A method for determining the direction of a fall of a human body

Abstract

A method for judging the falling direction of a human body, comprising the following steps: 1) establishing a falling monitoring database; 2) collecting real-time angular velocity and real-time acceleration of the human body as sample data and storing the sample data in the falling monitoring database; 3) performing feature extraction to form an original data feature space; 4) performing data standardization processing on all feature vectors in the original data feature space to obtain a standard data feature space; 5) performing dimension reduction processing on the standard data feature space, processing the standard data feature space after the dimension reduction processing by using a threshold method, screening out sample data meeting the requirements, and forming a falling data set; 6) judging whether the human body is in a falling state by using the falling data set through an XGBoost model, and when the human body is in the falling state, sample data screened out by the XGBoost model forming a falling data subset for judging the falling direction of the human body; and 7) judging the falling direction of the human body by using the acceleration components of the X axis and the Z axis of each sample data in the falling data subset.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring, and more specifically to a method for determining the direction of a human fall. Background Technology

[0002] Elderly people are a high-risk group for falls. Falls are associated with high hospitalization and mortality rates, placing a significant financial burden on families. If treatment is delayed, falls can directly lead to injuries such as fractures and paralysis, and may also indirectly trigger other health problems. As falls increasingly become a key factor threatening the quality of life and safety of the elderly, the prevention and protection of falls among the elderly are receiving increasing attention. Given the sudden and random nature of falls, the direction of the fall has become a key focus of research.

[0003] Large datasets are increasingly used in studies of human fall orientation. To interpret such datasets, their dimensionality needs to be significantly reduced in an interpretable way to retain most of the information. Many techniques have been developed for this purpose, with Principal Component Analysis (PCA) being one of the oldest and most widely used. Its idea is to reduce the dimensionality of the dataset while preserving as much "variability" as possible; preserving variability can be understood as allowing for a large variance in the data. In PCA, the dimensionality-reduced features need to reflect the information contained in the original data as much as possible, and these features should be as independent as possible. PCA extracts the most valuable information based on variance and ensures that the dimensionality-reduced features are linearly independent. PCA can be viewed as a parameter-free technique, meaning that without data cleaning, everyone will get the exact same dimensionality reduction result. While PCA is easy to implement universally due to the absence of subjective parameters, it also has the drawback of being unable to be personalized. Furthermore, using PCA for dimensionality reduction requires ensuring that the data follows a Gaussian distribution; otherwise, the obtained principal features may not be optimal.

[0004] The growth and update speed of datasets is accelerating, and data is evolving towards higher dimensions and unstructured formats. While massive amounts of complex data contain a wealth of useful information, they also increase the difficulty of effectively utilizing it. As data dimensionality increases, the computational time and storage space spent on data processing also rise. Effective information is often buried within complex data, making it difficult to discover the essential characteristics of the data. Processing this data requires significant time and manpower, greatly reducing the efficiency and accuracy of determining the direction of a fall. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for determining the direction of a human fall. This method performs dimensionality reduction processing on the feature space formed by real-time collected human body data, and achieves "low loss" in the feature dimensionality reduction process, preserving the properties of the original data and obtaining optimal low-dimensional data to ensure the accuracy of determining the direction of a human fall.

[0006] The objective of this invention is achieved through the following approach:

[0007] A method for determining the direction of a fall includes the following steps:

[0008] 1) Establish a fall monitoring database;

[0009] 2) An inertial sensor is installed on the waist of the human body to collect the real-time angular velocity and real-time acceleration of the human body as sample data and store them in the fall monitoring database;

[0010] The acceleration includes the acceleration components of the human body on the X, Y, and Z axes of the three-dimensional coordinate system, as well as the acceleration vector composed of these three acceleration components;

[0011] 3) Extract features from all peak segments of each sample data to obtain the feature vector corresponding to each sample data, forming the original data feature space;

[0012] 4) Standardize all feature vectors in the original data feature space to obtain the standard data feature space;

[0013] 5) Dimensionality reduction is performed on the standard data feature space, and the threshold method is used to process the dimensionality-reduced standard data feature space to select sample data that meet the requirements and form a fall dataset.

