CN115035686A - Real-time falling detection method, system and medium based on channel state information - Google Patents

Abstract

The invention relates to the field of fall detection, and particularly discloses a real-time fall detection method, a system and a medium based on channel state information, wherein the method comprises the steps of acquiring CS I data streams in real time and acquiring effective carriers from the CS I data streams; sequentially carrying out matrix extraction on the effective carrier waves corresponding to each data packet, and carrying out sliding window interception on the effective carrier waves according to the amplitude disturbance judgment condition to obtain effective action fragments; when an effective action fragment is intercepted, carrying out noise reduction on an effective carrier in a current window by utilizing wavelet transformation, and filtering by utilizing a Butterworth filter to obtain an initial available signal; extracting characteristic values of the initial available signals to obtain signal characteristics; normalizing the signal characteristics to obtain a characteristic value matrix; and importing the characteristic value matrix into an SVM classification model to obtain an action classification result. The invention has high detection efficiency, high detection accuracy and less occupation of operation resources, thereby reducing the equipment cost.

Description

Real-time fall detection method, system and medium based on channel state information

technical field

The invention relates to the field of fall detection, in particular to a real-time fall detection method, system and medium based on channel state information.

Background technique

As the aging process of China's population continues to accelerate, the number of empty nesters is gradually increasing, and the health of the elderly has become a hot spot of social concern. Falls are one of the main causes of injury-related death among the elderly, and the number of indoor falls for the elderly living alone is increasing every year; studies have shown that medical outcomes after a fall largely depend on the timely response and rescue time. In clinical medicine, the longer the delay in treatment after a fall, the greater the risk of death. Therefore, how to realize automatic and real-time fall detection technology has become an important requirement for elderly health monitoring.

So far, health detection methods for daily life have become commonplace, and various applications have been widely developed. There are three development directions for fall detection technology: sensor-based fall detection, mechanical vision-based fall detection, and WiFi signal-based channel state information (CSI) fall detection. Sensor-based fall detection technology has high accuracy in ideal conditions, but it is easily affected by the environment, has relatively low accuracy, and cannot be widely deployed; although vision-based fall detection technology can achieve high accuracy, its High cost, easy to be affected by factors such as light, camera position, background, etc., and will increase the risk of privacy leakage; the fall detection method based on WiFi-CSI has strong privacy protection, wide application base, and does not need to carry any sensors , non-visual perception, not affected by light, humidity, temperature, etc., and strong scalability. In recent years, the Internet of Things technology has developed rapidly and has been widely used, and the popularity of WiFi has been greatly improved, providing the realization environment and equipment support for the WiFi-based fall detection system.

However, the existing WiFi-CSI-based fall detection solutions, although capable of judging the occurrence of falls, lack the technology for real-time detection, the detection timeliness has a lot of room for improvement, and the computing resources occupied are also relatively high. In order to apply this technology to products, there is a lack of real-time application technology. At the same time, it is still necessary to improve detection accuracy, accuracy and computing speed, and reduce hardware deployment costs and energy consumption.

SUMMARY OF THE INVENTION

In order to overcome the above problems, the present invention provides a real-time fall detection method, system and medium based on channel state information.

The present invention provides a real-time fall detection method based on channel state information, including:

Use the TCP protocol to connect the wifi signal sending device, and obtain the CSI data stream in real time through the wifi signal receiving device; wherein, the data stream includes several continuous progressive data packets;

Analyze the data stream to obtain a valid carrier from it;

Preset amplitude disturbance judgment conditions; perform matrix extraction on the effective carrier corresponding to each data packet in turn to obtain matrix parameters corresponding to each data packet, and perform sliding window interception on the effective carrier according to the amplitude disturbance judgment condition to obtain A valid action segment; wherein, the amplitude disturbance judgment condition is used to judge whether the sliding window has several matrix parameters including corresponding human actions;

When the effective action segment is intercepted, the effective carrier in the current window is denoised by wavelet transform, and filtered by Butterworth filter to obtain the initial usable signal;

Extract the eigenvalues of the initial available signal to obtain the signal features;

Normalize the signal features to obtain an eigenvalue matrix;

Import the eigenvalue matrix into the SVM classification model to obtain the action classification result;

Determine whether the action classification result is a fall, and if so, send an alarm message.

Preferably, the signal features include: time-domain mean, time-domain standard deviation, time-domain maximum value, time-domain minimum value, time-domain range, number of over-average points, time-domain 1/4 quantile, time-domain 3/ Quartile and time domain quartile range, the first largest FFT, the frequency corresponding to the first largest FFT, the third largest FFT, the frequency corresponding to the third largest FFT, the fifth largest FFT, the corresponding frequency of the fifth largest FFT Frequency, Frequency Domain Mean, Frequency Domain Standard Deviation, Frequency Domain 1/4 Quantile, Frequency Domain 3/4 Quantile, Frequency Domain Interquartile Range, Amplitude Statistics Skewness, Amplitude Statistics Kurtosis, Shape Statistics Mean, shape statistic standard deviation, shape statistic skewness, and shape statistic kurtosis.

Preferably, the feature value extraction is performed on the initially available signal to obtain the signal feature, specifically:

Wavelet transform is performed on the initially available signal again to obtain two wavelet signals, and feature value extraction is performed on the two wavelet signals to obtain signal features;

Wherein, the signal features include the second time-domain mean value, the second time-domain standard deviation, the second time-domain range, the second time-domain over-average points, and the time-domain second quartile obtained from one of the wavelet signals range; and the third time-domain mean value, the third time-domain standard deviation, the third time-domain range, the third time-domain over-average points, and the third time-domain quartile range extracted for another wavelet signal.

