Contactless human identification method based on continuous wave radar signal

Abstract

The present invention discloses a non-contact human identity recognition method based on continuous radar signals. By collecting radar echo signals from the chest area of the human body, multiple features of the signal in the time domain and frequency domain are extracted, and these feature data are used for identity verification and identification. The steps are as follows: when the human body is in a sitting posture, a continuous wave radar device is used to collect radar echoes of the human body's breathing and heartbeat activities to obtain the human body's radar echo signal, and the signal is pre-processed to extract multiple time-frequency features from the radar echo signal. Each feature is input into the established identity recognition model. The model realizes the function of identity verification and identification of the user by continuously training on a large amount of data. The identity recognition method proposed by the present invention is not affected by light, has high stability, and can cope with identity verification and identification tasks in various real environments.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to a non-contact human identity recognition method based on continuous wave radar signals. Background Art

[0002] The need for personal identification is widespread in modern society, and its applications are extensive. It serves not only to identify individuals but also to indicate qualifications or permissions. Common applications include residential access control systems, unlocking mobile devices, and verifying financial and e-commerce transactions. Identity identification plays a vital role in social security management.

[0003] Currently, widely used human identity recognition methods are primarily based on inherent human biometrics, with fingerprint and facial recognition being the most common. While these methods offer a certain degree of reliability, they do have some shortcomings. The human verification information used in these methods is often static, and static biometrics are easily cracked, which can create difficulties for the recognition system and potentially lead to certain risks. Furthermore, fingerprint recognition is a contact-based method that requires direct contact between the hand and the sensor, posing a security risk due to indirect contact between people. Facial recognition, on the other hand, relies on cameras to capture facial images and then perform a comparison. This method has high lighting requirements, and obtaining clear images sufficient for recognition is often difficult in poor lighting conditions.

[0004] Utilizing continuous wave radar signals can overcome these shortcomings and identify individuals. Continuous wave radar detection produces unique Doppler spectrum signatures for each individual's vital signs. This method enables long-distance, contactless, and safer identification. Radar wave acquisition has no ambient light requirements, making it more versatile. Furthermore, radar wave information is more difficult to decipher, resulting in higher system reliability.

[0005] Existing radar-based identification methods typically use a single probe, which provides very limited information and makes it difficult to fully capture the characteristics of the human body. Using dual radar sensors can better address this problem. Multiple radar probes can collect more comprehensive information about human life activities. Summary of the Invention

[0006] The purpose of the present invention is to solve the above-mentioned defects in the prior art and provide a non-contact human identity recognition method based on continuous wave radar signals.

[0007] The purpose of the present invention can be achieved by taking the following technical solutions:

[0008] A non-contact human identity recognition method based on continuous wave radar signals, the detection method comprising the following steps:

[0009] S1. Aim the dual radar sensors at the chest area of a subject, where the subject is in a sitting position, and simultaneously collect four-channel human body radar echo signals;

[0010] S2. Preprocessing the radar echo signal of each channel, wherein the preprocessing operations include high-frequency noise removal, baseline removal, signal detection and screening, multi-channel signal fusion and normalization;

[0011] S3, extracting time domain features and frequency domain features from the preprocessed radar echo signal;

[0012] S4. Divide the extracted time domain features and frequency domain features into a training set and a test set, input the data in the training set into a previously established identity recognition model for training to obtain a trained model, input the test data into the trained identity recognition model to obtain the corresponding recognition results.

[0013] Furthermore, the process of step S1 is as follows:

[0014] S101. Use dual radar sensors to form a radar array. The two sensors are placed side by side to maintain the same horizontal line, and both are facing the subject to ensure that there is no angle offset that affects data collection. The distance between the center of the array and the chest of the subject needs to be kept fixed, and the distance is set to a uniform fixed value. The subject remains seated to receive signal collection. The seat is an adjustable chair. The seat height is adjusted according to the subject's body proportions to ensure that the array center of the radar sensor and the subject's chest remain at the same horizontal height. The subject breathes normally and needs to avoid large movements that interfere with the radar waves, causing the radar waves to be mixed with information other than life activities. During this process, the radar array emits continuous radar waves and collects radar echo signals from the human body in this state;

[0015] S102, the STM32-based main controller converts the collected human body radar echo signal from analog to digital. The main controller includes a high-precision digital-to-analog conversion module that can convert the analog signal from the sensor end into a digital signal that is easy for computer storage and processing;

[0016] S103, transmitting the digital signal using serial communication to obtain four-channel human body radar echo data. The format of the digital signal is conducive to efficient signal processing using software programs.

