CN103761832A - Wearable human body collision early warning and protection device and early warning and protection method thereof - Google Patents

Abstract

The invention relates to a wearable human body collision warning and protection device, which includes a belt body, a plurality of airbags are arranged at intervals in the middle of the belt body, and multi-sensor groups, microsilicon accelerometers, gyroscopes, and alarms are installed on the belt body above and below the airbags. The output terminals of the unit, the multi-sensor group, the microsilicon accelerometer and the gyroscope are connected to the input terminals of the main controller through the unit controller, and the output terminals of the main controller are respectively connected to the input terminals of the airbag and the alarm unit through the unit controller connected. The invention also discloses an early warning and protection method of the wearable human body collision early warning and protection device. In the present invention, the master controller compares the collected information with the empirical data set, gives the number of suspicious approaching objects and the approaching distance, and issues an execution command, sounds an alarm, and controls the opening of the airbag. The specified SMS information is sent to the specified mobile phone, and the main controller decides to select the secondary protection to execute the action.

Description

A wearable human body collision warning and protection device and its warning and protection method

technical field

The invention relates to the technical field of collision early warning and protection, in particular to a wearable human body collision early warning and protection device and an early warning and protection method thereof.

Background technique

Motion detection is about the data detection of the human body itself. The current research mainly focuses on athlete data measurement, fall detection, etc. The main means is to realize the application monitoring system through video or inertial measurement unit. Quick and effective acquisition of human body movement information can enable special groups to receive necessary assistance. However, human body collision detection, early warning and protection for pedestrians is still a blank. The road safety of children and special groups is a necessary link for safe production and social harmony. Avoiding pedestrian collisions has important economic and social effects.

With the development and application of sensor technology, from the initial single-channel sensor to modern multi-channel optical, electrical, acoustic and other sensors, the amount of data obtained correspondingly increases accordingly. However, any information acquisition technology has its accuracy and measurement limitations, and the signal is inevitably disturbed by the environment. For multiple sensors from the same scene, due to the different types of sensors and measurement principles, the scene itself changes and various Due to the existence of such interference, the obtained information is distorted or malfunctioned to varying degrees.

Contents of the invention

The primary purpose of the present invention is to provide a wearable human body collision warning and protection device that includes two-stage protection execution actions, is small in size, is easy to carry, and can be worn on the body.

In order to achieve the above object, the present invention adopts the following technical solutions: a wearable human body collision warning and protection device, including a belt body, a plurality of airbags are arranged at intervals in the middle of the belt body, and multiple airbags are installed on the belt body above and below the airbags. The sensor group, microsilicon accelerometer and gyroscope, alarm unit, the output terminals of multi-element sensor group, microsilicon accelerometer and gyroscope are connected with the input terminal of the main controller through the unit controller, and the output terminal of the main controller is connected with the input terminal of the main controller through the unit The controller is connected with the input ends of the air bag and the alarm unit respectively.

The multi-element sensor group is composed of an ultrasonic sensor, an air pressure sensor and an infrared sensor. The unit controller is composed of a distance measuring unit, an approach judgment unit, an instability judgment unit and an execution unit. The wireless information sending unit is installed at the bottom; the output end of the ultrasonic sensor is connected with the input end of the distance measuring unit, the output end of the infrared sensor is connected with the input end of the proximity judgment unit, and the output end of the air pressure sensor is respectively connected with the distance measurement unit and the proximity judgment unit The input terminals of the microsilicon accelerometer and gyroscope are connected with the input terminals of the instability judging unit, and the output terminals of the executive unit are respectively connected with the input terminals of the airbag, the alarm unit and the wireless information sending unit.

The alarm unit adopts a loudspeaker.

The number of the airbags is four to six, and each airbag is located on the same central line.

The output ends of the distance measuring unit, the proximity judging unit and the instability judging unit are all connected to the input end of the main controller, and the output end of the main controller is connected to the input end of the execution unit.

Another object of the present invention is to provide an early warning and protection method of a wearable human body collision early warning and protection device, which method includes the steps in the following order:

(1) Power-on initialization;

(2) The main controller collects the information output by the multi-sensor group, microsilicon accelerometer and gyroscope, and is in normal operation by default;

(3) The main controller extracts the feature quantity from the data sequence collected by the sensor, and judges whether a collision is about to occur according to the support vector machine SVM. If the judgment result is yes, it starts the alarm unit to give an alarm and continues to collect information; otherwise, returns step (2);

(4) After the alarm unit alarms, the main controller judges whether there has been an obvious collision state, if the judgment result is yes, then opens the airbag, and starts the wireless information sending unit for wireless transmission; if the judgment result is no, then judges whether For false alarm.

When judging the false alarm, if the judgment result is yes, it is in the normal operation state by default, otherwise, it is judged whether a collision situation is about to occur.