[0014] 6) The fall dataset is used to determine whether a person is in a fall state through the XGBoost model. When a person is in a fall state, the sample data selected by the XGBoost model forms a subset of fall data used to determine the direction of the fall.

[0015] 7) Determine the direction of a fall by analyzing the acceleration components of the X and Z axes of each sample in the fall data subset.

[0016] Preferably, the specific steps for dimensionality reduction processing of the standard data feature space are as follows:

[0017] (1) The XGBoost model was trained using a public dataset to obtain the XGBoost dimensionality reduction model, and a minimum threshold for feature dimensions was set.

[0018] (2) Use the standard data feature space as the current feature space;

[0019] (3) Use the XGBoost dimensionality reduction model to obtain the weights of each feature vector in the current feature space, which are used as the feature importance of each feature vector in the current feature space;

[0020] (4) Sort the corresponding feature vectors in the current feature space in descending order according to their importance.

[0021] (5) Delete the last feature vector in the current feature space to form a new feature space to be processed;

[0022] (6) Repeat step (5) until the dimension of the feature space to be processed is equal to the minimum threshold of the feature dimension, so as to obtain multiple feature spaces to be processed with different numbers of feature vectors, and calculate the accuracy of each feature space to be processed, and take the feature space with the highest accuracy as the optimal feature space.

[0023] (7) Compare the accuracy of the optimal feature space with that of the current feature space:

[0024] a) If the accuracy of the optimal feature space is greater than the accuracy of the current feature space, then use the optimal feature space as the new current feature space and repeat steps (3) to (7).

[0025] b) If the accuracy of the optimal feature space is less than or equal to the accuracy of the current feature space, then the current feature space is used as the standard data feature space after dimensionality reduction.

[0026] Preferably, the feature importance is the number of times the feature vector is used in all trees of the XGBoost dimensionality reduction model.

[0027] Preferably, the standardization method is used as a dimensionless means according to the following formula to perform data standardization processing on all feature vectors in the original data feature space:

[0028]

[0029] In the formula, mean is the mean of the eigenvectors, σ is the standard deviation, x is the eigenvector, and X is the standardized eigenvector.

[0030] Preferably, the fall dataset is obtained by setting a first judgment threshold and a second judgment threshold, and then processing the standard data feature space after dimensionality reduction using a thresholding method. The specific steps are as follows:

[0031] I. If the first time-domain feature value in the feature vector is less than or equal to the first judgment threshold, then the sample data corresponding to the feature vector does not meet the requirements;

[0032] II. When the first time-domain feature value in the feature vector is greater than the first judgment threshold, if the second time-domain feature value in the feature vector is less than or equal to the second judgment threshold, then the sample data corresponding to the feature vector does not meet the requirements; if the second time-domain feature value in the feature vector is greater than the second judgment threshold, then the sample data corresponding to the feature vector meets the requirements.

[0033] III. Repeat steps I to II to evaluate all feature vectors in the feature space of the standard data after dimensionality reduction, and select the sample data that meets the requirements to form the fall dataset.

[0034] Preferably, the angular velocity includes the angular velocity components of the human body on the X, Y, and Z axes of the three-dimensional coordinate system, and the angular velocity vector composed of these three angular velocity components;

[0035] The standard data feature space after dimensionality reduction includes the area of ​​angular velocity signal amplitude, the area of ​​acceleration signal amplitude, the cumulative change of angular velocity, the maximum value of the Y-axis acceleration component, the mean value of the Z-axis acceleration component, the standard deviation of the X-axis acceleration component, the numerical range of the Z-axis angular velocity component, the mean value of the Z-axis angular velocity component, the minimum value of the Z-axis acceleration component, the mean value of the X-axis angular velocity component, the average rate of change of angular velocity, the maximum value of acceleration, the minimum value of the X-axis angular velocity component, the standard deviation of the X-axis angular velocity component, and the mean value of acceleration.