Preferably, the TCP protocol is used to connect the wifi signal sending device, and the CSI data stream is obtained in real time through the wifi signal receiving device; wherein, the data stream includes several consecutive progressive data packets; specifically:

Set at least one wifi signal sending device and m wifi signal receiving devices in the space to be tested, and simultaneously obtain data sent by at least one of the wifi signal sending devices through the m wifi signal receiving devices, and obtain 1 dat file; Among them, m is a positive integer;

Extract the csi data stream from the acquired dat file; wherein, the data stream includes several consecutive progressive data packets, and each data packet corresponds to a 1×m×30 subcarrier matrix.

Preferably, the analysis of the data stream to obtain a valid carrier therefrom is specifically:

Every preset time interval, compare the sensitivities of 1×m×30 sub-carriers to the action, and select the sub-carrier that is most sensitive to the action as the effective carrier.

Preferably, the preset amplitude disturbance judgment condition; perform matrix extraction on the effective carrier corresponding to each data packet in turn to obtain the matrix parameter corresponding to each data packet, and according to the amplitude disturbance judgment condition, perform a matrix extraction on the effective carrier The sliding window is intercepted to obtain valid action segments; wherein, the amplitude disturbance judgment condition is used to judge whether the sliding window has several matrix parameters including corresponding human actions; specifically:

Preset amplitude disturbance judgment conditions, where the amplitude disturbance judgment conditions include a segment start condition and an amplitude disturbance threshold;

Perform matrix extraction on the effective carrier corresponding to each data packet in turn to obtain matrix parameters corresponding to each data packet; wherein, the amplitude disturbance judgment condition is used to judge whether there are several matrix parameters containing corresponding human actions in the sliding window ;

According to the segment start condition, intercept the first action segment in the data stream through the sliding window;

It is detected in real time whether the first action segment is greater than the amplitude disturbance threshold, and if so, it is output as a valid action segment.

Preferably, the amplitude disturbance judgment condition is preset; the effective carrier corresponding to each data packet is sequentially extracted by matrix to obtain the matrix parameter corresponding to each data packet, and a sliding window is performed on the effective carrier according to the amplitude disturbance judgment condition Intercept to obtain valid action segments; wherein, the amplitude disturbance judgment condition is used to judge whether the sliding window has several matrix parameters containing corresponding human actions, and also includes:

Collect data samples of different actions performed by the human body at different spatial positions, including at least one data sample of falling and at least one data sample of non-falling;

Mark the data samples, distinguish the amplitude disturbance of the falling data samples and the non-falling data samples, and obtain the amplitude disturbance judgment condition.

The present invention also provides a real-time fall detection device based on channel state information, including: a data acquisition module, a carrier acquisition module, an action detection module, a first data module, a second data module, a third data module, a classification module and an alarm module ;

The data acquisition module is used to connect the wifi signal sending device using the TCP protocol, and obtain the data stream of the CSI in real time through the wifi signal receiving device; wherein, the data stream includes several continuous progressive data packets;

The carrier acquisition module is used to analyze the data flow, and obtain a valid carrier therefrom;

The motion detection module is used to preset the amplitude disturbance judgment condition; perform matrix extraction on the effective carrier corresponding to each data packet in turn, and obtain the matrix parameter corresponding to each data packet, and according to the amplitude disturbance judgment condition, the effective carrier is obtained. The carrier is intercepted by a sliding window to obtain valid action segments; wherein, the amplitude disturbance judgment condition is used to judge whether the sliding window has several matrix parameters that include corresponding human actions;

The first data module is used for denoising the effective carrier in the current window by using wavelet transform and filtering by using Butterworth filter to obtain an initial available signal when a valid action segment is intercepted;

The second data module is used to perform feature value extraction on the initially available signal to obtain signal features;

The third data module is used to normalize the signal features to obtain an eigenvalue matrix;

The classification module is used to import the eigenvalue matrix into the SVM classification model to obtain the action classification result;

The alarm module is used for judging whether the action classification result is a fall, and if so, sending an alarm message.

The present invention provides a computer-readable storage medium, where the computer-readable storage medium includes a stored computer program, wherein when the computer program runs, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned channel state information-based real-time fall detection method.

The beneficial effects of the present invention are:

(1) By obtaining the CSI data stream in real time, and according to the amplitude disturbance judgment condition, the sliding window is used to intercept the effective action segments, and then put into the SVM classification model for analysis, and the current human body state results are obtained, which can detect falls in real time. , the data can be analyzed in real time, and conclusions can be drawn and output in a very short time.

(2) Real-time uninterrupted detection is carried out through the tcp protocol connection channel to realize real-time rapid fall detection. When there is a small amount of packet loss on the link, the tcp protocol can automatically retransmit without intervention, which helps to ensure the accuracy of information. Improve detection efficiency.

(3) Using the amplitude disturbance condition to judge the action generation can effectively segment the action segment, improve the detection efficiency, reduce the occupation of computing resources, and thus reduce the equipment cost.

Preferably, a variety of eigenvalues in the time domain and frequency domain are used for judgment, and 36 eigenvalues based on the time domain and frequency domain are obtained as the best judgment basis by inputting the SVM classification model to try, so that the classification accuracy rate is greatly improved. improve the detection accuracy.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings, wherein:

1 is a flow chart of a method according to an embodiment of the present invention;

2 is a schematic diagram of a data waveform after wavelet transformation according to one embodiment of the present invention;

FIG. 3 is a CSI amplitude diagram when bending over according to one embodiment of the present invention;

FIG. 4 is a CSI amplitude diagram when one embodiment of the present invention falls;

5 is a table of human motion frequency parameters according to one embodiment of the present invention;

FIG. 6 is a schematic diagram of a collection site according to another embodiment of the present invention.

