation results, and displays it on the smart mirror in real time.

[0014] In a second aspect, a computing device includes:

[0015] One or more processors;

[0016] A storage device for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the system.

[0017] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0018] The above-described solution of the present invention has at least the following beneficial effects:

[0019] By capturing the three-dimensional spatial information of user movements and constructing a real-time skeleton model, the system captures the spatial displacement and rotation of limbs in the direction perpendicular to the mirror surface. This comprehensively covers movement scenarios involving the swinging of core parts, such as bending over and turning, solving the problem of the system's difficulty in quantifying three-dimensional motion characteristics and making health assessments more comprehensive. Based on the construction of the main reference plane and dynamic feature plane, combined with the system calculation of spatial posture offset and motion stability indicators, the system can objectively reflect the postural standardization and limb stability of users during movement, avoiding the bias of assessments relying on subjective judgment or single-dimensional data, and improving the accuracy and reliability of health assessments. The pre-trained assessment model integrates the three-dimensional spatial coordinate data of the user's key joints, and the generated motion posture correction parameters are closely matched to the user's individual motion characteristics, making them highly targeted. This helps users locate the deviations in their movements, provides scientific guidance for motion posture optimization, helps users develop standardized exercise habits, and reduces the risk of sports injuries caused by non-standard movements. Real-time visualized spatial posture calibration guidance and motion stability optimization suggestions are presented intuitively on the smart mirror, allowing users to obtain feedback and adjust their movements immediately. The system is easy to operate and highly interactive, improving its usability and shortening the cycle of user motion posture improvement, thus enhancing the practicality of health assessments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a smart mirror multi-dimensional health assessment system provided by an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating how, based on the main reference plane and a real-time skeleton model, four feature points—the left and right shoulder joints, the left and right hip joints—are selected to construct a dynamic feature plane representing the limb's motion state. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] like Figure 1 As shown, an embodiment of the present invention proposes a smart mirror multi-dimensional health assessment system, comprising:

[0024] The acquisition module is used to perform human skeletonization processing on the acquired continuous three-dimensional image sequence of user movement and the position information of user limbs in space, extract the three-dimensional spatial coordinate data of key joints of the user's body, and obtain a real-time skeleton model of user movement.

[0025] The pointing module is used to establish a first reference vector pointing to the center points of the left and right shoulder joints and a second reference vector pointing to the center points of the left and right hip joints, based on the real-time skeleton model and using the center point of the user's torso as the spatial reference point, so as to determine the main reference plane of the user's movement.

[0026] The construction module is used to select four feature points—the left and right shoulder joints, the left and right hip joints—based on the main reference plane and the real-time skeleton model, in order to construct a dynamic feature plane that represents the limb movement state.

[0027] The calculation module is used to obtain the spatial attitude offset by calculating the angle between the normal vectors of the dynamic feature plane and the main reference plane, and to construct the minimum bounding rectangle based on the spatial positions of the four feature points to obtain the motion stability index.

[0028] The evaluation module is used to conduct a health assessment of the user's movement quality based on spatial posture offset and motion stability indicators, combined with the three-dimensional spatial coordinate data of the user's key joints, through a pre-trained evaluation model, and generate a comprehensive evaluation result including motion posture correction parameters.

[0029] The display module generates visual content, including spatial attitude calibration guidance and motion stability optimization suggestions, based on the comprehensive evaluation results, and displays it on the smart mirror in real time.

[0030] In this embodiment of the invention, by capturing the three-dimensional spatial information of user movement and constructing a real-time skeleton model, the spatial displacement and rotation state of limbs in the direction perpendicular to the mirror surface are captured, comprehensively covering movement scenarios involving the swinging of core parts such as bending over and turning over. This solves the problem that the system is difficult to quantify three-dimensional motion characteristics, making health assessment more comprehensive. Based on the construction of the main reference plane and the dynamic feature plane, combined with the system calculation of spatial posture offset and motion stability index, the system can objectively reflect the posture norms and limb stability of the user during the movement process, avoiding the bias of assessment relying on subjective judgment or single-dimensional data, and improving the accuracy and reliability of health assessment. The pre-trained assessment model integrates three-dimensional spatial coordinate data of key joints in the user's body, generating motion posture correction parameters that closely match the user's individual motion characteristics. This highly targeted approach helps users pinpoint movement deviations, providing scientific guidance for motion posture optimization and helping users develop standardized exercise habits, thus reducing the risk of sports injuries caused by non-standard movements. Real-time visualized spatial posture calibration guidance and motion stability optimization suggestions are presented intuitively on the smart mirror, allowing users to receive immediate feedback and adjust their movements. The system is easy to operate and highly interactive, improving its usability and shortening the cycle of motion posture improvement, thereby enhancing the practicality of health assessments.

[0031] In a preferred embodiment of the present invention, the acquired continuous three-dimensional image sequence of user movement and the spatial position information of the user's limbs are processed into a human skeleton, and the three-dimensional spatial coordinate data of the key joints of the user's body are extracted to obtain a real-time skeleton model of the user's movement, including:

[0032] The smart mirror uses an integrated depth camera to continuously capture multiple frames of depth images during the user's movement, forming a continuous 3D image sequence. Infrared sensors capture the spatial position information of the user's limbs, including the distance and orientation data of each limb relative to the sensor. Specifically, the process involves: first, activating the smart mirror's built-in depth camera and continuously capturing multiple frames of depth images during the user's movement at a preset fixed frame rate. Each frame contains the spatial distance information of each pixel relative to the camera's optical center, i.e., the depth value. Then, all the depth images captured chronologically are sequentially linked with timestamps to form a complete sequence. A continuous three-dimensional image sequence reconstructs the temporal trajectory of the user's limb movements. Simultaneously, the infrared sensor integrated into the smart mirror is activated. This sensor periodically emits infrared pulse signals and receives reflected signals from the limbs, capturing in real time the straight-line distance data of the user's main limbs, such as the upper limbs, lower limbs, and torso, relative to the sensor's detection center, as well as the horizontal deflection angle of each limb relative to the sensor's central axis and the vertical pitch angle relative to the sensor's reference plane. Then, the above data is classified and integrated according to the correlation logic of limb part, timestamp, distance, horizontal azimuth angle, and vertical azimuth angle to form position information that can accurately locate the spatial position of each limb part of the user.

[0033] Based on continuous 3D image sequences and positional information, multiple key joints of the user's body are identified. Specifically, after data acquisition, the synchronously acquired continuous 3D image sequences and limb spatial position information are used as joint input data and connected to a human joint recognition framework based on multi-feature fusion. First, inter-frame optical flow processing is performed on the continuous 3D image sequences to establish pixel matching relationships between adjacent frames. Then, the edge curvature features of the human body contour, the length ratio features of limb segments, and the directional trend features of the motion trajectory are extracted in each frame. At the same time, with limb spatial position information as constraints, the distance and orientation data of each limb part captured by infrared sensors are used to delineate the image candidate regions of the corresponding limbs. For example, based on the distance range and horizontal / vertical azimuth angle of the upper limb, the potential distribution areas of the shoulder joint and elbow joint are delineated in the image to eliminate background object interference from the source and avoid misidentification of irrelevant points.

[0034] Then, a pre-defined human joint feature template library is invoked. This template library is constructed based on standard human anatomical features and a large number of static / dynamic standard posture samples, containing core information such as the two-dimensional morphological features of each key joint, the relative positional relationship threshold of adjacent joints, and the length ratio range of limb segments. First, coarse matching of limb regions is performed on the extracted human features. Based on the defined candidate regions, a subset of joint templates corresponding to the limbs in the template library is selected. Then, fine matching of features is performed, comparing the extracted edge curvature, relative position, length ratio, and other features with the standard features of each joint in the template subset one by one, and calculating the scores of the three core indicators. The specific calculation process is as follows: The calculation of the morphological similarity score first extracts the edge curvature sequence within a preset range around the candidate joint, denoted as A=[a1, a2, ..., an] (n is the number of curvature sampling points), and at the same time, retrieves the standard edge curvature sequence of the corresponding joint from the template library, denoted as B=[b1, b2, ..., bn]. The similarity between the two is calculated using the cosine similarity formula. Since the cosine similarity result ranges from [-1, 1], it needs to be converted to a score between 0 and 1 using a linear mapping. The mapping formula is: morphological similarity score = (cos +1) / 2, the higher the similarity, the closer the score is to 1; the calculation of the position deviation score is as follows: first, obtain the relative position threshold D (preset fixed value) between the corresponding joint point and its adjacent joint points from the template library. This threshold is preset to a fixed value of 50 pixels, which can cover the reasonable position fluctuation range of the joint point in normal movement, and effectively distinguish abnormal deviations; then extract the two-dimensional pixel coordinates of the candidate joint point and the standard position coordinates of the joint point in the template, and calculate the position deviation Q between the two using the Euclidean distance formula; then calculate the deviation ratio k=Q / D. If k≥1, that is, the actual position deviation exceeds the preset threshold of 50 pixels, then the candidate point is judged to have extremely low matching reliability, and the position deviation score = 0; if k<1, that is, the actual position deviation is within the reasonable threshold range of 50 pixels, then the position deviation score = 1-k. The smaller the deviation, the closer the k value is to 0, and the closer the score is to 1, thus quantifying the degree of fit between the candidate joint point position and the standard position; the calculation of the proportional fit score is as follows: first, clarify the target limb segment where the candidate joint point is located and the corresponding For adjacent limb segments, such as the target limb segment being the shoulder-elbow joint segment and the adjacent limb segment being the elbow-wrist joint segment; or the target limb segment being the hip-knee joint segment and the adjacent limb segment being the knee-ankle joint segment, etc., the standard proportion range corresponding to this group of limb segments is retrieved from the template library. Based on the statistical settings of a large number of static / dynamic human posture samples, the standard proportion range of the core limb segments is specifically as follows: the proportion range between the shoulder-elbow joint segment and the elbow-wrist joint segment is [1.2, 1.5]; the proportion range between the hip-knee joint segment and the knee-ankle joint segment is [1.2, 1.5]; and the proportion range between the shoulder-hip joint segment and the hip-ankle joint segment is [0.8, 1.1]. The proportion range of other auxiliary limb segments, such as the proportion range between the cervical vertebra point and the lumbar vertebra point and the lumbar vertebra point and the hip center point, is preset to [0.9, 1.2]. Then, the actual length L1 of the target limb segment and the actual length L2 of the adjacent limb segments are calculated by the image pixel distance to obtain the actual length ratio P = L1 / L2. If P falls within the standard proportion range of the corresponding limb segment... Within this range, if the ratio of shoulder-elbow joint to elbow-wrist joint is between 1.2 and 1.5, then the ratio fit score = 1; if P < 1 / 2, the score is lower. (The actual proportion is less than the standard minimum value), calculate the deduction factor. Score = 1 - k1 (minimum is 0); if P > (The actual proportion is greater than the standard maximum value), calculate the deduction factor k2=(P- ) / The score is calculated as 1 - k2 (minimum 0). This linear deduction method quantifies the degree of fit between the actual limb proportions and the standard proportions.