The judgment method for judging whether a collision situation is about to occur is as follows:

(1) Extract the sensor features of the multi-sensor group, including time domain features, frequency domain features and time trend features

a, Extraction of temporal features

The time-domain characteristics are the time-domain statistical characteristics obtained according to the output waveforms of various sensors, and the following indicators are used:

Absolute mean: $\left|\stackrel{-}{x}\right|=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left|x_i\right|$

RMS value: $x_{\mathrm{rms}}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left|x_i\right|}$

Skewness: $a=\sqrt{\frac{1}{6N}}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(\frac{x_i-\mathrm{\μ}}{\mathrm{\σ}}\right)}^3$

Kurtosis: $\mathrm{\β}=\sqrt{\frac{N}{\mathrm{twenty\; four}}}[\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\frac{x_i-\mathrm{\μ}}{\mathrm{\σ}}\right)}^4-3]$

In the formula, _xi is the discrete time series of the sensor signal, N is the number of samples in the sequence, μ is the mean value, and σ is the variance;

b, Extraction of frequency domain features

The output sequences of various sensors have their own typical frequency characteristics, and the frequency spectrum is analyzed to obtain the amplitude-frequency sequence, and the amplitude values at several special frequencies on the amplitude-frequency sequence are extracted;

c, Extraction of time trend features

The original signal measured by the sensor is the superposition of the definite state characteristic signal and random disturbance. Assuming that the signal signal sequence of the original measurement is {x _t ,t=1,2,...,N}, the sequence can be expressed as a sum of a trend item and a random item:

x _t =d _t +ξ _t

In the formula, d _t is a trend item, which is a certain function of time t; ξ _t is a random item, which reflects the random component of the signal;

When extracting trend features, it is first necessary to denoise the time series, remove random items, retain trend items, and then use various trend analysis methods to extract features;

The Mann-Kendall test is used to extract the change trend of the sensor detection sequence. For the time series {x _t ,t=1,2,...,N}, the null hypothesis H ₀ of the Mann-Kendall test is that the random variable is independent of time, Calculate the statistic T

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

在零假设H₀下，如果时间序列中不存在趋势，当N比较大时，则统计量T近似服从正态分布，且有下面两式成立：In the formula,

Under the null hypothesis H ₀ , if there is no trend in the time series, when N is relatively large, the statistic T approximately obeys the normal distribution, and the following two formulas hold:

When n>10, calculate the test statistic Z

If a certain time series is calculated to get Z——serial trend feature quantity;

(2) Establish the basic model of SVM for dangerous pattern classification based on feature quantity

Assuming that the feature set obtained from the sensor input is composed of two categories - impending collision and normal, if the feature x[i] corresponds to the first category, then y[i]=1, if the feature x[i] corresponds to the second category class, then y[i]=-1, then there is a training sample set {x[i], y[i]}, i=1,2,3,...,n, and find the optimal classification surface wx-b=0 , when the case of linear inseparability, use the kernel inner product K(x[i],x[j]) to map to the inner product of the corresponding vector in the high-dimensional space through the kernel function, instead of x[i]x[j];

The SVM classification training steps are as follows:

a, the training set is selected as: T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _l ,y _l )}∈(R ⁿ ×Y) ^l , where, x _i ∈ R ⁿ is a batch of corresponding sensor data features marked safe or dangerous, y _i ∈ Y={-1,1}, i=1,...,l; y _i =+1 means that a collision is about to occur, y _i =-1 indicates the numerical identification of the normal driving state; i corresponds to the i-th moment, each moment has the feature quantity obtained by a fixed number of sensing sequences that are constantly updated, and each moment has a decision-making quantity y _i generated, the first The time interval between i and i+1 moments; ? indicates the number of elements in the training set, through l training samples, get the support vector machine SVM, and then make a decision through the SVM;

b. Select an appropriate penalty parameter C>0 and a kernel function K(x,x'), where the kernel function can choose any of several typical kernel functions;

c, Construct and solve a convex quadratic programming problem: $\underset{a}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_i{\mathrm{the\; y}}_j{a}_i{a}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i$

0≤αi≤C得解

satisfy

0≤αi≤C to get the solution

计算d, calculate b ^* : select the component of α ^* located in the open interval (0, C)

calculate

${\mathrm{<span\; class="notranslate">\; b</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; *</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

e, Construct decision function F(x)=sgn(g(x)), where $g\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i^{*}{\mathrm{the\; y}}_{i}K(x_i,x_j)+b^{*};$

After the training is completed, bring the current sensor data feature quantity into the decision function, judge whether the current situation is dangerous or safe according to the calculation result of the decision function, if the decision function is positive, it means safety, if the decision function is negative , indicating that a collision is imminent.

