A method for detecting a human posture

Abstract

The application discloses a human posture detection method, which considers physiological responses of a target to be detected, so that the human posture is more accurately detected. The detection method comprises the following steps: S1, collecting human posture data and vital sign data of a subject, and preparing posture-vital sign synchronous time series data set; S2, obtaining a body shaking score distribution model of a subject with normal human posture, obtaining a shaking abnormal score range from the body shaking score distribution model, training a head posture-sign-classification model, and training a fall detection model; S3, obtaining a pre-judgment model; S4, obtaining a pre-judgment model threshold, a similarity recognition model threshold and a human posture recognition model threshold; S5, collecting respiratory rate, heart rate vital sign time series data, human feature point position and head Euler angle human posture time series data of a human body to be detected, and obtaining a total score; comparing the total score with the human posture recognition model threshold, and obtaining human posture data of the human body to be detected.

Description

Technical Field

[0001] This invention relates to the field of human posture, and more specifically, to a method for detecting human posture. Background Technology

[0002] There are various existing methods for human posture detection. A common method involves capturing full-body, slightly occluded video data of the target using a camera, identifying the locations of human feature points using deep neural networks, and analyzing the human posture based on the coordinates of these feature points. However, this method only considers the physical characteristics of the target's motion and does not take into account its physiological responses. Summary of the Invention

[0003] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method for detecting human posture that takes into account the physiological response of the target and makes the detection of human posture more accurate.

[0004] To solve the above-mentioned technical problems, the embodiments of the present invention adopt the following technical solutions:

[0005] A method for detecting human posture, comprising:

[0006] S1. Collect human posture data and vital sign data of the subjects, create a posture-vital sign synchronous time series dataset, and perform action annotation and calculate body sway score on the dataset; the subjects include subjects with normal human posture and subjects with abnormal human posture.

[0007] S2, obtain the body sway score distribution model of the subject in normal human posture, obtain the abnormal sway score range from the body sway score distribution model, train the head posture-sign classification model, and train the fall detection model.

[0008] S3, based on the body sway fraction distribution model, head posture-sign classification model and fall detection model of S2, obtain the pre-judgment model σ(t) as shown in equation (1):

[0009]

[0010] In the formula, C1 represents the score obtained by the fall detection model, C2 represents the score obtained by the body sway score distribution model, and C3 represents the score obtained by the head posture-sign classification model.

[0011] Based on equation (1), the similarity recognition model Y(t) shown in equation (2) is obtained:

[0012] Y(t)=α·S tim (t)+β·S spa (t) Equation (2)

[0013] In the formula, S tim (t) represents the time similarity score, S spa (t) represents the spatial similarity score, and α represents S. tim The weighting coefficients of (t), β represents S sp The weighting coefficients of (t);

[0014] Based on equations (1) and (2), the human pose recognition model S shown in equation (3) is obtained:

[0015]

[0016] In the formula, w(t2-t1) is the time weight function, whose input is the duration, t1 is the start time of the duration, t2 is the end time of the duration, σ(t) is the pre-judgment model score at a certain time, and Y(t) is the similarity score calculated by the similarity recognition model at the corresponding time.

[0017] S4, obtain the threshold of the pre-judgment model, the threshold of the similarity recognition model, and the threshold of the human pose recognition model;

[0018] S5. Collect the respiratory rate, heart rate, vital signs time series data, human feature point positions, and head Euler angle human posture time series data of the human body to be tested, and substitute them into equation (3) to obtain the total score; compare the total score with the threshold of the human posture recognition model to obtain the human posture data to be tested.

[0019] Preferably, in step S1, the process of creating a posture-vital signs synchronized time series dataset is as follows: the two time series data of subject human posture data and vital signs data are aligned with the same sampling frequency, interpolation and timestamp to obtain a posture-vital signs synchronized time series dataset. The variables in the posture-vital signs synchronized time series dataset are heart rate, respiratory rate, human feature point position and head Euler angle.