[0035] After the scores of the three indicators are calculated, fixed weights are preset according to the importance of each indicator to the accuracy of joint point recognition (morphological similarity has the highest weight, followed by positional deviation, and proportional fit is an auxiliary weight). For example, weight 1 = 0.5, weight 2 = 0.3, and weight 3 = 0.2 (satisfying weight 1 + weight 2 + weight 3 = 1). The comprehensive score of feature matching, i.e., the confidence score, is calculated by weighted summation formula. The specific calculation method is: confidence score = morphological similarity score × 0.5 + positional deviation score × 0.3 + proportional fit score × 0.2. Finally, through multi-feature consistency verification, candidate points with confidence scores lower than the preset threshold, such as 0.6, are eliminated. Multiple key joint points of the user's body are identified, including core joint points such as the left shoulder joint, right shoulder joint, left hip joint, and right hip joint, as well as auxiliary joint points such as the elbow joint, knee joint, ankle joint, cervical spine point, and lumbar spine point. Each identified joint point corresponds to its two-dimensional pixel coordinates (u, v) in the current frame image, and the confidence score is stored synchronously.

[0036] Based on multiple key joints, the coordinate data of each joint in three-dimensional space is calculated. A real-time skeleton model is then constructed based on the three-dimensional coordinate data of each joint and the connection relationships of the human skeleton. Specifically, this includes: firstly, filtering the identified key joints using a confidence score (removing abnormal joints with a confidence score below 0.6; if a joint is missing, it is filled using interpolation based on the motion trajectory of the same joint in the previous two frames). After filtering, four types of key data are used for three-dimensional coordinate calculation: the two-dimensional pixel coordinates (u, v) of the joint in the current frame; the depth value d corresponding to the pixel in the depth image (i.e., the straight-line distance of the joint relative to the optical center of the camera); the distance L of the limb part to which the joint belongs, captured by the infrared sensor; the horizontal azimuth angle θ; the vertical azimuth angle φ; and the depth camera intrinsic parameters (focal length) calibrated and stored before the smart mirror leaves the factory. , The principal point coordinates are u0 and v0, where u0 is the pixel coordinate of the center point of the image in the horizontal direction and v0 is the pixel coordinate of the center point of the image in the vertical direction. Next, a world coordinate system is established with the plane where the smart mirror is located as a reference. The x-axis is parallel to the mirror surface and points horizontally to the right, the y-axis is parallel to the mirror surface and points vertically upward, and the z-axis is perpendicular to the mirror surface and points towards the user. That is, the positive direction of the z-axis is the opposite direction of the user facing the mirror. The origin O of the coordinate system is the projection point of the optical center of the depth camera on the mirror surface. Based on this world coordinate system, the three-dimensional spatial coordinates (x, y, z) of the key joint points are calculated according to the following specific process. When calculating the three-dimensional coordinates (xc, yc, zc) of the joint points in the camera coordinate system, based on the pinhole camera, they are obtained by converting pixel coordinates to depth values. The formula is xc = (u - u0) × d / yc=(v-v0)×d / zc=d, where (u-u0) and (v-v0) are used to correct the offset of the pixel coordinates relative to the principal point. After dividing by the focal length and multiplying by the depth value, the horizontal, vertical and depth coordinates in the camera coordinate system can be obtained.

[0037] The coordinates are corrected by combining infrared sensor data. This takes into account the slight deviation in depth values ​​from the depth camera caused by limb occlusion. Calibration is performed using the distance L and azimuth angle (θ, φ) from the infrared sensor, as shown in the formula: =xc×(L / d)×cosφ×sinθ、 =yc×(L / d)×sinφ、 =xc×(L / d)×cosφ×cosθ, where (L / d) is used to correct the deviation between the depth value and the infrared distance, and cosφ, sinφ, sinθ, and cosθ are used to combine the horizontal / vertical azimuth angles to calibrate the actual orientation offset of the joint points in space; since there is a fixed installation offset between the camera coordinate system and the world coordinate system (the offset Δx, Δy, Δz is recorded after factory calibration), the camera coordinate system coordinates are converted to world coordinate system coordinates (x, y, z), and the final conversion formula is x= +Δx, y= +Δy、z= +Δz.

[0038] After each frame of data is acquired and the 3D spatial coordinates of each key joint point are calculated according to the above process, an inter-frame Kalman filter mechanism is immediately activated for smoothing. Specifically, based on the joint point coordinates calibrated in the previous frame, the reasonable range of coordinates for the current frame is calculated using the Kalman filter state prediction equation. The equation is as follows: ,in Predict coordinates (including x, y, and z three-dimensional components) for key points in the current frame. The state transition matrix is ​​given by the assumption that the motion of the joints follows uniform inertia. Let I be a 3×3 identity matrix. These are the joint coordinates after calibration from the previous frame. For control input matrix (when there is no additional control input) =0), To control the input vector ( =0); the preset prediction error threshold is 5cm, that is, the actual coordinates of the current frame are calculated using the Euclidean distance formula. If the distance is greater than 5cm, it is considered an anomaly caused by occlusion or sensor noise. If the actual coordinates are within the threshold range, the predicted and actual values ​​are fused using a filtering update equation to output the calibration coordinates. If the distance exceeds the threshold, the actual coordinates are directly used. As calibration coordinates, real-time coordinate calibration is completed. After coordinate calibration, based on skeletal connections, such as the left shoulder joint and left elbow joint connected through the humerus, and the right hip joint and right knee joint connected through the femur, a connection link is established for each key joint point. Each link is constrained according to the actual length of the bone, with a preset ratio fluctuation threshold of ±10%. The ratio corresponding to the spatial distance between the connected joint points is calculated in real time. If it exceeds the threshold, the coordinates are finely adjusted to ensure that it conforms to the human physiological structure. Based on the above connections and ratio constraints, a real-time skeleton model is constructed. The model uses the joint point calibration coordinates as the core nodes and the bone connections as the associated edges. For each new frame of acquired data received, the three-dimensional coordinates of all relevant nodes and the spatial pose of the bone connections are updated synchronously. The update time is matched with the depth camera frame rate (e.g., ≤33ms at 30 frames / second) to ensure real-time synchronization with the user's movements.

[0039] This embodiment uses a depth camera and an infrared sensor to collaboratively acquire three-dimensional image sequences and limb spatial position information, identify key joints, and construct a real-time skeleton model. It can comprehensively capture the spatial displacement and rotation of the user's limbs in the direction perpendicular to the mirror surface, and fully acquire the three-dimensional spatial information of key joints, avoiding the problem of not being able to quantify three-dimensional motion features in two-dimensional image evaluation.

[0040] In a preferred embodiment of the present invention, based on a real-time skeleton model, the user's torso center point is used as a spatial reference point to establish a first reference vector pointing from the spatial reference point to the center points of the left and right shoulder joints, and a second reference vector pointing to the center points of the left and right hip joints, to determine the main reference plane for user movement, including:

[0041] Based on the 3D spatial coordinate data of the left and right hip joints in the real-time skeleton model, the spatial coordinates of the user's torso center point are calculated. Specifically, this includes: after constructing the real-time skeleton model, establishing the main body reference plane based on this model. First, the 3D spatial coordinates (x1, y1, z1) of the left hip joint and the 3D spatial coordinates (x2, y2, z2) of the right hip joint are extracted from the real-time updated skeleton model of the current frame. Before extraction, inter-frame data verification is performed (comparing the trend of the joint coordinates in the previous frame to exclude sudden outliers) to ensure accuracy. Coordinate accuracy: Considering the symmetrical structure of the human torso, the left and right hip joints are the core connection points between the torso and the lower limbs. Their center points can accurately represent the reference position of the torso. Therefore, the spatial coordinates of the user's torso center point are calculated by averaging the coordinates. Specifically, the x-coordinate of the torso center point is calculated as (x1 + x2) / 2, the y-coordinate of the torso center point is calculated as (y1 + y2) / 2, and the z-coordinate of the torso center point is calculated as (z1 + z2) / 2. After the calculation is completed, the coordinates of the torso center point are stored (denoted as x0, y0, z0).

[0042] The midpoint of the spatial coordinates of the center points of the left and right shoulder joints is calculated as the shoulder center point; the midpoint of the spatial coordinates of the center points of the left and right hip joints is calculated as the hip center point. Specifically, after obtaining the coordinates of the torso center point (x0, y0, z0), the three-dimensional spatial coordinates of the left shoulder joint (x3, y3, z3) and the three-dimensional spatial coordinates of the right shoulder joint (x4, y4, z4) are extracted from the real-time skeleton model of the current frame. After inter-frame data verification (excluding outliers), the midpoint of these two coordinates is calculated as the shoulder center point (used to represent the symmetry reference of the upper torso). The specific calculation method is: shoulder center point x-coordinate = (x3 + x...) / ( ... 4) / 2, shoulder center point y coordinate = (y3+y4) / 2, shoulder center point z coordinate = (z3+z4) / 2; at the same time, using the extracted and verified three-dimensional spatial coordinates of the left hip joint (x1, y1, z1) and right hip joint (x2, y2, z2), the midpoint of the coordinates of the two is calculated as the hip center point (used to characterize the symmetry reference of the lower part of the torso). The specific calculation method is as follows: hip center point x coordinate = (x1+x2) / 2, hip center point y coordinate = (y1+y2) / 2, hip center point z coordinate = (z1+z2) / 2. The entire calculation process is strictly based on the same frame skeleton data to avoid the impact of cross-frame errors on the accuracy of subsequent reference vectors.

