Personnel Feature Extraction and Personnel Identification Methods Based on Cardiac Radar Signals

Abstract

The technical problem this invention aims to solve is that existing personnel identification methods based on cardiac radar signals only use the raw cardiac displacement signal as the input to the authentication model, without extracting features to obtain representative characteristics. To address this problem, one technical solution of this invention is to provide a personnel feature extraction method based on cardiac radar signals. Another technical solution of this invention is to provide a personnel identification method based on cardiac radar signals, employing the aforementioned method. The personnel identification method proposed in this invention does not rely on visual features and is less susceptible to environmental changes and deliberate visual camouflage; this invention only requires radar signal acquisition, eliminating the need for audio and video recording, thus better protecting the privacy of the identified individual; this invention uses millimeter-wave radar to acquire personnel features, requiring no cooperation from the identified individual during the identification process, making it easier to achieve seamless and continuous identification.

Description

Technical Field

[0001] This invention relates to a method for extracting personnel features and a method for identifying personnel using the same method, and more particularly to a method for extracting personnel features based on cardiac radar signals and a method for identifying personnel using cardiac radar signals implemented using the same method. Background Technology

[0002] Personnel identification technology has numerous applications in scenarios such as behavior recognition, security, and surveillance. A stable, efficient, and privacy-inviolating method for identifying people within a specific area is of great significance.

[0003] The radar reflection signal from the heart contains characteristic information about a person. Compared to commonly used biometrics such as fingerprints, voiceprints, facial features, and iris features, the characteristics obtained from cardiac radar signals have many advantages: these characteristics can be detected in all surviving individuals, meaning that no additional liveness detection module is needed in the identification and detection system; cardiac activity is unconscious and uncontrollable, meaning that identification does not require the person's cooperation, making it easier to achieve non-invasive and continuous identification. Furthermore, compared to facial and voice features, the characteristics obtained from cardiac radar signals are less likely to involve personal privacy information.

[0004] Current personnel identification methods are mostly based on facial feature recognition, with few methods based on cardiac radar signals. A typical personnel identification method, as described in patent CN201710250082.9, usually uses surveillance cameras to capture video and then employs a machine learning model for personnel identification. The problem with this type of method is that it is susceptible to facial or clothing disguises, leading to incorrect judgments; additionally, such methods also raise privacy concerns. On the other hand, patent CN202110675375.8 proposes a personnel identification method using cardiac displacement radar signals. The drawback of this method is that it only uses the raw cardiac displacement signal as input to the authentication model, without extracting features to obtain representative characteristics. Summary of the Invention

[0005] The technical problem to be solved by this invention is that existing personnel identification methods based on cardiac radar signals only use the raw cardiac displacement signal as the input of the authentication model, without extracting features from it to obtain representative features.

[0006] To address the aforementioned technical problems, one technical solution of the present invention is to provide a method for extracting human features based on cardiac radar signals, characterized by comprising the following steps:

[0007] Step 1: The radar's signal synthesizer generates multiple linear frequency modulated (LFM) pulse signals. A pulse signal sequence composed of these LFM pulse signals is transmitted via the transmitting antenna, while ambient reflected signals are received via the receiving antenna. A mixer combines the transmitted and received signals to generate an intermediate frequency (IF) signal sequence if0(t), if1(t),...,if... k (t);

[0008] Step 2: For the intermediate frequency signal sequence if0(t), if1(t),...,if k The i-th intermediate frequency signal in (t) if i (t) is sampled to obtain a digital signal if i [n], and process the obtained digital signal to obtain the relative displacement x[i] of the heart when the i-th linear frequency modulated pulse signal is emitted, where 0≤i≤k, specifically including the following steps:

[0009] Step 2-1: Use a filter to filter the intermediate frequency signal if i [n] is filtered to remove electromagnetic interference and noise from points that are too close or too far away, resulting in the signal if i ′ [n];

