CN115127550A - A system for recognizing the motion state of the human body using a six-axis inertial measurement device - Google Patents

Abstract

The invention discloses a system for recognizing the motion state of a human body by using a six-axis inertial measurement device. Wherein, the system includes: an inertial measurement device, configured to collect axial inertial information, and obtain irregular motion data of the human body; a classification and identification device, configured to classify and identify the irregular motion through a pre-trained motion state identification model motion data to determine the motion state of the human body; wherein, the inertial measurement device includes: a sensor measurement circuit, consisting of three sets of single-axis accelerometers and gyroscopes whose accuracy is greater than a threshold; a signal processing circuit, which measures together with the sensor Rigid-flex integration of the circuit, wherein the rigid circuit part is assembled on the inner wall of the inertial measurement device with screws, the flexible circuit part is used for signal transmission, and the signal processing circuit is used for preprocessing the axial inertial information to obtain the preprocessing of the irregular motion data.

Description

A system for recognizing the motion state of the human body using a six-axis inertial measurement device

technical field

The present invention relates to the field of inertial navigation, in particular to a system for recognizing the motion state of a human body using a six-axis inertial measurement device.

Background technique

The human body may have some irregular behaviors in the motion state in some special spaces, such as crawling and squatting in a small space. The identification of irregular human motion states can be used in human safety monitoring, accident and disaster rescue and other fields. Taking accident and disaster rescue as an example, it is necessary to recognize the movement state of the human body to achieve rapid and efficient rescue of personnel, and to prevent secondary disasters from causing damage to rescuers. However, the current identification methods of irregular human motion state are not real-time and accurate enough to be applied in actual rescue scenarios.

For the above problems, no effective solution has been proposed yet.

SUMMARY OF THE INVENTION

The embodiment of the present invention provides a system for identifying the motion state of a human body by using a six-axis inertial measurement device, so as to at least solve the technical problem of inaccurate identification of the motion state of an irregular human body.

According to an aspect of the embodiments of the present invention, there is provided a system for identifying a motion state using an inertial measurement device, including: an inertial measurement device configured to collect axial inertial information and obtain irregular motion data of a human body; a classification and identification device, It is configured to classify and identify the irregular motion data through a pre-trained motion state recognition model to determine the motion state of the human body; wherein the inertial measurement device includes: a sensor measurement circuit, which consists of three sets of precision greater than a threshold value It is composed of a single-axis accelerometer and a gyroscope; the signal processing circuit is integrated with the rigid-flexible combination of the sensor measurement circuit, wherein the rigid circuit part is assembled on the inner wall of the inertial measurement device with screws, and the flexible circuit part is used for signal transmission. The signal processing circuit is used for preprocessing the axial inertia information to obtain the preprocessed irregular motion data.

In the embodiment of the present invention, an inertial measurement device with higher accuracy is used to collect and preprocess the irregular motion data of the human body, and the irregular motion data is identified by the classification and identification device, thereby realizing the accurate identification of the human body motion state. The technical effect is solved, and the technical problem of inaccurate identification of irregular human motion state is solved.

Description of drawings

The accompanying drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the attached image:

1 is a schematic structural diagram of a system for identifying a motion state by using an inertial measurement device according to an embodiment of the present invention;

2 is a schematic structural diagram of another system for identifying a motion state by using an inertial measurement device according to an embodiment of the present invention;

3 is a schematic diagram of the overall assembly of an inertial measurement device according to an embodiment of the present invention;

4 is a schematic diagram of a rigid-flex integrated circuit according to an embodiment of the present invention;

5 is a schematic structural diagram of an inertial measurement device base according to an embodiment of the present invention;

6 is a schematic structural diagram of an inertial measurement device housing according to an embodiment of the present invention;

7 is a schematic diagram of a bottom-level circuit assembly according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of a front-position circuit assembly according to an embodiment of the present invention;

9 is a schematic diagram of a top-level circuit assembly according to an embodiment of the present invention;

10 is a schematic diagram of a side position circuit assembly according to an embodiment of the present invention;

11 is a schematic diagram of a rigid-flex integrated circuit and a base assembly according to an embodiment of the present invention;

FIG. 12 is an overall schematic diagram of an inertial measurement device according to an embodiment of the present invention;

FIG. 13 is a flowchart of a method of identifying a motion state according to an embodiment of the present invention.

Detailed ways

In order to make those skilled in the art better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only Embodiments are part of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

It should be noted that the terms "first", "second" and the like in the description and claims of the present invention and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used may be interchanged under appropriate circumstances such that the embodiments of the invention described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having" and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those expressly listed Rather, those steps or units may include other steps or units not expressly listed or inherent to these processes, methods, products or devices.

Example 1

According to an embodiment of the present invention, a system for identifying the motion state of a human body using a six-axis inertial measurement device is provided. As shown in FIG. 1 , the system includes:

The inertial measurement device 100 is configured to collect axial inertial information and obtain irregular motion data of the human body;

The classification and identification device 200 is configured to classify and identify the irregular motion data through a pre-trained motion state identification model, so as to determine the motion state of the human body.