[0043] Based on the spatial coordinates of the torso center point and the shoulder center point, a first reference vector is calculated; based on the spatial coordinates of the torso center point and the hip center point, a second reference vector is calculated. Specifically, after obtaining the coordinates of the shoulder center point, hip center point, and torso center point, the reference vectors required to construct the main body reference plane are calculated. Using the obtained torso center point (x0, y0, z0) as the vector starting point and the obtained shoulder center point as the vector ending point, the first reference vector is calculated through the coordinate difference. This vector represents the spatial direction from the torso center point to the shoulder. Specifically, the first reference vector (a... (b1, b2, b3) = (x-coordinate of shoulder center point - x0, y-coordinate of shoulder center point - y0, z-coordinate of shoulder center point - z0). Similarly, with the torso center point (x0, y0, z0) as the vector starting point and the hip center point as the vector ending point, the second reference vector is calculated. This vector represents the spatial direction from the torso center point to the hip. Specifically, the second reference vector (b1, b2, b3) = (x-coordinate of hip center point - x0, y-coordinate of hip center point - y0, z-coordinate of hip center point - z0). The two reference vectors together constitute the core spatial reference of the torso's main posture.

[0044] Based on the first and second reference vectors, the normal vector of the main reference plane is determined through vector cross product operation, thus establishing the main reference plane for user movement. Specifically, after obtaining the first reference vector (a1, a2, a3) and the second reference vector (b1, b2, b3), a vector cross product operation is performed on them. Since the two reference vectors are based on the torso structure and are not collinear, the cross product result will be the normal vector perpendicular to the plane containing these two vectors. This normal vector is the core parameter of the main reference plane. The cross product operation formula is: c1 = a2b3 - a3b2, c2 = a3b1 - a1b 3. c3 = a1b2 - a2b1; After calculating the normal vector (c1, c2, c3), normalize it (by dividing each component of the vector by the magnitude of the vector, the length of the vector is converted to 1, eliminating the influence of the vector magnitude on the plane construction); After normalization, take the determined torso center point (x0, y0, z, 0) as the reference point, and construct the main body reference plane in combination with the normalized normal vector. The equation of this plane satisfies c1(x - x0) + c2(y - y0) + c3(z - z0) = 0. The final main body reference plane can represent the torso posture reference when the user moves.

[0045] This embodiment, based on the core joint coordinate data in the real-time skeleton model, calculates and determines the torso center point, reference vector, and main body reference plane, establishing a stable and unified reference benchmark for motion posture assessment, thus solving the assessment deviation problem caused by unclear benchmarks in the assessment; the main body reference plane can accurately reflect the user's main posture during movement.

[0046] like Figure 2 As shown, in another preferred embodiment of the present invention, based on the main reference plane and a real-time skeleton model, four feature points—the left and right shoulder joints, the left and right hip joints—are selected to construct a dynamic feature plane characterizing the limb's motion state, including:

[0047] Based on a real-time skeleton model, the real-time 3D spatial coordinates of four feature points are obtained, including the real-time 3D spatial coordinates of the center points of the left shoulder joint, right shoulder joint, left hip joint, and right hip joint. Specifically, based on the constructed and real-time updated real-time skeleton model of the current frame, the real-time 3D spatial coordinates of four core feature points are extracted, namely the coordinates of the center points of the left shoulder joint (xA, yA, zA), the right shoulder joint (xB, yB, zB), the left hip joint (xC, yC, zC), and the right hip joint (xD, yD, zD). Before extraction, inter-frame data verification is performed, calculating the Euclidean distance between the candidate feature point coordinates of the current frame and the corresponding feature point coordinates of the previous frame. If the distance exceeds 5cm (a preset abnormal threshold covering the normal range of joint movement fluctuations), it is judged as an abnormal value, and the linear interpolation result of the corresponding feature point coordinates of the previous two frames is used to replace it; if the distance is within the threshold, the candidate coordinates are confirmed to be valid, ensuring that the extracted coordinates of the four feature points are accurate and continuous.

[0048] A first reference triangle plane is established using the center points of the left shoulder joint, right shoulder joint, and left hip joint. A second reference triangle plane is then established using the same center points. Specifically, based on the verified three-dimensional coordinates of four core feature points, two reference triangle planes are constructed in two steps (both based on the core torso region to ensure the relevance of the plane representation). The first reference triangle plane is constructed by selecting the center points of the left shoulder joint (A), right shoulder joint (B), and left hip joint (C). Point (C) is a set of three non-collinear feature points. First, calculate two independent non-collinear vectors in the plane. Vector AB is the spatial vector pointing from A to B, calculated as AB = (xB - xA, yB - yA, zB - zA); vector AC is the spatial vector pointing from A to C, calculated as AC = (xC - xA, yC - yA, zC - zA). Based on the principle of vector cross product (the cross product result is perpendicular to the plane containing the two input vectors, i.e., the plane normal vector), calculate the normal vector n1 of the plane, specifically, n1 = AB × AC = ( × - × , × - × , × - × ),in , , Let x, y, and z be the x, y, and z components of vector AB, respectively. , , These are the x, y, and z components of vector AC, respectively. The second reference triangle plane construction first requires selecting three non-collinear feature points: the center point of the left shoulder joint (A), the center point of the right shoulder joint (B), and the center point of the right hip joint (D). Following the same vector calculation logic, first calculate two independent non-collinear vectors in the plane. Vector AB remains (xB-xA, yB-yA, zB-zA), and vector AD is the spatial vector pointing from A to D, calculated as AD = (xD-xA, yD-yA, zD-zA). Then, calculate the normal vector n2 of the plane using the cross product of the vectors, specifically n2 = AB × AD = (...). × - × , × - × , × - × ),in , , These are the x, y, and z components of vector AD, respectively.

[0049] Based on the normal vectors of the first and second reference triangle planes, the composite normal vector of the dynamic feature plane is calculated. Specifically, since both reference triangle planes are constructed around the shoulder-hip core region, their contribution weights to the representation of the dynamic feature plane are completely consistent. Therefore, an equal-weighted fusion method is used to calculate the composite normal vector. The specific process is as follows: To eliminate the influence of the amplitude difference between the two normal vectors on the fusion result, normal vectors n1 and n2 need to be normalized separately. When calculating the magnitude of n1, its x, y, and z components are squared respectively, and then the squared results are added together. Finally, the square root of the sum is taken. Then, the x, y, and z components of n1 are divided by this magnitude to obtain the normalized normal vector. After normalization, calculate the magnitude of n1 and n2 using the same logic. Then, divide each component of n2 by its magnitude to obtain the normalized n2, ensuring that the lengths of the two vectors are uniformly 1. According to the equal weight rule of 0.5 for each vector, multiply the corresponding components of the normalized n1 and n2 respectively, and then add the two products of each dimension to obtain the fusion result of the three dimensions, which is combined to form a preliminary composite normal vector. To ensure that the composite normal vector conforms to the standard form of a planar normal vector (length is 1), the preliminary composite normal vector needs to be normalized again. Calculate the magnitude of the vector (the calculation method is the same as above), and then divide each component by its magnitude to obtain the normalized composite normal vector.

[0050] Based on the comprehensive normal vector and the spatial distribution characteristics of the four feature points, a complete dynamic feature plane representing the limb movement state is constructed. Specifically, this includes: to accurately anchor the spatial position of the dynamic feature plane, firstly, calculating the average spatial center point coordinates of the four core feature points. During calculation, the x-coordinates of the four feature points are added together and divided by 4 to obtain the x-coordinate of the average center point; following the same logic, the y-coordinates and z-coordinates of the average center point are calculated respectively. This average center point accurately reflects the overall spatial distribution center of the four core feature points and serves as a fixed reference point for the dynamic feature plane; combined with the obtained normalized comprehensive normal vector... The x, y, and z components are denoted as n'x, n'y, and n'z, respectively. A dynamic characteristic plane is constructed based on the point-normal form equation of the spatial plane. Specifically, the point-normal form equation is: n'x × (x - ... )+n'y×(y- )+n'z×(z- =0, where (x, y, z) are the three-dimensional coordinates of any point on the dynamic feature plane. , , The coordinates of the spatial average center point calculated above are shown in the figure. The core logic of this equation is to normalize the synthetic normal vector. The plane is perpendicular to any vector within the dynamic feature plane. This constraint clarifies the spatial orientation of the plane, and the spatial position of the plane is anchored by the average center point. The resulting dynamic feature plane can completely and accurately represent the real-time spatial posture of the shoulder and hip core region during limb movement.

[0051] This embodiment extracts the three-dimensional coordinates of core feature points and constructs a dynamic feature plane based on the fusion of two triangular planes, breaking through the limitations of two-dimensional plane evaluation and enabling comprehensive capture of the movement posture features of limbs in three-dimensional space. The construction of the dynamic feature plane fully integrates the movement state of the core area of ​​the shoulder and hip, and the calculation of the comprehensive normal vector ensures the accuracy of the plane representation, effectively covering movement scenarios involving the swinging of core parts such as bending over and turning, making the representation of limb movement state more comprehensive and accurate.

[0052] In a preferred embodiment of the present invention, the spatial attitude offset is obtained by calculating the angle between the normal vectors of the dynamic feature plane and the main reference plane, and a minimum bounding rectangle is constructed based on the spatial positions of the four feature points to obtain the motion stability index, including:

[0053] The spatial angle between the composite normal vector of the dynamic feature plane and the normal vector of the main reference plane is calculated to obtain the spatial posture offset. Specifically, to quantify the degree of offset of limb movement posture, the normalized normal vector of the main reference plane (ensuring a length of 1) is retrieved first, and then combined with the normalized composite normal vector of the dynamic feature plane to calculate the spatial angle between the two. This angle is the spatial posture offset. The specific process is as follows: Based on the geometric definition of vector dot product, the dot product of two normalized vectors is equal to the cosine of their angle. Since both vectors are normalized, the calculation only requires multiplying the corresponding x, y, and z components of the two vectors respectively, and then adding the three product results to obtain the cosine of the angle. The cosine value obtained above is converted into the corresponding spatial angle through the inverse cosine function. The value of this angle is limited to between 0° and 90°. If the calculated result is an obtuse angle, the supplementary angle is taken to ensure that the offset is represented in a positive form. The final angle directly reflects the degree of deviation of the core area dynamic feature plane from the main reference plane during limb movement.