The method of judging whether an obvious collision state has occurred is as follows: the time-domain features and frequency-domain features are obtained according to the sensor time series, and then the time-domain features and frequency-domain features are used to obtain the decision-making judgment through the support vector machine SVM, which is mainly based on the micro-silicon acceleration time series of meters and gyroscopes;

(1) Extract the time series features of microsilicon accelerometers and gyroscopes, including time domain features and frequency domain features

a, Extraction of temporal features

The time-domain characteristics are the time-domain statistical characteristics obtained according to the output waveforms of various sensors, and the following indicators are used:

Absolute mean: $\left|\stackrel{-}{x}\right|=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|$

RMS value: $x_{\mathrm{rms}}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|}$

Skewness: $a=\sqrt{\frac{1}{6N}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{x_i-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)}^3$

In the formula, _xi is the discrete time series of the sensor signal, N is the number of samples in the sequence, μ is the mean value, and σ is the variance;

b, Extraction of frequency domain features

The output sequences of various sensors have their own typical frequency characteristics, and the frequency spectrum analysis is performed on them to obtain the frequency domain distribution and corresponding amplitude;

(2) Establish the basic model of SVM for dangerous pattern classification based on feature quantity

Assuming that the feature set obtained from the sensor input is composed of two categories - impending collision and obvious collision, if the feature x[i] corresponds to the first category, then y[i]=1, if the feature x[i] corresponds to The second category, then y[i]=-1, then there is a training sample set {x[i], y[i]}, i=1,2,3,...,n, find the optimal classification surface wx-b =0, when the case of linear inseparability, use the kernel inner product K(x[i],x[j]), through the kernel function to map to the inner product of the corresponding vector in the high-dimensional space, instead of x[i]x[j ];

The SVM classification training steps are as follows:

a, the training set is selected as: T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _l ,y _l )}∈(R ⁿ ×Y) ^l , where, x _i ∈ R ⁿ is a batch of corresponding sensor data features marked safe or dangerous, y _i ∈ Y={-1,1}, i=1,...,l; y _i =+1 means that there is an obvious collision, y _i =-1 indicates the numerical identification of the impending collision; i corresponds to the i-th moment, and each moment has the feature quantity obtained by a fixed number of sensing sequences that are continuously updated, and each moment has a decision-making quantity y _i generated. The time interval between i and i+1 is the time interval; l represents the number of elements in the training set, and the support vector machine (SVM) is obtained through l training samples, and then the decision is made through the SVM;

b. Select an appropriate penalty parameter C>0 and a kernel function K(x,x'), where the kernel function can choose any of several typical kernel functions;

c, Construct and solve a convex quadratic programming problem: $\underset{a}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_i{\mathrm{the\; y}}_j{a}_i{a}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i$

satisfy 0≤αi≤C to get the solution a*aI*...aI*h

计算d, calculate b ^* : select the component of α ^* located in the open interval (0, C)

calculate

${\mathrm{<span\; class="notranslate">\; b</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; *</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

e, Construct decision function F(x)=sgn(g(x)), where $g\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i^{*}{\mathrm{the\; y}}_{i}K(x_i,x_j)+b^{*};$

After the training is completed, bring the current sensor data feature quantity into the decision function, and judge whether the current situation is dangerous or safe according to the calculation result of the decision function. If the decision function is positive, it means that a collision is about to occur. If the decision function is A negative value indicates a significant collision.

It can be seen from the above technical scheme that the present invention obtains distance measurement information through ultrasonic sensors and air pressure sensors, and obtains proximity judgment basis through infrared sensors and air pressure sensors, and obtains instability information through microsilicon accelerometers and gyroscopes; distance measurement unit and proximity judgment There is a cyclic interaction process in the unit, and the ranging information and approach information are reported to the main controller. By comparing with the empirical data set, the planning and decision-making algorithm designed in advance will give the number of suspicious approaching objects and the approaching distance, and issue execution commands to control The airbag is opened, and a sound alarm is issued. At the same time, the pre-set SMS information is sent to the designated mobile phone through the wireless information sending unit. The main controller judges the state according to the sensor data, and issues corresponding protective actions. The invention adopts the belt design, can be conveniently worn on the human body, has small volume and is easy to carry.

Description of drawings

Fig. 1 is the wearing schematic diagram of the present invention worn on the human body, the airbag is not opened;

2 and 3 are structural representations of the present invention respectively;

Fig. 4 is a schematic diagram of the state in which the airbag in Fig. 1 is opened;

Fig. 5 is a collection and execution block diagram of the present invention;

Fig. 6 is the overall structural block diagram of the present invention;

Fig. 7 is a working flowchart of the present invention.

