Running posture continuous capturing and analyzing method for training

Abstract

The invention relates to the technical field of motion capture, in particular to a running posture continuous capture analysis method for training, which comprises the following steps of: predicting a joint position and speed of a current frame based on a joint state vector of a previous frame and a currently processed single-frame image, acquiring a priori motion posture vector, and acquiring a priori motion posture vector; performing 2D key point positioning on the current single-frame image in parallel, extracting an incomplete joint coordinate measurement value caused by limb swing shielding, and combining the prior motion attitude vector with the joint coordinate measurement value; according to the method, the joint position and speed of the current frame are predicted based on cooperative processing of the previous frame joint state and the current image feature information, continuous and stable prior attitude data are obtained, meanwhile, the missing joint position measurement value influenced by limb shielding in the single frame image is extracted in parallel, and the prediction data and the actual measurement value are fused and compensated. Therefore, the technical problem that joint coordinates are not complete due to shielding is solved, and the continuity and stability of attitude capture are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motion capture, in particular to a running posture continuous capture analysis method for training. BACKGROUND

[0002] The technical field of motion capture generally involves capturing and analyzing the motion state of a person or object through sensors, optical devices, or image processing techniques, including limb position, joint angle, posture change, and other spatial motion information, to achieve accurate recording and digital reproduction of real actions.

[0003] The prior art usually uses sensors or image acquisition devices to capture spatial information of human motion in real time, relying on the device to directly record joint position, angle and posture information, but is limited by problems such as limb occlusion and missing joint coordinate data in actual motion scenes. The data captured by images or a single device alone often results in breakpoints or distortions in the recording process, making it difficult to accurately and completely reproduce the real motion process. In addition, the prior art fails to effectively distinguish the differences in the contribution of different motion posture characteristic parameters to overall posture stability, resulting in posture evaluation results that are easily affected by individual differences, making it difficult to quantify minor posture distortions and reducing the effectiveness and practicality of motion posture evaluation and training guidance. Therefore, improvements are needed. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art and propose a running posture continuous capture analysis method for training.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solution: a running posture continuous capture analysis method for training, comprising the following steps:

[0006] Based on the joint state vector of the previous frame and the current single-frame image being processed, the joint position and velocity of the current frame are predicted, the prior motion posture vector is obtained, 2D key point positioning is performed on the current single-frame image in parallel, the incomplete joint coordinate measurement values due to limb swing occlusion are extracted, the prior motion posture vector and the joint coordinate measurement values are combined, and a multi-source posture observation set is established;

[0007] Based on the multi-source posture observation set, the prior motion posture vector therein is taken as a predicted state, the joint coordinate measurement values therein are taken as observation values of an updated state, the filtered corrected joint state is obtained, for joints with missing observation values, the state value inherits the predicted state, for joints with existing observation values, the state value fuses prediction and observation, the filtered corrected joint state of each frame is spliced, and a continuous time series three-dimensional skeleton sequence is generated;

[0008] Based on the data of the starting several gait cycles in the continuous time sequence three-dimensional skeleton sequence, four points of left shoulder, right shoulder, left hip and right hip are selected to construct a core quadrilateral, real-time core quadrilateral geometric parameters are obtained, all the real-time core quadrilateral geometric parameters in the target gait cycle are calculated, and a personal ideal core template is established.

[0009] Based on the continuous time sequence three-dimensional skeleton sequence of each subsequent frame and the personal ideal core template, the core quadrilateral geometric parameters of the current frame are extracted, and the corresponding parameters of the personal ideal core template are calculated item by item to obtain a core shape distortion value. The core shape distortion value is converted to quantify the stability of the current running posture and obtain a posture congruence score.

[0010] Preferably, the obtaining step of the multi-source posture observation set is:

[0011] Based on the joint state vector of the previous frame, the joint position values and velocity values of the joint state vector are analyzed, the joint position values and velocity values obtained by analysis are updated item by item using the image feature information of the single frame image currently processed to obtain joint position prediction values and joint velocity prediction values of the current frame, and a prior motion posture vector is formed.

[0012] According to the single frame image currently processed, 2D key point positioning of each joint is performed one by one, whether each joint is shielded due to limb swing is judged, the coordinate missing condition of the shielded joint is screened, and the joint coordinate values that are not shielded and can be directly measured in the coordinate missing condition are extracted to obtain joint coordinate measurement values.

[0013] Based on the prior motion posture vector and the joint coordinate measurement values, the corresponding joint position prediction values in the prior motion posture vector are mapped and merged with the joint coordinate measurement values one by one to form a multi-source posture observation set.

[0014] Preferably, the obtaining step of the filter-corrected joint state is:

[0015] Based on the multi-source posture observation set, the prior motion posture vector and the joint coordinate measurement values contained in the multi-source posture observation set are analyzed, the joint position prediction values and the joint velocity prediction values in the prior motion posture vector are defined as prediction states, and the joint coordinate measurement values are defined as observation values of updated states to form an initial state mapping relationship.

[0016] According to the initial state mapping relationship, whether the joint coordinate measurement values are missing is judged for each joint. If it is judged that the joint coordinate measurement values are missing, the prediction state of the corresponding joint is directly inherited as the filter-corrected state of the joint. If it is judged that the joint coordinate measurement values are not missing, the prediction state and the observation value are fused and calculated item by item to obtain the filter-corrected state, the filter correction of the single frame joint state is completed, and a filter-corrected joint state is generated.

[0017] Preferably, the step of acquiring the continuous time-series three-dimensional skeleton sequence is:

[0018] Based on the generated filtered correction joint state of each frame, the frames are spliced in sequence according to the acquisition sequence to establish a continuous time-series three-dimensional joint state data sequence, and a continuous time-series three-dimensional skeleton sequence is generated.

[0019] Preferably, the step of acquiring the real-time core quadrilateral geometric parameter is:

[0020] Based on the continuous time-series three-dimensional skeleton sequence, the left shoulder coordinate, the right shoulder coordinate, the left hip coordinate, and the right hip coordinate values in each frame of the three-dimensional skeleton sequence within the initial several gait cycles are extracted, and the coordinate values of the four joints are connected frame by frame to construct a core quadrilateral, and each frame of the core quadrilateral is generated.

[0021] According to each frame of the core quadrilateral, the distance values between the coordinates of the adjacent vertices of the core quadrilateral are calculated one by one to determine the length values of the four sides, and the distance values between the coordinates of the relative vertices in the core quadrilateral are calculated one by one to determine the length values of the two diagonals, and the length values of the four sides and the length values of the two diagonals are combined to generate the real-time core quadrilateral geometric parameter.

[0022] Preferably, the step of acquiring the personal ideal core template is:

[0023] Based on the real-time core quadrilateral geometric parameter, all real-time core quadrilateral geometric parameters are calculated one by one in time sequence within the target gait cycle, the average value of each geometric parameter within the gait cycle is obtained, the average values are combined according to the types of geometric parameters, and the personal ideal core template is generated.

