CN111209816B - Non-contact fatigue driving detection method based on regular extreme learning machine - Google Patents

Abstract

The invention discloses a non-contact fatigue driving detection method based on a regular extreme learning machine, which includes the following steps: S10, collecting the driver's physiological signals through a Doppler radar module; S20, classifying the physiological signals; S30, classifying the physiological signals The signal is subjected to discrete Fourier transform to obtain the spectral characteristics; S40, perform feature transformation on the spectral characteristics; S50, design a regular extreme learning machine model to train the data set, thereby obtaining an algorithm model for driver fatigue state detection, and use this model to detect fatigue Check the status. The invention can avoid affecting the driver's normal driving and can efficiently and accurately detect the driver's fatigue state.

Description

A non-contact fatigue driving detection method based on regular extreme learning machine

Technical field

The invention belongs to the field of modeling detection, and specifically relates to a non-contact fatigue driving detection method based on a regular extreme learning machine.

Background technique

Drowsy driving is one of the most common causes of traffic accidents in the world. According to the WHO (World Health Organization) report, more than 1.3 million people die in traffic accidents every year, and 20 million to 50 million people suffer non-fatal injuries due to traffic accidents, of which about 20% are fatal traffic accidents. It is caused by fatigue driving. Therefore, if a system that automatically detects fatigue driving can be developed and can warn drivers in advance of fatigue driving, a large number of traffic accidents can be effectively avoided and the incidence of traffic accidents can be reduced.

At present, the methods for detecting fatigue status are mainly divided into two categories: 1. Contact fatigue status detection; 2. Non-contact fatigue status detection.

The contact fatigue detection method mainly detects the driver's physiological state. Although the data obtained by this method are reliable, have small errors, and are less affected by external interference, this method requires the driver to install a corresponding device for detecting physiological signals, which causes too much interference to the driver. To this end, researchers use radio to measure physiological signals and obtain signals through ZigBee, Bluetooth, etc. These technologies are relatively mature, but the accuracy will be greatly reduced, and human interference will cause detection artifacts and errors.

The non-contact fatigue detection method mainly monitors the driver's facial features and vehicle parameter detection. There are large individual differences in the analysis of driver facial features, and changes in brightness or the driver wearing sunglasses, masks and other items that cover the face will cause great interference to the detection, and the cost of the entire device will also increase; for the vehicle status And the detection of driving trajectory requires high hardware support and is expensive. Moreover, the requirements for external conditions are relatively strict (such as road signs, climate and lighting conditions, etc.). A big limitation of this method is that it is a detection of the vehicle, not a direct detection of the driver, so the reliability and accuracy are greatly reduced.

To sum up, although there are many methods to measure the driver's fatigue state in real time, most of them are limited to the theoretical research level. The monitoring devices that have been released have many limitations and many problems need to be solved. Each fatigue driving detection method has its advantages and limitations, so the detection of driver fatigue driving should not only use a single method. Many studies have shown that mixed detection methods are more reliable and accurate than single detection methods. Therefore, to develop an effective fatigue driving detection system, various detection methods should be combined in a hybrid system for detection. The data obtained from physiological condition detection is highly reliable, but it will cause greater interference to the driver.

Contents of the invention

In view of the above existing technical problems, the present invention realizes non-contact detection of the driver's physiological state, avoids physical and driving interference to the driver, improves accuracy; can eliminate differences between different individuals; and can quickly and efficiently process data Processing; the algorithm model is simple, the learning efficiency is fast, the number of iterations is small, and the accuracy is high. It provides a non-contact fatigue driving detection method based on the regular extreme learning machine.

Includes the following steps:

S10, collects the driver's physiological signals through the Doppler radar module;

S20, classify physiological signals;

S30, perform discrete Fourier transform on the physiological signal to obtain the spectral characteristics;

S40, perform feature transformation on the spectral characteristics;

S50, design a regular extreme learning machine model to train the data set, thereby obtaining an algorithm model for driver fatigue state detection, and detect fatigue state through this model.

Preferably, the physiological signal includes at least the driver's breathing signal and heartbeat signal.

