Method for evaluating the normativity of a sports gesture based on a behavioral characteristic

Abstract

The application discloses a motion posture normative evaluation method based on behavior feature recognition, relates to the technical field of motion posture evaluation, and comprises the following steps: obtaining key point trajectories by processing visual sensor data, extracting space-time features and combining the space-time features into feature vectors. The feature vectors are mapped to a high-dimensional semantic space by a pre-trained posture semantic encoder to obtain semantic codes containing deep motion patterns; three types of internal attributes, i.e., an interpretable motion rhythm, joint coordination and posture stability, are decoded from the codes. Distance measurement is performed on the attributes in the semantic space based on standard templates, an evaluation grade is determined according to rules, and finally, a structured evaluation report combining the grade and each sub-item value is generated. The method realizes a leap from appearance features to semantic understanding, and improves the accuracy of evaluation and the interpretability of results.

Description

Technical Field

[0001] This invention belongs to the field of motion posture assessment technology, specifically a motion posture normative assessment method based on behavioral feature recognition. Background Technology

[0002] Current computer vision-based motion posture normative assessments commonly employ keypoint trajectory comparison or shallow machine learning classification methods. These methods rely on direct, low-level geometric features extracted from image sequences, such as joint coordinates, angles, or velocities, and use these to perform similarity matching with standard action templates or feed them into classifiers for judgment. Essentially, these techniques compare the geometric forms of motion appearances, failing to deeply understand and model the high-level semantic information and underlying dynamic patterns behind motion behavior. Due to the lack of deep semantic representation of motion patterns, existing methods tend to suffer from decreased robustness and accuracy when faced with individual physiological differences, variations in motion speed, or slight perspective shifts, making it difficult to generalize to complex real-world applications.

[0003] Existing technologies struggle to effectively decouple and quantify the multidimensional intrinsic attributes that constitute movement norms. Conventional methods typically output an overall similarity score or a simple classification label, failing to reveal specific deviations in rhythm control, joint coordination, or body stability. Such assessments lack interpretability and cannot provide detailed feedback with clear guidance for sports training. Therefore, a technological solution is needed that can automatically learn and parse interpretable intrinsic movement attributes from raw motion data, and then perform refined, semantic-based assessments. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, the present invention proposes a motion posture normativity evaluation method based on behavioral feature recognition, comprising: Raw motion image sequences are captured using a visual sensor, extracted and preprocessed to generate standardized keypoint trajectory data; Extracting local temporal features and global spatial features reflecting dynamic changes in motion posture from standardized keypoint trajectory data; The extracted local temporal features are combined with global spatial features to form a composite behavioral feature vector, which serves as the original feature representation of motion posture. The composite behavior feature vector is input into a pre-trained pose semantic encoder, which maps the composite behavior feature vector to a high-dimensional semantic space to generate a pose semantic code containing motion pattern semantic information. Three types of intrinsic attributes of motion posture are decoded from the posture semantic encoding. The three types of intrinsic attributes include motion rhythm pattern, joint coordination pattern and posture stability pattern. Based on three types of intrinsic attributes, the distance metric between the motion pose to be evaluated and the preset standard pose template is calculated in the semantic space; The corresponding evaluation level is queried in the preset evaluation rule set according to the distance metric, and the evaluation level represents the standardization level of the motion posture. The assessment level is linked to the values ​​of each component of the motion rhythm pattern, joint coordination pattern, posture stability pattern, and distance metric on which the assessment level is based, forming a structured motion posture standardization assessment report.

[0005] Furthermore, the generation of standardized keypoint trajectory data includes: The system uses a visual sensor to capture raw motion image sequences generated during the process of the target being evaluated performing a specific motion task. Target detection and skeleton key point extraction are performed on each frame of the original motion image sequence. The two-dimensional pixel coordinates of each key point are arranged in time sequence to construct a spatiotemporal motion feature matrix. The key point coordinate data in the spatiotemporal motion feature matrix are preprocessed, including coordinate normalization, missing key point interpolation and filling, and smoothing filtering based on moving average, to generate standardized key point trajectory data. The process of performing target detection and skeleton key point extraction on each frame of the original motion image sequence includes: Each frame of the original motion image is input into a pre-trained human pose estimation neural network, which outputs the confidence scores and two-dimensional pixel coordinates of a predetermined number of key points in the human skeleton model. Based on a confidence threshold, the coordinates of reliable key points in each frame of the image are selected, and key points with a confidence level lower than the confidence threshold are discarded. Based on the connection relationship of the human skeleton model, the two-dimensional pixel coordinates of the same key point in different frame images are connected in time sequence to initially form the key point trajectory.