[0024] Preferably, the step of acquiring the core shape distortion value is:

[0025] Based on the continuous time-series three-dimensional skeleton sequence of each subsequent frame, the left shoulder coordinate, the right shoulder coordinate, the left hip coordinate, and the right hip coordinate are extracted in sequence to construct the core quadrilateral structure of each frame, the length of the four sides and the length of the two diagonals are calculated respectively, and the core quadrilateral geometric parameter of the current frame is generated.

[0026] According to the core quadrilateral geometric parameter of the current frame and each corresponding parameter value in the personal ideal core template, a relative error calculation is performed, and at the same time, the standard deviation of all geometric parameters within the ideal gait cycle is extracted from the personal ideal core template as a volatility index to construct a geometric parameter relative error set and a standard deviation set.

[0027] Based on the geometric parameter relative error set and the standard deviation set, the core shape distortion value is calculated.

[0028] Preferably, the steps for obtaining the attitude congruence score are as follows:

[0029] Based on the core shape distortion value, a set of core shape distortion values ​​in historical training samples is collected, and the 95th percentile value of the core shape distortion value set in the distribution ranking is calculated and set as the attitude stability reference threshold.

[0030] Calculate the ratio of the core shape distortion value of the current frame to the attitude stability reference threshold to obtain the normalized distortion value;

[0031] Based on the normalized distortion value, calculate the pose congruence score.

[0032] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0033] This invention, based on the collaborative processing of joint state from the previous frame and current image feature information, predicts joint position and velocity in the current frame, obtaining continuous and stable prior posture data. Simultaneously, it extracts missing joint position measurements affected by limb occlusion from single-frame images in parallel, fusing and compensating for the predicted data with actual measurements. This solves the technical challenge of incomplete joint coordinates due to occlusion, improving the continuity and stability of posture capture. Furthermore, it alternates between filtered prediction and observation fusion. For joints with missing measurements, the predicted state is inherited; for joints with measurements, the predicted state is dynamically corrected, forming accurate and continuous three-dimensional skeletal sequence data, improving the accuracy of digital reconstruction of running posture. Further, for the core running region, a core quadrilateral structure is constructed. Geometric features are extracted based on data from multiple gait cycles, and an ideal core template is established for each individual. Normalization calculations are performed by combining core shape distortion values ​​with posture stability reference thresholds, ensuring that posture stability assessment is no longer affected by individual body size and range of motion differences. This achieves highly targeted and robust quantitative assessment of running posture, guaranteeing the individual adaptability and practicality of the assessment results, and improving posture adjustment and exercise efficiency optimization during running training. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0036] Please see Figure 1 This invention provides a technical solution: a method for continuous capture and analysis of running posture during training, comprising the following steps:

[0037] Based on the joint state vector of the previous frame and the current single frame image, the joint position and velocity of the current frame are predicted, the prior motion posture vector is obtained, the 2D key point positioning of the current single frame image is performed in parallel, the incomplete joint coordinate measurement value caused by limb swing shielding is extracted, the prior motion posture vector and the joint coordinate measurement value are combined, and a multi-source posture observation set is established;

[0038] Based on the multi-source posture observation set, the prior motion posture vector therein is taken as a predicted state, the joint coordinate measurement value therein is taken as an observation value of an updated state, a filtered corrected joint state is obtained, for the joints with missing observation values, the state value inherits the predicted state, for the joints with existing observation values, the state value fuses the prediction and the observation, the filtered corrected joint states of each frame are spliced, and a continuous time sequence three-dimensional skeleton sequence is generated;

[0039] Based on the data of the first several gait cycles in the continuous time sequence three-dimensional skeleton sequence, four points of left shoulder, right shoulder, left hip and right hip are selected to construct a core quadrilateral, real-time core quadrilateral geometric parameters are obtained, all real-time core quadrilateral geometric parameters in a target gait cycle are calculated, and a personal ideal core template is established;

[0040] Based on the continuous time sequence three-dimensional skeleton sequence of each subsequent frame and the personal ideal core template, the core quadrilateral geometric parameters of the current frame are extracted, the core shape distortion value is obtained by calculating the corresponding parameters of the personal ideal core template item by item, the core shape distortion value is converted, the stability degree of the current running posture is quantified, and a posture congruence score is obtained.

[0041] The obtaining steps of the multi-source posture observation set are as follows:

[0042] Based on the joint state vector of the previous frame, the joint position values and velocity values of the joint state vector are analyzed, the joint position values and velocity values analyzed are updated item by item based on the image feature information of the current single frame image, the joint position predicted values and joint velocity predicted values of the current frame are obtained, and the prior motion posture vector is formed;

[0043] According to the current single frame image, the 2D key point positioning of each joint is performed one by one, it is judged whether each joint is shielded due to limb swing, the coordinate missing condition of the shielded joint is screened, and the joint coordinate values not shielded and directly measurable in the coordinate missing condition are extracted, and the joint coordinate measurement value is obtained;

[0044] Based on the prior motion posture vector and the joint coordinate measurement value, the joint position predicted values corresponding to the non-missing joints in the prior motion posture vector are mapped and merged with the joint coordinate measurement value one by one, and the multi-source posture observation set is formed.

[0045] Specifically, based on the joint state vector of the last frame, the three-dimensional spatial position values and three-dimensional velocity values of all human joints recorded in the vector are first analyzed, which are used as the basis for motion state prediction. Specifically, for each joint, a basic uniform motion model is used for preliminary position prediction, i.e., the preliminary predicted position is equal to the position of the last frame plus the product of the velocity of the last frame and the time interval between frames, and the velocity is temporarily inherited from the velocity of the last frame and remains unchanged. Subsequently, to improve the accuracy of the prediction, the initial prediction is corrected using the image feature information of the current single frame image. This process is implemented through a lightweight convolutional neural network, which is specifically composed of five convolutional layers with convolution kernel sizes of 5x5, 3x3, 3x3, 3x3, and 1x1, respectively, and all use ReLU activation functions. The network input is a 64x64 pixel image block centered on the preliminary predicted position of each joint, which is cropped from the current single frame image. The network processes these image blocks to extract local motion feature vectors containing information such as motion blur, lighting changes, and background texture. The obtained feature vectors are input into a three-layer fully connected network along with the preliminary predicted position and velocity values of the corresponding joints. The fully connected network learns the non-linear mapping relationship between image features and motion state changes and outputs a six-dimensional adjustment vector, which corresponds to the correction amounts of three-dimensional position and three-dimensional velocity. The preliminary predicted position and velocity values are added to the adjustment vector item by item to obtain more accurate current frame joint position prediction values and joint velocity prediction values. The updated position and velocity prediction values of all joints are recombined into a vector with the same input format to form the prior motion posture vector.