[0054] Based on the three-dimensional spatial coordinates of the center points of the left and right shoulder joints, the center points of the left and right hip joints, the minimum bounding rectangle containing the spatial positions of all feature points is determined. Specifically, for subsequent calculation of motion symmetry and stability indices, the three-dimensional feature point coordinates (xA, yA, zA), (xB, yB, zB), (xC, yC, zC), and (xD, yD, zD) are converted into two-dimensional coordinates. This is achieved through an orthogonal projection algorithm. The process is as follows: the obtained three-dimensional coordinates of the four core feature points are projected onto the constructed main reference plane. The core principle of the projection is to retain the coordinate components of the feature points within the main reference plane and discard the components perpendicular to the plane. In practice, the distance from each feature point to the main reference plane is first calculated. The vertical distance is calculated, and then the components in that vertical direction are eliminated by vector subtraction to obtain the two-dimensional projected coordinates of each feature point. The u-axis and v-axis in the projection plane are kept parallel to the x-axis and y-axis of the world coordinate system. The two-dimensional projected coordinates of the four feature points are traversed, and the minimum and maximum values ​​of all coordinates in the u-axis direction and the v-axis direction are extracted to define the boundary range of the projection area. The intersection of the minimum and minimum values ​​of the u-axis and v-axis, the intersection of the maximum and minimum values ​​of the u-axis and v-axis, the intersection of the maximum and maximum values ​​of the u-axis and v-axis, and the intersection of the minimum and maximum values ​​of the u-axis and v-axis are used as the four vertices to construct the minimum bounding rectangle. The boundary of this rectangle is parallel to the u-axis and v-axis in the projection plane and can completely contain the projected positions of all feature points.

[0055] The aspect ratio of the minimum bounding rectangle is calculated to obtain the limb movement symmetry index; the tilt angle of the minimum bounding rectangle relative to the main reference plane is calculated to obtain the limb movement stability index. Specifically, based on the constructed minimum bounding rectangle, limb movement indices are calculated from two dimensions: symmetry and stability. The specific process is as follows: The limb movement symmetry index is calculated by first calculating the side lengths of the minimum bounding rectangle in the u-axis and v-axis directions, i.e., the difference between the maximum and minimum values ​​of u-axis and v-axis. The longer of these two side lengths is defined as the length, and the shorter is defined as the width. Then, the width is divided by the length to obtain the limb movement symmetry index. The closer the value of this index is to 1, the more symmetrical the spatial distribution of the left and right limbs (shoulder and hip areas); the farther the value deviates from 1, the more symmetrical the spatial distribution of the left and right limbs (shoulder and hip areas). The lower the degree of symmetry, the more stable the limb movement. The calculation of the limb movement stability index uses the u-axis of the main body reference plane as the reference benchmark. The tilt angle between the long side of the rectangle and the u-axis is calculated. If the long side is parallel to the u-axis, that is, the length is the side length in the u-axis direction, the tilt angle is 0°. If the long side is parallel to the v-axis, that is, the length is the side length in the v-axis direction, the tilt angle is 90°. If the long side is oblique (when the side lengths in the two directions are equal, the vector corresponding to the side length in the u-axis direction is used), the tilt angle is calculated by using the coordinate difference between the two endpoints of the long side in the u-axis and v-axis directions and the arctangent function. The final result is converted into an angle value from 0° to 180°. This tilt angle is the limb movement stability index. The smaller the fluctuation of this angle during the movement, the more stable the core area movement is; the larger the fluctuation, the worse the stability.

[0056] In this embodiment, the degree of attitude deviation is quantified into spatial attitude deviation by calculating the spatial angle between the dynamic feature plane and the main reference plane, thus avoiding subjective judgment bias in the assessment. At the same time, based on the feature point projection, the minimum bounding rectangle is constructed to further extract the two key indicators of symmetry and stability, reflecting the quality of motion from different dimensions, making the dimensions of health assessment richer. The quantitative calculation of various indicators ensures the objectivity and consistency of the assessment results.

[0057] In a preferred embodiment of the present invention, based on spatial posture offset and motion stability indices, and combined with three-dimensional spatial coordinate data of key joints of the user's body, a pre-trained evaluation model is used to perform a health assessment of the user's motion quality, generating a comprehensive evaluation result including motion posture correction parameters, comprising:

[0058] Spatial posture offset, limb movement symmetry index, limb movement stability index, and 3D spatial coordinate data of key joints are input into a pre-trained evaluation model for analysis and processing. Preliminary evaluation parameters, including shoulder joint angle deviation, hip joint angle deviation, and trunk stability coefficient, are obtained. Specifically, based on the 3D spatial coordinate data of key joints, the real-time movement angles of the left and right shoulder joints and the left and right hip joints are calculated. This includes calculating the real-time movement angles of the left and right shoulder joints, and the left and right hip joints, respectively, based on the 3D spatial coordinate data of key joints after real-time coordinate calibration. The specific process is as follows: When calculating the shoulder joint angle, taking the left shoulder joint as an example, three key reference points are first identified: the center point of the left shoulder joint, the center point of the left elbow joint, and the center point of the trunk. Two spatial vectors are constructed based on these three points. The first vector points from the center point of the left shoulder joint to the left elbow joint. The first vector points from the center of the left shoulder joint to the center of the torso, reflecting the spatial extension direction of the humerus. The second vector points from the center of the left shoulder joint to the center of the torso, reflecting the connection direction between the torso and the shoulder. Next, the angle is calculated. First, the dot product of the two vectors is calculated by multiplying the x, y, and z components of each vector, then summing the results. Then, the magnitudes of the two vectors are calculated separately. The magnitude of each vector is calculated by adding the squares of its x, y, and z components, taking the square root of the sum, and finally dividing the dot product by the product of the magnitudes of the two vectors using the inverse cosine function. The result is the real-time motion angle of the left shoulder joint. Following the same calculation logic, using the center of the right shoulder joint, the center of the right elbow joint, and the center of the torso as reference points, corresponding spatial vectors are constructed, and the dot product, magnitude, and inverse cosine calculations are performed to obtain the real-time motion angle of the right shoulder joint.

[0059] When calculating the hip joint angle, taking the left hip joint as an example, the center points of the left hip joint, left knee joint, and trunk are selected as three reference points to construct two spatial vectors. The first vector points from the center point of the left hip joint to the center point of the left knee joint, reflecting the spatial extension direction of the femur. The second vector points from the center point of the left hip joint to the center point of the trunk, reflecting the connection direction between the trunk and the hip. Following the method for calculating the shoulder joint angle, the dot product of the two vectors is first calculated (the corresponding components are multiplied and then summed). Then, the magnitudes of the two vectors are calculated separately (the square root of the sum of the squares of each component). Finally, the angle between the two vectors is calculated using the inverse cosine function. This angle is the real-time motion angle of the left hip joint. Following the same logic, the real-time motion angle of the right hip joint is calculated using the center points of the right hip joint, right knee joint, and trunk as reference points.

[0060] The system compares real-time motion angles with pre-stored standard motion angles to generate initial angle deviation data for the shoulder and hip joints. Specifically, the system pre-stores standard motion angles for basic human limb movements (covering common scenarios such as daily fitness and limb extension). These standard angles are determined based on human anatomy and exercise physiology, with specific values ​​of 90 degrees for the shoulder joint (corresponding to the core movement angle of horizontal limb elevation) and 60 degrees for the hip joint (corresponding to the key movement angle of natural limb flexion and extension). The same standard values ​​are used for both left and right joints. The calculated real-time motion angle of the left shoulder joint is then compared with the pre-stored standard motion angle of the left shoulder. By comparing the real-time motion angle with the standard motion angle and taking the absolute value of the difference, the initial angle deviation data of the left shoulder joint is obtained. In the same way, the real-time motion angle of the right shoulder joint is compared with the standard motion angle of the right shoulder to obtain the initial angle deviation data of the right shoulder joint. Similarly, the real-time motion angle of the left hip joint is compared with the pre-stored standard hip motion angle, and the initial angle deviation data of the left hip joint is obtained by taking the absolute value of the difference. The real-time motion angle of the right hip joint is compared with the standard hip motion angle to obtain the initial angle deviation data of the right hip joint. Finally, the initial angle deviation data of the left shoulder joint, right shoulder joint, left hip joint, and right hip joint are formed.

[0061] Initial angle deviation data, spatial posture offset, limb movement symmetry index, and limb movement stability index are fused at the feature level to construct a multi-dimensional feature vector. This multi-dimensional feature vector is then input into a pre-trained evaluation model. The feature extraction layer in the evaluation model performs feature mining on the multi-dimensional feature vector to obtain high-dimensional abstract features. Specifically, this includes: first, identifying the seven core indices to be fused: left shoulder joint deviation, right shoulder joint deviation, left hip joint deviation, right hip joint deviation, spatial posture offset, limb movement symmetry index, and limb movement stability index. Because these indices are measured in different ways, direct concatenation would lead to the model overemphasizing indices with large numerical ranges. Therefore, normalization must be performed first. This is done by statistically analyzing sample data from different populations and different movement postures to determine the indices for each category. Reasonable value ranges are defined, i.e., the minimum and maximum values ​​of the indicators. For example, the minimum value of spatial posture offset is 0 degrees and the maximum value is 45 degrees; the minimum value of limb movement symmetry indicator is 0.3 and the maximum value is 1.0; the minimum value of joint deviation is 0 degrees and the maximum value is 30 degrees. Then, the current value of each indicator is calculated using the formula (current indicator value - minimum value of the indicator) / (maximum value of the indicator - minimum value of the indicator). All indicators are uniformly mapped to the interval [0, 1] to completely eliminate the interference caused by differences in dimensions and numerical ranges. After normalization, the standardized indicators of the seven dimensions are sequentially connected in a fixed order: left shoulder joint deviation, right shoulder joint deviation, left hip joint deviation, right hip joint deviation, spatial posture offset, limb movement symmetry indicator, and limb movement stability indicator, forming a 7-dimensional multidimensional feature vector with a unified structure.