[0017] Furthermore, the detailed process of performing signal preprocessing on the collected raw radar wave data in step S2 is as follows:

[0018] S201. Use a threshold denoising method based on wavelet transform to remove high-frequency noise from the radar echo signal of each channel. After performing wavelet decomposition on the signal, wavelet coefficients are obtained. A reasonable threshold is set to filter out the coefficients caused by noise. The expression of the threshold λ selected here is as follows:

[0019]

[0020] Where σ is the standard deviation of the signal and N is the sequence length of the signal. After removing the noise coefficient, only the wavelet coefficients generated by the radar echo signal are retained. Finally, these wavelet coefficients are used to reconstruct the signal, thus removing the high-frequency noise.

[0021] S202. Baseline drift is removed from each channel radar echo signal after high-frequency noise removal. The presence of baseline drift will cause some radar echo signal segments to appear significantly upward. This interference will seriously affect subsequent recognition and needs to be removed. A recursive filter is used to perform baseline removal. The recursive filter can handle low-frequency noise removal and DC offset elimination of radar wave signals and is very suitable for baseline drift removal of radar wave signals. The input and output expressions of the recursive filter used are as follows:

[0022] w[n] = x[n] + αw[n-1] Formula (2)

[0023] y[n] = w[n]-w[n-1] Formula (3)

[0024] Where x[n] represents the input signal, y[n] represents the output signal, w[n] is the intermediate conversion amount, and α is the scale factor of the recursive filter;

[0025] S203. Perform signal detection and screening on the radar echo signal of each channel to exclude abnormal radar echo signals in the collected data. Normal radar echo signals are periodic, and the autocorrelation function is used to detect periodicity. Calculate the autocorrelation sequence of the radar echo signal. According to the symmetry of the autocorrelation sequence, only general information needs to be retained. The sequence of the right half of the symmetry axis of the autocorrelation sequence is intercepted and recorded as r[n]. According to the characteristics of the autocorrelation function, the maximum value of r[n] must be r[0]. Find the second maximum value from r[n] and record it as r[i]. Use the ratio of r[i] to r[0] as the threshold value as the evaluation standard for signal periodicity, and eliminate the corresponding channel signals whose ratio is less than the threshold value.

[0026] S204: Calculate the energy value of the radar echo signal of each channel and select the signal with the largest energy as the reference signal. Using the reference signal as the standard, calculate the amplitude difference between each time node and the corresponding time node of the reference signal. Find the amplitude adjustment factor that minimizes this amplitude difference. Then, perform amplitude adjustment on the signals of the other channels except the reference signal. Finally, perform signal fusion using a weighted average method. The relationship expression is as follows:

[0027]

[0028] Where xm[n] represents the fused output signal, x0[n] represents the reference signal, M represents the number of channels, ai represents the amplitude adjustment factor of the i-th channel signal, τi represents the time shift factor of the i-th channel signal, and xi[n] represents the original signal of the i-th channel;

[0029] S205. Normalize the radar echo signal after the fusion to obtain a standardized radar echo signal. Data normalization is performed to maintain the statistical characteristics of the sample, making the data of different groups comparable and preserving the relative positions of all sampling points. The normalization transformation formula used here is as follows:

[0030]

[0031] Where y[n] represents the normalized signal, x[n] is the input signal, xmin is the minimum value of the signal, and xmax is the maximum value of the signal.

[0032] Furthermore, the process of step S3 is as follows:

[0033] S301. Within each cycle of the standardized radar echo signal, peaks and troughs are found. Signal envelopes are constructed based on these two characteristic points. These envelopes are an upper envelope reconstructed from the peaks, and a lower envelope reconstructed from the troughs. The upper envelope reflects the amplitude variation of a person's exhalation, while the lower envelope reflects the amplitude variation of a person's inhalation. These characteristics vary from person to person and are therefore suitable for use as good features for identification.

[0034] S302: Using a sliding window approach, the standardized radar echo signal is processed as follows to obtain various time-domain features. The difference between the maximum and minimum values within each window is taken to obtain a range feature, which reflects the relative amplitude of a person's exhalation and inhalation. The first-order difference sequence of the signal is calculated to obtain a rate-of-change feature of the radar signal. The average value within each window is calculated to obtain an amplitude feature, reflecting the mean of the signal values. The standard deviation and root mean square of the window sequence are calculated as two discrete degree features of the signal.