[0036] Preferably, the specific steps for determining the direction of a fall by analyzing the acceleration components along the X and Z axes in each sample data set of the fall data subset are as follows:

[0037] A) Determine the relationship between the acceleration components of the X-axis and Z-axis in each sample data of the fall data subset, and the positive and negative relationship between the acceleration components of the X-axis and Z-axis;

[0038] B) Calculate the proportion of the following six types of sample data in the fall data subset:

[0039] ①The first type of sample data consists of sample data where the X-axis acceleration component is greater than the Z-axis acceleration component;

[0040] ②The second type of sample data consists of sample data where the X-axis acceleration component is less than or equal to the Z-axis acceleration component;

[0041] ③The third type of sample data consists of sample data where the X-axis acceleration component is greater than 0;

[0042] ④ The fourth type of sample data consists of sample data where the X-axis acceleration component is ≤0;

[0043] ⑤ The fifth type of sample data consists of sample data where the Z-axis acceleration component is greater than 0;

[0044] ⑥ The sixth type of sample data consists of sample data where the Z-axis acceleration component is ≤0;

[0045] C) When the proportion of the first type of sample data in the subset of fall data is greater than that of the second type of sample data, if the proportion of the third type of sample data in the subset of fall data is greater than that of the fourth type of sample data, then the direction of the fall is to the left; if the proportion of the third type of sample data in the subset of fall data is less than or equal to that of the fourth type of sample data, then the direction of the fall is to the right.

[0046] D) When the proportion of the first type of sample data in the subset of fall data is less than or equal to the proportion of the second type of sample data, if the proportion of the fifth type of sample data in the subset of fall data is greater than the proportion of the sixth type of sample data, then the direction of the fall is backward; if the proportion of the fifth type of sample data in the subset of fall data is less than or equal to the proportion of the sixth type of sample data, then the direction of the fall is forward.

[0047] Preferably, the specific steps for determining whether a person is in a fall using the XGBoost model and a fall dataset are as follows:

[0048] First, a portion of the sample data from the fall dataset is used as training samples to train the XGBoost model, thus obtaining the XGBoost fall judgment model.

[0049] Then, a portion of the sample data from the fall dataset is used as test samples, and the XGBoost fall detection model is used to determine whether a person is in a fall state.

[0050] Preferably, the standard data feature space includes several standard time-domain feature values ​​from each feature vector corresponding to all sample data.

[0051] Preferably, the feature vector includes time-domain feature values ​​corresponding to the angular velocity and acceleration of the human body at various times, and the time-domain feature values ​​include mean, maximum value, minimum value, standard deviation, numerical range, change, cumulative change, average rate of change, and signal amplitude area.

[0052] The advantages of this invention are as follows:

[0053] (1) Before using the XGBoost fall detection model to judge the fall status of the human body, the threshold method is used to perform preliminary classification of the sample data, which greatly reduces the data processing pressure of the XGBoost fall detection model.

[0054] (2) While reducing the feature space dimension formed by real-time human body data, it can also ensure the "low loss" of feature data, which can maintain the properties of the original data and obtain the optimal low-dimensional data to ensure the accuracy of judging the direction of human fall and improve the judgment efficiency. Attached Figure Description

[0055] Figure 1 This is a flowchart of the present invention;

[0056] Figure 2 This is a flowchart of the dimensionality reduction process for standard data feature space in this invention;

[0057] Figure 3 This is a schematic diagram of the importance ratio curve formed by arranging the corresponding feature vectors in the current feature space in descending order according to the importance of the features when performing dimensionality reduction processing on the standard data feature space in this invention. Detailed Implementation

[0058] like Figures 1 to 2 As shown, a method for determining the direction of a person's fall includes the following steps:

[0059] 1) Establish a fall monitoring database;

[0060] 2) An inertial sensor is installed on the waist of the human body to collect the real-time angular velocity and real-time acceleration of the human body as sample data and store them in the fall monitoring database;

[0061] The angular velocity includes the angular velocity components of the human body on the X, Y, and Z axes of the three-dimensional coordinate system, as well as the angular velocity vector composed of these three angular velocity components;