[0062] The pre-trained evaluation model was constructed and trained as follows: To uncover the complex relationships between multidimensional features, such as the potential negative correlation between left shoulder joint deviation and trunk stability, a deep learning architecture consisting of a feature extraction layer, a fully connected layer, and an output layer was built. The design of each layer and the training process were closely integrated. The feature extraction layer adopted a 3-layer one-dimensional convolutional neural network (CNN). Since the input is a sequential 7-dimensional feature vector, a one-dimensional CNN can efficiently capture the local correlation between adjacent indicators. The first convolutional layer was set with 32 convolutional kernels and a kernel size of 3. This kernel size can just cover the three adjacent joint-related indicators such as left shoulder joint deviation, right shoulder joint deviation, and left hip joint deviation, focusing on capturing the cooperative change patterns between joints, such as whether the left and right shoulder deviations are related. The first layer increases and decreases synchronously, outputting a preliminary 32-dimensional feature map. The second convolutional layer has 64 kernels with a kernel size of 2. Based on the joint association features of the first layer, it further focuses on key combinations of joint and trunk indicators such as right hip joint deviation and spatial posture shift, exploring their interaction relationship, such as whether hip deviation will cause trunk posture shift, and outputting a more targeted 64-dimensional interaction feature map. The third convolutional layer has 128 kernels with a kernel size of 1. Instead of capturing cross-indicator associations, it focuses on the core information of a single indicator, such as the specific degree of deviation of spatial posture shift, while integrating the association features of the first two layers, outputting a 128-dimensional fused feature map. To avoid gradient vanishing and feature loss during training... To address the issue of excessive feature map fluctuations, each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The batch normalization layer standardizes the feature map values ​​to a uniform range, stabilizing the training process. The ReLU activation function maps the feature values ​​to a non-negative range, introducing non-linear expressive power, allowing the model to learn complex patterns such as the non-linear decrease in trunk stability as joint deviation increases. The fully connected layer receives the 128-dimensional feature map from the feature extraction layer, first converting it into a 128-dimensional one-dimensional vector through a flattening operation, and then inputting it into two fully connected layers. The first fully connected layer has 256 neurons to expand the model's parameter space, enabling the model to learn the complex mapping relationship between high-dimensional features and evaluation parameters. To avoid underfitting due to insufficient parameters, the second fully connected layer has 128 neurons. Building upon the expanded parameter space of the first layer, it compresses the feature dimension, focusing on core information such as features directly related to joint regularity and trunk stability, while filtering out secondary interference. Both layers are connected to a Dropout layer with a dropout probability of 0.3, randomly disabling 30% of the neuron connections to prevent the model from over-relying on some features and causing overfitting. The output layer has 3 neurons, corresponding to these 3 parameters, and uses a linear activation function. Linear activation ensures that the range of the output value is consistent with the true label (0 to 1), avoiding deviation of the predicted value from the actual range due to the non-linear compression of the activation function.

[0063] The model training process involves collecting exercise data from individuals aged 18 to 60 at different skill levels (beginner / intermediate / professional). This data includes standard exercise postures (standard dumbbell lateral raises, squats) and non-standard postures (shrugging, hip deviation, trunk swaying). Each data point contains a complete 7-dimensional feature vector, including shoulder joint angle deviation (0 to 1, higher values ​​indicate less proper shoulder movement), hip joint angle deviation (0 to 1, similarly), and trunk stability coefficient (0 to 1, higher values ​​indicate greater trunk stability). The data is then divided into a training set (40,000 data points, used for model parameter updates) and a test set (10,000 data points, used to validate generalization ability) in an 8:2 ratio. This division ensures... Samples of different ages, exercise levels, and posture types are evenly distributed across the two datasets, with standard posture samples accounting for 30% each. Mean squared error (MSE) is used as the loss function to measure the deviation between continuous numerical predicted values ​​and true values. The Adam optimizer is employed with an initial learning rate of 0.001 (to ensure initial training efficiency) and a decay rate of 1e-5 (to reduce the learning rate with each iteration to avoid later oscillations). The maximum number of iterations is 100. After each training iteration, the MSE value is calculated using the test set. If the MSE value on the test set does not decrease for 10 consecutive iterations (indicating model convergence), training is stopped and the model parameters are saved, resulting in the pre-trained evaluation model (the model corresponding to the smallest MSE value on the test set).

[0064] The constructed 7-dimensional multidimensional feature vector is input into the pre-trained evaluation model. The feature extraction layer processes the features step by step according to the logic of local correlation capture, interaction feature mining, and core feature integration. The first convolutional layer (32 kernels, kernel size 3) captures the local correlation features of three adjacent indicators, such as the synergistic change patterns of left shoulder joint deviation, right shoulder joint deviation, and left hip joint deviation. Then, batch normalization and ReLU activation are performed to standardize feature values, activate nonlinear features, and filter noise. Subsequently, the second convolutional layer (64 kernels, kernel size 2) is based on the features of the first layer to mine key features such as right hip joint deviation and spatial pose offset. The interaction relationships between indicators are analyzed; feature quality is further optimized through batch normalization and ReLU activation to enhance effective information expression; finally, the third convolutional layer (128 kernels, kernel size 1) extracts the core information of each indicator, such as the specific value of spatial posture offset and the core proportion of limb movement symmetry indicators. At the same time, the associated features of the first two layers are integrated to form a 128-dimensional high-dimensional abstract feature vector. This feature not only retains the core information of the seven original indicators such as left shoulder joint deviation and spatial posture offset, but also integrates the potential relationships between indicators, such as the negative correlation between joint deviation and trunk stability and the synergistic relationship between left and right joints.

[0065] By performing nonlinear transformation and weight allocation on high-dimensional abstract features through the fully connected layer in the evaluation model, preliminary evaluation parameters are obtained. These preliminary evaluation parameters include shoulder joint angle deviation values ​​for quantifying shoulder movement standardization, hip joint angle deviation values ​​for quantifying hip movement standardization, and trunk stability coefficients for characterizing trunk stability during movement. Specifically, relying on the fully connected layer in the pre-trained evaluation model, a progressive processing of the 128-dimensional high-dimensional abstract feature vector is performed, including feature selection, weight allocation, nonlinear mapping, and parameter calculation, ultimately generating three standardized preliminary evaluation parameters (all mapped to the 0-1 interval). The specific implementation process is as follows: Since there may still be a small amount of redundant information in the 128-dimensional high-dimensional abstract features, and there may be linear correlations between features, which cannot directly support accurate evaluation, preliminary simplification and targeted selection are first performed through the first fully connected layer. 256 neurons are set. The core objective is to achieve dimensionality reduction and preliminary selection through dimensionality compression, reducing redundant information. While consuming computational resources, this approach focuses on key features strongly correlated with motion posture assessment. The ReLU activation function is used to perform a nonlinear transformation on the input features, breaking the linear correlation between features and enhancing the model's ability to express complex motion patterns. Based on fixed weight parameters obtained through backpropagation optimization during model training, the 128-dimensional input features are differentiated in weighting. Core features strongly correlated with joint normality, such as the collaborative features of left and right shoulder joint deviations and left and right hip joint deviations, are assigned higher weights (approximately 0.6) to ensure that core information is captured. Secondary features weakly correlated with motion quality, such as the slight fluctuations in stability indicators, are assigned lower weights (approximately 0.1) to reduce interference from irrelevant information. After ReLU activation and preliminary weighting, a 256-dimensional feature vector is output. This vector represents the first simplification and optimization of high-dimensional features, eliminating most redundant noise and providing a high-quality, highly relevant input foundation for the refined processing of the second fully connected layer.

[0066] Following the 256-dimensional feature vector output from the first fully connected layer (which retains core correlation information), this layer further splits the features and refines the weights to provide targeted support for the accurate calculation of the three core evaluation parameters. With 128 neurons, its core function is to accurately focus on target features and refine weight allocation. For the three core evaluation dimensions of shoulder joint, hip joint, and trunk stability, the 256-dimensional features are split in a targeted manner to avoid interference between features from different evaluation dimensions. Based on the core requirements of motion posture evaluation, the 256-dimensional input features are subsetted and weighted. The first feature subset (corresponding to shoulder joint evaluation) includes collaborative features of left and right shoulder joint deviations and correlation features between shoulder joint deviations and symmetry indices, assigned the highest weight (approximately 0.7). This weighting strengthens the expression of shoulder-related features and provides targeted input specifically for calculating shoulder joint angle deviation values. The second feature subset (corresponding to hip joint evaluation) includes collaborative features of left and right hip joint deviations and correlation features between hip joint deviations and spatial posture offsets, assigned the second highest weight (approximately 0.65). The first feature set emphasizes the influence of hip-related features, specifically serving the calculation of hip joint angle deviation values. The third feature subset (corresponding to trunk stability assessment) includes fused features of spatial posture offset, symmetry index, and stability index, assigned corresponding weights (approximately 0.6), focusing on stability-related features of trunk movement and providing dedicated input for the calculation of trunk stability coefficients. A Dropout layer (dropout rate of 0.2, optimized during training) is connected to the output of the second fully connected layer to avoid model overfitting by randomly deactivating some neurons, ensuring that parameter calculations are not interfered with by noisy features. After refined weighting and Dropout filtering, a 128-dimensional feature vector is output. The dimension division is determined based on the proportion of feature importance, i.e., the first 32 dimensions correspond to the first feature subset (shoulder joint assessment), the middle 48 dimensions correspond to the second feature subset (hip joint assessment), and the last 48 dimensions correspond to the third feature subset (trunk stability assessment). This vector provides directional, high-purity, and low-noise input data for the calculation of the three core preliminary assessment parameters.

[0067] Continuing from the three directional feature subsets output by the second fully connected layer, a two-step method of weighted summation and normalization is used to calculate the core parameters, ensuring that the results are uniformly mapped to the 0-1 interval. The calculation process for the shoulder joint angle deviation value (P1) is as follows: the first feature subset (32 dimensions) is weighted and summed according to the corresponding refinement weights (approximately 0.7), fusing the scattered shoulder-related features into a single comprehensive feature value T1; the Sigmoid function is then used to map T1 to the 0-1 interval for normalization, with the formula P1 = ... The closer P1 is to 1, the greater the deviation of the shoulder movement from the standard posture; the closer it is to 0, the closer the shoulder movement is to the standard. The calculation process for the hip joint angle deviation value (P2) is as follows: the second feature subset (48 dimensions) is weighted and summed according to the corresponding refinement weight (approximately 0.65) to obtain the hip feature fusion value T2; the Sigmoid function is used to map it to the interval between 0 and 1 for normalization processing, and the formula is P2 = The closer P2 is to 1, the greater the deviation of the hip movement from the standard posture; the closer it is to 0, the closer the hip movement is to the standard posture. The trunk stability coefficient (P3) is calculated by weighting and summing the third feature subset (48 dimensions) according to the corresponding refinement weights (approximately 0.6) to obtain the trunk feature fusion value T3. This T3 is then normalized by mapping it to the 0-1 interval using the Sigmoid function. The formula is P3 = The closer P3 is to 1, the more stable the trunk movement is, with no obvious forward tilting, lateral deviation or swaying; the closer it is to 0, the more unstable the trunk movement is, and there is a risk of abnormal posture.