[0046] Based on the currently processed single-frame image, a pre-trained 2D human keypoint localization model, such as MoveNet-Thunder, is invoked to process the image. This model directly outputs the 2D pixel coordinates (x, y) of all predefined joints in the image, along with a confidence score ranging from 0 to 1. This confidence score reflects the model's grasp of the accuracy of the keypoint localization. Next, each joint is assessed to determine whether its localization failure or inaccuracy is due to rapid movement or mutual occlusion during running. This assessment is based on comparing the confidence score with a preset occlusion threshold. The threshold is set as follows: First, a calibration dataset containing at least 5000 frames of running video is prepared, and each keypoint in each frame is manually labeled to determine its visibility. Then, the keypoint localization model is run on this dataset, and the clearness of each keypoint is collected. The confidence score distribution of clearly visible keypoints and occluded keypoints is analyzed. Typically, the confidence scores of clearly visible keypoints tend to be concentrated in a higher range, such as a mean of 0.88, while the scores of occluded or blurred keypoints tend to be concentrated in a lower range, such as a mean of 0.3. To minimize classification errors, a specific value in the overlapping region of the two distributions is selected as a threshold. For example, 0.65 can be set as the occlusion judgment threshold. For each joint in the current frame, if its confidence score is lower than 0.65, the joint is judged to be an occluded joint, and its coordinate values ​​are considered unreliable, so they are discarded and marked as missing. Conversely, if the confidence score is greater than or equal to 0.65, the joint coordinates are considered valid. After filtering all joints, the 2D coordinate values ​​and their corresponding confidence scores of all joints judged to be unoccluded and directly measurable are extracted to obtain the joint coordinate measurement values.

[0047] Based on the prior motion posture vector and joint coordinate measurements obtained in the previous process, data pairing and structured merging are performed to prepare regularized input data for subsequent filtering and correction steps. This process requires prior acquisition of the camera's intrinsic parameter matrix, which includes the camera's focal length f. x f y and principal point coordinates c x c y These parameters were obtained through a one-time calibration using the standard Zhang Zhengyou checkerboard calibration method before shooting. For each joint in the current frame, its corresponding three-dimensional position prediction value (p) was first extracted from the prior motion pose vector. x ,p y ,p z ) and three-dimensional velocity prediction numerical (v x ,v y ,v z Simultaneously, the corresponding 2D coordinate measurement value (m) of the joint is retrieved from the joint coordinate measurement values. x ,my ) and a confidence score, here the mapping merge is not a direct numerical calculation fusion, but a structured integration of information from two different dimensions, specifically, for each joint, a data structure is created, which explicitly contains two parts: one part is the complete six-dimensional state from kinematic prediction, that is, the corresponding item in the prior motion posture vector, the other part is the two-dimensional coordinate from image observation, that is, the joint coordinate measurement value, if a joint is judged to be occluded in the previous step, its joint coordinate measurement value is empty, then in the data structure of this joint, the observation part will be filled with an invalid identifier, but the data of the prediction part is still complete, this pairing process will traverse all the joints defined in the human body skeletal model, such as head, neck, shoulders, elbows, wrists, hips, knees and ankles, for each joint, a pair of data records containing prediction and observation are generated, finally, all joint data records are spliced into a set according to the predetermined order to form a multi-source posture observation set.

[0048] The acquisition step of the filter correction joint state is:

[0049] Based on the multi-source posture observation set, the prior motion posture vector and the joint coordinate measurement value contained in the multi-source posture observation set are analyzed, the joint position prediction value and the joint velocity prediction value in the prior motion posture vector are defined as the prediction state, and the joint coordinate measurement value is defined as the observation value of the updated state, to form an initial state mapping relationship;

[0050] According to the initial state mapping relationship, whether the joint coordinate measurement value is missing is judged for each joint, if it is judged to be missing, the prediction state of the corresponding joint is directly inherited as the filter correction state of the joint, if it is judged to be not missing, the prediction state and the observation value are calculated to obtain the filter correction state, the filter correction of the single frame joint state is completed, and the filter correction joint state is generated.

[0051] Specifically, based on the multi-source pose observation set, first, a state filter is initialized for each joint in the system, which is responsible for maintaining the motion state of the joint, then, two parts of core data corresponding to each joint are parsed from the multi-source pose observation set, the first part is the prior motion pose vector derived from the kinematic model, the three-dimensional joint position prediction value and the three-dimensional joint velocity prediction value contained in it are directly adopted and combined into a six-dimensional state vector, this vector is defined as the predicted state of the joint at the current time step, the second part is the joint coordinate measurement value derived from image analysis, the two-dimensional joint coordinate value contained in it is adopted and defined as the observation value for correcting the predicted state, at the same time, in order to quantify the uncertainty of prediction and observation, the system will configure the corresponding covariance matrix for the two parts, the setting of the process uncertainty covariance matrix reflects the inherent instability of the motion model, the value is determined based on the statistical analysis of a large amount of standard running motion data, for example, for the speed component, its variance can be set to 0.01, indicating that there is slight random acceleration in the running process, for the position component, the variance is set to 0.001, the setting of the observation uncertainty covariance matrix is directly linked with the confidence score in the joint coordinate measurement value, through a reciprocal function for conversion, for example, the observation noise variance is equal to a basic variance value (such as 5.0 pixel square) divided by the confidence score of the joint, which means the higher the confidence, the more reliable the observation value, and the smaller the noise variance, the predicted state of each joint, the observation value and their respective uncertainty covariance matrix are bound to form the initial state mapping relationship.

[0052] According to the initial state mapping relationship established in the previous step, an independent filter correction calculation is started for each joint. First, check if the joint coordinate measurement value of the current joint exists. This existence judgment is based on the screening result of the key point confidence score in the previous step. If it is judged to be missing, it means that the joint is blocked or blurred in the current image. At this time, the filter will not perform the update step, but directly adopt the predicted state as the final output of the current frame, that is, the filter correction state of the joint is equal in value to its predicted state, and at the same time, its state uncertainty covariance will increase accordingly according to the process noise model. If it is judged to be not missing, a fusion calculation process is started. This process first uses the calibrated camera intrinsic matrix to project the position component in the three-dimensional predicted state to the two-dimensional image plane to generate a two-dimensional predicted observation coordinate. Then, the difference between the predicted observation coordinate and the real joint coordinate measurement value is calculated. This difference is called innovation, which quantifies the deviation between prediction and actual observation. Then, according to the uncertainty covariance of the predicted state and the uncertainty covariance of the observation value, a fusion weight is calculated. This weight determines to what extent the innovation is adopted to correct the predicted state. Specifically, if the uncertainty of the observation value is much smaller than the uncertainty of the predicted state, the weight will be larger, and vice versa. Subsequently, the innovation is multiplied by the fusion weight to obtain a six-dimensional correction vector. Add this correction vector to the original predicted state vector item by item to obtain the filter correction state that fuses the observation information. Finally, update the state uncertainty covariance of the joint. After this calculation, all joints obtain their final state in this frame. Combine the filter correction states of all joints in the predetermined order to generate a single-frame joint state filter correction.