In the figure: 1. Sending antenna; 2. Receiving antenna; 3. Data collection point.

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

Referring to FIG. 1 to FIG. 5, as one of the implementations of the present invention, a real-time fall detection method based on channel state information is disclosed, including:

S1, use the TCP protocol to connect the wifi signal sending device, and obtain the data stream of the CSI in real time through the wifi signal receiving device; wherein, the data stream includes several continuous progressive data packets;

S2, analyze the data stream, and obtain a valid carrier from it;

S3. Preset the amplitude disturbance judgment condition; perform matrix extraction on the effective carrier corresponding to each data packet in turn to obtain the matrix parameter corresponding to each data packet, and perform sliding window interception on the effective carrier according to the amplitude disturbance judgment condition , to obtain a valid action segment; wherein, the amplitude disturbance judgment condition is used to judge whether the sliding window has several matrix parameters including corresponding human actions;

S4. When an effective action segment is intercepted, use wavelet transform to denoise the effective carrier in the current window, and use Butterworth filter to filter to obtain an initial available signal;

S5, extract the characteristic value of the initial available signal to obtain the signal characteristic;

S6, normalize the signal features to obtain an eigenvalue matrix;

S7, import the eigenvalue matrix into the SVM classification model to obtain the action classification result;

S8. Determine whether the action classification result is a fall, and if so, issue an alarm message.

Preferably, the signal features include: time-domain mean, time-domain standard deviation, time-domain maximum value, time-domain minimum value, time-domain range, number of over-average points, time-domain 1/4 quantile, time-domain 3/ Quartile and time domain quartile range, the first largest FFT, the frequency corresponding to the first largest FFT, the third largest FFT, the frequency corresponding to the third largest FFT, the fifth largest FFT, the corresponding frequency of the fifth largest FFT Frequency, Frequency Domain Mean, Frequency Domain Standard Deviation, Frequency Domain 1/4 Quantile, Frequency Domain 3/4 Quantile, Frequency Domain Interquartile Range, Amplitude Statistics Skewness, Amplitude Statistics Kurtosis, Shape Statistics Mean, shape statistic standard deviation, shape statistic skewness, and shape statistic kurtosis.

Preferably, in step S4, the sub-steps are as follows:

S41. Perform wavelet transformation on the initially available signal again to obtain two wavelet signals, and perform feature value extraction on the two wavelet signals to obtain signal features;

Wherein, the signal features include the second time-domain mean value, the second time-domain standard deviation, the second time-domain range, the second time-domain over-average points, and the time-domain second quartile obtained from one of the wavelet signals range; and the third time-domain mean value, the third time-domain standard deviation, the third time-domain range, the third time-domain over-average points, and the third time-domain quartile range extracted for another wavelet signal.

Referring to Figure 2, the first waveform from top to bottom in the figure is the waveform of the original data (without any processing); the second waveform is the waveform of the processed data (after wavelet transform noise reduction, after filter filtering); The third waveform is the low-frequency component signal of the second waveform, that is, the wavelet signal s(t)1 in the two wavelet signals; the fourth waveform is the high-frequency component signal of the second waveform, that is, the wavelet signal s(t)1 in the two wavelet signals. The wavelet signal s(t)2.

Aiming at the above two points, it is one of the innovative points of the present invention. It is to perform wavelet transform and noise reduction on the original CSI signal and filter it with a Butterworth filter to obtain the initial usable signal s(t). There are two methods for this signal. Feature extraction:

The first method is to directly use the initial available signal s(t) to extract the eigenvalues in the time domain and frequency domain, and obtain a total of 26 eigenvalues;

The second method is to perform wavelet transform on the initial available signal s(t) again, and decompose the initial available signal s(t) into two wavelet signals s(t)1 and s(t)2. The signal is extracted again in time domain, including mean, standard deviation, range, over-average points, and interquartile range. These five eigenvalues can only show signal characteristics, and there are 10 eigenvalues in total.

The above two methods have a total of 36 eigenvalues, both of which can perfectly display the characteristics of the signal, thereby improving the accuracy of fall recognition. Among them, the specific definition of each eigenvalue is as follows:

The time domain mean is the average characteristic of the data signal in the time window, and the average value of multiple samplings is taken as the real value of the data to reflect its basic characteristics. It mainly reflects the static characteristics of the received signal, and can distinguish the approximate position information of the human body; its calculation formula is:

Among them, n represents the size of a time window, and a _i represents the magnitude of the ith subcarrier, which corresponds to the available function mean() in Matlab.

The time-domain standard deviation refers to all the numbers in the data set minus the sum of the squares of the averages of these numbers, then divide the result by the total number, and then take the square root to get the standard deviation. The standard deviation reflects the spread of the mean in the data set. A large standard deviation indicates that the numbers are more different from their mean; a small standard deviation means that the numbers are less different from their mean. It can reflect the fluctuation of the received signal and estimate the change of human motion near the receiving end; its calculation formula is as follows:

It corresponds to the function available in Matlab: std().

The time domain maximum value and the time domain minimum value represent the maximum value and the minimum value in the data set respectively; the time domain range represents the difference between the maximum value and the minimum value, which can indicate the degree of signal change. The temporal maximum value and the temporal minimum value can better distinguish the changes of human motion.

Among them, the maximum value in the time domain: max=max(a _i ), i∈{1, 2,...,n}, corresponding to the function max();

The minimum value in the time domain: min=min(a _i ), i∈{1, 2,...,n}, corresponding to the function min();

Time domain range: range=|max-min|, corresponding function: abs(max()-min());

The number of over-average points refers to the number of data that exceeds the average point in a window. The calculation formula is as follows:

是指示函数(indicator function)，当括号里条件成立时取值为1，否则为0。in,

is an indicator function, which takes the value 1 when the condition in the parentheses holds, and 0 otherwise.