[0035] S303. Perform a fast Fourier transform on the standardized radar echo signal to obtain the corresponding discrete Fourier sequence. The Fourier transform can obtain the various frequency components contained in the signal, converting the time domain signal into a frequency domain signal, facilitating the extraction of frequency domain features. The discrete Fourier sequence obtained by directly performing a fast Fourier transform on the radar echo signal is a set of complex number sequences. It is necessary to take the modulus value of it to obtain the amplitude spectrum sequence X[n] of the radar echo signal. The position and height of the spectrum peak are obtained and combined to obtain the main peak feature in the frequency domain. Using a window sliding method, the mean of the amplitude spectrum sequence within the window is calculated to obtain the first spectrum feature, the standard deviation of the amplitude spectrum sequence within the window is calculated to obtain the second spectrum feature, and the root mean square of the amplitude spectrum sequence within the window is calculated to obtain the third spectrum feature.

[0036] Among them, the formulas for calculating the sequence standard deviation σ and the sequence root mean square Rms are as follows:

[0037]

[0038]

[0039] Where X[n] represents the sequence to be calculated, N represents the number of points in the sequence, and μ represents the average value of the sequence.

[0040] Furthermore, the process of step S4 is as follows:

[0041] S401. Normalize the time domain features and frequency domain feature values respectively. For the upper envelope features and lower envelope features obtained in step S301, the range features, the rate of change features, the amplitude features obtained in step S302, and the frequency domain main peak features, the first spectrum features, the second spectrum features, and the third spectrum features obtained in step S303, map the values of each feature to the interval [0,1] according to the maximum and minimum values of each feature to obtain normalized time domain features and frequency domain features. This step takes into account that the extreme maximum and minimum values appearing in the feature vector are likely to affect the training of the subsequent model, and normalizes the indicators in the feature vector to reduce the impact of the extreme maximum and minimum values on the model.

[0042] S402. Label the individuals collected according to their identities, and number the characteristic data of different human bodies as the basis for identification labels. The labels can be used as expected results in model training and fed back to the front-end network to optimize the model parameters. Split the data set into training sets and test sets in proportion. The training set is used to train the model, and the test set is used to verify the model; perform feature selection on the data set to obtain a feature subset; feature selection is to treat each indicator as a feature, eliminate irrelevant features from several indicators such as the envelope and the root mean square, select appropriate features, and form a feature subset. This step takes into account the problem that too many features are prone to dimensionality disasters. Feature selection can remove irrelevant features, reduce the difficulty of learning tasks in model training, and thus improve the efficiency of the model;

[0043] S403: Input the feature data in the feature subset into the identity recognition model, set a learning rate, and the identity recognition model learns through backpropagation based on the input feature data and the labels corresponding to the feature data, so that the output results of the identity recognition model match the labels corresponding to the feature data, thereby obtaining a trained identity recognition model;

[0044] S404. Extract the feature data to be identified from the test set, input the feature data to be identified into the trained identity recognition model, obtain the output result, compare the output result with the label of the feature data to be identified, and calculate the recognition accuracy as the evaluation indicator of the identity recognition model.

[0045] Furthermore, the identity recognition model structure includes a feature fusion unit, a residual convolution unit and an output conversion unit connected in sequence.

[0046] Among them, the feature fusion unit is composed of 9 feature convolution layers connected in parallel and cascaded with 1 fully connected layer. The 9 feature convolution layers respectively process the corresponding upper envelope features, lower envelope features, range features, rate of change features, amplitude features, frequency domain main peak features, first spectrum features, second spectrum features, and third spectrum features; the residual convolution unit is composed of a main path and a branch connected in parallel and superimposed. The main path is composed of the first activation layer, the first convolution layer, the second activation layer, and the second convolution layer cascaded in sequence, and the branch path is composed of the pooling layer and the zero-filling layer cascaded; the output conversion unit is composed of the RELU activation layer and the fully connected layer cascaded.

[0047] The present invention has the following advantages and effects compared to the prior art:

[0048] (1) The present invention uses human body radar wave signals to identify and detect human identity. This method is based on the differences in human physiological activities. Vital sign activities will form unique Doppler spectrum characteristics. These characteristics are the result of human body dynamics. Human body radar wave information is difficult to forge, which improves the reliability of identification and the security of the identification system will be higher.