[0053] The acquisition step of the continuous time sequence three-dimensional skeleton sequence is:

[0054] Based on the filter correction joint state generated for each frame, the frames are spliced frame by frame according to the acquisition order to establish a continuous time sequence data sequence of three-dimensional joint states, generating a continuous time sequence three-dimensional skeleton sequence.

[0055] Specifically, based on the filtered and corrected joint state generated after filtering and correcting each frame of image, the system maintains a dynamic data structure for storing continuous motion data, which logically represents a time sequence list, and each element of the list records all joint information of a frame of picture. Specifically, after processing a frame of image and obtaining the corresponding filtered and corrected joint state vector, the system obtains the original timestamp or unique frame number of the frame in the video sequence, then encapsulates the timestamp or frame number and the filtered and corrected joint state vector as a data pair, and appends the data pair to the end of the time sequence list in time increasing order. The filtered and corrected joint state vector itself has a fixed internal structure, for example, it is a one-dimensional array containing all N key points of the human body, and the arrangement order is [x coordinate of joint 1, y coordinate of joint 1, z coordinate of joint 1, x velocity of joint 1, …, z velocity of joint N]. This process is continuously executed throughout the video processing process, and the length of the time sequence list grows continuously as the video frames are processed one by one, thereby accumulating three-dimensional human motion data in a continuous time period. This ordered list dynamically growing over time and containing accurate three-dimensional joint position and velocity information of each frame finally constitutes a continuous time sequence three-dimensional skeleton sequence for subsequent gait analysis.

[0056] The real-time core quadrilateral geometric parameter acquisition step is:

[0057] Based on the continuous time sequence three-dimensional skeleton sequence, the left shoulder coordinate, right shoulder coordinate, left hip coordinate and right hip coordinate values in each frame of the three-dimensional skeleton sequence in the initial several gait cycles are extracted, the coordinate values of the above four joints are connected frame by frame to construct a core quadrilateral, and the core quadrilateral of each frame is generated.

[0058] According to each frame of core quadrilateral, the distance values between the coordinates of adjacent vertices of the core quadrilateral are calculated one by one to determine the length values of the four sides, the distance values between the coordinates of the opposite vertices in the core quadrilateral are calculated one by one to determine the length values of the two diagonals, and the length values of the four sides and the two diagonals are combined to generate real-time core quadrilateral geometric parameters.

[0059] Specifically, based on the continuous time sequence three-dimensional skeleton sequence, first, the starting several gait cycles for establishing the template need to be automatically segmented from it, and this segmentation process is achieved by analyzing the motion trajectory of a specific joint (such as the left ankle or the right ankle) in the vertical direction. Specifically, the system tracks the vertical component (usually the Y-axis coordinate) of the three-dimensional coordinates of the left ankle joint over time. When the vertical component reaches a local minimum, it is marked as a complete heel strike event. A complete gait cycle is defined as the time period from one heel strike event to the next heel strike event of the same foot. The system will continuously monitor and identify the starting three complete gait cycles. Three cycles are selected to ensure data representativeness while avoiding the instability caused by warm-up or adjustment movements in the initial stage. After determining the starting and ending frames of the three gait cycles, the system will traverse each frame of the three-dimensional skeleton sequence within this time period. For each frame, the three-dimensional coordinate values (x, y, z) of the left shoulder, right shoulder, left hip, and right hip are accurately extracted. Subsequently, with these four three-dimensional coordinate points as vertices, they are logically connected in three-dimensional space. The left shoulder is connected to the right shoulder, the right shoulder is connected to the right hip, the right hip is connected to the left hip, and the left hip is connected to the left shoulder to construct a spatial quadrilateral. This process is repeated for all frames within the selected gait cycle to generate the core quadrilateral of each frame.

[0060] Based on the core quadrilateral of each frame generated in the previous step, the system performs geometric parameter quantification calculation on the core quadrilateral of each frame. This calculation process includes the measurement of the lengths of the four sides and the two diagonals. For any two vertices in three-dimensional space, such as the left shoulder coordinate (x LS ,y LS ,z LS ) and the right shoulder coordinate (x RS ,y RS ,z RS ), the distance between them, i.e., the side length or the diagonal length, is obtained by calculating the three-dimensional Euclidean distance. The calculation method is to add the squares of the differences in x, y, and z coordinates, and then take the square root. This calculation is applied to the four adjacent vertex pairs of the core quadrilateral to determine the lengths of the four sides, including the distance between the left shoulder and the right shoulder, the distance between the right shoulder and the right hip, the distance between the right hip and the left hip, and the distance between the left hip and the left shoulder. Similarly, this calculation is also applied to the two opposite vertex pairs of the core quadrilateral, i.e., the left shoulder and the right hip, and the right shoulder and the left hip, to determine the lengths of the two diagonals. After the calculation is completed, for each frame, the system will obtain a one-dimensional vector containing six floating-point numbers. These six numbers are arranged in a fixed order, such as [upper side length, right side length, lower side length, left side length, diagonal 1 length, diagonal 2 length]. This vector is the real-time core quadrilateral geometric parameter of the frame.

[0061] The step of obtaining the personal ideal core template is:

[0062] Based on the real-time core quadrilateral geometric parameters, the average value of each geometric parameter in the gait cycle is calculated in time series form for all real-time core quadrilateral geometric parameters in the target gait cycle, and the average value is combined according to the type of geometric parameter to generate the personal ideal core template.

[0063] Specifically, based on all real-time core quadrilateral geometric parameters calculated in the first three gait cycles, the system aggregates these time series data to establish a stable and representative personal benchmark. First, all real-time core quadrilateral geometric parameters of all frames are collected and grouped according to parameter types to form six independent time series, corresponding to the changes of the four side lengths and two diagonal lengths of the core quadrilateral over time. Next, for each of the six time series, the average value in the entire target gait cycle (i.e. the selected first three gait cycles) is calculated. The specific calculation process is as follows: the values of a specific geometric parameter (e.g. "distance from left shoulder to right shoulder") in all frames are added together, and then divided by the total number of frames to obtain the average value of the geometric parameter. This calculation is performed for all six geometric parameters to obtain six average values, which represent the average lengths of the sides and diagonals of the core quadrilateral of the user in a stable running state. Finally, the six calculated average values are combined into a six-element vector in the same order as the real-time core quadrilateral geometric parameters. This vector is defined as the personal ideal core template of the user.

[0064] The step of obtaining the core shape distortion value is:

[0065] Based on the continuous time series three-dimensional skeleton sequence of each subsequent frame, the left shoulder coordinates, right shoulder coordinates, left hip coordinates, and right hip coordinates are extracted in sequence to construct the core quadrilateral structure of each frame, and the lengths of the four sides and two diagonals are calculated to generate the core quadrilateral geometric parameters of the current frame.