[0006] Furthermore, the extraction of local temporal features and global spatial features reflecting dynamic changes in motion posture from standardized keypoint trajectory data includes: The local temporal features include keypoint velocity sequences, acceleration sequences, and motion direction change sequences; the global spatial features include keypoint relative distance sets, joint angle sets, and posture symmetry indices. For each standardized keypoint trajectory data, the coordinate difference between adjacent time steps is calculated to obtain the velocity sequence of the keypoint; the difference of the velocity sequence is calculated to obtain the acceleration sequence of the keypoint; the angular change of the velocity direction in adjacent time steps is calculated to obtain the motion direction change sequence. At the same time step, calculate the Euclidean distance between all key points pairwise to form a set of relative distances between key points; calculate the joint angle values ​​formed by the corresponding key points according to the predefined joint angle definition to form a set of joint angles; calculate the average spatial position deviation of the left and right symmetrical key points of the human body in a specific movement stage as an indicator of posture symmetry.

[0007] Further, the step of inputting the composite behavior feature vector into the pre-trained pose semantic encoder includes: The pose semantic encoder is a deep neural network based on a multi-head self-attention mechanism, which is obtained through unsupervised training and learning using feature vectors of massive standard motion pose samples. The posture semantic encoder receives a composite behavioral feature vector, and through multi-layer nonlinear transformation and attention weighting, outputs a dense vector of fixed dimension as the posture semantic code. The posture semantic code can represent the deep pattern of motion posture.

[0008] Furthermore, the three types of intrinsic attributes of motion posture decoded from the posture semantic encoding include: The pose semantic encoding is input into three parallel attribute decoding subnetworks; The first attribute decoding subnetwork reconstructs the attitude semantic encoding into a temporal signal reflecting the motion cycle pattern through a temporal deconvolution structure, and analyzes the motion rhythm pattern from it. The second attribute decoding subnetwork maps pose semantic encoding to a joint co-activation coefficient vector through a fully connected layer and an activation function. The joint co-activation coefficient vector reflects the joint coordination mode. The third attribute decoding subnetwork maps attitude semantic encoding into several stability quantification indicators through regression analysis, including the center of mass swing amplitude and the support surface change rate. These stability quantification indicators together constitute the attitude stability mode.

[0009] Furthermore, the step of calculating the distance metric between the motion pose to be evaluated and a preset standard pose template in the semantic space based on three types of intrinsic attributes includes: The distance metric integrates rhythm difference, coordination deviation, and stability margin deviation. Multiple standard posture templates are pre-stored, each containing a corresponding standard motion rhythm pattern, standard joint coordination pattern, and standard posture stability pattern; The dynamic time warping distance between the motion rhythm pattern of the target to be evaluated and the standard motion rhythm pattern of the standard posture template is calculated and used as the rhythm difference degree. Calculate the cosine similarity between the joint coordination pattern of the target to be evaluated and the standard joint coordination pattern of the standard posture template, and take its complement as the coordination deviation. The weighted Euclidean distance between each quantitative index of the attitude stability mode of the target to be evaluated and the corresponding index of the standard attitude stability mode of the standard attitude template is calculated as the stability margin deviation. The rhythm difference, coordination deviation, and stability margin deviation are weighted and summed to obtain a comprehensive distance metric.

[0010] Further, the step of querying the corresponding evaluation level in a preset evaluation rule set based on the distance metric includes: The preset evaluation rule set defines multiple distance measurement intervals, and each distance measurement interval is associated with an evaluation level; The calculated distance metric is compared with the distance metric interval in the evaluation rule set to determine the interval to which the distance metric belongs; The evaluation level associated with the interval is extracted as the normative evaluation result of the motion posture to be evaluated.

[0011] Furthermore, before extracting local temporal features and global spatial features, the process also includes: The standardized keypoint trajectory data is segmented into motion stages to identify the start and end frames of a complete motion cycle. The trajectory data within the motion cycle is used as the basic unit for subsequent feature extraction and evaluation. If the captured motion image sequence contains multiple motion cycles, the subsequent steps are performed independently for each motion cycle.

[0012] Furthermore, the motion phase segmentation of the standardized keypoint trajectory data includes: Select velocity amplitude sequences from key points representing core body movements; Detect the local minima of the velocity amplitude sequence, and identify the data segment between two adjacent local minima as a potential motion stage; By combining prior knowledge of a specific motion task, the identified potential motion phases are merged or split and adjusted to ultimately determine the boundary of each complete motion cycle.