Time Domain 1/4 Quantile, Time Domain 3/4 Quantile, Time Domain Interquartile Range, Frequency Domain 1/4 Quantile, Frequency Domain 3/4 Quantile, and Frequency Domain Interquartile range, where the 1/4 quantile, 3/4 quantile, and interquartile range in the time domain and the frequency domain have the same meaning.

The above-mentioned quartile parameters all refer to the degree of dispersion of skewed data when describing the data through the quartile statistical description and analysis method, that is, arrange all the data from small to large, and the number that is exactly arranged in the lower 1/4 position is called the lower four. The quantile (according to the percentage, that is, the number in the 25% position) is also called the first quartile, and the number in the upper 1/4 position is called the upper quartile (according to the percentage ratio, that is, 75 The number in the % position) is also called the third quartile, and the interquartile range is the difference between the upper and lower quartiles. Exactly 50% of the data is contained between the upper and lower quartiles.

Quartile Characteristics:

The interquartile range can only describe the dispersion degree of the middle 50% of the data, and it still cannot fully reflect the dispersion of all the data. The larger the interquartile range, the greater the dispersion of the data in the middle 50%; the smaller the interquartile range, the smaller the dispersion of the data in the middle 50%; to a certain extent, the interquartile range can also reflect the median The representativeness of the number; the interquartile range is an order statistic, so the interquartile range is suitable for measuring the degree of dispersion of ordinal and quantitative data.

The first largest FFT, the frequency corresponding to the first largest FFT, the third largest FFT, the frequency corresponding to the third largest FFT, the fifth largest FFT and the frequency corresponding to the fifth largest FFT:

Convert the csi time domain signal in an action time window to the csi signal information in the frequency domain through fast Fourier transform (FFT), and extract the first five maximum FFT values of the signal and the frequencies corresponding to these five values. During the experiment, it was found that the second largest FFT and the third largest FFT, the fourth largest and the fifth largest FFT in the frequency domain have exactly the same characteristic change values between the two groups of eigenvalues. Therefore, only the second largest FFT and the third largest FFT are taken. One of the three major FFTs, one of the fourth and fifth largest FFTs.

For example, there are 8 numbers: 1, 2, 3, 4, 5, 6, 7, 8;

After FFT, the following 8 numbers are obtained:

36.0000+0.0000i;-4.0000+9.6569i;-4.0000+4.0000i;

-4.0000+1.6569i;-4.0000+0.0000i;-4.0000-1.6569i;

-4.0000-4.0000i; -4.0000-9.6569i.

It can be seen from these 8 numbers that the first number is the largest, and the remaining 7 numbers are symmetrical with the fifth number as the center. This is determined by the Fourier transform, so except for the first number alone, the remaining numbers can be taken in half.

The following is the power spectral density, which describes the energy distribution of the data in the frequency domain; the power spectral density is further divided into two types of characteristics: amplitude statistical characteristics and shape statistical characteristics

The first is the amplitude statistical feature. Let C(i) be the frequency amplitude value of the i-th window, and N represents the number of windows. The calculation methods of several quantities of the amplitude statistical feature are as follows:

Frequency Domain Mean (Amplitude Statistical Feature Mean):

Frequency Domain Standard Deviation (Standard Deviation of Amplitude Statistics):

Amplitude statistical skewness:

Amplitude statistical kurtosis:

则形状统计特征的几个量计算方式如下The following are the shape statistics, let C(i) be the frequency amplitude value of the i-th window, N is the number of windows,

Then several quantities of shape statistical features are calculated as follows

Shape Statistics Mean:

Shape Statistics Standard Deviation:

Shape Statistical Skewness:

Shape statistic kurtosis:

Preferably, in step S1, the sub-steps are as follows:

S11. Set at least one wifi signal sending device and m wifi signal receiving devices in the space to be tested, and simultaneously obtain data sent by at least one of the wifi signal sending devices through the m wifi signal receiving devices, and obtain 1 dat file; where m is a positive integer;

S12. Extract the csi data stream from the acquired dat file; wherein, the data stream includes several consecutive progressive data packets, and each data packet corresponds to a 1×m×30 subcarrier matrix.

Preferably, in step S2, the sub-steps are as follows:

S21. Compare the sensitivities of 1×m×30 sub-carriers to the action at every preset time interval, and select the sub-carrier that is most sensitive to the action as an effective carrier.

Preferably, in step S3, the sub-steps are as follows:

S31, preset an amplitude disturbance judgment condition, where the amplitude disturbance judgment condition includes a segment start condition and an amplitude disturbance threshold;

S32. Perform matrix extraction on the effective carriers corresponding to each data packet in turn, to obtain matrix parameters corresponding to each data packet; wherein, the amplitude disturbance judgment condition is used to judge whether there are several corresponding human motions in the sliding window matrix parameter;

S33, according to the segment start condition, intercept the first action segment in the data stream through the sliding window;

S34. Detect in real time whether the first action segment is greater than the amplitude disturbance threshold, and if so, output it as a valid action segment.

Preferably, in the step S3, it also includes the following steps:

S311. Collect data samples of human bodies at different spatial positions performing different actions, including at least one falling data sample and at least one non-falling data sample;

S312 , marking the data samples, distinguishing the amplitude disturbance difference between the falling data sample and the non-falling data sample, and obtaining the amplitude disturbance judgment condition.