[0066] According to the core quadrilateral geometric parameters of the current frame and each corresponding parameter value in the personal ideal core template, the relative error is calculated, and the standard deviation of all geometric parameters in the ideal gait cycle is extracted from the personal ideal core template as a volatility indicator to construct the relative error set and the standard deviation set of the geometric parameters.

[0067] Based on the relative error set and the standard deviation set of the geometric parameters, the core shape distortion value is calculated, and the calculation formula is:

[0068]

[0069] wherein, is the relative error value of the kth core quadrilateral geometry parameter in the current frame, T k is the average value of the kth core quadrilateral geometry parameter in the personal ideal core template, P k is the actual value of the kth core quadrilateral geometry parameter in the current frame, s k is the standard deviation of the kth core quadrilateral geometry parameter in the personal ideal core template in the complete gait cycle, S is the adaptive stability weight corresponding to the kth core quadrilateral geometry parameter, g is the total number of core quadrilateral geometry parameters in the current frame, λ is the imbalance penalty coefficient, D" z is the core shape distortion value.

[0070] Specifically, based on the continuous time sequence three-dimensional skeleton sequence of each subsequent frame, the system first locates and extracts the three-dimensional space coordinates of the four key points of left shoulder, right shoulder, left hip and right hip from the data record of each frame. Each coordinate is composed of three values of x, y and z. The four coordinate points are obtained through previous filtering correction and have high time sequence continuity and accuracy. After extracting the coordinates, the system logically constructs a spatial quadrilateral structure with the four points as vertices in the three-dimensional space according to the fixed connection order of "left shoulder-right shoulder-right hip-left hip-left shoulder". The structure is dynamically changing, and its shape and size will change in each frame during running. Then, in order to quantify the geometric shape of the quadrilateral, the system calculates the lengths of its four sides and two diagonals one by one. This calculation uses the three-dimensional Euclidean distance formula. Specifically, for any two vertex coordinates, the square of the difference in each coordinate axis (x, y, z) is calculated, and the square root of the sum of the three square values is taken to obtain the straight-line distance between the two points. This process will be repeated six times: calculating the distance between the left shoulder and the right shoulder (upper side length), the distance between the right shoulder and the right hip (right side length), the distance between the right hip and the left hip (lower side length), the distance between the left hip and the left shoulder (left side length), the distance between the left shoulder and the right hip (diagonal line one), and the distance between the right shoulder and the left hip (diagonal line two). The six calculated length values are combined into a six-element vector to generate the core quadrilateral geometry parameter of the current frame.

[0071] According to the core quadrilateral geometry parameters of the current frame consisting of six length values generated in the previous step for the current frame, and calling the previously established personal ideal core template, the system begins to analyze the deviations item by item, and the personal ideal core template stores the average values of the six geometry parameters of the core quadrilateral of the user in the ideal gait, and the system will pair the six geometry parameter values of the current frame with the corresponding six average values in the template one by one. For each pair of parameters, the relative error is calculated. The specific method is to subtract the average value in the template from the parameter value of the current frame, take the absolute value, and then divide by the average value in the template. This calculation will be performed for all six parameters, thereby obtaining a set of six relative error values. At the same time, the system also needs an index to measure the stability of each parameter itself. This index has been calculated and stored when the personal ideal core template is constructed, that is, the standard deviation of each geometry parameter in the ideal gait cycle as a template. The standard deviation reflects the natural fluctuation range of the parameter in the stable running state. The smaller the standard deviation, the more stable the parameter. The system extracts the six standard deviation values from the additional information of the personal ideal core template to form a standard deviation set. Finally, the calculated six relative error values are packaged into a geometry parameter relative error set, and the extracted six standard deviation values are packaged into a standard deviation set.

[0072] Formula: The benefit of the formula is that an adaptive weight w k based on the stability of the parameter itself is introduced k , which is calculated by the reciprocal square of the standard deviation s k of the parameter in the ideal gait cycle. This means that for those geometry parameters that should be very stable in the ideal running posture, such as shoulder width, even if a small relative error δ k is generated, it will be amplified because of its large weight w k , thereby having a significant impact on the total distortion value. Conversely, for those parameters that have a large natural fluctuation, such as the diagonal change caused by trunk twist, their weight is small, and the impact of their error on the total score is also correspondingly weakened. This design enables the evaluation model to intelligently focus on the abnormalities that "should not change but change", greatly improving the sensitivity and accuracy of the posture evaluation; secondly, the second part of the formula is a weighted coefficient of variation, which measures the dispersion or imbalance of the distribution of the six geometry parameter relative error values δ k , and its influence is adjusted by the imbalance penalty coefficient λ. This part captures the "coordination" of posture distortion. Even if the average error of all parameters is not large, if the error of one or two parameters is much larger than that of the other parameters, leading to a highly uneven error distribution, the value of this item will also increase significantly, thereby increasing the total distortion value, which effectively identifies hidden posture defects that have serious local problems but acceptable overall average error.

[0073] P k is the actual value of the kth core quadrilateral geometry parameter in the current frame, which is a basic data unit describing the core morphology of the current instantaneous posture, directly derived from the three-dimensional space measurement of the human body key points in the current video frame. The specific acquisition process is to extract the three-dimensional coordinates of the left shoulder, right shoulder, left hip, and right hip four joint points from the continuous time series three-dimensional skeleton sequence of the current frame through the foregoing steps, and calculate the lengths of the four sides and two diagonals of the core quadrilateral according to these coordinates. The six length values constitute the core quadrilateral geometry parameter set P k of the current frame, and P k is the kth element in the set, for example, if the six parameters are defined in order as follows: 1-upper side length (left shoulder-right shoulder), 2-right side length (right shoulder-right hip), 3-lower side length (right hip-left hip), 4-left side length (left hip-left shoulder), 5-diagonal 1 (left shoulder-right hip), and 6-diagonal 2 (right shoulder-left hip), then P1 represents the three-dimensional space distance between the left and right shoulders in the current frame, which is a real-time changing value directly reflecting the runner's body posture at the current time, for example, in a certain frame image, the coordinates of the runner's left and right shoulders are calculated as (0.2, 1.5, -0.1) meters and (-0.2, 1.5, -0.1) meters respectively, and the calculated value of P1 is 0.4 meters.

[0074] T k is the average value of the kth core quadrilateral geometry parameter in the individual ideal core template, which represents the reference morphology value under the user's individualized, most stable, and most efficient running posture. Its acquisition process is that when the user performs initial running posture capture, the system automatically identifies and intercepts several (for example, 3) complete and stable gait cycle data in the initial stage. For each frame image in these cycles, the system calculates the six geometry parameters of the core quadrilateral, thereby obtaining a series of time series data for each geometry parameter. Subsequently, the average value of each time series data (for example, the sequence of upper side length values of all frames) is calculated, which is defined as the ideal template value T k for the geometry parameter. This process is performed for all six parameters, and the final six average value set constitutes the core content of the individual ideal core template, which is the gold standard for all subsequent posture evaluations. For example, by analyzing the data of the user's initial 3 gait cycles (a total of 180 frames), the average value of the upper side length (left and right shoulder distance) in 180 frames is calculated as 0.395 meters, and the set value of T1 is 0.395 meters.