[0010] Step 2-2: Use an adaptive filter to process the signal if i ′ [n] is filtered to remove noise caused by random human movement, resulting in the signal if″ i [n];

[0011] Steps 2-3: For the signal if″ i [n] is subjected to a Fast Fourier Transform to obtain its spectrum IF i [n];

[0012] Steps 2-4: Calculate the relative displacement of the heart x[i] when the i-th linear frequency modulated pulse signal is transmitted using the following formula:

[0013]

[0014] In the formula, N is the intermediate frequency signal if i [n] is the number of sampling points included;

[0015] Step 3: Calculate the period T of the relative cardiac displacement, and segment the relative cardiac displacement data obtained in Step 2 according to the period T to obtain a series of single-period relative cardiac displacements x0[n], x1[n], ..., x m [n];

[0016] Step 4: For the series of single-cycle relative cardiac displacements x0[n], x1[n], ..., x obtained in Step 3... m The relative cardiac displacement x in the j-th single cycle of [n] j [n] Perform feature extraction to obtain the feature vector V of the relative displacement segment. j , 0≤j≤m, where, for the j-th single cycle, the relative displacement of the heart x j [n] Feature extraction specifically includes the following steps:

[0017] Step 4-1: Based on the relative displacement signal x of the heart j Find two regional maxima x in [n] max1 and x max2 The two regional maxima x max1 and x max2 The corresponding times are n max1 and n max2 ;

[0018] Step 4-2: Find the maximum value x in both regions max1 and x max2 Find a local minimum x between min Regional minimum x min The corresponding time is n min And there are n max1 ≤n min ≤n max2 ;

[0019] Step 4-3: Construct vector V j =[x max1 ,x max2 ,x min ,n max1 ,n max2 ,n min ], which serves as the feature vector of the relative displacement of that segment;

[0020] All feature vectors constitute the collected personnel feature vector sequence V = [V0, V1, ..., V...]. m ].

[0021] Preferably, in step 1, the expression for a single linear frequency modulated pulse signal is given by the following formula.

[0022] P(t) = sin(2π(f) c +St)t)

[0023] In the formula, f c Let S be the starting frequency of the pulse signal, and S be the frequency slope of the pulse signal, where 0 ≤ t ≤ t c , t c The duration of the pulse signal.

[0024] Preferably, in step 2-1, the filter is a Butterworth bandpass filter. Let the distance range between the heart of the person to be identified and the radar be (d). min ,d max If f is the lower cutoff frequency of the Butterworth bandpass filter, then f is the lower cutoff frequency of the Butterworth bandpass filter. l and upper limit cutoff frequency f h They are given by the following formulas respectively:

[0025] f l =2Sd min / c

[0026] f h =2Sd max / c

[0027] In the formula, c represents the speed of light, and S is the frequency slope of the pulse signal.

[0028] Preferably, in step 3, the variation period T is calculated using the following steps:

[0029] Step 3-1: Calculate the correlation coefficient corr(l) with respect to the relative cardiac displacement x[i] using the following formula.

[0030]

[0031] In the formula, hr min / 60×f s ≤l≤hr max / 60×f s hr min and hr max These are the preset minimum and maximum heart rate values, f. s This is the transmission frequency of the pulse signal;

[0032] Step 3-2: Calculate the period T of the relative displacement change of the heart.

[0033] Another technical solution of the present invention provides a person identification method based on cardiac radar signals, which employs the above-mentioned person feature extraction method based on cardiac radar signals, characterized by including the following steps:

[0034] Step S1: Within the area to be identified, obtain the personnel feature vector sequence V = [V0, V1, ..., V] using the personnel feature extraction method based on cardiac radar signals described above. m ];

[0035] Step S2: In the personnel feature vector sequence V = [V0, V1, ..., V...] mThe feature vector sequence segment is selected using two equal-length detection windows, one before and one after. The two equal-length detection windows are defined as the back window and the front window, respectively, with the starting position of the back window being the ending position of the front window.