In this embodiment, the starting condition of the segment is common to any scene, and the starting condition of the segment at least includes the action data set of falling and bending over; a dat data file of 6s is used, and the time interval between packets is 0.01s , that is, the sampling frequency is 100HZ, that is, a data file corresponds to 600 packets. When we measured the data set, whether we fell or bent over, we stipulated that the action should be stationary for the first 2s, and then the action was executed after 2s. Therefore, we selected the start of the judgment segment as the 250th under the condition of amplitude disturbance. The amplitude interval between the 1st and 300th packets to match the collected dataset.

Amplitude disturbance threshold, which is a preset threshold through manual observation; in different environments, the effects of multipath effects are also different, this threshold has not been able to achieve automatic configuration once and for all, so this embodiment When used in other scenarios, the amplitude disturbance threshold needs to be reconfigured.

Referring to Figure 3 and Figure 4, it can be seen that when testing the action data set, the amplitude change range is between 300-400. Here, due to human error (every person's time concept is inconsistent), it will lead to time shift, This embodiment uniformly judges the amplitude of the interval between the 250th packet and the 300th packet; and it can also be seen from FIG. 3 and FIG. 4 that both the fall and the stooping action have a sudden change in amplitude. This embodiment By setting the amplitude disturbance judgment conditions, the data streams that do not produce sudden changes in the judgment interval are filtered out. The setting of the amplitude disturbance threshold is related to the environment. It is necessary to simulate related actions during deployment, collect one or more sets of data files in advance, observe the range of amplitude mutation, and then set the threshold.

In other embodiments, after the deployment of the wifi device is completed, the user can be guided to perform several sets of actions through a preset program, so as to calculate the sudden change of amplitude according to the ratio or preset algorithm through one or more sets of data files collected. range, get the amplitude disturbance threshold and automatically complete the preset.

Referring to FIG. 5 , it can be seen that the action range of the human body is mainly concentrated in 0-80 Hz, so the Butterworth filter is selected as the low-pass filter in this embodiment.

As another embodiment of the present invention, it is a real-time detection system based on WIFI in an ideal static and unmanned environment, including a signal sending module, namely a wifi router, a data receiving module, namely a wifi receiving antenna 2, and a data processing module. And SVM classification model and real-time detection module, namely microcomputer. There is one transmitting antenna 1, and three receiving antennas 2.

The implementation method of this embodiment includes the following steps:

A1. Environment construction. In this example, the Ubuntu system version 14.04 is selected. After the system is installed on the microcomputer, the CSI tool is used to modify the driver and kernel according to the deployed environment. Carrier information; from the intel 5300 network card, this embodiment can extract 30 subcarriers, and each data packet contains 30 subcarriers;

A2. The system collects data. In this embodiment, two methods are adopted to collect data by sampling. One is to specify 6s to write data packets into files in dat format; the other is to specify 600 packets to carry out Write the data packet to the file in dat format; among them, both methods correspond to the 100HZ sampling rate, that is, the time interval between packets is 0.01s;

A3. Action collection, arrange the experimenter to go to the data collection point 3 to do the specified action and collect the corresponding data; the plan in this embodiment is as follows in 6 seconds: for the first 2s, stand still on the boundary of the experimental area, such as As shown in Figure 6, after two seconds, go to the data collection point 3, and then perform the corresponding action, and after the action is performed, stand still;

A4. Signal processing, including sub-steps as follows:

A4.1. Signal preprocessing, first read the dat file, and then use the extraction function contained in the open csi tool kit to extract the csi subcarrier from the dat file, because the csi extracts the frequency response, so the csi data The format is a complex number, in the form of a+bi, by taking the modulo (that is, taking the amplitude), take |a+bi| as the base signal of the system;

A4.2, Signal noise reduction and filtering, perform wavelet transform on data packets for noise reduction, and use Butterworth filter for filtering to obtain the initial available signal;

A4.3. Feature extraction, extract the feature value of the initial available signal to obtain the signal feature; this step is mainly through the optimization of the algorithm to improve the computing power, this formula is obtained by synthesizing a large number of papers;

A4.4. Feature normalization. At the beginning of this step, there are two feasible solutions, one is standardization and the other is normalization. We have proved through practice that the variance of some eigenvalues is extremely small, resulting in the data being enlarged , so this embodiment adopts normalization;

A4.5, SVM classification, using the code set about libsvm, the determination of the vector machine hyperparameters c and g in the SVM classification model, and using the optimal hyperparameters c and g functions to find the optimal hyperparameters c and g of each training set. g;

A5. Through the TCP protocol, connect the computer and the router to realize real-time data transmission, set the computer locally as the server server, then connect to the router, and set the router as the client client to realize TCP transmission;

A6. The signal processing model is formed into a function_model function, and an amplitude disturbance condition is set to ensure the execution of the system through double while loops. The first layer of while loop is to ensure that no data connection is disconnected after a long time, and can wait for the connection again , the second layer of while loop is to ensure that the data can be collected in real time, update the data, execute the model, and finally output the result of judging each action.

The sampling rate used in this embodiment is 100 Hz, and the data packet is taken as the time interval of 6s, so theoretically, one data packet should contain 600 packets. However, since the quality of the channel link cannot be guaranteed to meet the optimal conditions, the delay of some packets may be slightly longer, so that if the data is received within the specified time, 600 packets cannot be received. Therefore, in this embodiment, the Each time the data is received, the number of packets of the data packet will be counted. If the current link environment cannot receive 600 packets after 6s at the sampling rate of 100HZ, this embodiment will use until 600 packets are received. , even if the time exceeds 6s.