Detailed ways

A wearable human body collision warning and protection device, comprising a belt body 1, a plurality of airbags 2 are arranged at intervals in the middle of the belt body 1, and a multi-element sensor group 3 and a microsilicon accelerometer are installed on the belt body 1 above and below the airbags 2 And gyroscope 4, alarm unit 5, the output end of multi-element sensor group 3, microsilicon accelerometer and gyroscope 4 are connected with the input end of main controller 8 through unit controller 7, the output end of main controller 8 is through unit The controller 7 is connected to the input terminals of the airbag 2 and the alarm unit 5 respectively. As shown in FIG. A variety of sensors, so that the coverage of various sensors and airbags 2 can be expanded to 360 degrees, to achieve all-round monitoring and protection.

As shown in Figures 2, 3 and 4, the number of the airbags 2 is four to six, and each airbag 2 is located on the same central line. The four protective airbags are independent of each other and not connected to each other, and are distributed on the belt body 1 at equal intervals. Arrange the airbag 2 on the top of the sensor, at the waist belt of the human body. When the airbag 2 is opened, it can protect the torso of the human body from collision. At the same time, an alarm unit and a wireless information sending unit are used. When the airbag 2 is about to open, the alarm unit will Just start working, meanwhile, the wireless information sending unit sends information to predetermined mobile phone.

As shown in Figures 5 and 6, the distance measurement information is obtained through the ultrasonic sensor and the air pressure sensor, the proximity judgment basis is obtained through the infrared sensor and the air pressure sensor, the instability information is obtained through the microsilicon accelerometer and the gyroscope 4, and the distance measurement unit and There is a cyclic interaction process between the proximity judging units, and the distance measurement information and proximity information are reported to the main controller 8. By comparing with the empirical data set, the planning and decision-making algorithm designed in advance will give the number of suspicious approaching objects and the approaching distance, and Issue the execution intention, issue specific execution commands to the execution unit according to the current state and execution intention, and control the actions of the airbag 2, the alarm unit 5 and the wireless information sending unit 6. The multi-sensor group 3 and the human body intention signal are used as the input of multi-sensor data fusion, and the state is judged through multi-sensor data fusion, and then an acoustic warning is generated, and two levels of airbag protection execution and wireless transmission protection execution are performed.

As shown in Figure 6, the multi-element sensor group 3 is composed of an ultrasonic sensor, an air pressure sensor and an infrared sensor, and the unit controller 7 is composed of a distance measuring unit, an approach judgment unit, an instability judgment unit and an execution unit. A wireless information sending unit 6 is installed above or below the airbag 2 on the belt body 1; the output end of the ultrasonic sensor is connected with the input end of the distance measuring unit, the output end of the infrared sensor is connected with the input end of the proximity judgment unit, and the air pressure sensor is connected with the input end of the proximity judgment unit. The output ends are respectively connected with the input ends of the ranging unit and the proximity judgment unit, the output ends of the microsilicon accelerometer and the gyroscope 4 are connected with the input ends of the instability judgment unit, and the output ends of the execution unit are respectively connected with the airbag 2, the alarm The input ends of unit 5 and the wireless information sending unit 6 are connected, the output ends of the distance measuring unit, the proximity judgment unit, and the instability judgment unit are all connected with the input end of the main controller 8, and the output end of the main controller 8 is connected with the execution connected to the input of the unit. The alarm unit 5 adopts a loudspeaker, and the voice alarm "please pay attention to avoiding" is accompanied by a specific identification sound. This protective device can be divided into three layers: the first layer is the sensing and execution power, which is mainly responsible for the sensing of real-time data, and the execution of three major tasks according to the state discrimination results; the second layer is the unit controller 7; The third layer is the main controller 8 .

As shown in Figure 7, when the device is working, its specific workflow is as follows:

The first step is power-on initialization;

In the second step, the main controller 8 collects the information output by the multi-element sensor group 3, the microsilicon accelerometer and the gyroscope 4, and is in a normal operation state by default;

In the third step, the main controller 8 extracts the feature quantity from the data sequence collected by the sensor, and judges whether a collision situation is about to occur according to the support vector machine SVM, if the judgment result is yes, then starts the alarm unit 5 to give an alarm, and continues to collect information; Otherwise, return to the second step;

In the 4th step, after the alarm unit 5 reports to the police, the main controller 8 judges whether an obvious collision state has occurred, if the judgment result is yes, then open the airbag 2, and start the wireless information sending unit 6 for wireless transmission; if the judgment result is If not, judge whether it is a false alarm. When judging the false alarm, if the judgment result is yes, it is in the normal operation state by default, otherwise, it is judged whether a collision situation is about to occur.

In other words, when the proximity information and ranging information conform to the empirical function or exceed the predetermined threshold, the system enters state 2. When the instability judgment is established and the abnormal proximity information and abnormal ranging information remain for a period of time, the system enters state 3. Among them, the first execution is acoustic alarm; the second execution is airbag opening and wireless transmission; the first state is normal operation; the second state is the early warning state of an impending collision; the third state is an obvious collision state.