[0049] (2) The information collection of the present invention is non-sensing, non-invasive and contactless, which makes it more flexible and convenient to use and avoids the risks brought by indirect contact.

[0050] (3) The present invention utilizes multiple radar probes to acquire vital signs signals of human life activities, thereby improving the fault tolerance of data acquisition and improving the quality of the collected radar waves.

[0051] (4) The present invention proposes a set of signal preprocessing methods to remove high-frequency noise and baseline drift in the signal, and uses the autocorrelation method to screen out abnormal signals, which greatly enhances the reliability of the data and thus ensures the accuracy of the time domain characteristics and frequency domain characteristics of the proposed signal.

[0052] (5) The present invention extracts multiple features from human vital signs radar signals and performs feature screening on the extracted features, thereby avoiding interference from the correlation between the individual features and preventing overfitting of the identity recognition model. Furthermore, a residual convolutional network is used as the identity recognition model to establish a mapping relationship between the extracted signal features and individuals of different identities, thus achieving identity recognition through human radar signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0054] Figure 1 This is a flow chart of a non-contact human identity recognition method based on continuous wave radar signals disclosed in Example 1 of the present invention;

[0055] Figure 2 This is a flow chart of data preprocessing of collected radar echo signals in Example 1 of the present invention;

[0056] Figure 3 1 is a waveform diagram comparing the radar echo signal before and after high-frequency filtering in Example 1 of the present invention;

[0057] Figure 4 Schematic diagram of the radar echo signal itself and its autocorrelation function during the periodic detection of the radar echo signal in Example 1 of the present invention;

[0058] Figure 5This is a structural block diagram of the identity recognition model used in Example 1 of the present invention;

[0059] Figure 6 is a scatter plot of the classification of the test results in Example 1 of the present invention;

[0060] Figure 7 This is a scatter plot of the classification of test results in Example 2 of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0062] Example 1

[0063] Figure 1 This is a flow chart of a non-contact human identity recognition method based on continuous wave radar signals disclosed in an embodiment of the present invention;

[0064] S1. Aim the continuous wave radar sensor at the chest area of the human body. The subject sits in a static position for the test. During the test, the system collects multi-channel radar echo signals.

[0065] S101. Use dual radar sensors to form a radar array. The two sensors are placed side by side and kept on the same horizontal line. Both are facing a subject to ensure that there is no angle deviation that affects data collection. The distance between the center of the array and the subject's chest is fixed at 0.5m. The subject remains in a seated state to receive signal collection. The seat height is adjusted according to the subject's body proportions to ensure that the center of the radar sensor array is at the same horizontal height as the subject's chest. The subject breathes normally and needs to avoid large movements that interfere with the radar echo signal wave, causing the radar echo signal to be mixed with information other than life activities;

[0066] S102, using an STM32-based main controller to collect digital signals. The main controller includes a high-precision digital-to-analog conversion module that can convert the analog signals from the sensor end into digital signals that are easy for computer storage and processing;

[0067] S103. Use serial communication to transmit the digital signal. Use a serial port assistant to collect and save the data on a computer. Create a folder for the tester, which serves as their tag number. Save the corresponding radar echo signal data file in TXT format, with the digital signal in hexadecimal format.

[0068] S2. For the radar echo signal collected in step S1, the data file of the radar echo signal is read using pycharm software, and the radar echo signal is preprocessed;

[0069] Figure 2 is a flow chart of data preprocessing of collected radar echo signals in an embodiment of the present invention;

[0070] S201. Use a threshold denoising method based on wavelet transform to remove high-frequency noise from the radar echo signal of each channel. Select the db8 wavelet function as the wavelet basis, perform a 5-layer wavelet decomposition on the radar echo signal, and obtain wavelet coefficients. Set a threshold to filter out the coefficients caused by noise. The expression of the threshold λ selected here is as follows:

[0071]

[0072] Where σ is the standard deviation of the signal and N is the sequence length of the signal. After removing the noise coefficient, only the wavelet coefficients generated by the original radar echo signal are retained. Finally, these wavelet coefficients are used to reconstruct the signal, thus removing the high-frequency noise.