[0020] Preferably, in step S4, obtaining the threshold of the pre-judgment model includes: substituting the input sample into equation (1) to obtain the score, and statistically analyzing the mean and variance of the sample set, then the threshold of the pre-judgment model is... Where μ is the mean, σ is the standard deviation, and n is the sample size;

[0021] Obtaining the threshold of the similarity recognition model includes: substituting the input sample into equation (2) to obtain the score, and calculating the mean and variance of the sample set. Then, the threshold of the pre-judgment model is... Where μ is the mean, σ is the standard deviation, and n is the sample size;

[0022] To obtain the threshold of the human pose recognition model, the following steps are taken: substituting the input sample into equation (3) to obtain the score, and statistically analyzing the mean and variance of the sample set. Then, the threshold of the pre-judgment model is: Where μ is the mean, σ is the standard deviation, and n is the sample size.

[0023] Preferably, S5 specifically includes:

[0024] S51: Real-time acquisition of respiratory rate, heart rate, vital signs time series data, body feature point positions, and head Euler angle human posture time series data of the human body under test; and data preprocessing by stacking the data sequences; the time series is segmented according to the head posture in the preprocessed data to form time periods; the initial and final head postures of each time period are compared with abnormal head postures to filter the time periods with high similarity.

[0025] S52, substitute the data of each time period selected in S51 into formula (1) to obtain the pre-judgment model score; according to the pre-judgment model score, determine whether to perform similarity recognition. If yes, proceed to step S53; otherwise, the posture of the human body to be tested is normal.

[0026] S53, the time series data of the human body to be tested obtained in S52 is reordered and deduplicated according to the head posture to obtain a new sequence with the head posture data as the index and the coordinates of the body feature points and vital signs as variables. The above time series and the new sequence are compared with the abnormal posture data to identify the sequence similarity. The two similarity scores are substituted into formula (2) to obtain the similarity identification model score.

[0027] S54, count the duration for which the score of the pre-judgment model is greater than its model threshold, and substitute it into equation (3) to calculate the human posture recognition model score and determine the human posture.

[0028] Preferably, in step S52, determining whether to perform similarity recognition based on the pre-judgment model score includes:

[0029] The score of the pre-judgment model is compared with the threshold of the pre-judgment model. If the score of the pre-judgment model is greater than the threshold of the pre-judgment model, then similarity recognition is performed.

[0030] Preferably, in step S54, determining the human posture includes:

[0031] The human posture recognition model score calculated by S54 is compared with the human posture recognition model threshold. If the human posture recognition model score is greater than the human posture recognition model threshold, the human posture is abnormal. If the human posture recognition model score is less than or equal to the human posture recognition model threshold, the human posture is normal.

[0032] Preferably, in step S51, the time sequence is segmented based on the head pose in the preprocessed data, including:

[0033] The time series is divided into time series corresponding to various head movements based on head posture: a linear piecewise fitting algorithm is applied to the three-dimensional time series of head Euler angles to obtain three fitted polylines and the coordinates of the inflection points of the polylines. The x-coordinates of the time points corresponding to all inflection points are integrated into a new list, and duplicate values ​​are removed and sorted. The time points in the new list are used to divide the time periods of head movements.

[0034] Preferably, in step S3, the human pose recognition model...

[0035]

[0036] In the formula, x represents the duration, a represents the upper limit of the duration of the abnormal human posture, and b represents the lower limit of the duration of the abnormal human posture.