[0013] Furthermore, the weighted Euclidean distance between the quantitative indices of the attitude stability mode of the target to be evaluated and the corresponding indices of the standard attitude stability mode of the standard attitude template, as the stability margin deviation, includes: The system obtains multiple stability quantification indicators contained in the attitude stability mode of the target to be evaluated, and obtains multiple standard stability quantification indicators corresponding to the standard attitude stability mode of the standard attitude template, wherein the stability quantification indicators include the center of mass swing amplitude and the support surface change rate. Each stability quantification indicator is assigned a preset weight coefficient, and the sum of the weight coefficients of all stability quantification indicators is 1. For each stability quantification index, calculate the absolute difference between the index value of the target to be evaluated and the corresponding standard index value of the standard attitude template; Multiply the absolute difference of each stability quantification index by its corresponding weight coefficient to obtain the weighted deviation of the stability quantification index. The stability margin deviation is obtained by summing the weighted deviations of all stability quantification indicators and then taking the square root of the sum of squares.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By using a pre-trained pose semantic encoder, composite behavioral feature vectors representing motion representations are mapped to a high-dimensional semantic space, generating pose semantic codes that contain abstract motion pattern information. This process transforms motion data from a low-level geometric feature space to a high-level semantic feature space, enabling the model to capture the more generalized essence of motion patterns beyond simple trajectory similarity. This technical solution effectively improves the robustness of feature representation to individual differences, non-rigid changes in motion speed, and environmental interference, overcoming the problem of insufficient generalization ability caused by traditional methods relying on fixed geometric rules or shallow statistical features.

[0015] From the generated posture semantic code, three distinct intrinsic physical or physiological attributes—motor rhythm patterns, joint coordination patterns, and posture stability patterns—are decoded. Based on these interpretable attributes, the quantitative distance to the standard template is calculated within the semantic space. This technical solution allows the evaluation process to move beyond simply comparing the overall resemblance of the movement to delve into the core elements constituting a standardized movement for itemized and quantitative comparison. It directly generates interpretable evaluation criteria that point to specific areas for improvement, enhancing the professionalism and guiding value of the evaluation results. Furthermore, by measuring within a unified semantic space, it ensures consistency in evaluation standards across different attributes. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the motion posture normativity evaluation method based on behavioral feature recognition described in this invention. Figure 2 This is a flowchart for extracting local temporal features and global spatial features. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 The system captures raw motion image sequences using a visual sensor, extracts and preprocesses these sequences to generate standardized keypoint trajectory data. From this standardized keypoint trajectory data, it extracts local temporal features reflecting dynamic changes in motion posture and global spatial features reflecting overall distribution characteristics. The extracted local temporal features and global spatial features are combined to form a composite behavioral feature vector, which serves as the original feature representation of the motion posture. This composite behavioral feature vector is input into a pre-trained posture semantic encoder, which maps the vector to a high-dimensional semantic space, generating a posture semantic code containing rich semantic information about motion patterns. From the generated posture semantic code, three types of intrinsic attributes of the motion posture are decoded: motion rhythm pattern, joint coordination pattern, and posture stability pattern. Based on these three intrinsic attributes, a distance metric between the motion posture to be evaluated and a preset standard posture template is calculated in the semantic space. According to the calculated distance metric, the corresponding evaluation level is queried from a preset evaluation rule set; this level represents the standardization level of the motion posture. The evaluation level is bound to the values ​​of the motion rhythm pattern, joint coordination pattern, posture stability pattern, and various components of the distance metric used in its generation, forming a structured motion posture standardization evaluation report.

[0019] In one embodiment of the present invention, taking a squatting exercise as an example scenario, a sequence of raw motion images generated during a user's squatting motion is captured by a visual sensor. This sequence consists of multiple consecutive frames. In a specific implementation, each frame of the raw motion image is input into a pre-trained human pose estimation neural network. The network outputs the confidence scores and two-dimensional pixel coordinates of a predetermined number of key points in the human skeleton model. These key points include joints such as the shoulders, hips, knees, and ankles. Based on a preset confidence threshold, reliable key point coordinates in each frame are selected, discarding key points with confidence scores below the threshold (e.g., when a key point is occluded, its confidence score may be below the threshold). According to the connection relationships of the human skeleton model, the two-dimensional pixel coordinates of the same key point in different frames are sequentially connected to initially form a key point trajectory. All key point trajectories are then arranged sequentially to construct a spatiotemporal motion feature matrix. The keypoint coordinate data in the spatiotemporal motion feature matrix are preprocessed, including coordinate normalization, missing keypoint interpolation and filling, and smoothing filtering based on moving average. This generates standardized keypoint trajectory data. Coordinate normalization transforms pixel coordinates to a relative coordinate system centered on the human torso. Missing keypoint interpolation and filling uses linear interpolation to fill in coordinates lost due to occlusion. Smoothing filtering based on moving average uses a moving average filter with a window size of 5 to reduce data noise.