[0036] Let the initial time of the sliding window be t, and the window length be swl. Then the two feature vector sequence segments before and after the sliding window are V and V, respectively. a =[V t V t+1 ,...,V t+swl-1 ], V b =[V t+swl V t+swl+1 ,...,V t+2swl-1 ];

[0037] Step S3: Separate the feature vector sequence segments V a and V b The features are fed into a pre-trained machine learning model to perform feature aggregation, resulting in an aggregated feature vector V. af and V bf ;

[0038] Step S4: Aggregate the feature vector V af and V bf The cosine value of the included angle, cos(θ), can be calculated using the following formula:

[0039]

[0040] Step 5: Compare all the obtained cosine values ​​with the preset threshold. If a cosine value is less than the preset threshold, it is determined that there is an anomaly in the personnel identification result within the current detection window time range.

[0041] Preferably, in step S3, the acquisition of the machine learning model includes the following steps:

[0042] Step S3-1: Using the aforementioned method for extracting personnel features based on cardiac radar signals, collect a sufficiently long sequence of personnel feature vectors from a sufficient number of people at different times and under different environments;

[0043] Step S3-2: Divide the feature vector sequence obtained in step S3-1 into feature vector sequence segments according to the preset sliding window length;

[0044] Step S3-3: Establish a machine learning model with an autoencoder structure. Using the minimization of reconstruction error as the criterion, perform self-supervised training on the model using the collected feature vector sequence fragments. After training, use the encoder part of the model for feature extraction.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] 1) The personnel identification method proposed in this invention does not rely on visual features and is not easily affected by environmental changes and human visual camouflage;

[0047] 2) The method proposed in this invention only requires the collection of radar signals and does not require the recording of audio and video, which can better protect the privacy of the person being identified;

[0048] 3) This invention uses millimeter-wave radar to collect personnel characteristics. The identification process does not require the cooperation of the person being identified, making it easier to achieve non-intrusive and continuous identification. Attached Figure Description

[0049] Figure 1 This is a flowchart of the personnel feature extraction method based on millimeter-wave radar of the present invention;

[0050] Figure 2 This is a flowchart of the personnel identification method based on personnel characteristics according to the present invention;

[0051] Figure 3 This is a structural diagram of the feature aggregation machine learning model of the present invention. Detailed Implementation

[0052] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0053] like Figure 1 As shown in the figure, this embodiment discloses a method for extracting personnel features based on cardiac radar signals, which specifically includes the following steps:

[0054] Step 1: The radar's signal synthesizer generates multiple linear frequency modulated (LFM) pulse signals. A pulse signal sequence composed of these LFM pulse signals is transmitted via the transmitting antenna, while the receiving antenna receives ambient reflected signals. A mixer combines the transmitted and received signals to generate an intermediate frequency (IF) signal sequence if0(t), if1(t),...,if... k (t).

[0055] In this embodiment, the expression for a single linear frequency modulated pulse signal is given by equation (1).

[0056] P(t) = sin(2π(f) c +St)t) (1)

[0057] In equation (1), f c Let S be the starting frequency of the pulse signal, and S be the frequency slope of the pulse signal, where 0 ≤ t ≤ t c , tc The duration of the pulse signal.

[0058] Step 2: For the intermediate frequency signal sequence if0(t), if1(t),...,if k The i-th intermediate frequency signal in (t) if i (t) is sampled to obtain a digital signal if i [n], and process the obtained digital signal to obtain the relative displacement x[i] of the heart when the i-th linear frequency modulated pulse signal is emitted, where 0≤i≤k.

[0059] Step 2 specifically includes the following steps:

[0060] Step 2-1: Use a filter to filter the intermediate frequency signal if i [n] is filtered to remove electromagnetic interference and noise from points that are too close or too far away, resulting in the signal if i ′ [n].