[0068] After a complete process of feature vector construction, model feature extraction, fully connected layer processing, and parameter calculation, three standardized preliminary evaluation parameters are finally generated: shoulder joint angle deviation value P1 (0 to 1 range), hip joint angle deviation value P2 (0 to 1 range), and trunk stability coefficient P3 (0 to 1 range). These three parameters fully cover the two core evaluation dimensions of limb joint movement standardization and trunk stability, and all have clear physical meaning and quantitative attributes, which can reflect the key characteristics of the user's movement posture.

[0069] Based on preliminary evaluation parameters, motion posture correction parameters are generated for the user's shoulder joint motion angle, hip joint motion angle, and trunk stability. Specifically, this includes: combining pre-stored standard motion parameters (with a standard shoulder joint motion angle of 90 degrees and a standard hip joint motion angle of 60 degrees), and using a body-specific calculation logic to generate correction parameters for the shoulder joint, hip joint, and trunk, including adjustment direction, amplitude, and specific suggestions. The specific process is as follows: the shoulder joint correction parameter calculation first obtains the user's current real-time motion angles of the left and right shoulder joints, using pre-stored standard shoulder joint motion angles... Using the standard shoulder joint angle (90 degrees) as a baseline, calculate the difference between the standard shoulder joint angle and the current real-time shoulder joint angle. Based on this difference, determine the adjustment direction and the basic adjustment range. If the difference is positive, for example, 90 degrees - 85 degrees = 5 degrees, it indicates that the current real-time shoulder joint angle is less than the standard value, and the adjustment direction is to increase the shoulder joint angle. The absolute value of the difference (5 degrees) is used as the basic adjustment range. If the difference is negative, for example, 90 degrees - 93 degrees = -3 degrees, it indicates that the current real-time shoulder joint angle is greater than the standard value, and the adjustment direction is to decrease the shoulder joint angle. The absolute value of the difference (3 degrees) is used as the basic adjustment range. After adjusting the range of motion, it is necessary to calibrate based on the previously generated shoulder joint angle deviation value. This deviation value directly reflects the degree to which the overall shoulder movement deviates from the standard. The larger the deviation value, the more significant the difference between the movement and the standard. If the base range of motion is used directly for adjustment, it is easy to cause movement deformation or imbalance. Therefore, a unified calibration formula is used: Final adjustment range = Base adjustment range × (1 - Shoulder joint angle deviation value). For example, if the shoulder joint angle deviation value is 0.1 (the shoulder movement is close to the standard), and the base adjustment range is 5 degrees, the final adjustment range = 5 degrees × (1 - 0.1) = 4.5 degrees. This range of motion is close to the standard. The adjustment is designed to meet the requirements without disrupting the continuity of movement. If the shoulder joint angle deviation is 0.5 (significant deviation in shoulder movement), the basic adjustment range is 5 degrees, and the final adjustment range = 5 degrees × (1 - 0.5) = 2.5 degrees. This gentle adjustment helps users gradually correct their posture. Finally, the adjustment direction and the final adjustment range are integrated to generate clear standardized correction parameters for the left and right shoulder joints, respectively. For example, the left shoulder joint needs to be increased by 4.5 degrees and the right shoulder joint needs to be decreased by 2.7 degrees, ensuring that users can clearly understand the specific adjustment requirements for each shoulder joint.

[0070] To ensure consistency in limb movement assessment and adjustment standards, the calculation of hip joint correction parameters follows the core logic of the shoulder joint, processing the left and right hip joints separately. The core is related to the hip joint angle deviation value. First, the real-time movement angles of the user's left and right hip joints are obtained. Using the pre-stored standard hip joint movement angle (60 degrees) as a benchmark, the difference between the standard movement angle and the current real-time movement angle is calculated. If the difference is positive, the adjustment direction is to increase the hip joint angle, and the absolute value of the difference is used as the basic adjustment range. If the difference is negative, the adjustment direction is to decrease the hip joint angle, and the absolute value of the difference is used as the basic adjustment range. Then, the same logic as the shoulder joint is applied... The calibration formula is used to calibrate the adjustment range. The final adjustment range = basic adjustment range × (1 - hip joint angle deviation value). For example, if the hip joint angle deviation value is 0.3 (moderate hip movement deviation), the basic adjustment range is 4 degrees, and the final adjustment range = 4 degrees × (1 - 0.3) = 2.8 degrees. This range can ensure the adjustment effect and maintain the stability of the movement. Finally, clear standardized correction parameters for the left hip joint and the right hip joint are generated for the left and right hip joints respectively. For example, the left hip joint needs to be increased by 3.2 degrees and the right hip joint needs to be decreased by 1.8 degrees, clearly defining the adjustment direction and specific range for each hip joint.

[0071] The calculation of trunk stability correction parameters uses the previously output trunk stability coefficient as the core judgment basis, combined with the acquired spatial posture offset and limb movement symmetry indicators, to generate targeted adjustment suggestions. First, the system presets a judgment threshold of 0.6 for the trunk stability coefficient. This threshold is based on statistical analysis of exercise data from over 1000 healthy individuals, taking the lower quartile of the trunk stability coefficient distribution; over 90% of healthy exercisers have coefficients higher than this value. Standardized trunk stability correction parameters are generated in two cases: Case 1, if the trunk stability coefficient ≥ 0.6, it indicates that the trunk movement state is stable, with no obvious forward tilt, lateral deviation, or swaying, and the output trunk posture is stable, maintaining the current state of trunk stability. Standardized correction parameters guide users to maintain their current posture. Case 2: If the trunk stability coefficient is <0.6, it indicates that there is an abnormal trunk posture. A complete solution needs to be generated by integrating multiple dimensions of indicators. That is, the direction of correction is clearly defined based on the spatial posture offset. For example, if the forward tilt is 10 degrees, it needs to be adjusted backward. Balance suggestions are given with reference to the limb movement symmetry index. For example, if the left side amplitude is too large, the left side amplitude is reduced. The adjustment amplitude level is divided according to the degree to which the coefficient is lower than the threshold. For example, if it is lower than the threshold of 0.2, it is a large adjustment, and if it is lower than 0.05, it is a small adjustment. Finally, it is integrated into a complete trunk stability standardized correction parameter. For example, the trunk needs to be adjusted backward significantly, while the left limb movement amplitude is reduced to make the left and right amplitudes consistent and maintain the vertical stability of the trunk.

[0072] After the above targeted calculations, five types of standardized correction parameters were finally generated: standardized correction parameters for the left shoulder joint, standardized correction parameters for the right shoulder joint, standardized correction parameters for the left hip joint, standardized correction parameters for the right hip joint, and standardized correction parameters for trunk stability.

[0073] Based on the difference between the movement posture correction parameters and the user's current movement state, a comprehensive health score is calculated, comprising multiple dimensions. Specifically, this includes: dividing the standardized correction parameters for the left and right shoulder joints, the left and right hip joints, and the trunk stability coefficient into three core assessment dimensions: shoulder movement standardization, hip movement standardization, and trunk stability. The comprehensive health score is generated through a logic of calculating and weighting the scores by dimension. The specific process is as follows: The shoulder movement standardization score is calculated as follows: the angle adjustment range included in the standardized correction parameters for the left and right shoulder joints directly reflects the difference between the shoulder movement and the standard posture. The smaller the adjustment range, the closer the shoulder movement is to the standard, the higher the standardization, and the higher the score. A linear mapping formula is used to convert the angle adjustment range into a score from 0 to 100, with a coefficient set to 10 in the formula to ensure accurate adjustment. When the deviation range is within the common 0 to 10 degree range in daily exercise, the score can fully cover 0 to 100 points. The formula is: Single shoulder score = 100 - angle adjustment range of the standardized correction parameter of the single shoulder joint × 10. For example, if the angle adjustment range of the standardized correction parameter of the left shoulder joint is 0.5 degrees, the corresponding single shoulder score = 100 - 0.5 × 10 = 95 points; if the angle adjustment range of the standardized correction parameter of the right shoulder joint is 1.2 degrees, the corresponding single shoulder score = 100 - 1.2 × 10 = 88 points. Since the left and right shoulder joints have an equally important impact on the quality of movement, the arithmetic mean of the single shoulder scores of both sides is taken as the final dimension score of shoulder movement standardization. The formula is: Shoulder movement standardization score = (left shoulder score + right shoulder score) / 2. For example, combining the above single shoulder scores, the shoulder movement standardization score is (95 + 88) / 2 = 91.5 points, with a score range of 0 to 100 points.

[0074] The calculation logic for hip movement standardization score is completely consistent with that for shoulder movement standardization. The core correlation is the standardized correction parameters of the left and right hip joints. The unilateral hip score uses the same linear mapping formula as the unilateral shoulder score: unilateral hip score = 100 - angle adjustment range of the unilateral hip joint standardized correction parameter × 10. For example, if the angle adjustment range of the left hip joint standardized correction parameter is 0.8 degrees, the corresponding unilateral hip score is 100 - 0.8 × 10 = 92 points; if the angle adjustment range of the right hip joint standardized correction parameter is 0.6 degrees, the corresponding unilateral hip score is 100 - 0.6 × 10 = 94 points. The final dimension score is the arithmetic mean of the unilateral hip scores on both sides, and the formula is: hip movement standardization score = (left hip score + right hip score) / 2. For example, combining the above unilateral hip scores, the hip movement standardization score is (92 + 94) / 2 = 93 points, with a score range of 0 to 100 points.

[0075] The trunk stability score is calculated in relation to the trunk stability coefficient, which directly characterizes the stability of trunk movement. The larger the coefficient, the more stable the trunk, the lower the risk of movement imbalance, and the higher the score. A linear mapping formula is used to convert the trunk stability coefficient in the range of 0 to 1 into a score of 0 to 100. The formula is: Trunk stability score = Trunk stability coefficient × 100. For example, if the trunk stability coefficient is 0.85, the corresponding trunk stability score is 0.85 × 100 = 85 points, with a score range of 0 to 100.

