CN111209815B - Non-contact fatigue driving detection method based on BP neural network with momentum optimization - Google Patents

Abstract

The invention discloses a non-contact fatigue driving detection method based on a BP neural network with momentum optimization, which comprises the following steps: s10, acquiring physiological signals of a driver through a Doppler radar module; s20, classifying physiological signals; s30, filtering the physiological signals and decomposing the physiological signals by using a CEEMD algorithm to obtain signals of which the two sections of heart beat and respiration contain complete time-frequency domain characteristics; s40, designing a BP neural network model with optimized momentum to train the data set, so as to obtain an algorithm model for detecting the fatigue state of the driver; s50, detecting the fatigue state through an algorithm model for detecting the fatigue state of the driver. The invention can efficiently and accurately detect the fatigue state of the driver while avoiding influencing the normal driving of the driver.

Description

A Non-contact Fatigue Driving Detection Based on Momentum Optimized BP Neural Network method

technical field

The invention belongs to the field of modeling detection, and in particular relates to a non-contact fatigue driving detection method based on a momentum-optimized BP neural network.

Background technique

Drowsy driving is one of the most common causes of traffic accidents in the world. According to the report of WHO (World Health Organization), more than 1.3 million people die in traffic accidents every year, and 20 to 50 million people suffer non-fatal injuries because of traffic accidents, and about 20% of fatal traffic accidents are among them. caused by drowsy driving. Therefore, if a system for automatically detecting fatigue driving can be developed and the driver can be warned in advance that he is in a state 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 state are mainly divided into two categories: 1. Contact fatigue state detection; 2. Non-contact fatigue state detection.

The contact fatigue detection method is mainly to detect the physiological state of the driver. Although the data obtained by this method is reliable, the error is small, and it is less affected by external interference, but this method needs to install a corresponding device for detecting physiological signals on the driver, which interferes too much with 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 artificial interference will cause detection artifacts and errors.

The non-contact fatigue detection method mainly monitors the driver's facial features and detects vehicle parameters. The analysis of the driver's facial features has great individual differences, and the change of 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 whole device will also increase; for the vehicle status And the detection of the driving track, the required hardware support is relatively high, and the price is expensive. Moreover, the requirements for external conditions are relatively harsh (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, and the reliability and accuracy are greatly reduced.

To sum up, although there are many methods to measure the driver's fatigue status 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 the reliability and accuracy of mixed detection methods are higher 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 by physiological condition detection is highly reliable, but it interferes a lot with the driver. Therefore, it is necessary to design a device for non-contact detection of the driver's physiological signal, and then combine the neural network model to train and learn the physiological signal, and obtain an algorithm model for detecting the driver's fatigue state.

Contents of the invention

In view of the above technical problems, the present invention realizes the non-contact detection of the driver's physiological state, avoids causing physical and driving interference to the driver, and improves accuracy; it can process data quickly and efficiently; the algorithm model adopts momentum optimization The BP neural network model has faster learning efficiency and higher accuracy than the traditional BP neural network.

Based on the above purpose, the present invention provides a non-contact fatigue driving detection method based on a momentum-optimized BP neural network.

Include the following steps:

S10, collecting physiological signals of the driver through the Doppler radar module;

S20, classifying the physiological signal;

S30, filtering the physiological signal and decomposing it with the CEEMD algorithm to obtain two signals including heartbeat and respiration with complete time-frequency domain characteristics;

S40, designing a momentum-optimized BP neural network model to train the data set, thereby obtaining an algorithm model for driver fatigue state detection;

S50, detecting the fatigue state by using an algorithm model for detecting the fatigue state of the driver.

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

Preferably, the momentum-optimized BP neural network iteratively updates the weight matrix according to the error backpropagation algorithm, and uses gradient descent to minimize the mean square error between the actual output value of the network and the expected output value, thereby obtaining driver fatigue state detection algorithm model.

Preferably, in the BP neural network model of described design momentum optimization, the BP neural network weight iteration equation is:

Among them, α is the learning rate, m is the total sample size, β is the momentum coefficient, and β∈(0,1), is the momentum gradient value of the previous iteration, /> is the cumulative value of the weight error.