Preferably, the discrete Fourier transform is performed on the physiological signal to obtain the spectral characteristics, and then the amplitude B _A and period _BT of the respiratory signal and the period _HT of the heartbeat signal are obtained.

Preferably, the feature transformation is:

Among them, R _T represents the ratio of the respiratory cycle B _T to the heartbeat cycle H _T. h _θ (x) is a hypothetical function obtained through the gradient descent algorithm. Substitute B _T and H _T into h _θ (x) respectively, and h _θ The ratio of (B _T ) to h _θ (H _T ) is expressed as _RA .

Preferably, the regular extreme learning machine is trained according to the training set data and randomly sets the input layer weight matrix ω to obtain the output layer weight matrix.

Preferably, the regular extreme learning machine output layer weight calculation equation is:

Among them, H is the hidden layer activation matrix, C is the regularization coefficient, I is the identity matrix, and L is the expected output matrix, that is, the output label.

Preferably, the Doppler radar module adopts the microwave Doppler radar detector probe sensor HB100 module.

Compared with the existing technology, the Doppler radar module used in the present invention can achieve non-contact and accurate detection of the driver's physiological signals, and the physiological signals can accurately reflect the driver's fatigue state, and can effectively solve the problem of different fatigue conditions of different individuals. Differences in physiological signals between states. The regular extreme learning machine model used has fast learning efficiency, few iterations, high accuracy, and greatly reduced calculation amount.

Description of the drawings

Figure 1 is a step flow chart of a non-contact fatigue driving detection method based on a regular extreme learning machine according to an embodiment of the present invention;

Figure 2 is a standard diagram of the expert evaluation method of a non-contact fatigue driving detection method based on a regular extreme learning machine according to an embodiment of the present invention;

Figure 3 is a physiological signal diagram collected by a Doppler radar module in a non-contact fatigue driving detection method based on a regular extreme learning machine according to an embodiment of the present invention;

Figure 4 is a physiological signal amplitude-frequency characteristic diagram of a non-contact fatigue driving detection method based on a regular extreme learning machine according to an embodiment of the present invention;

Figure 5 is a scatter diagram of the respiratory cycle and amplitude and its linear fitting curve diagram of a non-contact fatigue driving detection method based on a regular extreme learning machine according to an embodiment of the present invention;

Figure 6 is a diagram showing the time domain changes of physiological signals after feature transformation of a non-contact fatigue driving detection method based on a regular extreme learning machine according to an embodiment of the present invention;

Figure 7 is a regular extreme learning machine network model diagram of a non-contact fatigue driving detection method based on a regular extreme learning machine according to an embodiment of the present invention;

Figure 8 is a diagram showing the network training results of the regular extreme learning machine of a non-contact fatigue driving detection method based on the regular extreme learning machine according to an embodiment of the present invention.

Detailed ways

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 part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

The method of the present invention includes data collection, data processing and data training. The data acquisition part is mainly composed of a Doppler radar module as the core and a simulated driving software system supporting the simulated driver, which is used to collect the driver's breathing and heartbeat signals. The data processing part mainly uses the expert evaluation method to classify the collected respiratory and heartbeat signals into different levels, then filter each group of signals, and then perform discrete Fourier transform on the signals to obtain a spectrogram, and then extract the respiratory cycle. , amplitude, heart rate and other characteristic values. The data training part mainly involves designing a regular extreme learning machine network model, training the collected data, and obtaining an algorithm model for driver fatigue state detection.

Example 1

Referring to Figure 1, a step flow chart of a non-contact fatigue driving detection method based on a regular extreme learning machine, S10, collects the driver's physiological signals through the Doppler radar module;

S20, use expert evaluation methods to classify physiological signals;

S30, perform discrete Fourier transform on the physiological signal to obtain the spectral characteristics;

S40, perform feature transformation on the spectral characteristics;

S50, design a regular extreme learning machine model to train the data set, thereby obtaining an algorithm model for driver fatigue state detection, and detect fatigue state through this model.

The above physiological signals include at least the driver's breathing signal and heartbeat signal.