[0076] To ensure that the comprehensive health score objectively reflects the impact of each dimension on exercise quality, a 5-fold cross-validation was used to optimize the weight allocation (the validation set consisted of over 500 user data points covering different ages and exercise levels, with the goal of minimizing the error of the comprehensive health score). The final weights for each dimension are as follows: shoulder movement standardization score: 0.4 (joint standardization is the core of exercise quality, directly affecting exercise efficiency and the risk of injury); hip movement standardization score: 0.4 (equally important as shoulder movement standardization, together constituting the core standardization indicators of limb movement); trunk stability score: 0.2 (provides support for limb movement, and is...). The stability auxiliary protection index is used to calculate the comprehensive health score using a weighted summation formula: Comprehensive Health Score = Shoulder Movement Standardization Score × 0.4 + Hip Movement Standardization Score × 0.4 + Trunk Stability Score × 0.2. The comprehensive health score ranges from 0 to 100 points. A higher score indicates more standard exercise posture, higher exercise quality, and lower risk of sports injury. It can comprehensively quantify and reflect the user's overall exercise health level. Finally, it outputs the comprehensive health score (0 to 100 points) and specific scores for the three dimensions of shoulder movement standardization, hip movement standardization, and trunk stability, fully presenting the user's strengths and weaknesses in exercise posture.

[0077] This embodiment calculates the real-time motion angles of key joints and obtains initial deviations by comparing with standard angles, providing basic data support for motion posture assessment and ensuring the relevance of the assessment. It integrates multi-dimensional indicators at the feature level, fully consolidating key information such as joint angle deviations, trunk posture, motion symmetry, and stability, avoiding the limitations of single-indicator assessments. The pre-trained assessment model, through deep feature mining and nonlinear transformation, can capture the complex correlations between various indicators, improving the accuracy and reliability of the assessment results. The generated motion posture correction parameters clearly define the specific adjustment direction and magnitude, providing users with actionable optimization suggestions. The comprehensive health score intuitively reflects the quality of exercise, helping users fully understand their own exercise status and achieving scientific and accurate motion posture assessment and optimization guidance.

[0078] In a preferred embodiment of the present invention, based on the comprehensive evaluation results, a visualization containing spatial attitude calibration guidance and motion stability optimization suggestions is generated and displayed in real time on the smart mirror, including:

[0079] Based on the shoulder joint angle correction parameters in the posture correction parameters, shoulder joint angle calibration guidance information is generated; based on the hip joint angle correction parameters in the posture correction parameters, hip joint angle calibration guidance information is generated; based on the trunk stability coefficient, suggestions for center of gravity distribution adjustment and core muscle activation guidance are generated. Specifically, the generation of shoulder joint angle calibration guidance information first extracts the adjustment direction and final adjustment range from the standardized correction parameters of the left and right shoulder joints, and combines them with the determined standard shoulder joint motion angle to generate complete calibration guidance. That is, the core logic of angle calibration is first clarified, based on the equation: current real-time shoulder joint angle + final adjustment range = standard shoulder joint motion angle. The core objective of the adjustment should be clearly understood by the user. The operational instructions should then be refined, categorized according to the direction of adjustment. For example, if the adjustment aims to increase the shoulder joint angle, the instructions should clearly state that the posterior shoulder muscles should be engaged appropriately, the upper arm should be slowly raised, and scapular stability should be maintained to avoid compensatory shrugging. If the adjustment aims to decrease the shoulder joint angle, the instructions should state that the anterior shoulder muscles should be relaxed, the upper arm should be slowly lowered, and the angle between the torso and upper arm should gradually approach the standard value, ensuring the user masters the correct technique. Finally, a range of motion control instruction should be added, explaining that each adjustment should be performed in steps of 50% of the final adjustment range. After completing the first step, pause for 2 seconds before proceeding to the second step to avoid large, one-time adjustments that could lead to imbalance and ensure a smooth and controllable adjustment process.

[0080] The generation of hip joint angle calibration guidance information requires first clarifying the core logic of angle calibration. Based on the equation that the current real-time hip joint angle + final adjustment range = standard hip joint movement angle, a clear calibration goal is conveyed to the user. Then, the action operation guidance is refined and described according to the adjustment direction. If the adjustment direction is to increase the hip joint angle, the user is prompted to engage the hip flexor muscles, slowly flex or abduct the lower limb, keep the waist straight, and avoid compensatory lumbar curvature. If the adjustment direction is to decrease the hip joint angle, the user is prompted to relax the hip extensor muscles, slowly retract the lower limb, and maintain the angle between the hip joint and the thigh gradually approaching the standard value. Finally, the amplitude control prompts are added, also adopting the step-by-step adjustment principle. The user is prompted to complete the calibration in steps of 50% of the final adjustment range, and hold the position for 3 seconds after each adjustment to ensure that the muscles around the hip adapt to the adjustment range, which echoes the amplitude control logic of the shoulder joint calibration.

[0081] The process for generating center of gravity distribution adjustment suggestions begins by setting a center of gravity offset adjustment coefficient of 0.8. This coefficient is optimized based on center of gravity distribution data from healthy individuals. Next, suggestions are calculated and generated for different ranges of trunk stability coefficient. If the trunk stability coefficient is ≥0.6, the allowable center of gravity offset range is calculated as: trunk stability coefficient × allowable center of gravity offset coefficient 0.1 (in centimeters). The allowable center of gravity offset coefficient of 0.1 transforms the trunk stability coefficient range from 0 to 1 into a centimeter-level allowable offset range consistent with human movement physiology. This coefficient is determined based on normal center of gravity fluctuation data during exercise in healthy individuals, ensuring that the allowable range is neither too wide (avoiding...). Ignoring minor imbalances and not being overly strict (avoiding excessive constraints that could affect the continuity of movement), a suggestion is generated that the center of gravity is evenly distributed, maintaining the current center of gravity position without additional adjustment. If the trunk stability coefficient is <0.6, first calculate the center of gravity offset correction amount = spatial posture offset amount × (1 - trunk stability coefficient) × center of gravity offset adjustment coefficient, and then determine the center of gravity adjustment direction according to the spatial posture offset direction. For example, if leaning forward, the center of gravity is adjusted backward; if leaning to the left, the center of gravity is adjusted to the right. Finally, a suggestion is generated to adjust the center of gravity in the XX direction. The adjustment distance is implemented step by step according to the center of gravity offset correction amount. After each adjustment, the suggestion is to pause for 2 seconds to ensure that the center of gravity adjustment is accurately adapted to the trunk stability state.

[0082] Core muscle activation guidance is generated by first setting the baseline core muscle activation intensity at 30%, which corresponds to the core activation benchmark for healthy individuals during daily exercise, providing a basic reference for activation guidance. Next, the target core muscle activation intensity is calculated as: Baseline Core Muscle Activation Intensity + (1 - Trunk Stability Coefficient) × Core Activation Intensity Adjustment Coefficient 50%. The core activation intensity adjustment coefficient dynamically adjusts the percentage of extra core muscle activation based on the degree of trunk stability deviation. This coefficient is determined based on core training benchmarks in exercise physiology to ensure that extra activation is not excessive. This exceeds the safe range of force exertion for the human body, while simultaneously achieving the adaptation logic that the worse the stability, the higher the core force requirement, making force exertion guidance more targeted; finally, it identifies the force exertion muscle groups based on the type of abnormal trunk posture: if the trunk leans forward / backward, it indicates that the transverse abdominis and erector spinae muscles contract synergistically according to the target force exertion intensity to maintain the stability of the trunk in the sagittal plane; if the trunk is tilted to the side, it indicates that the oblique abdominal muscles on both sides contract differently according to the target force exertion intensity, with the right oblique abdominal muscle exertion ratio increasing when tilted to the left and the left oblique abdominal muscle exertion ratio increasing when tilted to the right, ensuring that the core muscle exertion is accurately adapted to the trunk stability requirements.

[0083] The guidance on shoulder joint angle calibration, hip joint angle calibration, and center of gravity distribution adjustment is integrated to form a complete set of visual guidance content. Specifically, this includes: First, classifying and organizing the guidance information into three modules: shoulder calibration, hip calibration, and trunk stability adjustment. The shoulder calibration module includes guidance on shoulder joint angle calibration, the hip calibration module includes guidance on hip joint angle calibration, and the trunk stability adjustment module includes suggestions for center of gravity distribution adjustment and core muscle activation guidance. Each module is logically ordered according to target requirements, operational guidelines, and precautions, making the information hierarchy clearer. Second, standardizing the information presentation format, all guidance information is presented in a standardized manner, using a fixed structure of action instructions, body part prompts, and effect descriptions to avoid vague expressions and ensure users can quickly grasp the core information. Finally, adapting to visualization display needs, each module is assigned a dedicated display position based on the display area characteristics of the smart mirror (shoulder calibration information corresponds to the upper area of ​​the mirror, hip calibration information corresponds to the middle area, and trunk stability adjustment information corresponds to the lower-middle area), while also reserving a graphical annotation interface.

[0084] By using different colors to indicate the difference between the user's current movement state and the target movement state, visual guidance content is overlaid on the user's image on the smart mirror in real time. Specifically, the color marking rules are first set up. Based on the final adjustment range (for the shoulder and hip joints) and the center of gravity shift correction amount (for the torso), the difference levels are divided and matched with corresponding colors. When the difference is small (final adjustment range < 2 degrees or center of gravity shift correction amount < 1 cm), a yellow mark is used, indicating a slight deviation that requires minor adjustment; when the difference is moderate (2 degrees ≤ final adjustment range ≤ 5 degrees or 1 cm ≤ center of gravity shift correction amount ≤ 3 cm), an orange mark is used, indicating a moderate deviation that requires targeted adjustment; when the difference is large (final adjustment range > 5 degrees or center of gravity shift correction amount > 3 cm), a red mark is used, indicating a serious deviation that requires significant adjustment; and when there is no difference (final adjustment range = 0 degrees or center of gravity shift correction amount = 0 cm), a green mark is used, indicating a standard posture that should be maintained. The system visually distinguishes the degree of deviation through color. After setting the color labeling rules, the motion capture module tracks key points of the user's limbs (shoulder joint, hip joint, and trunk center) in real time, overlaying the corresponding visual guidance content of the module onto the corresponding parts of the user's image in the mirror: text-based guidance information is displayed next to the corresponding part in the form of a semi-transparent floating frame, with the color consistent with the difference level label; motion guidance graphics indicate the adjustment direction in the form of arrows, such as increasing the angle by pointing outwards, decreasing the angle by pointing inwards, and adjusting the center of gravity by pointing the arrow to the target center of gravity position, with the arrow color also consistent with the difference level label, allowing the user to clearly perceive the adjustment direction; to ensure that the guidance content is synchronized with the user's movement status in real time, a dynamic update mechanism is established, that is, the latest movement posture data is obtained every 0.5 seconds, and the difference level, color label, and guidance information content are updated synchronously, realizing a dynamic closed loop of status monitoring, difference labeling, and guidance adjustment, ensuring the real-time and accuracy of guidance.