The inertial measurement device includes: a sensor measurement circuit, which is composed of three sets of single-axis accelerometers and gyroscopes with a precision greater than a threshold; a signal processing circuit, which is integrated with the sensor measurement circuit by rigid-flexible combination, wherein the rigid circuit part The screws are assembled on the inner wall of the inertial measurement device, the flexible circuit part is used for signal transmission, and the signal processing circuit is used for preprocessing the axial inertial information to obtain the preprocessed irregular motion data.

In an exemplary embodiment, three groups of single-axis accelerometers and gyroscopes with an accuracy greater than a threshold are orthogonally combined according to the right-hand rule to measure the X-axis, Y-axis, Z-axis acceleration information and angular velocity information, respectively.

In an exemplary embodiment, signals between three sets of single-axis accelerometers and gyroscopes with an accuracy greater than a threshold value and the signal processing circuit are transmitted through the internal routing of the rigid-flex circuit.

In an exemplary embodiment, the rigid circuit at the bottom is fixed on the base by four limit screws, the flexible circuit is bent, and the rigid circuit at the front is fixed on the front formed by the base pillars by screws. and fix the other rigid circuits on the top and side surfaces of the base pillars. The data connectors are connected by wires to connect the rigid circuit, the flexible circuit and the base to the inertial measurement The housing of the device is secured with screws.

In an exemplary embodiment, the signal processing circuit is further configured to perform the following preprocessing on the irregular motion data: filtering, data normalization, feature extraction, and dimensionality reduction.

In an exemplary embodiment, the classification and identification device is configured to search for an optimal kernel function and its parameters by means of cross-validation according to different application scenarios, feature dimensions and data characteristics, and use the kernel function, Supervised learning is performed on the training samples, and a classifier is calculated to obtain the motion state recognition model.

In an exemplary embodiment, the classification and identification device is configured to: search for an optimal kernel function and its parameters by using an iterative grid search method according to different application scenarios, feature dimensions and data characteristics, and use the Kernel function, supervised learning is performed on the training samples, and a classifier is calculated to obtain the motion state recognition model.

In an exemplary embodiment, the classification and identification device is further configured to: set a search range of each parameter of the iterative grid search method, and divide the search range into a fixed number of grids; traverse each grid to perform cross-validation, and obtain The recognition rate of each grid, and the difference between the maximum recognition rate and the minimum recognition rate is calculated according to the recognition rate of each grid; when the difference is less than the preset threshold, the parameter combination corresponding to the current maximum recognition rate is output as the parameters of the kernel function, when the difference is greater than the preset threshold, reset the search range, and perform the above steps until the difference is less than the preset threshold.

In an exemplary embodiment, the signal processing circuit is further configured to: find the maximum value H and the minimum value L of the measurement point in the motion three-axis acceleration data; find the maximum value max and the minimum value min in the three-dimensional angular velocity; The maximum value H and minimum value L of the measurement point, and the maximum value max and the minimum value min in the three-dimensional angular velocity are normalized.

In an exemplary embodiment, the signal processing circuit is further configured to: perform singular value decomposition on the training matrix, and obtain its eigenvalues and singular values. According to the principle of principal components, the larger the singular value, the more information it contains , the contribution rate of each component in the training matrix is obtained, and then the cumulative contribution rate is obtained, and features are selected according to the contribution rate for dimensionality reduction processing.

In this embodiment, the classification and recognition device firstly preprocesses the collected irregular motion data by means of machine learning, converts it into an input vector of the motion state recognition model, and performs irregular human motion state according to the pretrained motion state recognition model. The determination of irregular human motion state improves the real-time and accuracy of the identification of irregular human motion state, and further solves the technical problem of inaccurate identification of irregular human motion state.

In addition, the present embodiment also improves the inertial measurement device.

In the prior art, the accelerometers and gyroscopes used in inertial measurement devices usually directly use a set of three-axis acceleration sensors and three-axis angular velocity sensors to form a six-axis inertial measurement unit, which is directly sensitive to inertial information in three directions, and is affected by the sensor. The error affects the measurement accuracy is low. In addition, the sensor measurement circuit and the signal processing circuit are usually connected by wire welding, the assembly operation is complicated, the space is occupied, and it is easy to cause problems such as wiring errors, electromagnetic interference of wires, and low structural reliability.

These problems will affect the overall stability of the inertial measurement device, thereby affecting the accuracy of the acquired data of the inertial measurement device.

In this embodiment, the inertial measurement device is improved. In order to solve the problem of low measurement accuracy of the six-axis inertial measurement unit directly formed by using a three-axis accelerometer and a three-axis angular velocity sensor, three groups of high-precision single-axis accelerometers and gyroscopes are respectively used to form a high-precision single-axis inertial measurement unit. Precision six-axis inertial measurement unit. In addition, the sensor measurement circuit and the signal processing circuit adopt an integrated design of rigid and flexible combination. The rigid circuit part is assembled on the inner wall of the measurement device with screws to ensure the overall structure is stable and reliable. Signal transmission is realized through the flexible circuit, which simplifies the assembly process and reduces the internal space occupation of the measurement device. .