[0037] Compared with existing technologies, the present invention has the following beneficial effects: the detection method of the present invention considers the physiological response of the target, thereby making the detection of human posture more accurate. The detection method of the present invention includes: S1, collecting human posture data and vital sign data of the subject, creating a posture-vital sign synchronous time-series dataset, and performing action annotation and calculating body sway score on the dataset; S2, obtaining a body sway score distribution model of the subject with normal human posture, obtaining the range of abnormal sway scores from the body sway score distribution model, training a head posture-vital sign classification model, and training a fall detection model; S3, obtaining a pre-judgment model based on the body sway score distribution model, head posture-vital sign classification model, and fall detection model in S2; S4, obtaining the threshold of the pre-judgment model, the threshold of the similarity recognition model, and the threshold of the human posture recognition model; S5, collecting the respiratory rate, heart rate, vital sign time-series data, human feature point positions, and head Euler angle human posture time-series data of the subject, and obtaining a total score; comparing the total score with the threshold of the human posture recognition model to obtain the human posture data to be tested. This invention captures changes in the subject's body posture and, based on the subject's physiological responses, can more accurately detect the subject's posture during free movement. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0039] The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. For those skilled in the art, several changes and modifications can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. The endpoints and any values ​​of the ranges disclosed herein are not limited to the precise ranges or values; these ranges or values ​​should be understood to include values ​​close to these ranges or values. For numerical ranges, the endpoint values ​​of various ranges, the endpoint values ​​of various ranges and individual point values, and individual point values ​​can be combined with each other to obtain one or more new numerical ranges, which should be considered as specifically disclosed herein.

[0040] The present invention will now be described in detail with reference to specific embodiments.

[0041] like Figure 1 As shown, a method for detecting human posture in this embodiment includes:

[0042] Step S1: Collect human posture data and vital sign data of the subjects, create a posture-vital sign synchronous time series dataset, and perform action annotation and calculate body sway score on the dataset; the subjects include subjects with normal human posture and subjects with abnormal human posture.

[0043] In step S1, the process of creating the posture-vital signs synchronized time-series dataset is as follows: The two time-series data sets, namely the subject's human posture data and vital signs data, are aligned using the same sampling frequency, interpolation, and timestamps to obtain the posture-vital signs synchronized time-series dataset. The variables in the posture-vital signs synchronized time-series dataset are heart rate, respiratory rate, human feature point positions, and head Euler angles.

[0044] In step S1, action labeling of the dataset includes: using existing time-series labeling tools, by examining actions in videos saved by video sensors, labeling the corresponding time points of the above time series with action categories, including falling, swaying, walking, etc., and using a specific integer to represent a unique action category. The final time series includes action label variables.

[0045] In step S1, the method for calculating the body sway score is as follows: the body sway score is obtained by accumulating the Euclidean distance between the normalized human body feature point position coordinates of the subject in the preceding and following frames.

[0046] Step S2: Obtain the body sway score distribution model of the subject in normal human posture, obtain the abnormal sway score range from the body sway score distribution model, train the head posture-sign classification model, and train the fall detection model.

[0047] In step S2, the head pose-sign classification model is shown in the following equation:

[0048]

[0049] st(w T φ(x i ))>ρ-ζ i

[0050] ζ i >0

[0051] In the formula: w and ρ are hyperplane parameters, where w is the normal vector of the hyperplane and ρ is a bias term used to represent the hyperplane w. T x+ρ=0, such that all positive samples lie on one side of the hyperplane; ζ i ν is a slack variable, a penalty term to prevent overfitting; ν is a regularization parameter, representing the proportion of samples not in the target class in the training set; n is the number of samples; φ is the sample x. i A mapping function that maps to a high-dimensional space. For the test sample x... test When the decision function w T φ(x test When -ρ ≥ 0, the sample is judged to be normal; otherwise, the decision function is negative, and the sample is judged to be abnormal. i = 1, ..., n.

[0052] The training process for the body sway fraction distribution model includes the following steps:

[0053] 1. Data preparation: Collect data on the location of body feature points of the test subject in a normal state, and calculate the sway score using the sway score formula.

[0054] 2. Data statistics: Statistically calculate the fractional probability density and cumulative probability of the shaking, fit the fitted cumulative probability function, and use it as the model output.

[0055] The training process for the head pose-sign classification model includes the following steps:

[0056] 1. Data Preparation: Collect a training set of head Euler readings, respiratory rate, and heart rate data from test subjects in normal conditions. Ensure the data is accurate, well-labeled, and free of anomalies.

[0057] 2. Feature engineering: Selecting and preprocessing features from data, including standardization or normalization and handling missing values.

[0058] 3. Model selection: Choose an appropriate multivariate classification algorithm, such as Support Vector Machine (SVM) or Isolation Forest.