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

Compared with the prior art, the Doppler radar module used in the present invention can realize the non-contact and accurate detection of the physiological signal of the driver, and the physiological signal can accurately reflect the fatigue state of the driver, and can effectively solve the problem of different fatigue conditions of different individuals. Differences in physiological signals across states. The momentum-optimized BP neural network model used has fast learning efficiency, fewer iterations, high accuracy, and greatly reduces the amount of calculation.

Description of drawings

Fig. 1 is a flow chart of the steps of a non-contact fatigue driving detection method based on a momentum-optimized BP neural network according to an embodiment of the present invention;

Fig. 2 is a kind of expert evaluation method standard diagram of the non-contact fatigue driving detection method based on the BP neural network of momentum optimization according to an embodiment of the present invention;

Fig. 3 is a kind of non-contact fatigue driving detection method based on the momentum optimized BP neural network according to an embodiment of the present invention, when the physiological signal of the driver is collected and processed by Doppler radar and decomposed by the CEEMD algorithm for the heartbeat signal and breathing signal frequency analysis chart;

Fig. 4 is the BP neural network model diagram of the momentum optimization of a non-contact fatigue driving detection method based on the momentum optimized BP neural network according to an embodiment of the present invention;

FIG. 5 is a diagram of a momentum-optimized BP neural network network training result of a non-contact fatigue driving detection method based on a momentum-optimized BP neural network according to an embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

The method of the invention includes data collection, data processing and data training. The data acquisition part is mainly composed of the Doppler radar module as the core and the simulation driving software system supporting the simulation driver, which is used to collect the driver's breathing and heartbeat signals. The data processing part is mainly to classify the collected breathing and heartbeat signals through the expert judgment method, and then filter and process each group of signals, and then decompose the signals by CEEMD algorithm to obtain heartbeat signals and breathing signals without phase distortion . The data training part is mainly to design the momentum-optimized BP neural network model, train the collected data, and obtain the algorithm model of driver fatigue state detection.

Example 1

Referring to Fig. 1, a flow chart of the steps of a non-contact fatigue driving detection method based on a momentum-optimized BP neural network, S10, collects the driver's physiological signal through the Doppler radar module;

S20, classifying the physiological signal;

S30, filtering the physiological signal and decomposing it with the CEEMD algorithm to obtain two signals x ₁ and x ₂ containing complete time-frequency domain features of heartbeat and respiration;

S40, designing a momentum-optimized BP neural network model to train the data set, thereby obtaining an algorithm model for driver fatigue state detection;

S50, detecting the fatigue state by using an algorithm model for detecting the fatigue state of the driver.

Physiological signals include at least the driver's breathing signal and heartbeat signal.

The momentum-optimized BP neural network iteratively updates the weight matrix according to the error back propagation algorithm, and uses gradient descent to minimize the mean square error between the actual output value and the expected output value of the network, so as to obtain the algorithm model of driver fatigue state detection.

In designing the BP neural network model optimized by momentum, the weight iteration equation of BP neural network is:

Among them, α is the learning rate, m is the total sample size, β is the momentum coefficient, and β∈(0,1), is the momentum gradient value of the previous iteration, /> is the cumulative value of the weight error.

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

See Figure 2 for the expert judgment criteria for fatigue ratings. There are 4 grades of fatigue state classification for drivers: awake state, level I fatigue state, II fatigue state, and level III fatigue state. Each fatigue level has corresponding characteristics, such as blinking frequency, breathing times, etc. Through the analysis of video information by experts, the fatigue level of the driver in this segment of video information can be judged, and then the driver’s fatigue level corresponding to this segment of video information can be obtained. Physiological signal of fatigue rating. All the collected physiological signals are classified by this method.

Refer to FIG. 3 for the physiological signal collected by the Doppler radar module and its spectrum characteristic diagram. It can be seen that the physiological signals of the human body can be successfully collected through the radar module. The physiological signal includes the breathing signal and the heartbeat signal, which are decomposed by the CEEMD algorithm to obtain two signals containing complete time-frequency domain characteristics of the heartbeat and breathing.