[0073] Figure 3 This is a waveform comparison before and after high-frequency filtering of the radar echo signal in Example 1 of the present invention. It can be clearly observed that the burrs on the waveform have disappeared, the curve has become smooth, and the efficiency effect is very good;

[0074] S202. Baseline drift is removed from the radar echo signals of each channel after high-frequency noise removal. A recursive filter is used to perform baseline removal. The recursive filter can handle low-frequency noise removal and DC offset elimination of radar wave signals and is very suitable for baseline drift removal of radar echo signals. The input and output expressions of the recursive filter used are as follows:

[0075] w[n]=x[n]+αw[n-1] Formula (2)

[0076] y[n]=w[n]-w[n-1] Formula (3)

[0077] Where x[n] represents the input signal, y[n] represents the output signal, w[n] is the intermediate conversion amount, and α is the scale factor of the recursive filter. Here, α is set to 0.97. After passing through the filter, the signal with baseline drift removed is obtained.

[0078] S203. Perform signal detection and screening on the signal of each channel to eliminate abnormal radar echo signals in the data. Calculate the autocorrelation sequence of the signal, and cut off the sequence of the right half of the symmetry axis of the autocorrelation sequence, record it as r[n]. The maximum value of r[n] must be r[0]. Find the second maximum value from r[n], record it as r[i]. Use the ratio of r[i] to r[0] as the evaluation standard for signal periodicity. Set the screening threshold of this ratio to 0.5. Calculate the above ratio for the radar echo signal of each channel. For channel signals with a ratio lower than 0.5, eliminate them.

[0079] Figure 4 This is the radar echo signal itself and its autocorrelation function image obtained by periodically detecting the radar echo signal of one of the channels in Example 1 of the present invention. It can be seen that the periodicity of the signal is poor, the maximum value of the autocorrelation function is 1, and the ratio of the submaximum value to the maximum value is less than 0.5. The data of this channel needs to be eliminated.

[0080] S204: Calculate the energy value of the radar echo signal of each channel and use the radar echo signal of the channel with the largest energy value as the reference signal. The radar echo signals of the remaining channels are compared with the reference signal at each time point, and the amplitude difference between them and the reference signal is calculated. The amplitude adjustment factor that minimizes the difference is then found. The radar echo signals of the channels other than the reference signal are then amplitude-adjusted. Finally, the signals are fused using a weighted average method. The relationship expression is as follows:

[0081]

[0082] Among them, x m [n] represents the fused output signal, x0[n] represents the reference signal, M represents the number of channels, a i represents the amplitude adjustment factor of the i-th channel signal, τ i represents the time shift factor of the i-th channel signal, x i [n] represents the original signal of the i-th channel;

[0083] S205. Normalize the radar echo signal after the fusion to obtain a standardized radar echo signal. Data normalization is performed to maintain the statistical characteristics of the sample, making the data of different groups comparable and preserving the relative positions of all sampling points. The normalization transformation formula used here is as follows:

[0084]

[0085] Among them, y[n] represents the normalized signal, x[n] is the input signal, and x min is the minimum value of the signal, xmax is the maximum value of the signal.

[0086] S3. Extract multiple features from the standardized radar echo signal, including time domain features and frequency domain features;

[0087] S301. For the standardized radar echo signal, find the peaks and troughs within each cycle of the signal. Based on these two characteristic points, construct the signal envelopes: an upper envelope reconstructed from the peaks, and a lower envelope reconstructed from the troughs. The upper envelope features are obtained from the upper envelope, and the lower envelope features are obtained from the lower envelope. The upper envelope can reflect the variation trend of the human exhalation amplitude, while the lower envelope reflects the variation trend of the human inhalation amplitude. These characteristics vary from person to person and are therefore suitable as good features for identity recognition.

[0088] S302: Using a sliding window approach, the standardized radar echo signal is processed as follows to obtain various time-domain features. The difference between the maximum and minimum values within each window is taken to obtain a range feature, which reflects the relative amplitude of a person's exhalation and inhalation. The first-order difference sequence of the signal is calculated to obtain a rate-of-change feature of the radar signal. The average value within each window is calculated to obtain an amplitude feature, which reflects the mean of the signal values.