[0075] s k is the standard deviation of the kth core quadrilateral geometry parameter in the individual ideal core template in a complete gait cycle, which is a key indicator for quantifying the inherent variability of each geometry parameter under the ideal running posture, and it is related to Tk At the same time in the process of building the personal ideal core template, the specific acquisition process is, on the basis of the same starting several stable gait cycle data as the calculation of T k , for each time series data of geometric parameters (for example, all 180 frame upper edge length value sequence), calculate its standard deviation, standard deviation can reflect the discrete degree of data points around its average value (that is, T k ), a smaller s k value indicates that the geometric parameter is very stable in the ideal running posture, and the shape changes little, while a larger s k value indicates that the parameter itself has a larger amplitude of periodic fluctuation. The six calculated standard deviation values and the six average values T k are stored in the personal ideal core template, which provides the basis for subsequent weighted calculation, for example, after analyzing the above 180 frame data, the standard deviation of the upper edge length sequence is 0.008 meters, and the value of s1 is 0.008 meters, while the standard deviation of the diagonal length sequence calculated at the same time may be 0.03 meters.

[0076] δ k is the relative error value of the kth core quadrilateral geometric parameter in the current frame, which is a dimensionless value, used to measure the relative degree of deviation of the current posture parameter from its personal ideal benchmark, which depends entirely on the aforementioned parameters P k and T k , the specific calculation formula is This calculation process is performed in the "building geometric parameter relative error set and standard deviation set" step, the system will traverse the six geometric parameters (k from 1 to 6), and apply this formula to calculate each parameter, thus obtaining a geometric parameter relative error set containing six relative error values, this value converts the percentage form of deviation into a value that can be directly involved in weighted summation, since it takes the absolute value, it only cares about the magnitude of deviation and not the direction of deviation, for example, if the upper edge length P1 of the current frame is 0.4 meters, and the template value T1 is 0.395 meters, then the calculation value of its relative error δ1 is

[0077] w k is the adaptive stability weight corresponding to the kth core quadrilateral geometric parameter, which is the core mechanism of the intelligent evaluation model, it is not a fixed empirical value, but is dynamically calculated according to the stability of each geometric parameter, the calculation formula is where the denominator part is the sum of the reciprocals of the variances of all parameters, which plays a normalizing role, so that the sum of all weights w k is 1, this calculation process uses the parameter s ki.e. the standard deviation of each geometric parameter under ideal running posture, the smaller the variance (square of the standard deviation) of the parameter, the larger the reciprocal of the parameter, and thus the larger the weight w k of the parameter in the total distortion calculation, which means that the system will automatically give more attention to those parameters that should be stable, for example, if the upper side length standard deviation s1 = 0.008 and the diagonal line standard deviation s5 = 0.03, then and

[0078] g is the total number of core quadrilateral geometric parameters of the current frame, which is a fixed integer constant determined by the core morphological model defined in the method, in the method, the core quadrilateral is defined as a morphological model composed of four joint nodes of left shoulder, right shoulder, left hip and right hip, and the description of its geometric characteristics is realized by calculating the lengths of its four sides and the lengths of two diagonal lines, therefore, the number of geometric parameters that need to be analyzed and calculated is 4 (side length) + 2 (diagonal line length) = 6, therefore, the value of g is set to 6.

[0079] λ is an imbalance penalty coefficient, which is a preset hyperparameter used to adjust the penalty strength of the "imbalance" term in the distortion calculation, the setting of this coefficient is not arbitrary, but an empirical value obtained by optimizing a labeled data set, the setting process is as follows: first, collect a running video data set containing at least 100 runners of different levels, and invite 3 experienced running coaches to independently score the posture stability of the runners in the video (for example, 1-10 points), and take the average score as the "expert label" of the sample, then, run the method on the entire data set, keep other parameters unchanged, and let λ take values in a predetermined range (for example, 0.1 to 3.0, step 0.1), for each λ value, calculate the distortion value D z sequence, then calculate the Pearson correlation coefficient between the sequence and the expert score sequence, finally, select the λ value that makes the correlation coefficient the largest as the final configuration of the system, for example, it is found through testing that when λ takes the value of 1.5, the negative correlation between the calculated distortion value and the coach's score is the strongest (the larger the distortion value, the lower the score), therefore, λ is set to 1.5.

[0080] Calculation process:

[0081] A set of core quadrilateral geometric parameters P k captured by the current frame and its corresponding individual ideal core template parameters T k and s k are shown in the following table (unit: meters):

[0082] Table 1 Parameter data

[0083] Parameter k Geometric parameter name P k ]]> T k ]]> s k ]]> 1 Upper side length 0.400 0.395 0.008 2 Right side length 0.520 0.525 0.015 3 Lower side length 0.280 0.282 0.010 4 Left side length 0.518 0.526 0.016 5 Diagonal 1 0.610 0.600 0.030 6 Diagonal 2 0.605 0.615 0.032

[0084] The parameters are calculated from Table 1:

[0085] Step 1: Calculate the relative error δ of each parameter k .

[0086] δ1= |0.400-0.395| / 0.395≈0.01266;

[0087] δ2= |0.520-0.525| / 0.525≈0.00952;

[0088] δ3= |0.280-0.282| / 0.282≈0.00709;

[0089] δ4= |0.518-0.526| / 0.526≈0.01521;

[0090] δ5= |0.610-0.600| / 0.600≈0.01667;

[0091] δ6= |0.605-0.615| / 0.615≈0.01626;

[0092] Step 2: Calculate the reciprocal of the variance of each parameter and the sum.

[0093]

[0094] Step 3: Calculate the adaptive stability weight w k .

[0095] w1=15625 / 36063.36≈0.4332;

[0096] w2=4444.44 / 36063.36≈0.1232;

[0097] w3=10000 / 36063.36≈0.2773;

[0098] w4=3906.25 / 36063.36≈0.1083;

[0099] w5=1111.11 / 36063.36≈0.0308;

[0100] w6=976.56 / 36063.36≈0.0271;

[0101] Step 4: Calculate the weighted average relative error ∑w k δ k .

[0102] ∑w k δ k ≈(0.4332×0.01266)+(0.1232×0.00952)+(0.2773×0.00709)+(0.1083×0.01521)+(0.0308×0.01667)+(0.0271×0.01626);

[0103] ≈0.00548+0.00117+0.00196+0.00165+0.00051+0.00044=0.01121;

[0104] Step 5: Calculate the imbalance penalty term.

[0105] First, calculate δ k = 0.01121. k Next, calculate the weighted sum of the squares of the differences from the weighted average: ∑w k (δ j - ∑w j δ 2 ).