See Figure 4 for the BP neural network model diagram. The BP neural network iteratively updates the weight matrix θ according to the error back propagation algorithm. The basic idea of BP neural network is the gradient descent algorithm, which uses gradient descent to minimize the mean square error between the actual output value of the network and the expected output value. The BP neural network is based on the training weight of the standard gradient descent algorithm, and there is often a problem of local optimal solutions. By doing momentum optimization on the basis of BP neural network, the local optimal solution problem can be effectively solved, and the training is more efficient. The specific model building process is as follows:

First, implement forward propagation. Set the weight matrix of each layer to H ^(l-1) . l is an integer, l∈[2,L]. The activations of the hidden and output layers are denoted by a ^(l) , so:

The activation function selected by this network is the sigmoid function, which is defined as:

Using the activation function, nonlinear features can be added to make learning faster and more efficient. For the entire sample set:

a ⁽²⁾ = g(H ⁽¹⁾ X ⁽ⁱ⁾ ) (3)

a ^(l) = g(H ^(l-1) a ^(l-1) ) (4)

Use the forward propagation algorithm to calculate the activation term a ^(l) of each layer. According to the back propagation algorithm, it is necessary to calculate the error term δ ^(l) of the hidden layer and the output layer. The calculation process is as follows:

δ ^(L) = a ^(L) -T ⁽ⁱ⁾ (5)

δ ^(l) = (H ^(l) ) ^(T) δ ^(l+1) × g'(H ^(l) a ^(l) ) (6)

δ ⁽²⁾ ＝(H ⁽¹⁾ ) ^(T) δ ⁽²⁾ ×g'(H ⁽¹⁾ X ⁽ⁱ⁾ ) (7)

where g'(H ^(l) a ^(l) ) is the first derivative of g(H ^(l) a ^(l) ):

g'(H ^(l) a ^(l) )＝a ^(l) ×(1-a ^(l) ) (8)

After calculating the activation term a ^(l) and the error term δ ^(l) , according to the gradient descent algorithm, it is necessary to iterate and update the weight matrix to obtain the desired output. Therefore, define a new variable Indicates the weight error of the rth row and cth column of the first layer of the neural network, that is, the cumulative value of the weight error.

According to the principle of backpropagation, the first-order partial derivative of the cost function J(θ) has the following properties:

According to the gradient descent algorithm, the weight matrix is obtained as follows:

From Equation (11), it can be seen that each weight update is only related to the current gradient value, not the previous gradient. The momentum gradient descent algorithm uses an exponentially weighted average of previous gradient values to obtain the current gradient value, and uses the current gradient value to update the weights.

use to represent the momentum gradient value. The specific algorithm is as follows:

initialization is 0:

With the increase of training times, Constantly updated:

Among them, β is the momentum coefficient, β∈(0,1); is the current momentum gradient value; /> is the momentum gradient value of the last iteration; /> is the accumulation of weight errors.

Using the momentum gradient algorithm instead of the standard gradient descent algorithm, the final weight iteration formula is obtained:

Through multiple trainings, the final network depth and the number of neurons in each layer are expressed in NN:

NN＝[2n 50 60 60 50 4] (15)

Learning rate α=0.3, momentum coefficient β=0.9. See Figure 5 for the training results of the momentum-optimized BP neural network. In Figure 5, the entire canvas is divided into four small sections, each showing the output for each fatigue level. The abscissa indicates the fatigue level, and the ordinate indicates the corresponding probability. In fact, for each input, the corresponding output is 4 points, for example, the output of the input signal is represented as follows:

T = (0.2, 0.1, 0.4, 0.3) (16)

The meaning of T is: the probability that the input signal X corresponds to fatigue level 0 is 0.2, the probability of level I is 0.1, the probability of level II is 0.4, and the probability of level III is 0.3, so the fatigue level corresponding to this signal Class II. The prediction results and total prediction results of each fatigue level test set sample are as follows:

1) Awake state: 0.950

2) I fatigue state: 0.897

3) II fatigue status: 0.917

4) III fatigue state: 0.943

Overall probability: 0.927.