[0062] The acceleration includes the acceleration components of the human body on the X, Y, and Z axes of the three-dimensional coordinate system, as well as the acceleration vector composed of these three acceleration components;

[0063] 3) Extract features from all peak segments of each sample data to obtain the feature vector corresponding to each sample data, forming the original data feature space. The feature vector includes the time-domain feature values ​​corresponding to the angular velocity and acceleration of the human body at various times. The time-domain feature values ​​include mean, maximum value, minimum value, standard deviation, numerical range, change, cumulative change, average rate of change, and signal amplitude area. The calculation formula for the time-domain feature value corresponding to the acceleration vector in the sample data is shown in Table 1.

[0064] Table 1

[0065]

[0066] In Table 1, mean is the average value of acceleration vector a in the sample data, max is the maximum value of acceleration vector a in the sample data, min is the minimum value of acceleration vector a in the sample data, and std is the mean value of acceleration vector a in the sample data. t Let be the standard deviation of the acceleration vector 'a' in the sample data, range be the range of values ​​for the acceleration vector 'a' in the sample data, Δa be the change in the acceleration vector 'a' in the sample data, cum be the cumulative change in the acceleration vector 'a' in the sample data, ARC be the average rate of change of the acceleration vector 'a' in the sample data, and SMA be the signal amplitude area of ​​the acceleration vector 'a' in the sample data.i Let be the i-th acceleration vector in the sample data, n be the total number of acceleration vectors in the sample data, a(t+T) be the acceleration vector at time (t+T), a(t) be the acceleration vector at time t, and a be the acceleration vector in the sample data. xi Let a be the acceleration component of the human body along the X-axis in a three-dimensional coordinate system. yi Let a be the acceleration component of the human body along the Y-axis in a three-dimensional coordinate system. zi Let T be the acceleration component of the human body on the Z-axis of the three-dimensional coordinate system, T be the time taken for the entire fall, and t be the time t during the entire fall.

[0067] 4) Perform data standardization on all feature vectors in the original data feature space to obtain a standard data feature space, which includes several standard time-domain feature values ​​in each feature vector corresponding to all sample data.

[0068] In this embodiment, the standardization method is used as a dimensionless means according to the following formula to perform data standardization processing on all feature vectors in the original data feature space:

[0069]

[0070] In the formula, mean is the mean of the time-domain eigenvalues ​​in the eigenvector, σ is the standard deviation, x is the time-domain eigenvalue in the eigenvector, and X is the time-domain eigenvalue of the eigenvector after standardization.

[0071] 5) Dimensionality reduction is performed on the standard data feature space, and the threshold method is used to process the dimensionality-reduced standard data feature space to select sample data that meet the requirements and form a fall dataset.

[0072] 5-1) The specific steps for dimensionality reduction of the standard data feature space are as follows:

[0073] (1) The XGBoost model was trained using a public dataset to obtain the XGBoost dimensionality reduction model, and a minimum threshold for feature dimensions was set.

[0074] Typically, XGBoost dimensionality reduction models can be obtained directly from the XGBoost standard library.

[0075] (2) Use the standard data feature space as the current feature space;

[0076] (3) Use the XGBoost dimensionality reduction model to obtain the weights of each feature vector in the current feature space, which are used as the feature importance of each feature vector in the current feature space;

[0077] The weighting method is a common way to evaluate feature importance in XGBoost models. The weight of a feature vector is how many times the feature is used when splitting nodes across all trees. This is the weight metric of the XGBoost model, which means that the feature importance is the number of times the feature vector is used across all trees in the XGBoost dimensionality reduction model.

[0078] (4) Sort the corresponding feature vectors in the current feature space in descending order according to their importance.

[0079] like Figure 3 As shown in the figure, the feature importance curve is steep in the first half. Taking point A as the dividing point, the sum of the importance of the first 12 features accounts for 50% of the total. Taking point B as the dividing point, the sum of the importance of the first 22 features accounts for 80%, and the remaining half accounts for only 20%. These two data indicate that most of the feature importance is distributed among a small number of features, and the remaining features are unimportant or replaceable.