[0106] ≈0.4332(0.01266-0.01121) 2 +0.1232(0.00952-0.01121) 2 +...;

[0107] ≈9.05×10 -7 +3.53×10 -7 +4.69×10 -6 +1.74×10 -6 +9.24×10 -7 +6.64×10 -7 =9.3×10 -6 ;

[0108] Take the square root:

[0109] Step 6: Calculate the final core shape distortion value D″ z , setting λ = 1.5.

[0110] D″ z = 0.01121 + 1.5 × 0.00305 = 0.01121 + 0.004575 = 0.015785.

[0111] The result shows that the core shape distortion value of the current frame is 0.015785, which integrates the average deviation and the unevenness of the deviation between the current posture and the ideal template. A lower distortion value (such as close to 0) means that the current running posture is highly consistent with the personal best running posture, and the posture is stable. A higher value indicates that there is a significant posture deformation or incoordination, and the posture stability is poor.

[0112] The posture congruence score acquisition step is:

[0113] Based on the core shape distortion value, a set of core shape distortion values in the historical training samples is collected, and the 95th percentile value of the set in the distribution order is calculated, which is set as the posture stability reference threshold;

[0114] The ratio of the current frame core shape distortion value to the posture stability reference threshold is calculated to obtain the normalized distortion degree value;

[0115] Based on the normalized distortion degree value, the posture congruence score is calculated, and the calculation formula is:

[0116] F′ m = 100 x exp (-ln(2)·A r );

[0117] Where F′ m is the posture congruence score, which is used to measure the consistency of the current frame running posture with the ideal core shape, and A r is the normalized distortion degree value, which is equal to the ratio of the current frame core shape distortion value to the posture stability reference threshold.

[0118] Specifically, based on the core shape distortion value obtained in the previous calculation step, the system needs to establish an evaluation scale with practical significance for it. The establishment of this scale depends on the statistical analysis of the user's historical training data. Specifically, the system will access and retrieve the core shape distortion value calculated for each frame in all past training records of the user, and collect these values into a large data set, i.e. the core shape distortion value set. This set reflects the user's posture stability performance at different times and in different states. In order to extract a benchmark that represents the "excellent" and "needs to be improved" posture dividing line from it, the system performs statistical processing on the set. First, all the values in the set are sorted from small to large, and then the 95th percentile value is calculated. The 95th percentile is selected as the threshold because it represents that only 5% of the user's historical performance has a posture distortion exceeding this value. Therefore, it can be considered that the posture below this value belongs to the "normal" or "better" performance within the user's personal ability range, while the posture exceeding this value can be considered as a more obvious posture instability. The calculated 95th percentile value is set as the user's personal posture stability reference threshold.

[0119] After the personal posture stability reference threshold is determined, the system can normalize the core shape distortion value of the current frame. The purpose of this step is to eliminate individual differences in the absolute size of the distortion value and convert it into a relative index with universal comparison significance. The specific calculation process is very direct, that is, the core shape distortion value calculated for the current frame is divided by the posture stability reference threshold just calculated from the historical data. The result of this division operation is the normalized distortion degree value. For example, if the core shape distortion value of the current frame is 0.015785 and the posture stability reference threshold of the user obtained through historical data analysis is 0.025, then the normalized distortion degree value of the current frame is calculated as 0.015785 / 0.025 = 0.6314. This normalized value directly reflects how much the current posture distortion degree is compared to the personal “bad posture” benchmark. When the value is equal to 1, it means that the stability of the current posture is exactly at the level of the reference threshold. When the value is less than 1, it means that the current posture is better than the benchmark. When the value is greater than 1, it means that the stability of the current posture is significantly lower than the poor performance in the historical average level. The normalized distortion degree value is obtained.

[0120] Formula: F′ m = 100 * exp(-ln(2) * A r ), the advantage of the formula is to nonlinearly map a linear index measuring the degree of “bad” (the normalized distortion degree value A r ) into an intuitive scoring system with 100 as the full score of “good” (the posture full score F′ m ). By introducing the constant -ln(2) as a decay factor, when the normalized distortion degree value A r is equal to 1, that is, the current posture distortion degree exactly reaches the personal historical poor level reference threshold, the posture full score F′ m is exactly equal to 50 points. Defining the 95th percentile of personal historical performance as the “passing line” (50 points), this provides a clear, stable and personalized evaluation benchmark for the user. Scores higher than 50 points represent better than the personal “passing” level, and scores lower than 50 points represent room for improvement. In addition, the exponential function form ensures that when A r increases from 0, the rate of score decline from 100 points starts fast and then slows down. This means that for excellent postures (A r is small), even a small increase in distortion will cause a relatively significant score drop, thereby maintaining high sensitivity to subtle posture deterioration. For already poor postures (A r is large), the score decline will become flat, avoiding the cliff-like drop in score to negative or extremely low values due to extremely poor postures.

[0121] A r To normalize the distortion degree value, which is a dimensionless intermediate variable, the original core shape distortion value is standardized relative to the user's personal historical performance, so that the score has individual adaptability and comparability. The acquisition process has been described in detail. First, the system needs a set of historical core shape distortion values of the user, which is constructed and updated by continuously recording the core shape distortion value D" z calculated by the system for each frame of each training of the user. When scoring is needed, the system will sort this set and calculate its 95th percentile to obtain a dynamically updated posture stability reference threshold. Then, divide the core shape distortion value calculated in the current frame by this posture stability reference threshold to obtain A r , which directly reflects the position of the current posture stability in the user's personal ability spectrum. For example, by analyzing the training data of 500,000 frames of a user in the past month, it is calculated that the 95th percentile of the core shape distortion value is 0.025. This value becomes the user's current posture stability reference threshold. If the core shape distortion value of the current frame is 0.015785, the value of A r is 0.015785 / 0.025=0.6314.

[0122] Calculation process:

[0123] Here, the parameter values obtained in the previous steps will be used to calculate the final posture congruence score.

[0124] Given that the normalized distortion degree value A r is 0.6314.

[0125] Step 1: Substitute the value of A r into the formula.

[0126] F′ m =100×exp(-ln(2)·0.6314);

[0127] Step 2: Calculate the product of the exponential part.

[0128] -ln(2)≈-0.693147;

[0129] -ln(2)·0.6314≈-0.693147×0.6314≈-0.43765;

[0130] Step 3: Calculate the value of the natural exponential function exp.

[0131] exp(-0.43765)≈0.6455;

[0132] Step 4: Multiply by 100 to get the final posture congruence score.

[0133] F′ m = 100 x 0.6455 = 64.55;

[0134] The result shows that the pose congruence score of the current frame is 64.55, which is an intuitive evaluation between 0 and 100, which measures the consistency of the current running posture with the user's personal ideal core shape and the stability of the posture. The higher the score, the closer the posture to the personal best state, and the better the stability. According to the design of the formula, 50 is the "pass line" based on the user's historical data. Therefore, the result of 64.55 means that the running posture of the current frame is better than the average level of the user's historical performance (i.e. 95% of the time), and belongs to a good and stable posture category. This score can be fed back to the user in real time, helping them intuitively understand the quality of their posture at every moment, and adjust their running posture according to the score changes.