[0089] S303. Perform fast Fourier transform on the standardized radar echo signal to obtain the corresponding discrete Fourier sequence. The various frequency components contained in the signal can be obtained through Fourier transform, and the time domain signal can be converted into a frequency domain signal, which is convenient for extracting frequency domain features. Directly performing fast Fourier transform on the signal results in a set of complex number sequences. It is necessary to take the modulus value of it to obtain the amplitude spectrum sequence X[n] of this signal. The peak position of the amplitude spectrum sequence and the height of the peak point are combined to obtain the main peak feature in the frequency domain. Using the window sliding method, the mean value of the amplitude spectrum sequence in the window is calculated to obtain the first spectrum feature, the standard deviation of the amplitude spectrum sequence in the window is calculated to obtain the second spectrum feature, and the root mean square of the amplitude spectrum sequence in the window is calculated to obtain the third spectrum feature. The formulas for calculating the sequence standard deviation σ and the sequence root mean square Rms are as follows:

[0090]

[0091]

[0092] Where X[n] represents the sequence to be calculated, N represents the number of points in the sequence, and μ represents the average value of the sequence;

[0093] S4. Input the extracted features into the established neural network model as the learning parameters of each feature. Use enough data to allow the model to fully learn. Input the test data into the trained model to obtain the corresponding recognition results.

[0094] S401. For the nine time-domain and frequency-domain features obtained above, namely, upper envelope feature, lower envelope feature, range feature, rate of change feature, amplitude feature, frequency-domain main peak feature, first spectrum feature, second spectrum feature, and third spectrum feature, each feature is normalized. Based on the maximum and minimum values in the feature vector of that feature, the numerical values of that feature are uniformly mapped to the interval [0, 1] to obtain the normalized time-domain and frequency-domain features. This step takes into account that the extreme maximum and minimum values in the feature vector are likely to affect the subsequent model training, and thus the indicators in the feature vector are normalized to reduce the impact of the extreme maximum and minimum values on the model.

[0095] Figure 5 This is a structural block diagram of the identity recognition model used in Example 1 of the present invention;

[0096] S402: Label the feature data based on the differences between the test subjects. The feature data for different subjects is numbered and used as the basis for identification labels. During model training, the labels can be used as expected results and fed back to the front-end network to optimize model parameters. The dataset is split proportionally into training and test sets, with 70% of the total data used as the training set and 30% as the test set. Feature selection is performed on the dataset to obtain a feature subset. Feature selection treats each indicator as a feature, eliminates irrelevant features from several indicators such as the envelope and root mean square, selects appropriate features, and forms a feature subset.

[0097] S403. Input each type of feature vector in the feature subset into the identity recognition model. The large amount of data allows the model to spontaneously adjust its internal structural parameters so that the model matches the input sample and the expected result. A learning rate of 0.001 is used during initial training. When the loss drops to 10, the learning rate is reduced by 10% after each epoch. The optimizer uses the stochastic gradient descent method, and the loss value used is the mean square error.

[0098] S404: Extract the feature data to be identified from the test set. To verify the accuracy of the model, extract a portion of the feature data set from other categories of people as a false sample. Input this data set into the trained model to obtain the output. Compare the output with the label of the data set to evaluate the accuracy of the model. The output results are divided into four categories:

[0099] TP (True Positive): The input sample is this person, and the output recognition result is true;

[0100] FN (False Negative): The input sample is this person, and the output recognition result is false;

[0101] FP (False Positive): The input sample is not this person, and the output recognition result is true;

[0102] TN (True Negative): The input sample is not this person, and the output recognition result is false;

[0103] The results of Example 1 are shown in Table 1:

[0104] Table 1. Test results of Example 1

[0105]

[0106] As can be seen from Table 1, the comprehensive recognition rate of the identity recognition model performs well.

[0107] Figure 6 This is a scatter plot of the test results classification in Example 1 of the present invention. The 180 points in the figure are all from positive samples. The output results use 0.5 as the dividing line, and values above 0.5 are considered true predictions. It can be seen that most of the points are distributed above the dividing line, reflecting the good recognition ability of positive samples.

[0108] Example 2

[0109] Based on the non-contact distributed radar-based objective detection method for brain cognitive load disclosed in Example 1, this example further provides an implementation process of the non-contact distributed radar-based objective detection method for brain cognitive load. The specific process is as follows:

[0110] S1. Refer to the corresponding steps in Example 1 and do not repeat them here;

[0111] S2. For the radar echo signal collected in step S1, the data file of the radar echo signal is read using pycharm software, and the radar echo signal is preprocessed; in Example 2, only the radar echo signal of the first channel is used, and the radar echo signals of other channels are discarded;

[0112] S201: Using a threshold denoising method based on wavelet transform to remove high-frequency noise from a single-channel radar echo signal. The high-frequency noise removal process is similar to the corresponding steps in Example 1 and will not be described again here.