Through the above improvements, the sensor measurement circuit adopts a rigid-flexible integrated design, which greatly simplifies the assembly operation process. The use of high-precision single-axis devices is respectively sensitive to the three-axis inertial information, which improves the information measurement accuracy. The measurement circuits are connected by a flexible method to realize signal transmission, which reduces space occupation and internal interference. The rigid circuit part is fixedly connected with the measurement device, which improves the reliability of the overall structure, is convenient for the human body to carry, and improves the stability of signal transmission.

Example 2

According to an embodiment of the present invention, a system for recognizing a motion state by using an inertial measurement device is also provided. As shown in FIG. 2 , the system includes an inertial measurement device 100 and a classification and identification device 200 . The structure and function of the inertial measurement device in this embodiment are the same as those in Embodiment 1, which will not be repeated here.

The difference between this embodiment and Embodiment 1 is the classification and identification apparatus 200. The classification and identification apparatus of this embodiment includes a model training module 202. The following will describe in detail how the model training module performs model training.

Support vector machine (support vector machine, SVM for short, in terms of classification principle, it is a model for classifying data into two categories, defined as a linear classifier that maximizes the classification interval in the feature space. Its learning strategy is is to maximize the classification interval, and convert the classification hyperplane search problem into a convex quadratic programming problem to solve. SVM is a supervised machine learning algorithm, based on the principle of minimum structural risk, to minimize the confidence range and empirical risk, thereby It improves its generalization ability, and can achieve better classification results even when there are not many training samples. The core of the algorithm is how to find an optimal hyperplane that can divide the samples into two categories. The main idea is to Generalizing in-plane linearly separable problems to high-dimensional classification problems

However, the problems that need to be classified in daily life are often of many types, such as the classification of the motion state of the human body. It is a multi-classification problem, mainly by combining multiple binary SVMs to realize the construction of multi-class SVMs. Common methods are 1-v-1 SVMs and 1-v-r SVMs. 1-v-r SVMs algorithm is the earliest method of SVM to solve multi-classification problems. For a classification problem with a total of k categories, k sub-classifiers will be constructed. All classes are used as another class to construct a classifier. For a data to be classified, it is necessary to perform k classification operations, calculate the classification function value of each classifier, and select the class corresponding to the largest function value as the class to which it belongs. When there are many categories, the training data sets corresponding to a single category and all the remaining categories have a large asymmetry, and there is a complex data distribution. , in addition, when the classification function values are repeated, there will be an inseparability problem. Based on the above points, this multi-classification algorithm has been rarely used in practical applications. The 1-v-1 SVMs method is mainly used in the examples of this application.

In one example, the 1-v-1 SVMs classification method works as follows:

个子分类器，每个子分类器的训练数据仅需要这两类所对应的数据集，因此其数据复杂度要小于1-v-r算法。在数据验证时，遍历所有子分类器，每个子分类器的分类结果作为该类别的一票，统计最终所有类别的票数，选择票数最多的类别作为最终分类结果。对于将第i和第j类区分开来的分类器，定义第i类样本为正类样本，第j类样本为负类样本。For a classification problem with a total of k classes, it constructs a total of

The training data of each sub-classifier only needs the corresponding data sets of the two categories, so the data complexity is smaller than the 1-vr algorithm. During data verification, all sub-classifiers are traversed, the classification result of each sub-classifier is used as a vote for the category, the final votes of all categories are counted, and the category with the most votes is selected as the final classification result. For the classifier that distinguishes the i-th and j-th classes, the i-th class sample is defined as a positive class sample, and the j-th class sample is a negative class sample.

The SVM classification process is as follows:

(1) Feature extraction is performed on sample data, and the optimal subset is selected.

(2) Divide each action feature set into a training set and a test set, and combine the action sample sets in pairs (here, it is assumed that there are six action classifications, and 15 combinations are obtained, which are set as k1-k15.

(3 Starting from k1, the training set and the test set are used as data sources, the grid search algorithm is used to search for the optimal parameters, and the optimal parameters are used to create a sub-classifier, and its optimal recognition rate is recorded.

(4 Perform the operations in step (3) on k2, k3...k15 respectively.

其中：α_i为拉格朗日乘子；K(x_i，x)为计算样本点x_i与x之间的内积，n代表参与建模的n个带标签的训练数据，i为第i个训练数据；αi为拉格朗日乘子；K(xi，x)代表计算训练数据xi与x之间的内积；b为常数；yi为训练数据xi对应的标签。After high-dimensional transformation of the feature vector through the kernel function, supervised learning is performed on the training samples, and the functional formula for calculating the classifier is:

Among them: α _i is the Lagrange multiplier; K(x _i , x) is the inner product between the calculated sample points x _i and x, n represents the n labeled training data participating in the modeling, and i is the first i training data; αi is the Lagrange multiplier; K(xi, x) represents the inner product between the training data xi and x; b is a constant; yi is the label corresponding to the training data xi.

Another kernel function will be described in detail below.