[0059] 4. Model Training: Train the selected single-classification model using the training data. Input the training data into the single-classification model training program and iterate a specified number of times.

[0060] 5. Model Evaluation: Use methods such as validation sets or cross-validation to evaluate the accuracy of the model.

[0061] 6. Parameter tuning: Adjust the hyperparameters of the model based on the evaluation results to improve the model's performance.

[0062] Step S3: Based on the body sway fraction distribution model, head posture-sign classification model, and fall detection model from step S2, obtain the pre-judgment model σ(t) as shown in equation (1):

[0063]

[0064] In the formula, C1 represents the score obtained by the fall detection model, C2 represents the score obtained by the body sway score distribution model, and C3 represents the score obtained by the head posture-sign classification model.

[0065] Based on equation (1), the similarity recognition model Y(t) shown in equation (2) is obtained:

[0066] Y(t)=α·S tim (t)+β·S sp (t) Equation (2)

[0067] In the formula, S tim (t) represents the time similarity score, S spa (t) represents the spatial similarity score, and α represents S. tim The weighting coefficients of (t), β represents S spa The weighting coefficients of (t).

[0068] Time series whose scores exceed the threshold of the pre-judgment model are selected, reordered and deduplicated according to head pose, resulting in new sequences with head pose data as the index and human feature point coordinates and vital signs as variables. The time series and the new sequences are then compared with abnormal pose data to obtain a time similarity score S. tim (t) and spatial similarity score S spa (t).

[0069] Based on equations (1) and (2), the human pose recognition model S shown in equation (3) is obtained:

[0070]

[0071] In the formula, w(t2-t1) is the time weight function, whose input is the duration, t1 is the start time of the duration, t2 is the end time of the duration, σ(t) is the pre-judgment model score at a certain time, and Y(t) is the similarity score calculated by the similarity recognition model at the corresponding time. The time weight function satisfies equation (4):

[0072]

[0073] In the formula, x represents the duration, a represents the upper limit of the duration of the abnormal human posture, and b represents the lower limit of the duration of the abnormal human posture.

[0074] Step S4: Obtain the threshold values ​​for the pre-judgment model, the similarity recognition model, and the human pose recognition model.

[0075] In step S4, the threshold of the pre-judgment model is obtained, including: using the training data after head action screening as input samples, substituting them into equation (1) to obtain the score, and statistically analyzing the mean and variance of the sample set, then the threshold of the pre-judgment model is... Where μ is the mean, σ is the standard deviation, and n is the sample size.

[0076] Obtaining the threshold of the similarity recognition model includes: substituting the input sample into equation (2) to obtain the score, and calculating the mean and variance of the sample set. Then, the threshold of the pre-judgment model is... Where μ is the mean, σ is the standard deviation, and n is the sample size;

[0077] To obtain the threshold of the human pose recognition model, the following steps are taken: substituting the input sample into equation (3) to obtain the score, and statistically analyzing the mean and variance of the sample set. Then, the threshold of the pre-judgment model is: Where μ is the mean, σ is the standard deviation, and n is the sample size.

[0078] Step S5: Collect the respiratory rate, heart rate, vital signs time series data, human feature point positions, and head Euler angle human posture time series data of the human body to be tested, substitute them into equation (3) to obtain the total score; compare the total score with the threshold of the human posture recognition model to obtain the human posture data to be tested.

[0079] Step S5 specifically includes steps S51-S54:

[0080] Step S51: Real-time acquisition of respiratory rate, heart rate, vital signs time-series data, body feature point locations, and head Euler angle human posture time-series data of the subject. The data is then stacked and preprocessed. Data preprocessing utilizes existing technologies, including missing value completion and normalization. The time series is segmented based on the head posture in the preprocessed data, forming various time periods. Each time period is then filtered based on whether its initial and final head postures match abnormal head postures. The similarity between each time period and the initial and final head postures of the abnormal head posture data is compared, and time periods with high similarity are selected. Similarity refers to the similarity between the segmented head movements and the head movements that induce dizziness. This is evaluated using Euler angle data representing head posture. A high similarity is considered when the similarity exceeds a similarity comparison threshold.