[0113] S202: removing baseline drift from the single-channel radar echo signal after high-frequency noise removal, using a recursive filter to perform baseline removal processing. The baseline drift removal process refers to the corresponding steps in Example 1 and is not repeated here.

[0114] S203. Perform signal detection on the radar echo signal of a single channel to ensure that it is not an abnormal radar echo signal. Calculate the signal's autocorrelation sequence, find the maximum and second maximum values of the autocorrelation sequence, and calculate their ratio. This ratio is a criterion for evaluating signal periodicity. A ratio below 0.5 is considered an unqualified signal with poor periodicity. The ratio of the maximum and second maximum values of the autocorrelation sequence of this channel is 0.83, indicating that the signal is normal and usable.

[0115] S204: Since it is a single-channel radar echo signal, there is no need to perform signal fusion of multiple channels;

[0116] S205 : Perform normalization processing on the signal to obtain a standardized radar echo signal.

[0117] S3. Extract multiple features from the standardized radar echo signal, including time domain features and frequency domain features;

[0118] S4. Input the extracted features into the established neural network model as the learning parameters of each feature. Use enough data to allow the model to fully learn. Input the test data into the trained model to obtain the corresponding recognition results.

[0119] S401, refer to the corresponding steps in Example 1, which will not be repeated here;

[0120] S402: Refer to the corresponding steps in Example 1 and do not repeat them here;

[0121] S403: Refer to the corresponding steps in Example 1 and will not be repeated here;

[0122] S404: Extract the feature data to be identified from the test set. To verify the accuracy of the model, extract a portion of the feature data set of other types of people as a false sample. Input this data set into the trained model to obtain the output result.

[0123] The results of Example 2 are shown in Table 2:

[0124] Table 2. Test results of Example 2

[0125]

[0126] As can be seen from Table 2, the overall recognition rate of the identity recognition model is acceptable, but compared with the recognition results in Table 1, it is not as good as the recognition results in Example 1, and the overall accuracy is 8.6% lower. The use of a single-channel radar echo signal, compared with the signal collected by a single radar sensor, performs worse in this recognition model than the four-channel radar echo signal obtained by a dual radar array. This shows that the information obtained by the single-channel radar echo signal is limited and cannot fully exert the effect of the recognition model. The method proposed in the present invention obtains a multi-channel radar echo signal. Through a series of preprocessing methods such as signal fusion, it can fully utilize the collected information and improve the accuracy of identity recognition.

[0127] Figure 7 This is a scatter plot of the test results classification in Example 2 of the present invention. The 180 points in the figure are all from positive samples. The output results are divided by 0.5. Above 0.5, the prediction is true. It can be seen intuitively that the distribution position of most points is not as good as Figure 6 The recognition result.