[0135] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application shall still fall within the protection scope of the present application.

Claims

1. A method for continuous capture and analysis of running posture for training, characterized in that, Includes the following steps: Based on the joint state vector of the previous frame and the single-frame image being processed, the joint position and velocity of the current frame are predicted, the prior motion posture vector is obtained, 2D key point localization is performed on the current single-frame image in parallel, the incomplete joint coordinate measurement values ​​due to limb swing occlusion are extracted, and the prior motion posture vector and joint coordinate measurement values ​​are combined to establish a multi-source posture observation set. Based on the multi-source attitude observation set, the prior motion attitude vectors are used as the predicted state, and the joint coordinate measurements are used as the observations for updating the state. The filtered and corrected joint states are obtained. For joints with missing observations, the state values ​​inherit the predicted state. For joints with existing observations, the state values ​​are fused with the predictions and observations. The filtered and corrected joint states of each frame are stitched together to generate a continuous temporal three-dimensional skeleton sequence. Based on the data of the first few gait cycles in the continuous temporal three-dimensional skeleton sequence, the coordinates of four points, namely the left shoulder, right shoulder, left hip and right hip, are selected to construct a core quadrilateral. The real-time geometric parameters of the core quadrilateral are obtained, and all the real-time geometric parameters of the core quadrilateral within the target gait cycle are calculated to establish an ideal core template for an individual. Based on the continuous temporal 3D skeleton sequence of each subsequent frame and the personal ideal core template, the core quadrilateral geometric parameters of the current frame are extracted and calculated item by item with the corresponding parameters of the personal ideal core template to obtain the core shape distortion value. The core shape distortion value is converted to quantify the stability of the current running posture and obtain the posture integrity score.

2. The running posture continuous capture and analysis method for training according to claim 1, characterized in that, The steps for obtaining the multi-source attitude observation set are as follows: Based on the joint state vector of the previous frame, the position and velocity values ​​of each joint in the joint state vector are parsed. Using the image feature information of the current single frame image, the parsed joint position and velocity values ​​are predicted and updated item by item to obtain the predicted joint position and velocity values ​​of the current frame, forming a priori motion posture vector. Based on the current single-frame image, perform 2D key point localization for each joint one by one, determine whether each joint is occluded due to limb movement, filter out the missing coordinates of occluded joints, and extract the joint coordinate values ​​that are not occluded and can be directly measured from the missing coordinates to obtain the joint coordinate measurement values. Based on the prior motion posture vector and the joint coordinate measurement values, the predicted values ​​of the corresponding joint positions that are not missing in the prior motion posture vector are mapped and merged one by one with the joint coordinate measurement values ​​to form a multi-source posture observation set.

3. The method for continuous capture and analysis of running posture for training according to claim 1, characterized in that, The steps for obtaining the filtered and corrected joint state are as follows: Based on the multi-source attitude observation set, the prior motion attitude vector and joint coordinate measurement values ​​contained in the multi-source attitude observation set are analyzed. The predicted values ​​of joint position and joint velocity in the prior motion attitude vector are defined as the predicted state, and the joint coordinate measurement values ​​are defined as the observation values ​​of the updated state, thus forming an initial state mapping relationship. Based on the initial state mapping relationship, it is determined whether the joint coordinate measurement value is missing for each joint. If it is determined to be missing, the predicted state of the corresponding joint is directly inherited as the filtering correction state of the joint. If it is determined not to be missing, the predicted state and the observed value are fused one by one to calculate the filtering correction state, thus completing the filtering correction of the joint state of a single frame and generating the filtering correction joint state.

4. The method for continuous capture and analysis of running posture for training according to claim 1, characterized in that, The steps for obtaining the continuous temporal three-dimensional skeleton sequence are as follows: Based on the filtered and corrected joint states generated in each frame, the data are stitched together frame by frame according to the acquisition order to establish a continuous temporal data sequence of three-dimensional joint states, thereby generating a continuous temporal three-dimensional skeleton sequence.

5. The method for continuous capture and analysis of running posture for training according to claim 1, characterized in that, The steps for obtaining the real-time core quadrilateral geometric parameters are as follows: Based on the continuous temporal three-dimensional skeleton sequence, the coordinates of the left shoulder, right shoulder, left hip, and right hip in each frame of the three-dimensional skeleton sequence within the initial several gait cycles are extracted. The coordinates of the above four joints are connected frame by frame to construct a core quadrilateral, generating the core quadrilateral for each frame. Based on the core quadrilateral in each frame, the distance between the coordinates of adjacent vertices of the core quadrilateral is calculated one by one to determine the side length of the four sides. The distance between the coordinates of opposite vertices in the core quadrilateral is calculated one by one to determine the length of the two diagonals. The side lengths of the four sides and the lengths of the two diagonals are combined to generate the real-time geometric parameters of the core quadrilateral.

6. The method for continuous capture and analysis of running posture for training according to claim 1, characterized in that, The steps to obtain the core template of personal ideals are as follows: Based on the real-time core quadrilateral geometric parameters, all real-time core quadrilateral geometric parameters are calculated in time series form within the target gait cycle to obtain the average value of each geometric parameter within the gait cycle. The average values ​​are then combined according to the geometric parameter type to generate a personal ideal core template.

7. The method for continuous capture and analysis of running posture for training according to claim 1, characterized in that, The steps for obtaining the core shape distortion value are as follows: Based on the continuous temporal 3D skeleton sequence of each subsequent frame, the coordinates of the left shoulder, right shoulder, left hip, and right hip are extracted in sequence to construct the core quadrilateral structure of each frame. The lengths of the four sides and the lengths of the two diagonals are calculated respectively to generate the core quadrilateral geometric parameters of the current frame. Based on the core quadrilateral geometric parameters of the current frame, a relative error is calculated with each corresponding parameter value in the personal ideal core template. At the same time, the standard deviation of all geometric parameters within the ideal gait cycle is extracted from the personal ideal core template as a volatility index, and a set of relative errors and standard deviations of geometric parameters are constructed. The core shape distortion value is calculated based on the set of relative errors and standard deviations of the geometric parameters.

8. The method for continuous capture and analysis of running posture for training according to claim 1, characterized in that, The steps for obtaining the attitude congruence score are as follows: Based on the core shape distortion value, a set of core shape distortion values ​​in historical training samples is collected, and the 95th percentile value of the core shape distortion value set in the distribution ranking is calculated and set as the attitude stability reference threshold. Calculate the ratio of the core shape distortion value of the current frame to the attitude stability reference threshold to obtain the normalized distortion value; Based on the normalized distortion value, calculate the pose congruence score.

Images