[0085] This embodiment, based on motion posture correction parameters, generates shoulder and hip joint angle calibration guidance that directly addresses the core requirements of standardized correction parameters. It clarifies the key points of movement and the force exertion points, avoiding blind adjustments by users and improving the targetedness and accuracy of posture correction. Combining trunk stability coefficient and spatial posture offset, it generates suggestions for center of gravity distribution adjustment and core muscle group force exertion guidance through quantitative calculation, helping users improve trunk stability from both the center of gravity and muscle force exertion aspects, reducing the risk of movement imbalance. Through classification integration and color coding, complex guidance information is transformed into clear and easy-to-understand visual content. Combined with real-time image overlay on a smart mirror, users can quickly identify the difference between their own movement posture and the target state, intuitively obtain the adjustment direction, and reduce the information understanding cost. The visual guidance content is dynamically updated with the user's movement status, realizing real-time monitoring and immediate guidance of movement posture, helping users to correct deviations in a timely manner during exercise, avoiding the solidification of incorrect postures, and improving exercise quality and safety. It connects the previous parameter calculation and correction parameter generation stages, forming a complete process of parameter output, guidance generation, and visual feedback.

[0086] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0087] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0088] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A smart mirror multi-dimensional health assessment system, characterized in that, include: The acquisition module is used to perform human skeletonization processing on the acquired continuous three-dimensional image sequence of user movement and the position information of user limbs in space, extract the three-dimensional spatial coordinate data of key joints of the user's body, and obtain a real-time skeleton model of user movement. The pointing module is used to establish a first reference vector pointing to the center points of the left and right shoulder joints and a second reference vector pointing to the center points of the left and right hip joints, based on the real-time skeleton model and using the center point of the user's torso as the spatial reference point, so as to determine the main reference plane of the user's movement. The construction module is used to select four feature points—the left and right shoulder joints, the left and right hip joints—based on the main reference plane and the real-time skeleton model, in order to construct a dynamic feature plane that represents the limb movement state. The calculation module is used to obtain the spatial attitude offset by calculating the angle between the normal vectors of the dynamic feature plane and the main reference plane, and to construct the minimum bounding rectangle based on the spatial positions of the four feature points to obtain the motion stability index. The evaluation module is used to conduct a health assessment of the user's movement quality based on spatial posture offset and motion stability indicators, combined with the three-dimensional spatial coordinate data of the user's key joints, through a pre-trained evaluation model, and generate a comprehensive evaluation result including motion posture correction parameters. The display module generates visual content, including spatial attitude calibration guidance and motion stability optimization suggestions, based on the comprehensive evaluation results, and displays it on the smart mirror in real time.

2. The intelligent mirror multi-dimensional health assessment system according to claim 1, characterized in that, The acquired continuous 3D image sequence of user movement and the spatial position information of the user's limbs are processed into a human skeleton, and the 3D spatial coordinate data of the key joints of the user's body are extracted to obtain a real-time skeleton model of the user's movement, including: The smart mirror uses a depth camera integrated with it to continuously capture multiple frames of depth images during the user's movement, forming a continuous three-dimensional image sequence; it also uses an infrared sensor to capture the position information of the user's limbs in space, including the distance and orientation data of each limb part relative to the sensor. Based on continuous three-dimensional image sequences and location information, multiple key joints of the user's body are identified; Based on multiple key joints, the coordinate data of each joint in three-dimensional space is calculated, and a real-time skeleton model is constructed based on the three-dimensional spatial coordinate data of each joint and the connection relationship of the human skeleton.

3. The intelligent mirror multi-dimensional health assessment system according to claim 2, characterized in that, Based on a real-time skeleton model, the user's torso center point is used as the spatial reference point. A first reference vector pointing from the spatial reference point to the center points of the left and right shoulder joints, and a second reference vector pointing to the center points of the left and right hip joints, are established to determine the main reference plane for user movement, including: Based on the three-dimensional spatial coordinate data of the left and right hip joints in the real-time skeleton model, the spatial coordinates of the user's torso center point are calculated. Calculate the midpoint of the spatial coordinates of the center points of the left and right shoulder joints, and use it as the center point of the shoulder; calculate the midpoint of the spatial coordinates of the center points of the left and right hip joints, and use it as the center point of the hip. The first reference vector is calculated based on the spatial coordinates of the center point of the torso and the center point of the shoulder; the second reference vector is calculated based on the spatial coordinates of the center point of the torso and the center point of the hip. Based on the first and second reference vectors, the normal vector of the main reference plane is determined by the vector cross product operation, and the main reference plane for user motion is established.

4. The intelligent mirror multi-dimensional health assessment system according to claim 3, characterized in that, Based on the main reference plane and the real-time skeleton model, four feature points—the left and right shoulder joints, the left and right hip joints—are selected to construct a dynamic feature plane representing the limb's motion state, including: Based on the real-time skeleton model, the real-time three-dimensional spatial coordinates of four feature points are obtained, including the real-time three-dimensional spatial coordinates of the center point of the left shoulder joint, the center point of the right shoulder joint, the center point of the left hip joint, and the center point of the right hip joint. A first reference triangle plane is established using the center points of the left shoulder joint, the right shoulder joint, and the left hip joint; a second reference triangle plane is established using the center points of the left shoulder joint, the right shoulder joint, and the right hip joint. Based on the normal vectors of the first and second reference triangle planes, the combined normal vector of the dynamic feature plane is calculated. Based on the comprehensive normal vector and the spatial distribution characteristics of the four feature points, a complete dynamic feature plane representing the limb movement state is constructed.

5. The intelligent mirror multi-dimensional health assessment system according to claim 4, characterized in that, The spatial attitude offset is obtained by calculating the angle between the normal vectors of the dynamic feature plane and the main reference plane, and a minimum bounding rectangle is constructed based on the spatial positions of the four feature points to obtain motion stability indices, including: The spatial attitude offset is obtained by calculating the spatial angle between the composite normal vector of the dynamic feature plane and the normal vector of the main reference plane. Based on the three-dimensional spatial coordinates of the center points of the left shoulder joint, right shoulder joint, left hip joint, and right hip joint, determine the minimum bounding rectangle that contains the spatial positions of all feature points. Calculate the aspect ratio of the minimum bounding rectangle to obtain the limb movement symmetry index; calculate the tilt angle of the minimum bounding rectangle relative to the main reference plane to obtain the limb movement stability index.

6. The intelligent mirror multi-dimensional health assessment system according to claim 5, characterized in that, Based on spatial posture offset and motion stability indices, combined with three-dimensional spatial coordinate data of key joints in the user's body, a pre-trained evaluation model is used to conduct a health assessment of the user's motion quality, generating a comprehensive evaluation result that includes motion posture correction parameters, including: Spatial posture offset, limb movement symmetry index, limb movement stability index, and three-dimensional spatial coordinate data of key joints are input into the pre-trained evaluation model for analysis and processing to obtain preliminary evaluation parameters including shoulder joint angle deviation value, hip joint angle deviation value, and trunk stability coefficient. Based on the initial evaluation parameters, motion posture correction parameters are generated for the user's shoulder joint motion angle, hip joint motion angle, and trunk stability. Based on the degree of difference between the motion posture correction parameters and the user's current motion state, a comprehensive health score containing multiple dimensions is calculated.

7. The intelligent mirror multi-dimensional health assessment system according to claim 6, characterized in that, Spatial posture offset, limb movement symmetry index, limb movement stability index, and three-dimensional spatial coordinate data of key joints are input into a pre-trained evaluation model for analysis and processing. Preliminary evaluation parameters, including shoulder joint angle deviation, hip joint angle deviation, and trunk stability coefficient, are obtained: Based on the three-dimensional spatial coordinate data of key joints, the real-time motion angles of the left and right shoulder joints and the real-time motion angles of the left and right hip joints are calculated. The real-time motion angle is compared with the pre-stored standard motion angle to generate initial angle deviation data for the shoulder and hip joints; The initial angle deviation data, spatial posture offset, limb movement symmetry index and limb movement stability index are fused at the feature level to construct a multi-dimensional feature vector. The multi-dimensional feature vector is then input into a pre-trained evaluation model. The feature extraction layer in the evaluation model performs feature mining on the multi-dimensional feature vector to obtain high-dimensional abstract features. By performing nonlinear transformation and weight allocation on high-dimensional abstract features through the fully connected layer in the evaluation model, preliminary evaluation parameters are obtained. These preliminary evaluation parameters include shoulder joint angle deviation values ​​for quantifying the standardization of shoulder movements, hip joint angle deviation values ​​for quantifying the standardization of hip movements, and trunk stability coefficients for characterizing the stability of the trunk during movement.

8. The intelligent mirror multi-dimensional health assessment system according to claim 7, characterized in that, Based on the comprehensive evaluation results, a visualization containing spatial attitude calibration guidance and motion stability optimization suggestions is generated and displayed in real time on the smart mirror, including: Based on the corrections for shoulder joint angles in the posture correction parameters, shoulder joint angle calibration guidance information is generated; based on the corrections for hip joint angles in the posture correction parameters, hip joint angle calibration guidance information is generated; based on the trunk stability coefficient, suggestions for center of gravity distribution adjustment and core muscle activation guidance are generated. The shoulder joint angle calibration guide, hip joint angle calibration guide, and center of gravity distribution adjustment suggestions are integrated to form a complete visual guide. By using different colors to indicate the difference between the user's current movement state and the target movement state, the visual guidance content is overlaid and displayed in real time on the user's image on the smart mirror.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to perform the system as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, performs the system as described in any one of claims 1 to 8.

Images