[0081] In step S51, the time series is segmented according to the head pose in the preprocessed data, including: dividing the time series into time series corresponding to each head action according to the head pose; applying a linear piecewise fitting algorithm to the three-dimensional time series of head Euler angles to obtain three fitted polylines and the coordinates of the inflection points of the polylines; integrating the abscissas of all inflection points into a new list, removing duplicate values ​​and sorting them; and using the time points in the new list to divide the head action time period.

[0082] Step S52: Substitute the data from each time period selected in step S51 into formula (1) to obtain the pre-judgment model score; based on the pre-judgment model score, determine whether to perform similarity recognition. If yes, proceed to step S53; otherwise, the posture of the human body to be tested is normal.

[0083] In step S52, determining whether to perform similarity recognition based on the pre-judgment model score includes: comparing the pre-judgment model score with the pre-judgment model threshold; if the pre-judgment model score is greater than the pre-judgment model threshold, then performing similarity recognition.

[0084] Step S53: The time series data of the human body to be tested obtained in step S52 is reordered and deduplicated according to head posture to obtain a new sequence with head posture data as index and body feature point position coordinates and vital signs as variables. The above time series and the new sequence are compared with the abnormal posture data for sequence similarity identification. The two similarity scores are substituted into formula (2) to obtain the similarity identification model score.

[0085] Step S54: Calculate the duration during which the score of the pre-judgment model is greater than its model threshold, and substitute it into equation (3) to calculate the human posture recognition model score and determine the human posture.

[0086] In step S54, determining the human posture includes: comparing the human posture recognition model score calculated in step S54 with the human posture recognition model threshold; if the human posture recognition model score is greater than the human posture recognition model threshold, the human posture is abnormal; if the human posture recognition model score is less than or equal to the human posture recognition model threshold, the human posture is normal.

[0087] Compared to existing posture detection and fall detection methods, the method of this invention adds analysis of respiratory rate and heart rate, thereby detecting human posture during sudden dizziness and predicting the fall trend of the test target from more dimensions. An example is given below.

[0088] A method for detecting human posture, including

[0089] First, the posture detection system is activated to acquire time-series data of the breathing rate and heart rate of the healthy target, as well as time-series data of body feature point coordinates and head Euler angles.

[0090] Secondly, linear interpolation was used to unify the time axis of time-series data acquired from different sensors, and the time series were stacked to obtain time-series sequences with respiratory rate, heart rate, body feature point coordinates, and head Euler angles as variables. Subsequently, piecewise linear fitting was applied to the head Euler angle time series (three-axis rotation angle time series), and the set of x-coordinates of the inflection points of each rotation angle time series after fitting them into a polyline was extracted. The three sets of x-coordinates were combined, deduplicated, and sorted to obtain the time points of head movements. By analyzing the head Euler angles at the beginning and end of each time period and comparing head movements with vestibular function postural tests, the time periods that induce balance perception deficits were identified, and the time index of the time series was filtered accordingly.

[0091] Time segments are filtered by the change in Euler angles between the front and back of the head, i.e., the change in Euler angles at the beginning and end of each of the above time periods is calculated.

[0092] 1. Body swaying: This was obtained from a large number of samples.

[0093] Formula for the fractional body sway C1(t):

[0094]

[0095] Where i represents a human feature point. d i,t This represents the Euclidean distance between human feature point i at time t and the previous time.

[0096]

[0097] Where, x i,t The x-coordinate of human feature point i at time t; y i,tThe z-axis represents the y-coordinate of human feature point i at time t; i,t The z-axis coordinate of human feature point i at time t; x i,t-1 The x-coordinate of human feature point i at time t-1 is represented by y; i,t-1 The z-axis represents the y-coordinate of human feature point i at time t-1; i,t-1 This represents the z-axis coordinate of human feature point i at time t-1.