[0080] Table 2 shows the distribution of feature importance percentages across different intervals. As can be seen from Table 2, the importance percentages of many features are between 1% and 2%, a significant difference compared to the highest point of 7.96%. However, as the number of features gradually decreases, the importance percentage of some features will increase. On one hand, as the total number of features decreases, the importance percentage of each feature will increase slightly. On the other hand, some features can replace deleted features, thus increasing their own importance percentage.

[0081] Table 2

[0082]

[0083] (5) Delete the last feature vector in the current feature space to form a new feature space to be processed. That is, delete the feature vector with the least feature importance. In the new feature space to be processed, the influence of this feature vector on the final calculation result can be ignored.

[0084] (6) Repeat step (5) until the dimension of the feature space to be processed is equal to the minimum threshold of the feature dimension, so as to obtain multiple feature spaces to be processed with different numbers of feature vectors, and calculate the accuracy of each feature space to be processed, and take the feature space with the highest accuracy as the optimal feature space.

[0085] Table 3 shows the dimensionality reduction process and the results after each reduction step. After the first dimensionality reduction, the dimensionality of the standard data feature space was significantly reduced from 44 dimensions to 18 dimensions. After the second dimensionality reduction, the dimensionality of the standard data feature space decreased from 18 dimensions to 16 dimensions. The first two dimensionality reduction processes were accompanied by a slight improvement in accuracy. When the third dimensionality reduction was performed, the dimensionality of the standard data feature space decreased again, but this also led to a decrease in accuracy. Considering the performance of the three dimensionality reductions, this embodiment selects the result after the second dimensionality reduction as the optimal feature space. When the XGBoost model uses the data with the optimal feature space, the accuracy on the test set reached 97.14%, which is higher than the 96.78% on the validation set, indicating that using the optimal feature space has good generalization ability.

[0086] Table 3

[0087]

[0088] As the dimensionality of the standard data feature space decreases, the number of loops involved in building the tree structure in the XGBoost model also decreases, thus reducing training time. To verify this, this example tests the training time before and after dimensionality reduction. The results show that training the XGBoost model before dimensionality reduction took 0.47 seconds, while training after dimensionality reduction only took 0.3 seconds, a reduction of 0.17 seconds. This demonstrates that removing some less important features not only improves model accuracy but also reduces training time.

[0089] (7) Compare the accuracy of the optimal feature space with that of the current feature space:

[0090] a) If the accuracy of the optimal feature space is greater than the accuracy of the current feature space, then use the optimal feature space as the new current feature space and repeat steps (3) to (7).

[0091] b) If the accuracy of the optimal feature space is less than or equal to the accuracy of the current feature space, then the current feature space is used as the standard data feature space after dimensionality reduction.

[0092] Accuracy is typically calculated by using the XGBoost dimensionality reduction model to classify the sample data corresponding to the feature space to be processed. The accuracy is the proportion of correctly classified samples out of the total number of classified samples.

[0093]

[0094] In the formula, TP is the number of sample data where a fall occurred and was predicted as a fall, TN is the number of sample data where a fall did not occur and was predicted as a non-fall, FP is the number of sample data where a fall did not occur but was predicted as a fall, and FN is the number of sample data where a fall occurred but was predicted as a non-fall.

[0095] After dimensionality reduction, the dimension of the standard data feature space is reduced from 44 dimensions to 16 dimensions. The dimensionality-reduced standard data feature space includes the area of ​​angular velocity signal amplitude, the area of ​​acceleration signal amplitude, the cumulative change of angular velocity, the maximum value of the Y-axis acceleration component, the mean value of the Z-axis acceleration component, the standard deviation of the X-axis acceleration component, the numerical range of the Z-axis angular velocity component, the mean value of the Z-axis angular velocity component, the minimum value of the Z-axis acceleration component, the mean value of the X-axis angular velocity component, the average rate of change of angular velocity, the maximum value of acceleration, the minimum value of the X-axis angular velocity component, the standard deviation of the X-axis angular velocity component, and the mean value of acceleration.