Perform discrete Fourier transform on the physiological signal to obtain the spectrum characteristics, and then obtain the amplitude B _A and period B _T of the respiratory signal and the period _HT of the heartbeat signal.

The feature transformation is:

Among them, R _T represents the ratio of the respiratory cycle B _T to the heartbeat cycle H _T. h _θ (x) is a hypothetical function obtained through the gradient descent algorithm. Substitute B _T and H _T into h _θ (x) respectively, and h _θ The ratio of (B _T ) to h _θ (H _T ) is expressed as _RA .

The regular extreme learning machine is trained based on the training set data and randomly setting the input layer weight matrix ω to obtain the output layer weight matrix.

The calculation equation of the output layer weight of the regular extreme learning machine is:

Among them, H is the hidden layer activation matrix, C is the regularization coefficient, I is the identity matrix, and L is the expected output matrix, that is, the output label.

The Doppler radar module uses the microwave Doppler radar detector probe sensor HB100 module.

See Figure 2 for the expert evaluation criteria for fatigue levels. There are four levels of driver fatigue status classification: awake state, level I fatigue state, II fatigue state, and level III fatigue state. Each fatigue level has corresponding characteristics, such as blink frequency, number of gasps, etc. Through expert analysis of video information, the driver's fatigue level in the video information can be judged, and then the driver's fatigue level corresponding to the video information can be obtained. Physiological signals of fatigue levels. All collected physiological signals are classified by this method.

See Figures 3 and 4, which respectively show the physiological signals collected by the Doppler radar module and their spectrum characteristics. It can be seen that the physiological signals of the human body can be successfully collected through this radar module. The physiological signal includes respiratory signal, heartbeat signal and noise. By filtering the signal and discrete Fourier transform (DFT), the noise can be effectively filtered, and the spectrum characteristic map is obtained, and then the frequency and amplitude of the signal are extracted.

See Figure 5 for a scatter plot of respiratory cycles and amplitudes of different individuals and their linear fitting curves. Breathing and heartbeat signals vary from person to person. Excessive individual differences are not conducive to the classification of data by the neural network and will greatly reduce the classification accuracy. Under normal conditions, the respiratory rate is generally 1/5-1/4 of the heart rate. As the degree of fatigue deepens, life activities slow down, energy consumption decreases, and the ratio of respiratory frequency to heart rate decreases. The ratio of respiratory rate to heart rate is defined as R _T . On this basis, by modeling the respiratory frequency and respiratory amplitude of each individual, the relationship between respiratory frequency and respiratory amplitude can be found, and then the relationship between respiratory amplitude and heart rate can be found. The left picture in Figure 5 shows a scatter plot of the respiratory frequency and respiratory amplitude of two different testers under different fatigue states. It can be seen that the linear relationship between the respiratory frequency and respiratory amplitude of different testers is Different. A gradient descent algorithm was used to fit each tester's data. The fitted curve can be seen on the right side of Figure 5. The specific method of training the regular extreme learning machine model is as follows:

According to the principle of gradient descent algorithm, first define the hypothesis function h _θ (X) as:

Among them, θ is the weight, x ⁽ⁱ⁾ is the i-th input data.

Secondly, define the cost function J(θ). The cost function, also known as the squared error function, is the most common method for solving regression problems. The definition is as follows:

Among them, m is the total number of samples, x ⁽ⁱ⁾ is the i-th input data, and y ⁽ⁱ⁾ is the i-th target output.

According to the principle of the gradient algorithm, a set of θ values needs to be found to make the cost function converge. According to equation (3), the θ value that makes the cost function converge can be obtained.

θ _j is defined as the j-th parameter of the hypothesis function h _θ (x), and α is the learning rate. Substituting equation (2) into equation (3) we can get:

Finally, assuming that the relationship between respiratory amplitude and respiratory cycle is linear, then the function is assumed to be h _θ (X):

h _θ (x)＝θ ₀ +θ ₁ x (5)

Substituting equation (5) into equation (4), the θ value can be obtained through gradient descent. Thus, the relationship between respiratory amplitude and respiratory cycle is found. Note: Each set of data obtains a different θ value.