[0098] In this embodiment, the shaking score is calculated for a large number of samples, and its 95th percentile is 744.38.

[0099] 2. Fall detection

[0100] The confidence level of the target falling is calculated by using a fall detection (action detection) model: the body posture data of the above time period is input into the model to obtain the output of the corresponding neurons for each action, and the fall confidence level is calculated by using the softmax function.

[0101]

[0102] Among them, z c z represents the output of the neurons corresponding to each action in the action detection model. fall This is the output of the neuron corresponding to the fall action in the action detection model.

[0103] In this example, the calculated fall confidence level is 0.02%.

[0104] 3. Anomaly Detection

[0105] The extracted time-segment data from normal subjects are used as samples and input into the one-class SVM model for training. The model detects data that differs too much from the training set data and outputs 1, while outputting 0 for data that is similar to the training set data.

[0106] 4. Setting thresholds for the pre-judgment model

[0107] The normal subject sample data extracted above were used to calculate the shaking score, quantile, fall confidence, and anomaly detection, and then substituted into formula (1):

[0108]

[0109] Substituting a large sample of normal subjects into formula (1) yields the σ statistical distribution, with a mean of 0.683, a variance of 0.114, and a sample size of 1000. The threshold for the pre-judgment model is set as follows:

[0110] 4. Similarity calculation:

[0111] The time periods in which the output score of the pre-judgment model is greater than its model threshold are selected. The selected time series are then reordered according to the sum of Euler angles, and the sum of Euler angles is set as an index. The index is then resampled to obtain the sequence of vital signs with respect to changes in head posture.

[0112] The original time series and the reordered sequence were compared with the abnormal data sequence of vertigo patients. The time similarity was calculated using dynamic time warping and cross-correlation algorithms.

[0113] This example yields the time similarity S. tim (t) = -3260.6, S spa (t) = 388074.

[0114] With the time similarity weight α = 0.1 and the spatial similarity weight β = 0.001, the similarity recognition model score is Y = 0.1 × (-3260.6) + 0.001 × 388074 = 62.014.

[0115] For samples whose pre-judgment scores are greater than the pre-judgment model threshold, input them into the similarity model according to the time period to obtain similarity scores, and substitute them into formula (3) to calculate the total score.

[0116] The total score of 1000 samples was calculated, and its mean was found to be 38.360, with a variance of 18.804. Therefore, we set... The threshold is set, and if the value exceeds the threshold, the human posture is judged as a hidden fall.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting human posture, characterized in that, include: S1. Collect human posture data and vital sign data of the subjects, create a posture-vital sign synchronous time series dataset, and perform action annotation and calculate body sway score on the dataset; the subjects include subjects with normal human posture and subjects with abnormal human posture. S2, obtain the body sway score distribution model of the subject in normal human posture, obtain the abnormal sway score range from the body sway score distribution model, train the head posture-sign classification model, and train the fall detection model. S3, based on the body sway fraction distribution model, head posture-sign classification model and fall detection model of S2, obtain the pre-judgment model σ(t) as shown in equation (1): In the formula, C1 represents the score obtained by the fall detection model, C2 represents the score obtained by the body sway score distribution model, and C3 represents the score obtained by the head posture-sign classification model. Based on equation (1), the similarity recognition model Y(t) shown in equation (2) is obtained: Y(t) = α·S tim (t) + β·S spa (t) Equation (2) In the formula, S tim (t) represents the time similarity score, S spa (t) represents the spatial similarity score, and α represents S. tim The weighting coefficients of (t), β represents S spa The weighting coefficients of (t); Based on equations (1) and (2), the human pose recognition model S shown in equation (3) is obtained: In the formula, w(t2-t1) is the time weight function, whose input is the duration, t1 is the start time of the duration, t2 is the end time of the duration, σ(t) is the pre-judgment model score at a certain time, and Y(t) is the similarity score calculated by the similarity recognition model at the corresponding time. S4, obtain the threshold of the pre-judgment model, the threshold of the similarity recognition model, and the threshold of the human pose recognition model; S5. Collect the respiratory rate, heart rate, vital signs time series data, human feature point positions, and head Euler angle human posture time series data of the human body to be tested, and substitute them into equation (3) to obtain the total score; compare the total score with the threshold of the human posture recognition model to obtain the human posture data to be tested.