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 drawings, it will be appreciated by those of ordinary skill in the art that various changes 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 (3)

1. a non-contact fatigue driving detection method based on the BP neural network optimized by momentum, is characterized in that, comprises the following steps: S10, collecting physiological signals of the driver through the Doppler radar module; S20, classifying the physiological signal; S30, filtering the physiological signal and decomposing it with the CEEMD algorithm to obtain two signals including heartbeat and respiration with complete time-frequency domain characteristics; S40, designing a momentum-optimized BP neural network model to train the data set, thereby obtaining an algorithm model for driver fatigue state detection; S50, detecting the fatigue state through an algorithm model for driver fatigue state detection; The momentum-optimized BP neural network iteratively updates the weight matrix according to the error backpropagation algorithm, and uses gradient descent to minimize the mean square error between the actual output value and the expected output value of the network, thereby obtaining an algorithm model for driver fatigue state detection ; In the BP neural network model of described design momentum optimization, the BP neural network weight iteration equation is: Among them, α is the learning rate, m is the total sample size, β is the momentum coefficient, and β∈(0,1), is the momentum gradient value of the previous iteration, /> is the cumulative value of weight error; The model building process is as follows: First, to implement forward propagation, set the weight matrix of each layer to H ^(l-1) , l is an integer, l∈[2,L], the activation item of the hidden layer and the output layer is represented by a ^(l) , so : The activation function selected by this network is the sigmoid function, which is defined as: Using the activation function and adding nonlinear features, for the entire sample set: a ⁽²⁾ = g(H ⁽¹⁾ X ⁽ⁱ⁾ ) a ^(l) = g(H ^(l-1) a ^(l-1) ) Use the forward propagation algorithm to calculate the activation term a ^(l) of each layer. According to the back propagation algorithm, it is necessary to calculate the error term δ ^(l) of the hidden layer and the output layer. The calculation process is as follows: δ ^(L) = a ^(L) -T ⁽ⁱ⁾ δ ^(l) = (H ^(l) ) ^(T) δ ^(l+1) × g'(H ^(l) a ^(l) ) δ ⁽²⁾ ＝(H ⁽¹⁾ ) ^(T) δ ⁽²⁾ ×g'(H ⁽¹⁾ X ⁽ⁱ⁾ ) where g'(H ^(l) a ^(l) ) is the first derivative of g(H ^(l) a ^(l) ): g'(H ^(l) a ^(l) )＝a ^(l) ×(1-a ^(l) ) After calculating the activation term a ^(l) and the error term δ ^(l) , according to the gradient descent algorithm, it is necessary to iterate and update the weight matrix to obtain the desired output, therefore, define a new variable Represents the weight error of the rth row and cth column of the first layer of the neural network, that is, the cumulative value of the weight error, According to the principle of backpropagation, the first-order partial derivative of the cost function J(θ) has the following properties: According to the gradient descent algorithm, the weight matrix is obtained as follows: Among them, α is the learning rate, m is the total number of samples, the momentum gradient descent algorithm uses the exponentially weighted average of the previous gradient values to obtain the current gradient value, and uses the current gradient value to update the weight, use To represent the momentum gradient value, the specific algorithm is as follows: initialization is 0: With the increase of training times, Constantly updated: Among them, β is the momentum coefficient, β∈(0,1); is the current momentum gradient value; /> is the momentum gradient value of the last iteration; /> is the cumulative value of weight error, Using the momentum gradient algorithm instead of the standard gradient descent algorithm, the final weight iteration formula is obtained: Through multiple trainings, the final network depth and the number of neurons in each layer are expressed in NN: NN=[2n 50 60 60 50 4]. 2. a kind of non-contact fatigue driving detection method based on the BP neural network of momentum optimization as claimed in claim 1, is characterized in that, described physiological signal comprises driver's breathing signal and heartbeat signal at least. 3. a kind of non-contact fatigue driving detection method based on the BP neural network optimized by momentum as claimed in claim 1, is characterized in that, described Doppler radar module adopts microwave Doppler radar detector probe sensor HB100 module .