According to the above method, by appropriately discarding the measurement accuracy, the relationship between each person's respiratory cycle and amplitude can be described by a linear curve. Therefore, the breathing signal can be converted as follows:

f(t)＝Asin(B _T t+b)→f(t)＝h _θ (B _T )sin(B _T t+b) (6)

In the formula, A is the amplitude and b is the phase. The ratio R _T of respiratory frequency to heart rate has been defined previously. The heart rate and respiratory frequency of normal people are consistent, that is, the ratio of heart rate to respiratory frequency changes within a certain range.

Define a new variable R _A. R _A is defined as follows:

Among them, h _θ (x) is h _θ (x) in equation (5).

As the degree of fatigue deepens, R _T will change gradually within a certain range. Then, it only needs to be proved that as the degree of fatigue deepens, _RA will change regularly within a certain range. For equation (8), the expansion is as follows:

For the function V(x), V(x) is defined as follows:

Differentiate V(x):

Therefore, according to equation (9) and equation (11), the following conclusions can be drawn:

1) Function R _A decreases;

2) For each tester, that is, if θ ₁ remains unchanged, the higher the heart rate, the smaller the _RA ;

3) _RA <=1.

Combining the above conclusions, through eigenvalue transformation, a new variable R is obtained, called the ratio point:

R=( _RA , _RT ) (12)

R is unique to each tester. In this way, the breathing signal can be mapped to a new two-dimensional space.

f(t)＝Asin(Tt+b)→f(t)＝R _A sin(R _T t+b) (13)

See Figure 6 for a diagram of the time domain change effect of the physiological signal after feature transformation. In Figure 6(a), these two curves respectively represent the breathing signal of tester A in the level I fatigue state and the breathing signal of tester B in the level II fatigue state. It can be seen that the periods and amplitudes of these two curves are not much different. At this time, the result obtained by converting the respiratory signal according to equation (13) is shown in Figure 6(b). It can be seen that the two signals at this time are very different in period and amplitude. Therefore, differences in respiratory amplitude between individuals can be reduced by appropriately discarding measurement accuracy.

By classifying, filtering, discrete Fourier transform and eigenvalue conversion of the data, the following eigenvalues are finally determined as training sample data: R _A , R _T , _HT . Where _HT represents the heartbeat cycle.

The final sample library used for neural network training is as follows:

X ⁽ⁱ⁾ = (R _A R _T H _T ) ⁽ⁱ⁾ (14)

L ⁽ⁱ⁾ = (h(s) h(s-1) h(s-2) h(s-3)) ⁽ⁱ⁾ (15)

S＝{(X ⁽ⁱ⁾ T ⁽ⁱ⁾ )},i∈[1,m] (16)

Among them, X is the input data, L is the output label corresponding to X, S is the sample data set, and h(x) is the impact function, which is defined as follows:

s is the fatigue state, and the s value can be 0, 1, 2 or 3, which represent the awake state, level I fatigue state, level II fatigue state and level III fatigue state respectively, i is the sample index, and m is the total sample size.

See Figure 7 for the regular extreme learning machine network model. Extreme Learning Machine (ELM) is an algorithm proposed by Professor Huang Guangbin to solve single hidden layer neural networks. ELM has the advantages of high learning efficiency and strong generalization ability, and is widely used in classification, regression, clustering, feature learning and other problems. For the single hidden layer neural network in Figure 7, the total number of input samples is m, the input data is X, the network output is O, the hidden layer activation matrix is H, the hidden layer input weight matrix is ω, and the hidden layer output weight matrix is β , b is the threshold, and the activation function is the sigmoid function. The mathematical expression of the sigmoid function is as follows:

Using the activation function, nonlinear features can be added to make learning faster and more efficient. Then the calculation equation of the activation term H is as follows:

The output O can be expressed as:

βH=O (20)

The goal of single hidden layer neural network learning is to minimize the output error, that is:

||OL|| _n×m =0 (21)