[0020] In some embodiments, the standardized keypoint trajectory data is segmented into motion phases, and velocity amplitude sequences representing key points of core body motion are selected, such as the velocity amplitude sequences of hip key points. The formula for calculating the velocity amplitude sequence is: in: Indicates at time step velocity amplitude, and This indicates that the key point is at the time step. Two-dimensional coordinates, and This indicates that the key point is at the time step. The system uses two-dimensional coordinates. Local minima are detected in the velocity amplitude sequence, and the data segment between two adjacent local minima is identified as a potential motion phase. For example, in a squat exercise, a local minima may correspond to the start or end posture of the movement. Based on prior knowledge of the squat exercise, the identified potential motion phases are merged or split and adjusted to ultimately determine the boundary of each complete motion cycle. Prior knowledge includes that the squat exercise consists of two main phases: squatting and standing up. The trajectory data within a single complete motion cycle is used as the basic unit for subsequent feature extraction and evaluation. If the captured motion image sequence contains multiple motion cycles, subsequent steps are performed independently for each motion cycle. For example, if a squat training exercise contains ten repetitions, ten independent evaluation units are generated.

[0021] Optionally, in terms of data comparison, the velocity amplitude sequences of different movement cycles may show differences. For example, the peak value of the hip keypoint velocity amplitude sequence in the first squat cycle is lower than that in subsequent cycles, reflecting changes in movement amplitude due to fatigue. In specific implementations, the stability of movement phase segmentation can be verified by comparing the positions of local minima of the velocity amplitude sequence of the same keypoint in consecutive cycles. A timestamp deviation of the local minima within an acceptable range indicates consistent segmentation. It is understood that standardized keypoint trajectory data eliminates the influence of individual body shape and sensor position, making the movement trajectories of different users or multiple executions by the same user comparable. For example, after normalization, the hip keypoint trajectories of all users originate from the pelvic center. In some embodiments, missing keypoint interpolation filling ensures trajectory continuity and avoids missegmentation of movement phases due to temporary occlusion. For example, hand swings during squats may temporarily occlude shoulder keypoints, but the interpolated trajectory remains complete. Optionally, smoothing filtering based on moving averages reduces high-frequency noise, making the detection of local minima in the velocity amplitude sequence more robust. For example, after smoothing out small fluctuations in the original velocity sequence, local minima more clearly correspond to action transitions. It can be understood that motion phase segmentation combined with prior knowledge adjustments improves the accuracy of cycle boundaries. For example, prior knowledge of squatting motion stipulates that a complete cycle includes the entire process from a standing position to the lowest point and then standing back up; potential motion phases need to be merged accordingly.

[0022] See Figure 2In one embodiment of the present invention, taking the forehand swing in tennis as an example scenario, local temporal features reflecting dynamic changes in motion posture and global spatial features reflecting overall distribution characteristics are extracted from standardized keypoint trajectory data. In specific implementation, the local temporal features include keypoint velocity sequences, acceleration sequences, and motion direction change sequences. For each standardized keypoint trajectory data, the coordinate difference between adjacent time steps is calculated to obtain the keypoint velocity sequence; for example, calculating the coordinate difference of the wrist keypoint in consecutive frame images forms the wrist keypoint velocity sequence. The difference between the keypoint velocity sequences is calculated to obtain the keypoint acceleration sequence; for example, performing a difference operation on the wrist keypoint velocity sequence yields the acceleration change sequence of the wrist keypoint. The angular change of velocity direction in adjacent time steps is calculated to obtain the motion direction change sequence; for example, calculating the angle sequence between the motion vectors of the wrist keypoint in adjacent frames describes the directional change of the swing path.

[0023] In some embodiments, global spatial features include a set of relative distances between keypoints, a set of joint angles, and a posture symmetry index. At the same time step, the Euclidean distances between all pairwise keypoints are calculated to form a set of relative distances between keypoints. For example, the distances between multiple keypoint pairs, such as shoulder keypoints and elbow keypoints, hip keypoints and knee keypoints, are calculated in the same frame. Based on predefined joint angle definitions, the joint angle values ​​formed by the corresponding keypoints are calculated to form a set of joint angles. For example, the elbow joint angle is defined as being calculated from the coordinates of the shoulder keypoint, elbow keypoint, and wrist keypoint. The average spatial position deviation of the left and right symmetrical keypoints of the human body during a specific movement phase is calculated as a posture symmetry index. For example, during the preparation phase for swinging the racket to hit the ball, the average spatial offset between the left and right shoulder keypoints is calculated.