[0061] In this embodiment, the filter is a Butterworth bandpass filter. Let the distance range between the heart of the person to be identified and the radar be (d). min ,d max If f is the lower cutoff frequency of the Butterworth bandpass filter, then f is the lower cutoff frequency of the Butterworth bandpass filter. l and upper limit cutoff frequency f h The results are given by equations (2) and (3) respectively.

[0062] f l =2Sd min / c (2)

[0063] f h =2Sd max / c (3)

[0064] In equations (2) and (3), c represents the speed of light.

[0065] Step 2-2: Use an adaptive filter to process the signal if′ i [n] is filtered to remove noise caused by random human movement, resulting in the signal if″ i [n].

[0066] In this embodiment, the adaptive filter is a normalized minimum mean square filter;

[0067] Steps 2-3: For the signal if″ i [n] is subjected to a Fast Fourier Transform to obtain its spectrum IF i [n].

[0068] Steps 2-4: Calculate the relative cardiac displacement x[i] when the i-th linear frequency modulated pulse signal is transmitted using equation (4):

[0069]

[0070] In equation (4), N is the intermediate frequency signal if i [n] is the number of sampling points.

[0071] Step 3: Calculate the period T of the relative cardiac displacement, and segment the relative cardiac displacement data obtained in Step 2 according to the period T to obtain a series of single-period relative cardiac displacements x0[n], x1[n], ..., x m [n].

[0072] In step 3, the period of change T is calculated using the following steps:

[0073] Step 3-1: Calculate the correlation coefficient corr(l) for the relative cardiac displacement x[i] using equation (5).

[0074]

[0075] In equation (5), hr min / 60×f s ≤l≤hr max / 60×f s hr min and hr max These are the preset minimum and maximum heart rate values, respectively. In this embodiment, hr min =48,hr max =120, f s This is the transmission frequency of the pulse signal.

[0076] Step 3-2: Calculate the period T of relative cardiac displacement change using equation (6).

[0077]

[0078] Step 4: For the series of single-cycle relative cardiac displacements x0[n], x1[n], ..., x obtained in Step 3... m The relative cardiac displacement x in the j-th single cycle of [n] j [n] Perform feature extraction to obtain the feature vector V of the relative displacement segment. j Where 0 ≤ j ≤ m. All feature vectors constitute the collected personnel feature vector sequence V = [V0, V1, ..., V...]. m ].

[0079] In step 4, the relative cardiac displacement x in the j-th single cycle is... j[n] Feature extraction specifically includes the following steps:

[0080] Step 4-1: Based on the relative displacement signal x of the heart j Find two regional maxima x in [n] max1 and x max2 Their corresponding times are n max1 and n max2 The maximum values ​​in these two regions represent the relative displacement of the heart when the atria and ventricles are fully filled, respectively.

[0081] Step 4-2: Find the maximum value x in both regions max1 and x max2 Find a local minimum x between min Its corresponding time is n min And there are n max1 ≤n min ≤n max2 The minimum value in this region represents the relative displacement of the heart when the atrioventricular valves open.

[0082] Step 4-3: Construct vector V j =[x max1 ,x max2 ,x min ,n max1 ,n max2 ,n min ], which serves as the characteristic vector of the relative displacement of that segment.

[0083] Combination Figure 2 This embodiment also discloses a personnel identification method based on personnel characteristics, specifically including the following steps:

[0084] Step 1: Within the area to be identified, use the aforementioned method for extracting personnel features based on cardiac radar signals to obtain the personnel feature vector sequence V = [V0, V1, ..., V...]. m ].

[0085] Step 2: In the personnel feature vector sequence V = [V0, V1, ..., V...] m The feature vector sequence segments are selected using two equal-length sliding windows, one before and one after. These two windows are defined as the rear window and the front window, respectively, with the start position of the rear window being the end position of the front window. Let the start time of the front window be t, and the window length be swl. Then the two feature vector sequence segments are V... a =[V t V t+1 ,...,V t+swl-1 ], V b =[V t+swl V t+swl+1 ,...,Vt+2swl-1 ].