In this embodiment, since the actual number of transmitting antennas of the router is 1, while the number of antennas at the receiving end in this embodiment is 3, after the csi is extracted from the received dat file, one csi corresponds to a matrix, and a 1*3*30 sub-antenna carrier matrix. The system model of this embodiment does not need to use all the subcarriers received by the three antennas, so we compare the sensitivity of each antenna to the action through observation, and we choose the antenna that is most sensitive to our action as the one we need. After many tests, we chose the third antenna; and the purpose of this embodiment is to analyze the data with time as the axis, so we only need to use one of the 30 subcarriers. Therefore, the correct action is taken here. The most sensitive sub-carrier, that is, the first sub-carrier is used. This step is the key to improving the computing efficiency. Most people use the distribution variance of 30 sub-carriers instead, which greatly increases the computing cost and reduces the efficiency.

This embodiment has 9 data collection points 3, and each point collects 30 dat files. There are two actions in total, one is bending over and the other is falling down (because the bending action, from the data graph, is Roughly similar), 5 experimenters, with a total of 2700 datasets, which well reduces the data error caused by geographic location.

I also mentioned the use of wavelet transform for noise reduction. After a large amount of data research and analysis and actual test experiments, we found that using wavelet transform to reduce noise is the most suitable way to reduce noise in csi signals, and then we know that different parts of the human body The frequency of the movements is different. The hand movements have a short cycle and a high frequency; falls have a long cycle and a low frequency; therefore, we calculate the approximate frequency range by calculating the action cycle, and the frequency range is 0~80HZ, and the frequency is reduced in a small range. The range is expanded to avoid loss of motion details, and then the Butterworth filter is used to filter out irrelevant signals outside the frequency range to improve accuracy.

This embodiment has two versions of the optimization function, one is SVMcg and the other is SVMcgPP; SVMcgPP is an enhanced version of SVMcg, both SVMcg and SVMcgPP use the grid search method, that is, the stepping method (enumeration method) to search The optimal hyperparameters c and g are obtained. This embodiment uses an enumeration function to obtain the best SVM classification model. By setting a certain parameter range and setting the gradient, and comparing the results through continuous training, the maximum accuracy and the best corresponding parameter values are obtained. The SVM classification model. Although this method is relatively slow in generating models and requires certain computing resources, the final SVM classification model is generated and configured in the actual application scenario. in most scenarios of the present invention.

As another embodiment of the present invention, its specific implementation steps are as follows:

B1. Perform real-time data collection, use the TCP protocol, open the local 8090 port through the tcpip function, establish it as a server, and set the server input cache size and waiting time; enter the command line at the computer terminal, connect to port 8090, and wait for the system to receive data; by sending the packet command, the CSI Tool sends and receives data simultaneously in two-way CSI acquisition in monitoring mode;

B2. To judge whether there is an action, first extract the csi matrix, and judge whether an action is generated through the amplitude disturbance judgment condition, and according to the amplitude disturbance judgment condition, carry out the sliding window to intercept the effective action segment; if the amplitude disturbance judgment condition is satisfied , then proceed to the next step, if not satisfied, continue to collect the number;

B3. Perform data processing, call the function_model function, perform data preprocessing, use Butterworth filter for noise reduction and filtering, and then extract eigenvalues. The eigenvalues include 26 time-domain eigenvalues and frequency-domain eigenvalues. ; After getting about the 26 eigenvalue matrices, normalize them;

B4, action classification, put the normalized eigenvalue matrix into the SVM classification model for classification, and quickly obtain the action type.

This embodiment mainly collects data for training the model in the following situations:

Case 1: Turn on the router and the microcomputer, expose the experimental scene to the WIFI environment, ensure that the data receiving module can receive the WIFI signal stably, open the matlab on the microcomputer and execute the real-time detection program to prepare for receiving data.

The specific operations include the following steps:

①Open matlab by command on Linux system;

②Use the self-written read_quietprocess.m file for port connection and data transmission;

③ In the TCP protocol, use the local IP port 8090 as the server, and set the port parameters;

④In the program, put the CSI data stream obtained in real time into the ap1_mp1 array, and write each into the last bit of the ap1_mp1 array, and advance each bit of the ap1_mp1 array by one bit through the for loop to reach each The received data is continuously passed in the ap1_mp1 array;

⑤Create a panduan array, and put the 250th data value of ap1_mp1 into the last digit of panduan, so that the data of the panduan array and the ap1_mp1 array are constantly updated and transmitted, and the panduan array is a variable applied to the condition of the amplitude disturbance judgment condition;

⑥And set the amplitude disturbance judgment conditions. If the conditions are met, put the valid action fragments under the current window into the SVM classification model for classification and judgment. If they do not meet the conditions, they will continue to collect data. This step greatly reduces the amount of calculation and improves the calculation efficiency. ;

⑦ Finally, if the valid segment has been classified and judged, the current valid segment is cleared, that is, the panduan array and ap1_mp1 are cleared to avoid interference caused by repeated judgments.

Open a new terminal in the microcomputer, perform the log_to_server operation, enter the local IP address to achieve connection with port 8090, and execute the ping command, so that the local server side of the computer receives the channel information from the router and transmits it to the matlab program. .

The specific operations are as follows:

①Enter the netlink file, log_to_server, and enter the local IP address to connect with the local port 8090;

②Enter the ping command, so that the local server server receives the router channel information and transmits it to the matlab program;

③According to the method of step 2, the signal received by matlab is processed and the amplitude disturbance condition is judged, and finally classified;

④ If the classification model judges that it is a fall, the display output is a fall, and if it is judged not to be a fall, the display output is a non-fall.

Since the ideal environment is static and unmanned in this case, the amplitude disturbance judgment condition will not be triggered.

Case 2: This case is a real-time detection system based on WIFI in an environment where the target is bent over. The operation steps of case 2 are basically the same as those of case 1. The difference is that case 1 is a static and unmanned environment. , while the second case is the environment where the target bends over.