[0128] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A non-contact human identity recognition method based on continuous wave radar signals, characterized in that: The contactless human body identification method comprises the following steps: S1. Align the dual radar sensors with the subject's chest. The dual radar sensors form a radar array, with the two sensors placed side by side and aligned horizontally, both facing the subject. The distance between the center of the radar array and the subject's chest remains constant. The subject maintains a sitting posture and breathes normally. Four channels of human body radar echo signals are collected while the subject is in a sitting posture. S2. Preprocessing the radar echo signal of each channel, wherein the preprocessing operations include high-frequency noise removal, baseline removal, signal detection and screening, multi-channel signal fusion and normalization; S3. Extract time domain features and frequency domain features from the preprocessed radar echo signal; the process is as follows: S301. Finding peaks and troughs in each cycle of the standardized radar echo signal, and constructing envelopes of the signal based on the peaks and troughs found in each cycle, namely, an upper envelope reconstructed from the peaks and a lower envelope reconstructed from the troughs, to obtain upper and lower envelope features of the signal; S302. Process the standardized radar echo signal using a sliding window method to obtain multiple time domain features. The difference between the maximum and minimum values in each window is taken to obtain a range feature. A first-order difference sequence of the standardized radar echo signal is calculated to obtain a rate of change feature. An average value is calculated in each window to obtain an amplitude feature. S303: Perform a fast Fourier transform on the standardized radar echo signal to obtain a corresponding discrete Fourier sequence, take a modulus on the discrete Fourier sequence to obtain an amplitude spectrum sequence of the corresponding signal, obtain the peak position and the peak height of the amplitude spectrum sequence, and combine them to obtain a frequency domain main peak feature. Using a window sliding method, calculate the mean of the amplitude spectrum sequence within the window to obtain a first spectrum feature, calculate the standard deviation of the amplitude spectrum sequence within the window to obtain a second spectrum feature, and calculate the root mean square of the amplitude spectrum sequence within the window to obtain a third spectrum feature. S4. Divide the extracted time domain features and frequency domain features into a training set and a test set. Input the data in the training set into the previously established identity recognition model for training to obtain a trained model. Input the test data into the trained identity recognition model to obtain the corresponding recognition results. The process is as follows: S401, normalizing the time domain features and frequency domain feature values respectively, for the upper envelope feature and the lower envelope feature obtained in step S301, the range feature, the rate of change feature, the amplitude feature obtained in step S302, and the frequency domain main peak feature, the first spectrum feature, the second spectrum feature, and the third spectrum feature obtained in step S303, mapping the value of each feature to the interval [0,1] according to the maximum value and the minimum value of each feature, to obtain normalized time domain features and frequency domain features; S402: Label the individuals collected according to their identities, number the characteristic data of different human bodies, obtain a dataset containing specified labels, split the dataset into a training set and a test set in proportion, perform feature selection on the training set, and obtain a feature subset; S403: Input the feature data in the feature subset into the identity recognition model, set a learning rate, and the identity recognition model learns through backpropagation based on the input feature data and the labels corresponding to the feature data, so that the output results of the identity recognition model match the labels corresponding to the feature data, thereby obtaining a trained identity recognition model; S404. Extract the feature data to be identified from the test set, input the feature data to be identified into the trained identity recognition model, obtain the output result, compare the output result with the label of the feature data to be identified, and calculate the recognition accuracy as the evaluation indicator of the identity recognition model.

2. The non-contact human identity recognition method based on continuous wave radar according to claim 1, characterized in that: The process of step S1 is as follows: S101, the radar array transmits continuous radar waves and collects radar echo signals of human bodies in this state; S102, the main controller based on STM32 converts the collected human body radar echo signal from an analog signal to a digital signal; S103: Transmit the digital signal using serial communication to obtain four-channel human body radar echo data.

3. The non-contact human identity recognition method based on continuous wave radar according to claim 1, characterized in that: The process of step S2 is as follows: S201, using a threshold denoising method based on wavelet transform to remove high-frequency noise from the radar echo signal of each channel, specifically as follows: performing wavelet decomposition to obtain wavelet coefficients, filtering out coefficients caused by high-frequency noise using a pre-set threshold, retaining only the wavelet coefficients generated by the original radar echo signal, and finally reconstructing the radar echo signal using the wavelet coefficients; S202, using a recursive filter to remove baseline drift from the radar echo signals of each channel after high-frequency noise is removed; S203, performing signal detection and screening on the radar echo signal of each channel, detecting the periodicity of the signal by calculating the autocorrelation sequence, and eliminating channel signals with poor periodicity according to a pre-set threshold; S204, calculating the energy value of the radar echo signal of each channel, taking the radar echo signal of the channel with the maximum energy value as the reference signal, and performing time shift and amplitude adjustment on the radar echo signals of the remaining channels based on the reference signal. Finally, performing signal fusion using a weighted average method to obtain a fused radar echo signal; S205 , performing normalization processing on the fused radar wave signal to obtain a standardized radar echo signal.

4. The non-contact human identity recognition method based on continuous wave radar according to claim 1, characterized in that: The identity recognition model structure includes a feature fusion unit, a residual convolution unit and an output conversion unit connected in sequence. Among them, the feature fusion unit is composed of 9 feature convolution layers connected in parallel and cascaded with 1 fully connected layer. The 9 feature convolution layers respectively process the corresponding upper envelope features, lower envelope features, range features, rate of change features, amplitude features, frequency domain main peak features, first spectrum features, second spectrum features, and third spectrum features; the residual convolution unit is composed of a main path and a branch connected in parallel and superimposed. The main path is composed of the first activation layer, the first convolution layer, the second activation layer, and the second convolution layer cascaded in sequence, and the branch path is composed of the pooling layer and the zero-filling layer cascaded; the output conversion unit is composed of the RELU activation layer and the fully connected layer cascaded.

Images