[0096] In this embodiment, Principal Component Analysis (PCA) is also used to reduce the dimensionality of the standard data feature space under the same conditions. The data during the dimensionality reduction process are shown in Table 4 below:

[0097] Table 4

[0098]

[0099]

[0100] Feature importance percentage represents the proportion of feature information to be retained in the feature space after dimensionality reduction. As shown in Table 4, each application of PCA for dimensionality reduction is accompanied by a decrease in accuracy. The rate of accuracy decrease is relatively slow when reducing from 44 dimensions to 12 dimensions, with only a 0.36% reduction after three dimensionality reductions. Further increasing the dimensionality reduction effort leads to a significant drop in accuracy. When using the dimensionality reduction method of this invention to reduce the standard data feature space to 18 dimensions, the accuracy reaches 96.51%, and when reduced to 16 dimensions, the accuracy further increases to 96.78%. However, when using PCA to reduce the standard data feature space to 18 dimensions, the accuracy is only 96.07%, 0.44% lower than the accuracy of the dimensionality reduction method of this invention.

[0101] Therefore, it can be seen that whether comparing the highest accuracy of the dimensionality reduction process or the accuracy of reducing to the same dimension, the dimensionality reduction method of the present invention is always superior to PCA in performance.

[0102] 5-2) In this embodiment, the fall dataset is obtained by setting a first judgment threshold and a second judgment threshold, and then processing the standard data feature space after dimensionality reduction using a thresholding method. The specific steps are as follows:

[0103] I. If the first time-domain feature value in the feature vector is less than or equal to the first judgment threshold, then the sample data corresponding to the feature vector does not meet the requirements;

[0104] II. When the first time-domain feature value in the feature vector is greater than the first judgment threshold, if the second time-domain feature value in the feature vector is less than or equal to the second judgment threshold, then the sample data corresponding to the feature vector does not meet the requirements; if the second time-domain feature value in the feature vector is greater than the second judgment threshold, then the sample data corresponding to the feature vector meets the requirements.

[0105] III. Repeat steps I to II to evaluate all feature vectors in the feature space of the standard data after dimensionality reduction, and select the sample data that meets the requirements to form the fall dataset.

[0106] In this embodiment, the first time-domain feature value is the mean value of the Y-axis acceleration component, and the second time-domain feature value is the cumulative change in angular velocity. The sample data that meets the requirements are all sample data of suspected falls.

[0107] 6) Using the fall dataset, determine whether a person is in a fall using the XGBoost model according to the following method:

[0108] First, a portion of the sample data from the fall dataset is used as training samples to train the XGBoost model, thus obtaining the XGBoost fall judgment model.

[0109] Then, a portion of the sample data from the fall dataset is used as test samples, and the XGBoost fall detection model is used to determine whether a person is in a fall state.

[0110] When a person is in a fall, the sample data selected by the XGBoost model constitutes a subset of fall data used to determine the direction of the fall:

[0111] 7) Determine the direction of a fall by analyzing the X-axis and Z-axis acceleration components of each sample in the fall data subset. The specific steps are as follows:

[0112] A) Determine the relationship between the acceleration components of the X-axis and Z-axis in each sample data of the fall data subset, and the positive and negative relationship between the acceleration components of the X-axis and Z-axis;

[0113] B) Calculate the proportion of the following six types of sample data in the fall data subset:

[0114] ①The first type of sample data consists of sample data where the X-axis acceleration component is greater than the Z-axis acceleration component;

[0115] ②The second type of sample data consists of sample data where the X-axis acceleration component is less than or equal to the Z-axis acceleration component;

[0116] ③The third type of sample data consists of sample data where the X-axis acceleration component is greater than 0;

[0117] ④ The fourth type of sample data consists of sample data where the X-axis acceleration component is ≤0;

[0118] ⑤ The fifth type of sample data consists of sample data where the Z-axis acceleration component is greater than 0;

[0119] ⑥ The sixth type of sample data consists of sample data where the Z-axis acceleration component is ≤0;

[0120] C) When the proportion of the first type of sample data in the subset of fall data is greater than that of the second type of sample data, if the proportion of the third type of sample data in the subset of fall data is greater than that of the fourth type of sample data, then the direction of the fall is to the left; if the proportion of the third type of sample data in the subset of fall data is less than or equal to that of the fourth type of sample data, then the direction of the fall is to the right.