Using SVM to classify nonlinear separable problems, the first thought is to map the linear inseparable data to a new linear separable data set through a certain nonlinear mapping relationship, and then classify it as a linear problem, assuming its mapping relationship is φ, the process is divided into two steps:

(1) Establish a mapping relationship φ, and map the data to the new data space.

(2) Linearly classify the data in the new data space.

(3 At this time, the classification function can be expressed as the following form:

其中，α_i为拉格朗日乘子；(φ(x_i)，φ(x))为计算样本点x_i与x之间的内积。只是先对x进行映射。现在找出一种方法可以将映射和内积同时进行，就可以将对线性不可分问题的分类简化为和线性可分问题同样的步骤，从而建立一个非线性学习机，而这种直接计算的方法就是本申请实施例所提到的核函数法。此处所说的″核″是一个函数，对所有的x和z，满足如下关系：After high-dimensional transformation of the feature vector through the kernel function, supervised learning is performed on the training samples, and the functional formula for calculating the classifier is:

Among them, α _i is the Lagrange multiplier; (φ(x _i ), φ(x)) is the inner product between the calculated sample points x _i and x. Just map x first. Now find a way to perform mapping and inner product at the same time, you can simplify the classification of linear inseparable problems to the same steps as linearly separable problems, so as to build a nonlinear learning machine, and this direct calculation method It is the kernel function method mentioned in the embodiments of the present application. The "kernel" referred to here is a function that, for all x and z, satisfies the following relationship:

k(x, z) = (φ(x), φ(z))

That is, the result obtained by the function k(x, z) and the inner product operation (φ(x), φ(z)) after the mapping have the same effect. Usually, the high-dimensional mapping is accompanied by an explosive growth of dimensions, The inner product operation on high-dimensional data requires a lot of calculations, so the kernel function can save a lot of inner product calculations and achieve the same inner product effect. So far, the classification function for linearly inseparable problems can be expressed as follows:

The application of the kernel function not only solves the linear inseparability problem but also avoids a large number of inner product operation problems caused by the increase of dimensions in the dimension mapping. There is no same standard for the selection of the kernel function. , and generally find the optimal kernel function and its parameters by means of cross-validation. The following introduces several commonly used kernel functions and analyzes their application scenarios, and finally compares the performance of each kernel function for action recognition through experiments.

Selection of kernel function:

The Radial Basis Function Kernel (RBF) is used in the method of the embodiment of the present application.

K(x _i , x _j )=exp(-γx _i -x _j ² ) when γ>0

In the formula: x _i and x _j are the eigenvectors; γ is the parameter of the kernel function.

The SVM obtained by this kernel function is a classifier of the corresponding radial basis function. The most critical point is that this kernel function is a function that does not exhibit large deviations among all kernel functions.

The process of finding the kernel function by the grid search method will be described in detail below.

The selection of the kernel function needs to be tried and determined according to the classification effect, and different kernel functions have different parameters, and the parameters have a very important influence on the quality of the classification effect. Therefore, in the embodiment of the present application, it is necessary to optimize the parameters of each kernel function. , and compare the optimal effects to select the most suitable kernel function. The present application adopts a grid search method to select the kernel function.

This application uses an iterative grid search method to search for function parameters and penalty factor c. The specific process is as follows:

(1. Set the search range sum of each parameter, and divide the search range into a fixed number of cells.

(2. Traverse each grid for cross-validation, and record its corresponding recognition rate.

(3. Calculate the difference between the maximum recognition rate and the minimum recognition rate after traversing once. If the difference is less than 0.01, the process is terminated, and the parameter combination corresponding to the current maximum recognition rate is output. If the difference is greater than 0.01, jump to (1) Step 1, iteratively subdivide the grid corresponding to the current maximum recognition rate until the difference is less than 0.01.

The parameters of the kernel function will be specifically described below.

In this process, the parameters in the penalty factor C and the kernel function are determined, and the accuracy of recognition can be improved by adjusting these parameters.

C is the penalty coefficient, which is the tolerance for error. The higher the C, the less tolerance for errors, and the easier it is to overfit. The smaller C is, the easier it is to underfit. If C is too large or too small, the generalization ability will become poor, so finding a suitable C plays an important role in the optimization of the network.

Gamma is a parameter that comes with the RBF function after selecting it as the kernel. Implicitly determines the distribution of the data after mapping to the new feature space, the larger the gamma, the less the support vector, the smaller the gamma value, the more the support vector. The number of support vectors affects the speed of training and prediction.

Example 3

According to an embodiment of the present invention, there is also provided a system for recognizing a motion state by using an inertial measurement device, and the system includes an inertial measurement device and a classification and identification device. The structure and function of the classification and identification device in this embodiment are the same as those in Embodiments 1 and 2, and are not repeated here.

The difference between this embodiment and Embodiment 1 is the inertial measurement device. As shown in FIGS. 3 to 12 , the high-precision six-axis inertial measurement device in this embodiment is composed of a rigid-flex integrated circuit, a base, a housing, and a data connector. Among them, the inertial information measurement circuit adopts the integrated design method of rigid-flex circuit boards, including five rigid circuit boards, which are connected by four-part flexible circuits. Compared with the traditional wire connection method, the rigid-flex integrated design method can greatly reduce the overall space occupation and simplify the assembly operation process, avoid wiring errors and reduce electromagnetic interference, and improve the overall reliability of the system, so as to be suitable for special spaces such as rescue and disaster relief and other harsh environments. It provides the possibility to accurately detect the motion data of the human body.