In this case, the target makes a bending motion between the receiving end antenna and the transmitting end antenna to create a disturbance, and bending over makes the image fluctuate greatly, and then the curve returns to the original position, as shown in Figure 3.

Since the environment in which the target bends over is used in this case, the amplitude disturbance judgment condition will be triggered, but the classification result belongs to yes or no fall.

Case 3: This case is a real-time detection system based on WIFI in the environment where the target falls. The operation steps of case 3, case 2 and case 1 are basically the same, the difference is that case 3 is the environment where the target falls.

In this case, the target is used to make a fall between the receiving end antenna and the sending end antenna to create disturbance. The fall causes the image to fluctuate greatly, but the image of the falling action is inconsistent with the bending action image trend. The difference is that bending over is a Return to the original position of the curve after perturbation, while falling is to maintain the curve position after perturbation, as described in Figure 4.

Since the environment in which the target bends over is used in this case, the amplitude disturbance judgment condition is triggered, and the classification result is a fall, and an alarm message is issued at this time.

The invention also discloses a real-time fall detection device based on channel state information, comprising: a data acquisition module, a carrier acquisition module, an action detection module, a first data module, a second data module, a third data module, a classification module, and a Alarm module;

The data acquisition module is used to connect the wifi signal sending device using the TCP protocol, and obtain the data stream of the CSI in real time through the wifi signal receiving device; wherein, the data stream includes several continuous progressive data packets;

The carrier acquisition module is used to analyze the data flow, and obtain a valid carrier therefrom;

The motion detection module is used to preset the amplitude disturbance judgment condition; perform matrix extraction on the effective carrier corresponding to each data packet in turn, and obtain the matrix parameter corresponding to each data packet, and according to the amplitude disturbance judgment condition, the effective carrier is obtained. The carrier is intercepted by a sliding window to obtain valid action segments; wherein, the amplitude disturbance judgment condition is used to judge whether the sliding window has several matrix parameters that include corresponding human actions;

The first data module is used for denoising the effective carrier in the current window by using wavelet transform and filtering by using Butterworth filter to obtain an initial available signal when a valid action segment is intercepted;

The second data module is used to perform feature value extraction on the initially available signal to obtain signal features;

The third data module is used to normalize the signal features to obtain an eigenvalue matrix;

The classification module is used to import the eigenvalue matrix into the SVM classification model to obtain the action classification result;

The alarm module is used for judging whether the action classification result is a fall, and if so, sending an alarm message.

The present invention also discloses a terminal device, comprising a processor and a storage device, wherein the storage device is used to store one or more programs; when the one or more programs are executed by the processor, the processor implements the above-mentioned channel state information-based Real-time fall detection method. The processor may be a central processing unit (Central Processing Unit, CPU), or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), off-the-shelf processors Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The processor is the control center of the test equipment, and uses various interfaces and lines to connect various parts of the entire test equipment.

The storage device can be used to store computer programs and/or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and/or modules stored in the storage device and calling the data stored in the storage device. The storage device may mainly include a stored program area and a stored data area, wherein the stored program area may store an operating system, an application program required for at least one function, and the like; the storage data area may store data created according to the use of the terminal device. In addition, the storage device may include high-speed random access memory, and may also include non-volatile memory such as hard disk, internal memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device.

Wherein, if the integrated modules/units of the real-time fall detection device based on the channel state information are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the processes in the methods of the above embodiments, and can also be completed by instructing relevant hardware through a computer program, and the computer program can be stored in at least one computer-readable storage medium. When executed by the processor, the steps of the foregoing method embodiments can be implemented. Wherein, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file or some intermediate forms, and the like. The computer readable medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, removable hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access Memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

It should be noted that the above-described embodiments of the equipment and devices are only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be A physical unit, which can be located in one place or distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the protection scope of the present invention. . It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included within the protection scope of the present invention.

Claims (9)

1. The real-time fall detection method based on the channel state information is characterized by comprising the following steps:

connecting wifi signal sending equipment by using a TCP protocol, and acquiring a data stream of CSI in real time through wifi signal receiving equipment; wherein the data stream comprises a number of consecutively progressive data packets;

analyzing the data stream to obtain effective carriers;

presetting an amplitude disturbance judgment condition; sequentially carrying out matrix extraction on the effective carrier wave corresponding to each data packet to obtain a matrix parameter corresponding to each data packet, and carrying out sliding window interception on the effective carrier wave according to the amplitude disturbance judgment condition to obtain an effective action fragment; the amplitude disturbance judgment condition is used for judging whether a plurality of matrix parameters containing corresponding human body actions exist in the sliding window or not;

when an effective action fragment is intercepted, carrying out noise reduction on an effective carrier in a current window by using wavelet transformation, and carrying out filtering by using a Butterworth filter to obtain an initial available signal;

extracting characteristic values of the initial available signals to obtain signal characteristics;

normalizing the signal characteristics to obtain a characteristic value matrix;

importing the characteristic value matrix into an SVM classification model to obtain an action classification result;

and judging whether the action classification result is a fall or not, and if so, sending alarm information.

2. The real-time fall detection method based on channel state information according to claim 1, wherein: the signal characteristics include: the method comprises the steps of time domain mean, time domain standard deviation, time domain maximum, time domain minimum, time domain range, number of over-mean points, time domain 1/4 quantiles, time domain 3/4 quantile and time domain quartile range, first large FFT, frequency corresponding to the first large FFT, third large FFT, frequency corresponding to the third large FFT, fifth large FFT, frequency corresponding to the fifth large FFT, frequency domain average, frequency domain standard deviation, frequency domain 1/4 quantile, frequency domain 3/4 quantile, frequency domain quartile range, amplitude statistical skewness, amplitude statistical kurtosis, shape statistical mean, shape statistical standard deviation, shape statistical skewness and shape statistical kurtosis.