[0121] D) When the proportion of the first type of sample data in the subset of fall data is less than or equal to the proportion of the second type of sample data, if the proportion of the fifth type of sample data in the subset of fall data is greater than the proportion of the sixth type of sample data, then the direction of the fall is backward; if the proportion of the fifth type of sample data in the subset of fall data is less than or equal to the proportion of the sixth type of sample data, then the direction of the fall is forward.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications made to the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method of determining the direction of a fall of a human body, characterized by, The method comprises the following steps: 1) establishing a fall monitoring database; 2) setting an inertial sensor on the waist of a human body, collecting real-time angular velocity and real-time acceleration of the human body as sample data, and storing the sample data in the fall monitoring database; The acceleration comprises acceleration components of the human body on X, Y and Z axes of a three-dimensional coordinate system, and an acceleration vector composed of the three acceleration components; The angular velocity comprises angular velocity components of the human body on X, Y and Z axes of a three-dimensional coordinate system, and an angular velocity vector composed of the three angular velocity components; 3) performing feature extraction on all peak segment data of each sample data respectively to obtain a feature vector corresponding to each sample data, and forming an original data feature space; 4) performing data standardization processing on all feature vectors in the original data feature space to obtain a standard data feature space; 5) performing dimension reduction processing on the standard data feature space in the following manner: 5-1) training an XGBoost model by using a public data set to obtain an XGBoost dimension reduction processing model, and setting a minimum threshold of feature dimension; 5-2) taking the standard data feature space as a current feature space; 5-3) obtaining weights of feature vectors in the current feature space by using the XGBoost dimension reduction processing model, taking the weights as feature importance of the feature vectors in the current feature space, and the feature importance being a number of times that the feature vector is used in all trees of the XGBoost dimension reduction processing model; 5-4) arranging the feature vectors in descending order in the current feature space according to the feature importance; 5-5) deleting a last feature vector in the current feature space to form a new to-be-processed feature space; 5-6) repeating step 5-5) until a dimension of the to-be-processed feature space is equal to the minimum threshold of feature dimension, obtaining a plurality of to-be-processed feature spaces with different numbers of feature vectors, and calculating an accuracy rate of each to-be-processed feature space, and taking a to-be-processed feature space with the highest accuracy rate as an optimal feature space; 5-7) comparing the accuracy rates of the optimal feature space and the current feature space: 5-7-1) if the accuracy rate of the optimal feature space is greater than the accuracy rate of the current feature space, taking the optimal feature space as a new current feature space, and repeating steps 5-3) to 5-7); 5-7-2) if the accuracy rate of the optimal feature space is less than or equal to the accuracy rate of the current feature space, taking the current feature space as a dimension reduction processed standard data feature space, and the dimension reduction processed standard data feature space comprising angular velocity signal amplitude area, acceleration signal amplitude area, angular velocity cumulative change, maximum value of Y-axis acceleration component, mean value of Z-axis acceleration component, mean value of Y-axis acceleration component, standard deviation of X-axis acceleration component, numerical range of Z-axis angular velocity component, mean value of Z-axis angular velocity component, minimum value of Z-axis acceleration component, mean value of X-axis angular velocity component, average change rate of angular velocity, maximum value of acceleration, minimum value of X-axis angular velocity component, standard deviation of X-axis angular velocity component, and mean value of acceleration. 