The rigid-flex integrated circuit integrates three independent high-precision single-axis accelerometers, three independent high-precision single-axis gyroscopes and signal processing circuits. High-precision measurement of inertial information. Compared with the solution using a single three-axis sensor or six-axis inertial sensor, it can improve the measurement accuracy and system stability. The signal processing circuit can realize high-speed real-time transmission of sensor data, and send it to external equipment through the data connector of the measuring device.

The rigid-flex circuit board includes the bottom rigid circuit board 1, the front rigid circuit board 2, the side rigid circuit board 1 3, the side rigid circuit board 2 4, and the top rigid circuit board 5. Five rigid circuits are connected to the four parts of the rigid circuit board. part of the flexible circuit. Among them, the bottom rigid circuit board and the front rigid circuit board are connected through flexible circuit board one 6, the front rigid circuit board and the top rigid circuit board are connected through flexible circuit board two 7, and the side rigid circuit board one and the top rigid circuit board are connected through The flexible circuit board three 8 is connected, and the side rigid circuit board two and the top rigid circuit board are connected through the flexible circuit board four 9 .

The base of the inertial measurement device has four rigid limit struts of equal height, including two front struts 10 and two rear struts 11. The tops of the four limit struts each have an M1.6x6 screw hole 13, and four limit struts. There are two 1.6mm through holes 14 on the left and right sides of the bit support, and two 1.6mm through holes 15 at the front and rear with different heights from the side through holes, which are used to limit and fix the rigid-flex integrated circuit. The base of the measuring device also has four screw holes 12 for fixing and supporting the rigid-flex circuit. The rigid-flex integrated circuit board can be folded and assembled on the base and fixed by screws, and then the combination of the base and the rigid-flex integrated circuit can be assembled with the housing of the measuring device.

螺母通过前位刚性电路板预留孔位与后位支柱侧面预留的通孔对齐固定。底部刚性电路板装配示意图如图7所示。第三步弯折柔性电路板二7，保持顶部刚性电路板5与底座限位支柱上侧面构成平面平行，使用四根M1.6x6螺钉通过顶部刚性电路板预留螺钉孔与四个底座限位支柱顶面预留的螺钉孔位对齐固定。底部刚性电路板装配示意图如图8所示。第四步弯折柔性电路板三8与柔性电路板四9，保持侧位刚性电路板一3和侧位刚性电路板二4分别与底座限位支柱构成的两个侧面平行，并使用四根M1.6x8螺钉与四个

螺母通过两块侧位刚性电路板预留螺钉孔与底座限位支柱侧面预通孔对齐固定。底部刚性电路板装配示意图如图9所示。In the assembly process, the first step is to align the grooves on both sides of the bottom rigid circuit board 1 with the two front pillars 10 of the base to assemble on the base, and use four M1. The four screw holes 12 of the base are aligned and fixed. A schematic diagram of the bottom rigid circuit board assembly is shown in Figure 6. The second step is to bend the flexible circuit board 1 6, and attach the front rigid circuit board 2 to the two rear pillars 11 of the base. The plane formed by the front rigid circuit board and the rear pillars is kept parallel. Use four M1.6x8 screws with four

The nuts are aligned and fixed through the reserved holes of the front rigid circuit board and the through holes reserved on the side of the rear pillar. The schematic diagram of the bottom rigid circuit board assembly is shown in Figure 7. The third step is to bend the flexible circuit board 27, keep the top rigid circuit board 5 and the upper side of the base limit support to form a plane parallel, use four M1.6x6 screws through the top rigid circuit board reserved screw holes and the four base limit The screw holes reserved on the top surface of the pillars are aligned and fixed. A schematic diagram of the bottom rigid circuit board assembly is shown in Figure 8. The fourth step is to bend the flexible circuit board 3 8 and the flexible circuit board 4 9 , keep the lateral rigid circuit board 1 3 and the lateral rigid circuit board 2 4 parallel to the two sides formed by the limit pillars of the base, and use four M1.6x8 screws with four

The nut is aligned and fixed with the pre-through holes on the side of the base limit pillar through the reserved screw holes of the two lateral rigid circuit boards. A schematic diagram of the bottom rigid circuit board assembly is shown in Figure 9.

The schematic diagram of the assembly after the rigid-flex integrated circuit and the base are assembled is shown in Figure 11.

The data connector is J30J-15ZKP type connector, cut the connector wire to about 5cm, and solder it to the output point of the integrated circuit board data processing module according to the designed line sequence. Remove the original screws on both sides of the connector J30J-15ZKP, insert it into the trapezoidal hole 16 of the upper cover from the inside to the outside, and fix it with the special screws and the original spring washers, washers and nuts. Close the base and the case, and take eight M1.2x4 screws to fasten the base and the case. The overall schematic diagram after the assembly is completed is shown in Figure 13.