Among them, n represents the number of feature values, and m represents the total number of samples. L is the expected output matrix, which is the output label. Then there exist β, ω, and b such that:

βH=L (22)

According to the principle of extreme learning machine: as long as the activation function g(·) satisfies infinite differentiability in any interval, then the input weight matrix ω and bias b can be randomly generated, which means that the single hidden layer feedforward neural network does not need to and b are adjusted; and because the continuous probability distribution randomly generates ω and b, and assuming that the number of hidden layer neurons is k, then k ≤ N exists, so that ||L-βH|| _n×m ≤ε must be true , at this time it is found that the bias of the output layer is no longer needed. Therefore, when the hidden layer input weight matrix ω and offset b are determined, the hidden layer activation term H can be determined. After determining the activation term H, only the hidden layer output matrix β is required. According to the theory proposed by Professor Huang Guangbin: when the number k of hidden layer neurons is consistent with the number m of training samples, the hidden layer activation matrix H is an invertible matrix (there is a very small possibility that H is irreversible), then through the formula (22) can calculate the hidden layer output weight β that enables the neural network to fit the output with 0 error. However, in most cases, the number k of hidden layer neurons is much smaller than the number m of training set samples, and in this case the matrix H is irreversible. At this time, it is necessary to find the solution that minimizes the loss function, that is:

According to the minimum norm criterion, the solution is obtained by taking least squares:

in It is the Moore-Penrose generalized inverse of the hidden layer output matrix H, which is called the pseudo-inverse. Solve/> Methods as below:

In order to obtain more practical and reliable regression coefficients, a biased estimation regression method is needed, which is called ridge regression. When X is not a full-rank matrix or the linear correlation between some columns is large, its determinant is close to 0, which will lead to large errors in calculation, so a regular term needs to be added to the loss function.

where I is the identity matrix and C is the regularization coefficient. Research shows that relatively small weight coefficients can improve the stability and generalization ability of single hidden layer feedforward neural networks (SLFN), which means that regularization of ELM is necessary under complex problems.

To sum up, when training data is input and the input weight matrix is randomly initialized, the output weight matrix can be obtained through Equation (36).

After multiple trainings, the number of hidden layer neurons was finally determined to be 500, and the regularization coefficient C=1e5. See Figure 8 for the regular extreme learning machine network training results. In Figure 8, the entire canvas is divided into four small sections, each section showing the output for each fatigue level. The abscissa represents the fatigue level, and the ordinate represents the corresponding probability. In fact, for each input, the corresponding output is 4 points. For example, the output of the input signal is expressed as follows:

Test＝((0,0.2),(1,0.1),(2,0.4),(3,0.3)) (37)

The meaning of ^Test is: the input signal Fatigue level II. The prediction results and total prediction results of each fatigue level test set sample are as follows:

1) Awake state: 0.983

2)I fatigue state: 0.867

3) II fatigue state: 0.883

4) III fatigue state: 0.967

Overall probability: 0.925.

It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the present invention have been described in conjunction with the accompanying drawings, it will be understood by those of ordinary skill in the art that various modifications may be made without departing from the spirit and scope of the invention as defined by the appended claims. Changes in form and detail.

Claims (1)