6) The threshold method is used to process the standard data feature space after dimensionality reduction in the following manner, and the sample data meeting the requirements are screened out, and all the sample data meeting the requirements are regarded as sample data of suspected falls, and a fall data set is formed: 6-1) The mean value of the Y-axis acceleration component is taken as the first time domain characteristic value in the feature vector, and the angular velocity cumulative change is taken as the second time domain characteristic value in the feature vector; 6-2) A first judgment threshold is set for the first time domain characteristic value, and a second judgment threshold is set for the second time domain characteristic value; 6-3) When the first time domain characteristic value in the feature vector is ≤ the first judgment threshold, the sample data corresponding to the feature vector does not meet the requirements; 6-4) When the first time domain characteristic value in the feature vector is > the first judgment threshold, if the second time domain characteristic value in the feature vector is ≤ the second judgment threshold, the sample data corresponding to the feature vector does not meet the requirements; if the second time domain characteristic value in the feature vector is > the second judgment threshold, the sample data corresponding to the feature vector meets the requirements; 6-5) Repeat steps 6-3) to 6-4), judge all feature vectors in the standard data feature space after dimensionality reduction, and screen out sample data meeting the requirements to form a fall data set; 7) The fall data set is judged by the XGBoost model whether the human body is in a falling state, and when the human body is in a falling state, the sample data selected by the XGBoost model form a fall data subset for judging the falling direction of the human body; 8) The falling direction of the human body is judged by the X-axis and Z-axis acceleration components of each sample data in the fall data subset in the following manner: 8-1) The relationship between the X-axis and Z-axis acceleration components of each sample data in the fall data subset is judged, as well as the positive and negative relationship between the X-axis and Z-axis acceleration components; 8-2) Calculate the proportion of the following six types of sample data in the fall data subset: 8-2-1) The first type of sample data is the sample data with X-axis acceleration component > Z-axis acceleration component; 8-2-2) The second type of sample data is the sample data with X-axis acceleration component ≤ Z-axis acceleration component; 8-2-3) The third type of sample data is the sample data with X-axis acceleration component > 0; 8-2-4) The fourth type of sample data is the sample data with X-axis acceleration component ≤ 0; 8-2-5) The fifth type of sample data is the sample data with Z-axis acceleration component > 0; 8-2-6) The sixth type of sample data is the sample data with Z-axis acceleration component ≤ 0; 8-3) When the proportion of the first type of sample data in the fall data subset is > the second type of sample data, if the proportion of the third type of sample data in the fall data subset is > the fourth type of sample data, the falling direction of the human body is left; if the proportion of the third type of sample data in the fall data subset is ≤ the fourth type of sample data, the falling direction of the human body is right; 8-4) When the proportion of the first type of sample data in the fall data subset is ≤ the second type of sample data, if the proportion of the fifth type of sample data in the fall data subset is > the sixth type of sample data, the falling direction of the human body is rear; if the proportion of the fifth type of sample data in the fall data subset is ≤ the sixth type of sample data, the falling direction of the human body is front.

2. The method of claim 1, wherein the direction of the fall is determined by: The standardized method is used as a dimensionless means according to the following formula, and all feature vectors in the original data feature space are processed by data standardization: ； In the formula, is the mean value of the feature vector, is the standard deviation, is the feature vector, is the feature vector after normalization processing.

3. The method of claim 1, wherein the direction of the fall is determined by: The specific steps of using the XGBoost model and the fall data set to determine whether the human body is in a falling state are as follows: First, a part of the sample data in the fall data set is used as a training sample to train the XGBoost model, and an XGBoost fall judgment model is obtained; Then, a part of the sample data in the fall data set is used as a test sample, and the XGBoost fall judgment model is used to determine whether the human body is in a falling state.

4. The method of claim 1, wherein the step of determining the direction of the fall comprises: The standard data feature space includes a plurality of standard time domain feature values in each feature vector corresponding to all sample data.

5. The method of claim 1, wherein the step of determining the direction of the fall comprises: The feature vector includes time domain feature values corresponding to the angular velocity and acceleration of the human body at each time, and the time domain feature values include mean, maximum, minimum, standard deviation, numerical range, change amount, cumulative change amount, average change rate, and signal amplitude area.

Images