In the embodiment of the present application, the surface of the circuit board in the contact part between the rigid-flex integrated circuit and the base is free of components and contacts to prevent problems such as line short circuit, thereby providing the possibility for accurate collection of motion data under severe conditions.

The process of preprocessing the collected motion data by the signal processing circuit will be described in detail below.

Step 1, filter processing.

Usually, the original signal data, that is, the collected irregular motion data of the human body, has a long sequence and also contains a lot of noise. Therefore, in order to extract the eigenvalues and improve the accuracy of the algorithm subsequently, the original signal data needs to be processed, and the method of moving median filtering is adopted in this application. The moving average filtering algorithm is a typical linear filtering algorithm, which has a good inhibitory effect on periodic noise interference. Its main principle is to use the method of moving the window and realize the filtering of the original data through the neighborhood average method. Assuming that the input is X and the output is y, the calculation method of the moving average filter is shown in the following formula, where M is the window size of the moving average filter.

where n is the window length and y(n) is the output after filtering.

Step 2, data normalization

Since this system uses the acceleration data and angular velocity data of the inertial sensor, however, due to the different physical meanings of these two kinds of data, the numerical range is quite different, and the final eigenvalue range will also be larger when the features are extracted separately. It will cause some interference to the action classification algorithm based on feature value, and the feature with large value has a greater impact on the classification result. Therefore, normalization processing is required, and the specific algorithm steps are as follows:

(1) Find the maximum value H and the minimum value L of the measurement point in the motion triaxial acceleration data.

(2), also find the maximum value max and minimum value min in the three-dimensional angular velocity.

(3) Calculate the following formulas for the three-axis angular velocity data:

where L is the normalized minimum value, H is the normalized maximum value, _ui represents the original data point, and vi _represents the normalized data point.

Step 3, feature extraction.

In theory, more eigenvalues can express the characteristics of action data, but too many eigenvalues will greatly increase the amount of calculation, so it is necessary to select appropriate eigenvalues for the algorithm. The following are commonly used data feature extraction methods.

其中N为动作数据点数。Mean:

where N is the number of action data points.

其中N为动作数据点数，xi表示第i个输入，

表示输入值的平均值。Standard deviation:

where N is the number of action data points, xi represents the ith input,

Represents the average of the input values.

其中N为动作数据点数。Root mean square:

where N is the number of action data points.

Step 4, PCA dimensionality reduction processing

Principal Component Analysis (PCA), Principle Component Analysis is a commonly used dimensionality reduction method and an _{unsupervised dimensionality} reduction processing method. The value _ui , according to the principal component idea, the larger the singular value, the more information it contains, and then the contribution rate of each component in D _m×n is obtained, and then the cumulative contribution rate is obtained, and the feature is selected according to the contribution rate.

Among them, U and V are orthogonal matrices, Λ is a non-negative diagonal matrix of m×n, let Λ=Diag[λ1, λ2, ..., λm], the relationship between its eigenvalue λi and singular value ui is as follows: λ _i =(u _i ) ²

λ _i is the characteristic root of the training matrix D. According to the idea of principal components, the larger the singular value, the more information it contains. Therefore, the feature space composed of the first principal component corresponds to the new feature space D′:

D′ _m×l =U(:,1:l)×Λ _l×l

U(:, 1:l) is the matrix corresponding to the first column vector, and Λ _l×l is the diagonal matrix corresponding to the first larger singular value. From this, the contribution rate of each principal component in D is obtained as C _i :

Among them, λi and λj represent the i-th and j-th eigenvalues.

Correspondingly, the cumulative contribution rate C

where m is the number of columns of the matrix and k is the number of rows.

Usually, the cumulative contribution rate of features is required to reach more than 90%, so as to ensure that the selected features contain most of the motion information. However, after dimension reduction, if the number of dimensions is too small, there will be insufficient eigenvalues and the identification cannot be accurately performed. If the number of dimensions is too large, problems such as excessive computation and long identification time will occur. Therefore, other suitable methods are needed to select features, so that the recognition time can be shortened and the classification accuracy can be improved while reasonably reducing the dimension.

In this embodiment, the implementation steps of the system for identifying the motion state of the inertial measurement device are as follows:

1. Use the inertial measurement device worn on the chest to collect irregular motions in small and sheltered spaces, and complete the classification of the collected data, mark the time stamp, and transmit it to the classification and identification device through the wireless communication module.

2. Perform data processing and analysis on the collected target walking cadence and sensor information, and label different behaviors with corresponding labels to form a data label pair set.

3. Build an irregular action recognition network model, and adjust the parameters according to the recognition effect and the category of the action.

4. Divide the obtained data label pairs into a training set and a test set, in which the training set is sent to the built network model for training, and the test set is used to evaluate the classification effect of the model after the training is completed.

5. Import the real-time data of wearable personnel into the trained model for online real-time classification and evaluation.