[0024] Optionally, in terms of data comparison, the key point velocity sequences of correct and incorrect swing movements show significantly different patterns. In the correct movement, the wrist key point velocity sequence exhibits a single-peak shape of initial acceleration followed by deceleration, while in the incorrect movement, the wrist key point velocity sequence may show multiple irregular peaks. The formula for calculating the joint angle set is: in: Representing joint angles, key points A, B, and C represent the proximal, central, and distal key points constituting the joint, respectively. Vector This represents the vector pointing from keypoint B to keypoint A. This represents the vector pointing from keypoint B to keypoint C, with the symbol […]. The dot product operation of vectors is represented by the symbol. This represents the magnitude of the vector. It can be understood that the set of elbow angles calculated using this formula will exhibit a continuous, smooth curve from extension to flexion and back to extension in a standard swing motion. However, in incorrect movements with stiff elbows, the range of change in the set of elbow angles is small and the curve is flat. In some embodiments, the posture symmetry index can quantify the balance of the left and right sides of the body during movement. For example, in an ideal ready stance, the line connecting the left and right shoulder key points should be approximately horizontal, with a mean spatial position deviation close to zero. In an incorrect ready stance where the center of gravity is shifted to one side, the mean spatial position deviation of the shoulder key points will be significantly greater than zero. Optionally, the set of relative distances between key points can characterize the relative positional relationship between the limbs and the trunk during movement. For example, in a complete follow-through phase, the distance between the wrist key point on the racket side and the hip key point on the opposite side will significantly increase, while in incorrect movements where the follow-through is incomplete, this distance changes less. It can be understood that the local temporal features and global spatial features extracted synchronously from the same standardized key point trajectory data together constitute a comprehensive mathematical description of the swing motion in terms of temporal dynamics and spatial configuration.

[0025] In one embodiment of the present invention, taking the transition from "Downward-Facing Dog" to "Upward-Facing Dog" in yoga practice as an example scenario, a composite behavioral feature vector is input into a pre-trained posture semantic encoder. The posture semantic encoder is a deep neural network based on a multi-head self-attention mechanism. It is obtained through unsupervised training using feature vectors from a massive number of standard motion posture samples. The massive number of standard motion posture samples used for training includes standardized feature data from various yoga postures. The posture semantic encoder receives the composite behavioral feature vector describing the transition from "Downward-Facing Dog" to "Upward-Facing Dog," and through multi-layer nonlinear transformations and attention weighting, outputs a dense vector of fixed dimensions as the posture semantic code. The posture semantic code can represent the deep motion pattern semantic information of this transition movement.

[0026] In some embodiments, three types of intrinsic attributes of the motion posture are decoded from the posture semantic code, and the posture semantic code is input into three parallel attribute decoding sub-networks. The first attribute decoding sub-network reconstructs the posture semantic code into a temporal signal reflecting the periodic pattern of motion through a temporal deconvolution structure, and analyzes the motion rhythm pattern from it. For example, the decoded temporal signal can show the periodic fluctuation pattern of the body's center of gravity in the vertical direction. The second attribute decoding sub-network maps the posture semantic code into a joint co-activation coefficient vector through fully connected layers and activation functions. The joint co-activation coefficient vector reflects the joint coordination pattern. For example, the coefficients corresponding to the shoulder, hip, and ankle joints in the vector represent the degree of coordinated activity of these joints in the movement. The formula for generating the joint co-activation coefficient vector can be expressed as: in: This represents the output vector of joint co-activation coefficients. This represents the input pose semantic encoding vector. This represents the weight matrix of the fully connected layer. This represents the bias vector. This represents the activation function. The third attribute decoding subnetwork maps the posture semantic encoding into several stability quantification indicators through regression analysis, including the center of mass swing amplitude and the rate of change of the support surface. These stability quantification indicators together constitute the posture stability pattern. For example, the center of mass swing amplitude indicator describes the range of sway of the body's center of gravity during the movement.

[0027] Optionally, in terms of data comparison, the movement rhythm patterns presented by the posture semantic encoding of experienced yoga practitioners and beginners after decoding by the first attribute decoding subnetwork differ. The temporal signals of experienced practitioners show stable and highly repetitive periodic waveforms, while the temporal signals of beginners may show uneven period lengths or irregular waveforms. Joint co-activation coefficient vectors can quantify the coordination relationship of different joint groups. In the standard transition from "downward-facing dog" to "upward-facing dog," the joint co-activation coefficient vector elements corresponding to the shoulder girdle and core muscle groups have higher values, indicating that these areas co-dominate the movement. In compensatory errors, this may manifest as abnormally high coefficients corresponding to the wrists or lower back. It can be understood that the posture semantic encoder compresses and encodes high-dimensional composite behavioral feature vectors into dense vectors containing semantics. This process abstracts and preserves the essential pattern information of the movement posture, allowing subsequent attribute decoding to focus on intrinsic attributes such as rhythm, coordination, and stability. In some embodiments, the stability quantification index decoded by the third attribute decoding subnetwork can be directly compared. For example, in the same yoga pose, a smaller value in the center of mass swing amplitude index in the posture stability pattern usually indicates more stable body control during the movement. Optionally, three parallel attribute decoding subnetworks share the same pose semantic code as input, but each learns different mapping functions, specifically for reconstruction, classification, and regression tasks, thereby decoupling different motion attributes from a unified semantic representation. It can be understood that the three types of intrinsic attributes decoded from the pose semantic code provide a multi-perspective interpretation of motion posture: the motion rhythm pattern describes regularity in the time dimension, the joint coordination pattern describes synergy in the spatial dimension, and the posture stability pattern describes controllability in the state dimension.