The judgment method for judging whether a collision situation is about to occur is as follows:

(1) Extract the sensor features of the multi-sensor group, including time domain features, frequency domain features and time trend features

a, Extraction of temporal features

The time-domain characteristics are the time-domain statistical characteristics obtained according to the output waveforms of various sensors, and the following indicators are used:

Absolute mean:

RMS value: $x_{\mathrm{rms}}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|}$

Skewness: $a=\sqrt{\frac{1}{6N}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{x_i-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)}^3$

Kurtosis: $\mathrm{\&beta;}=\sqrt{\frac{N}{\mathrm{twenty\; four}}}[\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\frac{x_i-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)}^4-3]$

In the formula, _xi is the discrete time series of the sensor signal, N is the number of samples in the sequence, μ is the mean value, and σ is the variance;

b, Extraction of frequency domain features

The output sequences of various sensors have their own typical frequency characteristics, and the frequency spectrum is analyzed to obtain the amplitude-frequency sequence, and the amplitude values at several special frequencies on the amplitude-frequency sequence are extracted;

c, Extraction of time trend features

The original signal measured by the sensor is the superposition of the definite state characteristic signal and random disturbance. Assuming that the signal signal sequence of the original measurement is {x _t ,t=1,2,...,N}, the sequence can be expressed as a sum of a trend item and a random item:

x _t =d _t +ξ _t

In the formula, d _t is a trend item, which is a certain function of time t; ξ _t is a random item, which reflects the random component of the signal;

When extracting trend features, it is first necessary to denoise the time series, remove random items, retain trend items, and then use various trend analysis methods to extract features;

The Mann-Kendall test is used to extract the change trend of the sensor detection sequence. For the time series {x _t ,t=1,2,...,N}, the null hypothesis H ₀ of the Mann-Kendall test is that the random variable is independent of time, Calculate the statistic T

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

In the formula, Under the null hypothesis H ₀ , if there is no trend in the time series, when N is relatively large, the statistic T approximately obeys the normal distribution, and the following two formulas hold:

When n>10, calculate the test statistic Z

If a certain time series is calculated to get Z——serial trend feature quantity;

(2) Establish the basic model of SVM for dangerous pattern classification based on feature quantity

Assuming that the feature set obtained from the sensor input is composed of two categories - impending collision and normal, if the feature x[i] corresponds to the first category, then y[i]=1, if the feature x[i] corresponds to the second category class, then y[i]=-1, then there is a training sample set {x[i], y[i]}, i=1,2,3,...,n, and find the optimal classification surface wx-b=0 , when the case of linear inseparability, use the kernel inner product K(x[i],x[j]) to map to the inner product of the corresponding vector in the high-dimensional space through the kernel function, instead of x[i]x[j];

The SVM classification training steps are as follows:

a, the training set is selected as: T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _l ,y _l )}∈(R ⁿ ×Y) ^l , where, x _i ∈ R ⁿ is a batch of corresponding sensor data features marked safe or dangerous, y _i ∈ Y={-1,1}, i=1,...,l; y _i =+1 means that a collision is about to occur, y _i =-1 represents the numerical identification of the normal driving state; i corresponds to the i-th moment, each moment has a feature quantity obtained by a fixed number of sensing sequences that are constantly updated, and each moment has a decision-making quantity yi generated, the i-th The time interval between and i+1 is the time interval; l represents the number of elements in the training set, and the support vector machine SVM is obtained through l training samples, and then the decision-making judgment is made through the SVM;

b. Select an appropriate penalty parameter C>0 and a kernel function K(x,x'), where the kernel function can choose any of several typical kernel functions;

c, Construct and solve a convex quadratic programming problem: $\underset{a}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_i{\mathrm{the\; y}}_j{a}_i{a}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i$

0≤αi≤C得解

satisfy

0≤αi≤C to get the solution

计算d, calculate b ^* : select the component of α ^* located in the open interval (0, C)

calculate

${\mathrm{<span\; class="notranslate">\; b</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; *</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

e, Construct decision function F(x)=sgn(g(x)), where $g\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i^{*}{\mathrm{the\; y}}_{i}K(x_i,x_j)+b^{*};$

After the training is completed, bring the current sensor data feature quantity into the decision function, judge whether the current situation is dangerous or safe according to the calculation result of the decision function, if the decision function is positive, it means safety, if the decision function is negative , indicating that a collision is imminent.