Example 4

The embodiment of the present invention provides a system for recognizing the motion state of the human body by using a six-axis inertial measurement device and a method for identifying the motion state of an irregular human body by a classification and identification device. As shown in FIG. 13 , the method includes:

Step S402, performing data preprocessing according to the irregular motion data to obtain preprocessed irregular motion data;

Step S404: Input the preprocessed irregular motion data into a motion state recognition model to obtain an irregular human body motion state recognition type output by the motion state recognition model.

Wherein, the motion state recognition model is a multi-class Support Vector Machine, or SVM model for short, which is trained based on the irregular motion data corresponding to the irregular human motion state pre-collected.

In an example, the multi-class SVM model includes one-versus-one Support Vector Machines, referred to as 1-v-1 SVMs or a pairwise model; the 1-v-1 SVMs model includes a first A plurality of two-category SVM models of irregular human motion states calculated by the formula, wherein the first formula is: N=K(K-1)/2; wherein, N is the two-category of the irregular human motion states The number of SVM models, K is the total number of categories of irregular human motion states to be identified by the 1-v-1 SVMs model.

In an example, the irregular motion data includes: acceleration data and angular velocity data; and performing data preprocessing according to the irregular motion data to obtain preprocessed irregular motion data includes: according to the irregular motion data, perform data trimming, and obtain trimmed irregular motion data; wherein, the data trimming includes at least one of the following processes: abnormal data processing, repeated data processing, and missing value filling; Motion data, perform data normalization to obtain the preprocessed irregular motion data.

In an example, the acceleration data is acquired by an acceleration acquisition device, and the angular velocity is acquired by an angular velocity acquisition device; the acceleration acquisition device includes an accelerometer; the angular velocity acquisition device includes a gyroscope; the acceleration data is, according to three-dimensional The three-axis acceleration data determined by the spatial position includes: first acceleration data, second acceleration data, and third acceleration data; the angular velocity data is the three-axis angular velocity data determined according to the three-dimensional spatial position, including: first angular velocity data, The second angular velocity data and the third angular velocity data.

In one example, the method further includes: acquiring training data, the training data including the irregular motion data marked and preprocessed by the data; and dividing the training data into a training set and a test construct a classifier model, and define a plurality of binary classification models; and, input the training set into the classifier, and establish an initial classifier model by calling the instantiation method; input the test set into the current classifier Model, according to the output result of the current classifier model and the irregular human motion state type mark in the test set, the accuracy of the current classifier model is obtained; According to the accuracy of the current classifier model, the model parameters of the classifier model according to The grid search method is used to optimize parameters until the multi-class SVM model whose accuracy reaches a predetermined requirement is obtained through training.

In one example, the classifier model is described by a second formula, wherein the second formula is:

Among them, n represents the n labeled training data participating in the modeling, i is the ith training data; αi is the Lagrange multiplier; K(xi, x) represents the calculation of the internal value between the training data xi and x product; b is a constant; yi is the label corresponding to the training data xi.

In one example, the grid search method includes: setting a number of parameters to be adjusted in the classifier model, and a search range for a penalty factor; and, in the search range, determining a plurality of fixed values to form Data grid; according to the data grid, set the classifier model parameters and the penalty factor traversal to the parameters corresponding to the data grid, and record the accuracy of the classifier model under different parameters If under the different parameters, in the accuracy of the classifier model, the difference between the maximum value and the minimum value is not greater than the preset value, then the model parameters corresponding to the maximum accuracy and the described model parameters will be obtained. The penalty factor is determined as the value of the classifier model; if the difference between the maximum value and the minimum value in the accuracy of the classifier model under the different parameters is greater than the preset value, then according to the obtained For the model parameter and the penalty factor corresponding to the maximum accuracy, the grid search method is repeatedly performed until the difference is not greater than a preset value.

In an example, the six-axis inertial measurement device is fixed to the foot of the human body, the zero-speed correction method combined with the extended Kalman filter can be used to estimate the system state vector error, and then the state vector of the strapdown solution can be corrected, The position, velocity, and attitude of the personnel after the filter estimation are obtained. The zero-speed correction method can effectively suppress the accumulation of inertial measurement device errors caused by acceleration integration, but requires that the inertial measurement device can only be installed on the feet, and the feet are required to be stationary relative to the ground within a certain period of time to ensure timely correction of the state vector . The zero-speed correction method algorithm can be roughly divided into two parts: zero-speed detection and zero-speed correction.

When a pedestrian walks, the motion of the foot can be divided into two modes: static and motion. The main task of zero-speed detection is to identify the stationary phase of the person's feet and determine it as the zero-speed interval in the gait cycle. Given a sequence of observations, the zero-speed detection is carried out through the sequence of observations, and the accuracy of identifying the zero-speed interval is one of the key factors affecting the navigation effect of the ZUPT algorithm. The requirement of zero-speed detection is to reduce the probability of misjudging the motion state as the zero-speed state and increase the probability of accurately judging the zero-speed state, because if the motion state is misjudged as the zero-speed state, it will introduce new errors to the navigation system. Misjudging the static state as the moving state will reduce the number of corrections to the navigation system. The formula for zero-speed detection using the generalized likelihood ratio detection method is:

In the formula: W is the size of the zero-speed detection window; Z _n is the sensor output in the zero-speed window,

分别为加速度计和陀螺仪的噪声方差；

为零速探测窗口内加速度计输出各轴比力平均值矢量，

表示k时刻陀螺仪输出，n表示采样数量，

表示k时刻加速度计输出，g表示重力加速度。

are the noise variances of the accelerometer and gyroscope, respectively;

In the zero-speed detection window, the accelerometer outputs the average vector of the specific force of each axis,

represents the gyro output at time k, n represents the number of samples,

represents the accelerometer output at time k, and g represents the acceleration of gravity.