[0028] In one embodiment of the present invention, taking the standard squat movement in fitness as an example scenario, a distance metric between the movement posture to be evaluated and a preset standard posture template is calculated in the semantic space based on three types of intrinsic attributes. This distance metric integrates rhythm difference, coordination deviation, and stability margin deviation. In a specific implementation, the system pre-stores a standard posture template describing a perfect squat movement. The standard posture template includes a corresponding standard movement rhythm pattern, a standard joint coordination pattern, and a standard posture stability pattern. The dynamic time warping distance between the movement rhythm pattern of the user's squat movement and the standard movement rhythm pattern of the standard posture template is calculated. This dynamic time warping distance is used as the rhythm difference, which can align and compare two rhythm sequences that may have speed differences. The cosine similarity between the joint coordination pattern of the user's squat movement and the standard joint coordination pattern of the standard posture template is calculated. The complement of the cosine similarity is taken as the coordination deviation. A larger coordination deviation indicates a greater difference between the joint coordination relationship and the standard pattern.

[0029] In some embodiments, the weighted Euclidean distance between each quantitative index of the posture stability pattern of the target to be evaluated and the corresponding index of the standard posture stability pattern of the standard posture template is calculated as the stability margin deviation. Multiple stability quantitative indices included in the posture stability pattern of the user's squatting motion to be evaluated are obtained, as well as multiple standard stability quantitative indices corresponding to the standard posture stability pattern of the standard posture template. These stability quantitative indices include the amplitude of the center of mass swing and the rate of change of the support surface. A preset weight coefficient is assigned to each stability quantitative index, with the sum of all weight coefficients being 1. For example, the weight coefficient for the amplitude of the center of mass swing is set to 0.6, and the weight coefficient for the rate of change of the support surface is set to 0.4. The formula for calculating the weighted Euclidean distance of the stability margin deviation is: in: This represents the calculated stability margin deviation. This represents the total number of stability quantification indicators. Indicates the first Preset weighting coefficients for each stability quantification indicator. The first objective to be evaluated The value of a stability quantification indicator. The first one represents the standard attitude template. The values ​​of each standard stability quantification index are calculated. For each stability quantification index, the absolute difference between the index value of the target to be evaluated and the corresponding standard index value of the standard attitude template is calculated. The absolute difference of each stability quantification index is multiplied by its corresponding weight coefficient to obtain the weighted deviation of this stability quantification index. The weighted deviations of all stability quantification indices are summed in squares and the result is squared to obtain the final stability margin deviation. The rhythm difference degree, coordination deviation degree and stability margin deviation are weighted and summed to obtain a comprehensive distance metric.

[0030] Optionally, in terms of data comparison, there are differences in the distance metric values ​​calculated for squat movements of different qualities and the same standard posture template. It can be understood that the standard motion rhythm pattern in the standard posture template defines the standard time proportions for each phase of an ideal squat. However, if the user's actual motion rhythm pattern involves squatting too quickly or standing up too slowly, the calculated dynamic time warping distance value, i.e., the rhythm difference, will increase. In some embodiments, the standard joint coordination pattern defines the ideal coordination relationship of hip, knee, and ankle joint angle changes during the squat. If the user's movement exhibits a coordination deviation such as knee valgus, the cosine similarity between their joint coordination activation coefficient vector and the standard vector will decrease, leading to an increase in coordination deviation. Refer to Table 1 for specific numerical comparison examples of the calculated values ​​for three different squat movements.

[0031] Table 1: Comparison of Distance Measurement Components and Evaluation Results for Different Squat Movements Optionally, as shown in Data Comparison Table 1, User B's "squatting too fast" movement is mainly reflected in a significantly higher rhythm difference than the standard movement, while User C's "knee valgus" movement is mainly reflected in a significantly higher coordination deviation. It is understandable that stability margin deviation comprehensively reflects the stability of body center of gravity control and foot support. In User C's movement, knee valgus may be accompanied by body swaying and unstable support, therefore its stability margin deviation is also higher than the standard movement. The comprehensive distance metric integrates the differences of the three components through weighted summation, providing a scalarized metric for the overall difference between user movements and the standard template.

[0032] In one embodiment of the present invention, taking the "cloud hands" movement of Tai Chi as an example scenario, the corresponding evaluation level is queried from a preset evaluation rule set based on the calculated distance metric. The preset evaluation rule set defines multiple distance metric intervals, each associated with an evaluation level, which represents the standardization level of the movement posture. In specific implementations, the preset evaluation rule set is stored in the form of a data structure, such as a lookup table containing upper and lower bounds of the intervals and level identifiers. The distance metric intervals are either continuous or discontinuous numerical ranges. The calculated distance metric of the user's "cloud hands" movement is compared with the distance metric intervals defined in the evaluation rule set to determine the interval to which the distance metric belongs. The evaluation level associated with this interval is extracted, and this evaluation level is used as the standardization evaluation result of the "cloud hands" movement posture to be evaluated.