3. The real-time fall detection method based on channel state information according to claim 1, wherein the feature value extraction is performed on the initial available signal to obtain signal features, specifically:

performing wavelet transformation on the initial available signals again to obtain two wavelet signals, and performing characteristic value extraction on the two wavelet signals to obtain signal characteristics;

the signal characteristics comprise a time domain second mean value, a time domain second standard deviation, a time domain second range, a time domain second mean value point number and a time domain second quartile range, which are extracted from one wavelet signal; and extracting a time domain third mean value, a time domain third standard deviation, a time domain third range, a time domain third mean value point number and a time domain third quartile range from another wavelet signal.

4. The real-time fall detection method based on the channel state information is characterized in that the method uses a TCP protocol to connect with a wifi signal sending device and obtains a CSI data stream in real time through a wifi signal receiving device; wherein the data stream comprises a number of consecutively progressive data packets; the method specifically comprises the following steps:

setting at least one wifi signal sending device and m (since n is a variable hereinafter, m is repeatedly changed for avoiding the phenomenon) wifi signal receiving devices in a space to be detected, and simultaneously obtaining data sent by at least one wifi signal sending device through the m wifi signal receiving devices to obtain 1 dat file; wherein m is a positive integer;

extracting the data stream of the csi from the acquired dat file; the data stream comprises a plurality of data packets which are continuously progressive, and each data packet corresponds to a1 x m x 30 subcarrier matrix.

5. The method for fall detection in real time based on channel state information as claimed in claim 4, wherein the analyzing the data stream to obtain the effective carrier is specifically:

at preset time intervals, the sensitivities of the 1 × m × 30 subcarriers to the motion are compared, and the subcarrier most sensitive to the motion is selected as the effective carrier.

6. The real-time fall detection method based on the channel state information according to claim 1, wherein the preset amplitude disturbance judgment condition is set; sequentially carrying out matrix extraction on the effective carrier wave corresponding to each data packet to obtain a matrix parameter corresponding to each data packet, and carrying out sliding window interception on the effective carrier wave according to the amplitude disturbance judgment condition to obtain an effective action fragment; the amplitude disturbance judgment condition is used for judging whether a plurality of matrix parameters containing corresponding human body actions exist in the sliding window or not; the method specifically comprises the following steps:

presetting an amplitude disturbance judgment condition, wherein the amplitude disturbance judgment condition comprises a segment starting condition and an amplitude disturbance threshold;

sequentially carrying out matrix extraction on the effective carrier wave corresponding to each data packet to obtain a matrix parameter corresponding to each data packet; the amplitude disturbance judgment condition is used for judging whether a plurality of matrix parameters containing corresponding human body actions exist in the sliding window or not;

according to the segment starting condition, a first action segment in the data stream is intercepted through a sliding window;

and detecting whether the first action segment is larger than the amplitude disturbance threshold value in real time, and if so, outputting an effective action segment.

7. The real-time fall detection method based on the channel state information according to claim 1, wherein the preset amplitude disturbance judgment condition is set; sequentially carrying out matrix extraction on the effective carrier wave corresponding to each data packet to obtain a matrix parameter corresponding to each data packet, and carrying out sliding window interception on the effective carrier wave according to the amplitude disturbance judgment condition to obtain an effective action fragment; wherein, the amplitude disturbance judgment condition is used for judging whether a plurality of matrix parameters containing corresponding human body actions exist in the sliding window, and the method further comprises the following steps:

acquiring data samples of different movements of a human body at different spatial positions, wherein the data samples comprise at least one fallen data sample and at least one non-fallen data sample;

and marking the data samples, and distinguishing amplitude disturbance differences of fallen data samples and non-fallen data samples to obtain amplitude disturbance judgment conditions.

8. Real-time fall detection equipment based on channel state information is characterized by comprising: the device comprises a data acquisition module, a carrier acquisition module, an action detection module, a first data module, a second data module, a third data module, a classification module and an alarm module;

the data acquisition module is used for connecting wifi signal sending equipment by using a TCP protocol and acquiring CSI data stream in real time through wifi signal receiving equipment; wherein the data stream comprises a number of consecutively progressing data packets;

the carrier acquisition module is used for analyzing the data stream and acquiring an effective carrier from the data stream;

the action detection module is used for presetting amplitude disturbance judgment conditions; sequentially performing matrix extraction on the effective carrier wave corresponding to each data packet to obtain a matrix parameter corresponding to each data packet, and performing sliding window interception on the effective carrier wave according to the amplitude disturbance judgment condition to obtain an effective action fragment; the amplitude disturbance judgment condition is used for judging whether a plurality of matrix parameters containing corresponding human body actions exist in the sliding window or not;

the first data module is used for denoising the effective carrier wave in the current window by utilizing wavelet transformation and filtering by utilizing a Butterworth filter to obtain an initial available signal when an effective action segment is intercepted;

the second data module is used for extracting characteristic values of the initial available signals to obtain signal characteristics;

the third data module is used for normalizing the signal characteristics to obtain a characteristic value matrix;

the classification module is used for importing the characteristic value matrix into an SVM classification model to obtain an action classification result;

the alarm module is used for judging whether the action classification result is falling down, and if yes, sending alarm information.

9. A computer-readable storage medium, comprising a stored computer program, wherein when the computer program runs, the computer-readable storage medium controls an apparatus to execute the method for real-time fall detection based on channel state information according to any one of claims 1 to 7.

Images