The method of judging whether an obvious collision state has occurred is as follows: the time-domain features and frequency-domain features are obtained according to the sensor time series, and then the time-domain features and frequency-domain features are used to obtain the decision-making judgment through the support vector machine SVM, which is mainly based on the micro-silicon acceleration time series of meters and gyroscopes;

(1) Extract the time series features of microsilicon accelerometers and gyroscopes, including time domain features and frequency domain features

a, Extraction of temporal features

The time-domain characteristics are the time-domain statistical characteristics obtained according to the output waveforms of various sensors, and the following indicators are used:

Absolute mean: $\left|\stackrel{-}{x}\right|=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|$

RMS value: $x_{\mathrm{rms}}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|}$

Skewness: $a=\sqrt{\frac{1}{6N}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{x_i-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)}^3$

In the formula, _xi is the discrete time series of the sensor signal, N is the number of samples in the sequence, μ is the mean value, and σ is the variance;

b, Extraction of frequency domain features

The output sequences of various sensors have their own typical frequency characteristics, and the frequency spectrum analysis is performed on them to obtain the frequency domain distribution and corresponding amplitude;

(2) Establish the basic model of SVM for dangerous pattern classification based on feature quantity

Assuming that the feature set obtained from the sensor input is composed of two categories - impending collision and obvious collision, if the feature x[i] corresponds to the first category, then y[i]=1, if the feature x[i] corresponds to The second category, then y[i]=-1, then there is a training sample set {x[i], y[i]}, i=1,2,3,...,n, find the optimal classification surface wx-b =0, when the case of linear inseparability, use the kernel inner product K(x[i],x[j]), through the kernel function to map to the inner product of the corresponding vector in the high-dimensional space, instead of x[i]x[j ];

The SVM classification training steps are as follows:

a. The training set is selected as: T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…，(x _l ,y _l )}∈(R ⁿ ×Y) ^l , where, x _i ∈ R ⁿ is a batch of corresponding sensor data features marked safe or dangerous, y _i ∈ Y={-1,1}, i=1,...,l; y _i =+1 means that there is an obvious collision, y _i =-1 indicates the numerical identification of the impending collision; i corresponds to the i-th moment, and each moment has the feature quantity obtained by a fixed number of sensing sequences that are continuously updated, and each moment has a decision-making quantity y _i generated. The time interval between i and i+1 is the time interval; l represents the number of elements in the training set, and the support vector machine (SVM) is obtained through l training samples, and then the decision is made through the SVM;

b. Select an appropriate penalty parameter C>0 and a kernel function K(x,x'), where the kernel function can choose any of several typical kernel functions;

c, Construct and solve a convex quadratic programming problem: $\underset{a}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_i{\mathrm{the\; y}}_j{a}_i{a}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i$

0≤αi≤C得解

satisfy

0≤αi≤C to get the solution

计算d, calculate b ^* : select the component of α ^* located in the open interval (0, C)

calculate

${\mathrm{<span\; class="notranslate">\; b</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; *</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

e, Construct decision function F(x)=sgn(g(x)), where $g\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i^{*}{\mathrm{the\; y}}_{i}K(x_i,x_j)+b^{*};$

After the training is completed, bring the current sensor data feature quantity into the decision function, and judge whether the current situation is dangerous or safe according to the calculation result of the decision function. If the decision function is positive, it means that a collision is about to occur. If the decision function is Negative values indicate a significant collision

In a word, the present invention obtains the ranging information through the ultrasonic sensor and the air pressure sensor, the infrared sensor and the air pressure sensor obtain the proximity judgment basis, and obtain the instability information through the microsilicon accelerometer and the gyroscope 4; the distance measuring unit and the proximity judgment unit have a loop During the interaction process, the ranging information and approach information are reported to the main controller 8, and compared with the empirical data set, the planning and decision-making algorithm designed in advance will give the number of suspicious approaching objects and the approaching distance, and issue an execution command to control the airbag 2 Open, and give a sound alarm, and at the same time send the preset SMS information to the designated mobile phone through the wireless information sending unit, and the two kinds of protection execution actions are judged and decided by the main controller according to the on-site situation. The invention adopts the belt design, can be conveniently worn on the human body, has small volume and is easy to carry.

Claims (9)