[0086] Step 3: Separate the feature vector sequence segments V a and V b The features are fed into a pre-trained machine learning model to perform feature aggregation, resulting in an aggregated feature vector V. af and V bf These two feature vectors represent the characteristics of people within the entire sliding window area.

[0087] Step 3 involves obtaining the machine learning model through the following steps:

[0088] Step 3-1: Using the aforementioned method for extracting personnel features based on cardiac radar signals, collect a sufficiently long sequence of personnel feature vectors from a sufficient number of people at different times and under different environments.

[0089] Step 3-2: Divide the feature vector sequence obtained in Step 3-1 into feature vector sequence segments according to the preset sliding window length.

[0090] Step 3-3: Establish a machine learning model with an autoencoder structure. Using the minimization of reconstruction error as the criterion, perform self-supervised training on the model using the collected feature vector sequence fragments. After training, use the encoder part of the model for feature extraction.

[0091] In this embodiment, the machine learning model used can be a deep sequence model such as LSTM or GRU.

[0092] Step 4: Aggregate the feature vector V af and V bf The cosine value of the included angle is calculated using equation (7):

[0093]

[0094] Step 5: Compare all the obtained cosine values ​​with the preset threshold. If a cosine value is less than the preset threshold, it is determined that there is an anomaly in the personnel identification result within the current detection window time range.

Claims

1. A method for extracting personnel features based on cardiac radar signals, characterized in that, Includes the following steps: Step 1: The radar's signal synthesizer generates multiple linear frequency modulated (LFM) pulse signals. A pulse signal sequence composed of these LFM pulse signals is transmitted via the transmitting antenna, while ambient reflected signals are received via the receiving antenna. A mixer combines the transmitted and received signals to generate an intermediate frequency (IF) signal sequence if0(t), if1(t),...,if... k (t); Step 2: For the intermediate frequency signal sequence if0(t), if1(t),...,if k The i-th intermediate frequency signal in (t) if i (t) is sampled to obtain a digital signal if i [n], and process the obtained digital signal to obtain the relative displacement x[i] of the heart when the i-th linear frequency modulated pulse signal is emitted, where 0≤i≤k, specifically including the following steps: Step 2-1: Use a filter to filter the intermediate frequency signal if i [n] is filtered to remove electromagnetic interference and noise from points that are too close or too far away, resulting in the signal if i ′ [n]; Step 2-2: Use an adaptive filter to process the signal if i ′ [n] is filtered to remove noise caused by random human movement, resulting in the signal if″ i [n]; Steps 2-3: For the signal if″ i [n] is subjected to a Fast Fourier Transform to obtain its spectrum IF i [n]; Steps 2-4: Calculate the relative displacement of the heart x[i] when the i-th linear frequency modulated pulse signal is transmitted using the following formula: In the formula, N is the intermediate frequency signal if i [n] is the number of sampling points included; Step 3: Calculate the period T of the relative cardiac displacement, and segment the relative cardiac displacement data obtained in Step 2 according to the period T to obtain a series of single-period relative cardiac displacements x0[n], x1[n], ..., x m [n]; Step 4: For the series of single-cycle relative cardiac displacements x0[n], x1[n], ..., x obtained in Step 3... m The relative cardiac displacement x in the j-th single cycle of [n] j [n] Perform feature extraction to obtain the feature vector V of the relative displacement segment. j , 0≤j≤m, where, for the j-th single cycle, the relative displacement of the heart x j [n] Feature extraction specifically includes the following steps: Step 4-1: Based on the relative displacement signal x of the heart j Find two regional maxima x in [n] max1 and x max2 The two regional maxima x max1 and x max2 The corresponding times are n max1 and n max2 ; Step 4-2: Find the maximum value x in both regions max1 and x max2 Find a local minimum x between min Regional minimum x min The corresponding time is n min And there are n max1 ≤n min ≤n max2 ; Step 4-3: Construct vector V j =[x max1 ,x max2 ,x min ,n max1 ,n max2 ,n min ], which serves as the feature vector of the relative displacement of that segment; All feature vectors constitute the collected personnel feature vector sequence V = [V0, V1, ..., V...]. m ].