2. The method according to claim 1, characterized in that, In S1, the process of creating a posture-vital signs synchronized time series dataset is as follows: the two time series data of subject human posture data and vital signs data are uniformly sampled, interpolated, and timestamped to obtain a posture-vital signs synchronized time series dataset. The variables in the posture-vital signs synchronized time series dataset are heart rate, respiratory rate, human feature point position, and head Euler angle.

3. The method according to claim 1, characterized in that, In step S4, obtaining the threshold of the pre-judgment model includes: substituting the input sample into equation (1) to obtain the score, and calculating the mean and variance of the sample set. Then, the threshold of the pre-judgment model is... Where μ is the mean, σ is the standard deviation, and n is the sample size; Obtaining the threshold of the similarity recognition model includes: substituting the input sample into equation (2) to obtain the score, and calculating the mean and variance of the sample set. Then, the threshold of the pre-judgment model is... Where μ is the mean, σ is the standard deviation, and n is the sample size; To obtain the threshold of the human pose recognition model, the following steps are taken: substituting the input sample into equation (3) to obtain the score, and statistically analyzing the mean and variance of the sample set. Then, the threshold of the pre-judgment model is: Where μ is the mean, σ is the standard deviation, and n is the sample size.

4. The method according to claim 1, characterized in that, S5 specifically includes: S51: Real-time acquisition of respiratory rate, heart rate, vital signs time series data, body feature point positions, and head Euler angle human posture time series data of the human body under test; and data preprocessing by stacking the data sequences; the time series is segmented according to the head posture in the preprocessed data to form time periods; the initial and final head postures of each time period are compared with abnormal head postures to filter the time periods with high similarity. S52, substitute the data of each time period selected in S51 into formula (1) to obtain the pre-judgment model score; according to the pre-judgment model score, determine whether to perform similarity recognition. If yes, proceed to step S53; otherwise, the posture of the human body to be tested is normal. S53, the time series data of the human body to be tested obtained in S52 is reordered and deduplicated according to the head posture to obtain a new sequence with the head posture data as the index and the coordinates of the body feature points and vital signs as variables. The above time series and the new sequence are compared with the abnormal posture data to identify the sequence similarity. The two similarity scores are substituted into formula (2) to obtain the similarity identification model score. S54, count the duration for which the score of the pre-judgment model is greater than its model threshold, and substitute it into equation (3) to calculate the human posture recognition model score and determine the human posture.

5. The method according to claim 4, characterized in that, In step S52, determining whether to perform similarity recognition based on the pre-judgment model score includes: The score of the pre-judgment model is compared with the threshold of the pre-judgment model. If the score of the pre-judgment model is greater than the threshold of the pre-judgment model, then similarity recognition is performed.

6. The method according to claim 4, characterized in that, In step S54, determining the human posture includes: The human posture recognition model score calculated by S54 is compared with the human posture recognition model threshold. If the human posture recognition model score is greater than the human posture recognition model threshold, the human posture is abnormal. If the human posture recognition model score is less than or equal to the human posture recognition model threshold, the human posture is normal.

7. The method according to claim 4, characterized in that, In step S51, the time sequence is segmented based on the head pose in the preprocessed data, including: The time series is divided into time series corresponding to various head movements based on head posture: a linear piecewise fitting algorithm is applied to the three-dimensional time series of head Euler angles to obtain three fitted polylines and the coordinates of the inflection points of the polylines. The x-coordinates of the time points corresponding to all inflection points are integrated into a new list, and duplicate values ​​are removed and sorted. The time points in the new list are used to divide the time periods of head movements.

8. The method according to claim 1, characterized in that, In S3, in the human posture recognition model, In the formula, x represents the duration, a represents the upper limit of the duration of the abnormal human posture, and b represents the lower limit of the duration of the abnormal human posture.

Images