A detection threshold γ _T is set for zero-speed detection, which is used to determine the zero-speed detection flag bit C _k of the data. The formula is as follows:

Among them, T represents the test statistic of zero-speed detection, and U _n represents the observed amount of zero-speed detection. The main task of the zero-speed correction is to use the zero-speed in the zero-speed interval as the observational quantity, filter and estimate the state variable error through the extended Kalman filter, and then correct the result of the strapdown inertial solution. The mathematical model of the system is established based on the navigation error model. The inertial navigation update is completed in the time update, and the measurement update is performed in the zero-speed interval. Immediately after the measurement update, the state variable error is fed back to the navigation update result for error correction.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims (10)

1. a system utilizing six-axis inertial measurement device to identify the state of motion of human body, is characterized in that, comprises: an inertial measurement device, configured to collect axial inertial information and obtain irregular motion data of the human body; A classification and identification device configured to classify and identify the irregular motion data through a pre-trained motion state identification model to determine the motion state of the human body; Wherein, the inertial measurement device includes: The sensor measurement circuit consists of three sets of single-axis accelerometers and gyroscopes whose accuracy is greater than the threshold; The signal processing circuit is integrated with the rigid-flexible combination of the sensor measurement circuit, wherein the rigid circuit part is assembled on the inner wall of the inertial measurement device with screws, the flexible circuit part is used for signal transmission, and the signal processing circuit is used for the axial The inertial information is preprocessed to obtain the preprocessed irregular motion data. 2. system according to claim 1 is characterized in that, three groups of single-axis accelerometers and gyroscopes whose accuracy is greater than threshold are combined according to right-hand rule orthogonality, respectively measure X-axis, Y-axis, Z-axis acceleration information and angular velocity information. 3. The system according to claim 1, wherein the signals between three sets of single-axis accelerometers and gyroscopes with an accuracy greater than a threshold value and the signal processing circuit pass through the rigid circuit portion and the flexible circuit portion of internal trace transmission. 4. The system according to claim 1, wherein the rigid circuit at the bottom is fixed on the base by four limit screws, the flexible circuit is bent, and the rigid circuit at the front is screwed It is fixed on the front surface formed by the base pillars, and the other rigid circuits are fixed on the top and side surfaces formed by the base pillars. The data connectors are connected by wires to connect the rigid circuit, the flexible circuit and the base. The combination body and the housing of the inertial measurement device are fixed by screws. 5. The system according to any one of claims 1 to 4, wherein the signal processing circuit is further configured to: perform the following preprocessing on the irregular motion data: filtering, data normalization processing , feature extraction and dimensionality reduction. 6. The system according to any one of claims 1 to 4, characterized in that, the classification and identification device is configured to search for the optimal one through cross-validation according to different application scenarios, feature dimensions and data characteristics. A kernel function and its parameters, and using the kernel function, supervised learning is performed on the training samples, and a classifier is calculated to obtain the motion state recognition model. 7. The system according to any one of claims 1 to 4, wherein the classification and identification device is configured to: according to different application scenarios, feature dimensions and data characteristics, by using an iterative grid search method to find The optimal kernel function and its parameters, and using the kernel function, supervised learning is performed on the training samples, and a classifier is calculated to obtain the motion state recognition model. 8. The system according to claim 7, wherein the classification and identification device is further configured to: Set the search range of each parameter of the iterative grid search method, and divide the search range into a fixed number of cells; Traverse each grid for cross-validation, obtain the recognition rate of each grid, and calculate the difference between the maximum recognition rate and the minimum recognition rate according to the recognition rate of each grid; When the difference is less than the preset threshold, output the parameter combination corresponding to the current maximum recognition rate as the parameter of the kernel function, when the difference is greater than the preset threshold, reset the search range, And perform the above steps until the difference is less than the preset threshold. 9. The system of claim 5, wherein the signal processing circuit is further configured to: Find the maximum value H and the minimum value L of the measurement point in the motion triaxial acceleration data; Find the maximum value max and minimum value min in the three-dimensional angular velocity; A normalization process is performed based on the maximum value H and the minimum value L at the measurement point, and the maximum value max and the minimum value min in the three-dimensional angular velocity. 10 . The system according to claim 5 , wherein the signal processing circuit is further configured to: perform singular value decomposition on the training matrix to obtain its eigenvalues and singular values, and according to the principle of principal components, the larger the singular value is, 10 . , the more information it contains, the contribution rate of each component in the training matrix is obtained, and then the cumulative contribution rate is obtained, and features are selected according to the size of the contribution rate for dimensionality reduction processing.

Images