2. The method for extracting personnel features based on cardiac radar signals as described in claim 1, characterized in that, In step 1, the expression for a single linear frequency modulated pulse signal is given by the following equation. P(t)=sin(2π(f c +St)t) In the formula, f c Let S be the starting frequency of the pulse signal, and S be the frequency slope of the pulse signal, where 0 ≤ t ≤ t c , t c The duration of the pulse signal.

3. The method for extracting personnel features based on cardiac radar signals as described in claim 1, characterized in that, In step 2-1, the filter is a Butterworth bandpass filter. Let the distance range between the heart of the person to be identified and the radar be (d). min ,d max If f is the lower cutoff frequency of the Butterworth bandpass filter, then f is the lower cutoff frequency of the Butterworth bandpass filter. l and upper limit cutoff frequency f h They are given by the following formulas respectively: f l =2Sd min / c f h =2Sd max / c In the formula, c represents the speed of light, and S is the frequency slope of the pulse signal.

4. The method for extracting personnel features based on cardiac radar signals as described in claim 1, characterized in that, In step 3, the period of change T is calculated using the following steps: Step 3-1: Calculate the correlation coefficient corr(l) with respect to the relative cardiac displacement x[i] using the following formula. In the formula, hr min / 60×f s ≤l≤hr max / 60×f s hr min and hr max These are the preset minimum and maximum heart rate values, f. s This is the transmission frequency of the pulse signal; Step 3-2: Calculate the period T of the relative displacement change of the heart.

5. A method for personnel identification based on cardiac radar signals, employing the personnel feature extraction method based on cardiac radar signals as described in claim 1, characterized in that, Includes the following steps: Step S1: Within the area to be identified, obtain the personnel feature vector sequence V = [V0, V1, ..., V] using the personnel feature extraction method based on cardiac radar signals as described in claim 1. m ]; Step S2: In the personnel feature vector sequence V = [V0, V1, ..., V...] m The feature vector sequence segment is selected using two equal-length detection windows, one before and one after. The two equal-length detection windows are defined as the back window and the front window, respectively, with the starting position of the back window being the ending position of the front window. Let the initial time of the sliding window be t, and the window length be swl. Then the two feature vector sequence segments before and after the sliding window are V and V, respectively. a =[V t V t+1 ,...,V t+swl-1 ], V b =[V t+swl V t+swl+1 ,...,V t+2swl-1 ]; Step S3: Separate the feature vector sequence segments V a and V b The features are fed into a pre-trained machine learning model to perform feature aggregation, resulting in an aggregated feature vector V. af and V bf ; Step S4: Aggregate the feature vector V af and V bf The cosine value of the included angle, cos(θ), can be calculated using the following formula: Step 5: Compare all the obtained cosine values ​​with the preset threshold. If a cosine value is less than the preset threshold, it is determined that there is an anomaly in the personnel identification result within the current detection window time range.

6. The person identification method based on cardiac radar signals as described in claim 5, characterized in that, In step S3, the acquisition of the machine learning model includes the following steps: Step S3-1: Using the personnel feature extraction method based on cardiac radar signals as described in claim 1, collect a sufficiently long personnel feature vector sequence from a sufficient number of people at different times and under different environments; Step S3-2: Divide the feature vector sequence obtained in step S3-1 into feature vector sequence segments according to the preset sliding window length; Step S3-3: Establish a machine learning model with an autoencoder structure. Using the minimization of reconstruction error as the criterion, perform self-supervised training on the model using the collected feature vector sequence fragments. After training, use the encoder part of the model to extract features.

Images