[0033] In some embodiments, the preset evaluation rule set includes four distance measurement intervals, each associated with one of the four evaluation levels: "Excellent," "Good," "Pass," and "Needs Improvement." The boundary values ​​of the distance measurement intervals are set based on statistical analysis of distance measurements from a large number of standard actions and common incorrect actions. The formula for determining which interval a distance measurement belongs to can be expressed as: in: This represents the calculated distance metric value for the action to be evaluated. Indicates the first in the set of evaluation rules The lower bound of a distance metric interval. Indicates the first The upper bound of a distance metric interval. Indicates the first The evaluation level associated with each distance metric interval, symbol and This indicates a comparison operation. Optionally, in terms of data comparison, different users or different numbers of "cloud hand" gestures by the same user will have different distance metrics calculated due to their differences from the standard template, thus resulting in different evaluation levels.

[0034] It's understandable that a "cloud hand" movement with a distance metric of 0.10 would be rated "excellent" within the 0.00-0.15 range of the evaluation rule set, while a "cloud hand" movement with a distance metric of 0.30 would be rated "needs improvement" within the 0.25-0.40 range. In data comparison, a standard but slightly stiff movement might have a distance metric of 0.20, and within the 0.15-0.25 range, its evaluation level would be "good." This demonstrates the role of the evaluation rule set in grading and classifying movement quality. It's understandable that the evaluation rule set maps continuously quantified distance metrics to discrete, semantically meaningful evaluation levels. This is more intuitive for users or coaches than directly using specific distance metric values, and it also facilitates statistical analysis of historical evaluation results.

[0035] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A motion posture normativity evaluation method based on behavioral feature recognition, characterized in that, include: Raw motion image sequences are captured using a visual sensor, extracted and preprocessed to generate standardized keypoint trajectory data; Extracting local temporal features and global spatial features reflecting dynamic changes in motion posture from standardized keypoint trajectory data; The extracted local temporal features are combined with global spatial features to form a composite behavioral feature vector, which serves as the original feature representation of motion posture. The composite behavior feature vector is input into a pre-trained pose semantic encoder, which maps the composite behavior feature vector to a high-dimensional semantic space to generate a pose semantic code containing motion pattern semantic information. Three types of intrinsic attributes of motion posture are decoded from the posture semantic encoding. The three types of intrinsic attributes include motion rhythm pattern, joint coordination pattern and posture stability pattern. Based on three types of intrinsic attributes, the distance metric between the motion pose to be evaluated and the preset standard pose template is calculated in the semantic space; The corresponding evaluation level is queried in the preset evaluation rule set according to the distance metric, and the evaluation level represents the standardization level of the motion posture. The assessment level is linked to the values ​​of each component of the motion rhythm pattern, joint coordination pattern, posture stability pattern, and distance metric on which the assessment level is based, forming a structured motion posture standardization assessment report.

2. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 1, characterized in that, The generation of standardized key point trajectory data includes: The system uses a visual sensor to capture raw motion image sequences generated during the process of the target being evaluated performing a specific motion task. Target detection and skeleton key point extraction are performed on each frame of the original motion image sequence. The two-dimensional pixel coordinates of each key point are arranged in time sequence to construct a spatiotemporal motion feature matrix. The key point coordinate data in the spatiotemporal motion feature matrix are preprocessed, including coordinate normalization, missing key point interpolation and filling, and smoothing filtering based on moving average, to generate standardized key point trajectory data. The process of performing target detection and skeleton key point extraction on each frame of the original motion image sequence includes: Each frame of the original motion image is input into a pre-trained human pose estimation neural network, which outputs the confidence scores and two-dimensional pixel coordinates of a predetermined number of key points in the human skeleton model. Based on a confidence threshold, the coordinates of reliable key points in each frame of the image are selected, and key points with a confidence level lower than the confidence threshold are discarded. Based on the connection relationship of the human skeleton model, the two-dimensional pixel coordinates of the same key point in different frame images are connected in time sequence to initially form the key point trajectory.

3. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 1, characterized in that, The extraction of local temporal features and global spatial features reflecting dynamic changes in motion posture from standardized keypoint trajectory data includes: The local temporal features include keypoint velocity sequences, acceleration sequences, and motion direction change sequences; the global spatial features include keypoint relative distance sets, joint angle sets, and posture symmetry indices. For each standardized keypoint trajectory data, the coordinate difference between adjacent time steps is calculated to obtain the velocity sequence of the keypoint; the difference of the velocity sequence is calculated to obtain the acceleration sequence of the keypoint; the angular change of the velocity direction in adjacent time steps is calculated to obtain the motion direction change sequence. At the same time step, calculate the Euclidean distance between all key points pairwise to form a set of relative distances between key points; calculate the joint angle values ​​formed by the corresponding key points according to the predefined joint angle definition to form a set of joint angles; calculate the average spatial position deviation of the left and right symmetrical key points of the human body in a specific movement stage as an indicator of posture symmetry.

4. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 1, characterized in that, The step of inputting the composite behavior feature vector into the pre-trained pose semantic encoder includes: The pose semantic encoder is a deep neural network based on a multi-head self-attention mechanism, which is obtained through unsupervised training and learning using feature vectors of massive standard motion pose samples. The posture semantic encoder receives a composite behavioral feature vector, and through multi-layer nonlinear transformation and attention weighting, outputs a dense vector of fixed dimension as the posture semantic code. The posture semantic code can represent the deep pattern of motion posture.

5. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 1, characterized in that, The three types of intrinsic attributes of motion posture decoded from posture semantic encoding include: The pose semantic encoding is input into three parallel attribute decoding subnetworks; The first attribute decoding subnetwork reconstructs the attitude semantic encoding into a temporal signal reflecting the motion cycle pattern through a temporal deconvolution structure, and analyzes the motion rhythm pattern from it. The second attribute decoding subnetwork maps pose semantic encoding to a joint co-activation coefficient vector through a fully connected layer and an activation function. The joint co-activation coefficient vector reflects the joint coordination mode. The third attribute decoding subnetwork maps attitude semantic encoding into several stability quantification indicators through regression analysis, including the center of mass swing amplitude and the support surface change rate. These stability quantification indicators together constitute the attitude stability mode.

6. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 5, characterized in that, The method of calculating the distance metric between the motion pose to be evaluated and a preset standard pose template in semantic space based on three types of intrinsic attributes includes: The distance metric integrates rhythm difference, coordination deviation, and stability margin deviation. Multiple standard posture templates are pre-stored, each containing a corresponding standard motion rhythm pattern, standard joint coordination pattern, and standard posture stability pattern; The dynamic time warping distance between the motion rhythm pattern of the target to be evaluated and the standard motion rhythm pattern of the standard posture template is calculated and used as the rhythm difference degree. Calculate the cosine similarity between the joint coordination pattern of the target to be evaluated and the standard joint coordination pattern of the standard posture template, and take its complement as the coordination deviation. The weighted Euclidean distance between each quantitative index of the attitude stability mode of the target to be evaluated and the corresponding index of the standard attitude stability mode of the standard attitude template is calculated as the stability margin deviation. The rhythm difference, coordination deviation, and stability margin deviation are weighted and summed to obtain a comprehensive distance metric.

7. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 1, characterized in that, The step of querying the corresponding evaluation level in a preset evaluation rule set based on the distance metric includes: The preset evaluation rule set defines multiple distance measurement intervals, and each distance measurement interval is associated with an evaluation level; The calculated distance metric is compared with the distance metric interval in the evaluation rule set to determine the interval to which the distance metric belongs; The evaluation level associated with the interval is extracted as the normative evaluation result of the motion posture to be evaluated.

8. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 1, characterized in that, Before extracting local temporal features and global spatial features, the method further includes: The standardized keypoint trajectory data is segmented into motion stages to identify the start and end frames of a complete motion cycle. The trajectory data within the motion cycle is used as the basic unit for subsequent feature extraction and evaluation. If the captured motion image sequence contains multiple motion cycles, the subsequent steps are performed independently for each motion cycle.

9. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 8, characterized in that, The motion phase segmentation of the standardized key point trajectory data includes: Select velocity amplitude sequences from key points representing core body movements; Detect the local minima of the velocity amplitude sequence, and identify the data segment between two adjacent local minima as a potential motion stage; By combining prior knowledge of a specific motion task, the identified potential motion phases are merged or split and adjusted to ultimately determine the boundary of each complete motion cycle.

10. The motion posture normativity evaluation method based on behavioral feature recognition according to claim 6, characterized in that, The weighted Euclidean distance between the quantitative indicators of the attitude stability mode of the target to be evaluated and the corresponding indicators of the standard attitude stability mode of the standard attitude template is used as the stability margin deviation, including: The system obtains multiple stability quantification indicators contained in the attitude stability mode of the target to be evaluated, and obtains multiple standard stability quantification indicators corresponding to the standard attitude stability mode of the standard attitude template, wherein the stability quantification indicators include the center of mass swing amplitude and the support surface change rate. Each stability quantification indicator is assigned a preset weight coefficient, and the sum of the weight coefficients of all stability quantification indicators is 1. For each stability quantification index, calculate the absolute difference between the index value of the target to be evaluated and the corresponding standard index value of the standard attitude template; Multiply the absolute difference of each stability quantification index by its corresponding weight coefficient to obtain the weighted deviation of the stability quantification index. The stability margin deviation is obtained by summing the weighted deviations of all stability quantification indicators and then taking the square root of the sum of squares.

Images