1. A wearable human body collision warning and protection device, characterized in that: it comprises a belt body, a plurality of airbags are arranged at intervals in the middle of the belt body, and multi-element sensor groups and microsilicon accelerometers are installed on the belt body above and below the airbags And gyroscope, alarm unit, multi-element sensor group, microsilicon accelerometer and output end of gyroscope are connected with the input end of the main controller through the unit controller, and the output end of the main controller is respectively connected with the airbag, The input terminal of the alarm unit is connected. 2. The wearable human body collision warning and protection device according to claim 1, characterized in that: the multi-element sensor group is composed of an ultrasonic sensor, an air pressure sensor and an infrared sensor, and the unit controller is composed of a distance measuring unit, a proximity judgment unit, an instability judging unit and an execution unit. The belt body is located above or below the airbag to install a wireless information sending unit; the output of the ultrasonic sensor is connected to the input of the distance measuring unit, and the output of the infrared sensor is connected to the proximity judgment The input end of the unit is connected, the output end of the air pressure sensor is connected with the input end of the ranging unit and the proximity judgment unit, the output end of the microsilicon accelerometer and the gyroscope are connected with the input end of the instability judgment unit, and the execution unit The output ends of the airbag, the alarm unit, and the input ends of the wireless information sending unit are respectively connected. 3. The wearable human body collision warning and protection device according to claim 1, wherein the alarm unit adopts a loudspeaker. 4. The wearable human body collision warning and protection device according to claim 1, wherein the number of the airbags is four to six, and each airbag is located on the same center line. 5. The wearable human body collision warning and protection device according to claim 2, characterized in that: the output ends of the distance measuring unit, the proximity judgment unit, and the instability judgment unit are all connected to the input end of the main controller, and the main controller The output of the controller is connected to the input of the execution unit. 6. The early warning protection method of the wearable human body collision early warning protection device according to any one of claims 1 to 5, the method comprising the steps in the following order: (1) Power-on initialization; (2) The main controller collects the information output by the multi-sensor group, microsilicon accelerometer and gyroscope, and is in normal operation by default; (3) The main controller extracts the feature quantity from the data sequence collected by the sensor, and judges whether a collision is about to occur according to the support vector machine SVM. If the judgment result is yes, it starts the alarm unit to give an alarm and continues to collect information; otherwise, returns step (2); (4) After the alarm unit alarms, the main controller judges whether there has been an obvious collision state, if the judgment result is yes, then opens the airbag, and starts the wireless information sending unit for wireless transmission; if the judgment result is no, then judges whether For false alarm. 7. The early warning and protection method of the wearable human body collision early warning protection device according to claim 6, characterized in that: when judging the false alarm, if the judgment result is yes, then it is in the normal operation state by default; otherwise, it is judged whether A collision situation is imminent. 8. The early warning protection method of the wearable human body collision early warning protection device according to claim 6, characterized in that: the judging method for judging whether a collision situation is about to occur is specifically: (1) Extract the sensor features of the multi-sensor group, including time domain features, frequency domain features and time trend features a, Extraction of temporal features The time-domain characteristics are the time-domain statistical characteristics obtained according to the output waveforms of various sensors, and the following indicators are used: Absolute mean: $\stackrel{-}{x}\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|$ RMS value: $x_{\mathrm{rms}}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|}$ Skewness: $a=\sqrt{\frac{1}{6N}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{x_i-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)}^3$ Kurtosis: $\mathrm{\&beta;}=\sqrt{\frac{N}{\mathrm{twenty\; four}}}[\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\frac{x_i-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)}^4-3]$ In the formula, _xi is the discrete time series of the sensor signal, N is the number of samples in the sequence, μ is the mean value, and σ is the variance; b, Extraction of frequency domain features The output sequences of various sensors have their own typical frequency characteristics, and the frequency spectrum is analyzed to obtain the amplitude-frequency sequence, and the amplitude values at several special frequencies on the amplitude-frequency sequence are extracted; c, Extraction of time trend features The original signal measured by the sensor is the superposition of the definite state characteristic signal and random disturbance. Assuming that the signal signal sequence of the original measurement is {x _t ,t=1,2,...,N}, the sequence can be expressed as a sum of a trend item and a random item: x _t =d _t +ξ _t In the formula, d _t is a trend item, which is a certain function of time t; ξ _t is a random item, which reflects the random component of the signal; When extracting trend features, it is first necessary to denoise the time series, remove random items, retain trend items, and then use various trend analysis methods to extract features; The Mann-Kendall test is used to extract the change trend of the sensor detection sequence. For the time series {x _t ,t=1,2,...,N}, the null hypothesis H ₀ of the Mann-Kendall test is that the random variable is independent of time, Calculate the statistic T $\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$ In the formula,

Under the null hypothesis H ₀ , if there is no trend in the time series, when N is relatively large, the statistic T approximately obeys the normal distribution, and the following two formulas hold: When n>10, calculate the test statistic Z