1. A non-contact fatigue driving detection method based on a regular extreme learning machine, which is characterized by including the following steps: S10, collects the driver's physiological signals through the Doppler radar module; S20, classify physiological signals; S30, perform discrete Fourier transform on the physiological signal to obtain the spectral characteristics; S40, perform feature transformation on the spectral characteristics; S50, design a regular extreme learning machine model to train the data set, thereby obtaining an algorithm model for driver fatigue state detection, and detect fatigue state through this model; The physiological signals include at least the driver's breathing signal and heartbeat signal; The discrete Fourier transform is performed on the physiological signal to obtain the spectrum characteristics, and then the amplitude B _A and period _BT of the respiratory signal and the period _HT of the heartbeat signal are obtained; The feature transformation is: Among them, R _T represents the ratio of the respiratory cycle B _T to the heartbeat cycle H _T. h _θ (x) is a hypothetical function obtained through the gradient descent algorithm. Substitute B _T and H _T into h _θ (x) respectively, and h _θ The ratio of (B _T ) to h _θ (H _T ) is expressed by R _A ; The regular extreme learning machine is trained according to the training set data and randomly sets the input layer weight matrix ω to obtain the output layer weight matrix. The regular extreme learning machine output layer weight calculation equation is: Among them, H is the hidden layer activation matrix, C is the regularization coefficient, I is the identity matrix, and L is the expected output matrix, that is, the output label; The Doppler radar module uses the microwave Doppler radar detector probe sensor HB100 module; The specific method of training the regular extreme learning machine model is as follows: According to the principle of gradient descent algorithm, first define the hypothesis function h _θ (X) as: Among them, θ is the weight, x ⁽ⁱ⁾ is the i-th input data; Secondly, define the cost function J(θ). The cost function, also called the squared error function, is defined as follows: Among them, m is the total number of samples, x ⁽ⁱ⁾ is the i-th input data, and y ⁽ⁱ⁾ is the i-th target output; According to the principle of the gradient algorithm, a set of θ values is found to make the cost function converge. According to equation (3), the θ value that makes the cost function converge can be obtained; θ _j is defined as the j-th parameter of the hypothesis function h _θ (x), α is the learning rate; substituting equation (2) into equation (3) we can get: Finally, assuming that the relationship between respiratory amplitude and respiratory cycle is linear, then the function is assumed to be h _θ (x): h _θ (x)＝θ ₀ +θ ₁ x (5) Substitute Equation (5) into Equation (4), and obtain the θ value through gradient descent, thus finding the relationship between respiratory amplitude and respiratory cycle. The θ value obtained for each set of data is different; According to the above method, the relationship between each person's respiratory cycle and amplitude can be described by a linear curve. Therefore, the respiratory signal can be converted as follows: f(t)＝Asin(B _T t+b)→f(t)＝h _θ (B _T )sin(B _T t+b) (6) In the formula, A is the amplitude, b is the phase, defining the ratio R _T of the respiratory cycle to the heartbeat cycle. The respiratory cycle of a normal person is consistent with the heartbeat cycle, that is, the ratio of the respiratory cycle to the heartbeat cycle changes within a certain range. Define a new variable R _A. R _A is defined as follows: Among them, h _θ (x) is h _θ (x) in formula (5), As the degree of fatigue deepens, R _T will change progressively within a certain range. Then, it only needs to be proved that as the degree of fatigue deepens, R _A will change regularly within a certain range. For equation (8), expand as follows: For the function V(x), V(x) is defined as follows: Differentiate V(x): Therefore, according to equation (9) and equation (11), the following conclusions can be drawn: 1) Function R _A decreases; 2) For each tester, that is, if θ ₁ remains unchanged, the higher the heart rate, the smaller the _RA ; 3) _RA ＜=1; Combining the above conclusions, through eigenvalue transformation, a new variable R is obtained, called the ratio point: R=( _RA , _RT ) (12) R is unique for each tester; therefore, the breathing signal can be mapped to a new two-dimensional space; f(t)＝Asin(Tt+b)→f(t)＝R _A sin(R _T t+b) (13) By classifying, filtering, discrete Fourier transform and eigenvalue conversion of the data, the following eigenvalues are finally determined as training sample data: R _A , R _T , _HT , where _HT represents the heartbeat cycle; The final sample library used for neural network training is as follows: X ⁽ⁱ⁾ = (R _A R _T H _T ) ⁽ⁱ⁾ (14) L ⁽ⁱ⁾ = (h(s) h(s-1) h(s-2) h(s-3)) ⁽ⁱ⁾ (15) S＝{(X ⁽ⁱ⁾ T ⁽ⁱ⁾ )},i∈[1,m] (16) Among them, X is the input data, L is the output label corresponding to X, S is the sample data set, and h(x) is the impact function, which is defined as follows: s is the fatigue state, and the s value can be 0, 1, 2 or 3, which represent the awake state, level I fatigue state, level II fatigue state and level III fatigue state respectively, i is the sample index, and m is the total number of samples.