If a certain time series is calculated to get Z——serial trend feature quantity; (2) Establish the basic model of SVM for dangerous pattern classification based on feature quantity Assuming that the feature set obtained from the sensor input is composed of two categories - impending collision and normal, if the feature x[i] corresponds to the first category, then y[i]=1, if the feature x[i] corresponds to the second category class, then y[i]=-1, then there is a training sample set {x[i], y[i]}, i=1,2,3,...,n, and find the optimal classification surface wx-b=0 , when the case of linear inseparability, use the kernel inner product K(x[i],x[j]) to map to the inner product of the corresponding vector in the high-dimensional space through the kernel function, instead of x[i]x[j]; The SVM classification training steps are as follows: a, the training set is selected as: T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _l ,y _l )}∈(R ⁿ ×Y) ^l , where, x _i ∈ R ⁿ is a batch of corresponding sensor data features marked safe or dangerous, y _i ∈ Y={-1,1}, i=1,...,l; y _i =+1 means that a collision is about to occur, y _i =-1 indicates the numerical identification of the normal driving state; i corresponds to the i-th moment, each moment has the feature quantity obtained by a fixed number of sensing sequences that are constantly updated, and each moment has a decision-making quantity y _i generated, the first The time interval between i and i+1 is the time interval; l represents the number of elements in the training set, and the support vector machine (SVM) is obtained through l training samples, and then the decision is made through the SVM; b. Select an appropriate penalty parameter C>0 and a kernel function K(x,x'), where the kernel function can choose any of several typical kernel functions; c, Construct and solve a convex quadratic programming problem: $\underset{a}{\min }\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_i{\mathrm{the\; y}}_j{a}_i{a}_{j}K(x_i,x_j)-\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i$ satisfy

0≤α _i ≤C to get the solution

d, calculate b ^* : select the component of α ^* located in the open interval (0, C)

calculate ${\mathrm{<span\; class="notranslate">\; b</span>}}^{<span\; class="notranslate">\; *</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; *</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$ e, Construct decision function F(x)=sgn(g(x)), where $g\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{a}_i^{*}{\mathrm{the\; y}}_{i}K(x_i,x_j)+b^{*};$ After the training is completed, bring the current sensor data feature quantity into the decision function, judge whether the current situation is dangerous or safe according to the calculation result of the decision function, if the decision function is positive, it means safety, if the decision function is negative , indicating that a collision is imminent. 9. The early warning and protection method of the wearable human body collision early warning protection device according to claim 6, characterized in that: the method for judging whether an obvious collision state has occurred is specifically: obtaining time domain features and frequency domain features according to the sensor time series , and then use the time-domain features and frequency-domain features to obtain decision-making judgments through the support vector machine SVM, which is mainly based on the time series of micro-silicon accelerometers and gyroscopes; (1) Extract the time series features of microsilicon accelerometers and gyroscopes, including time domain features and frequency domain features a, Extraction of temporal features The time-domain characteristics are the time-domain statistical characteristics obtained according to the output waveforms of various sensors, and the following indicators are used: Absolute mean: $\left|\stackrel{-}{x}\right|=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|$ RMS value: $x_{\mathrm{rms}}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|x_i\right|}$ Skewness: $a=\sqrt{\frac{1}{6N}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{x_i-\mathrm{\&mu;}}{\mathrm{\&sigma;}}\right)}^3$ In the formula, _xi is the discrete time series of the sensor signal, N is the number of samples in the sequence, μ is the mean value, and σ is the variance; b, Extraction of frequency domain features The output sequences of various sensors have their own typical frequency characteristics, and the frequency spectrum analysis is performed on them to obtain the frequency domain distribution and corresponding amplitude; (2) Establish the basic model of SVM for dangerous pattern classification based on feature quantity Assuming that the feature set obtained from the sensor input consists of two categories - impending collision and obvious collision, if the feature x[i] corresponds to the first category, then y[i]=1, if the feature x[i] corresponds to The second category, then y[i]=-1, then there is a training sample set {x[i], y[i]}, i=1,2,3,...,n, find the optimal classification surface wx-b =0, when the case of linear inseparability, use the kernel inner product K(x[i],x[j]), through the kernel function to map to the inner product of the corresponding vector in the high-dimensional space, instead of x[i]x[j ]; The SVM classification training steps are as follows: a, the training set is selected as: T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _l ,y _l )}∈(R ⁿ ×Y) ^l , where, x _i ∈ R ⁿ is a batch of corresponding sensor data features marked safe or dangerous, y _i ∈ Y={-1,1}, i=1,...,l; y _i =+1 means that there is an obvious collision, y _i =-1 indicates the numerical identification of the impending collision; i corresponds to the i-th moment, and each moment has the feature quantity obtained by a fixed number of sensing sequences that are continuously updated, and each moment has a decision-making quantity y _i generated. The time interval between i and i+1 is the time interval; l represents the number of elements in the training set, and the support vector machine (SVM) is obtained through l training samples, and then the decision is made through the SVM; b. Select an appropriate penalty parameter C>0 and a kernel function K(x,x'), where the kernel function can choose any of several typical kernel functions; c, Construct and solve a convex quadratic programming problem:

satisfy 0≤α _i ≤C to get the solution

d, calculate b ^* : select the component of α ^* located in the open interval (0, C) calculate

e, Construct decision function F(x)=sgn(g(x)), where

After the training is completed, bring the current sensor data feature quantity into the decision function, and judge whether the current situation is dangerous or safe according to the calculation result of the decision function. If the decision function is positive, it means that a collision is about to occur. If the decision function is A negative value indicates a significant collision.

Images