Doppler radar positioning method based on complex convolutional neural network and time-frequency analysis

Abstract

The application discloses a Doppler radar positioning method based on a complex convolutional neural network and time-frequency analysis, comprising the following steps: constructing a multi-person moving Doppler through-wall demodulated echo signal, adopting the constructed echo signal, constructing a corresponding ideal time-frequency spectrum, and simultaneously constructing a training data set; taking a deep learning network as a basic framework, constructing a time-frequency spectrum preliminary prediction model, simultaneously adopting the training data set, training the model, and obtaining a final time-frequency spectrum prediction model; adopting the obtained model, predicting the real-time acquired radar echo signal, and acquiring a predicted time-frequency spectrum; adopting the acquired predicted time-frequency spectrum, predicting a trajectory of a target in a wall, and realizing radar positioning; the method combines deep learning with time-frequency analysis, suppresses the frequency ambiguity problem, introduces a network with a skip residual structure, and further improves the time-frequency resolution; the method improves the precision, reduces the offset, enhances the robustness, and improves the anti-noise performance.

Description

Technical Field

[0001] This invention belongs to the field of multi-target tracking technology, specifically relating to a Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis. Background Technology

[0002] Currently, through-wall radar, which calculates target trajectories using the Doppler effect, is considered an efficient positioning tool. Due to its portability, accurate positioning, and ease of use, through-wall radar has been widely used in many fields, including earthquake rescue, military operations, and security.

[0003] There are some limitations when using through-wall radar technology to track targets. When performing time-frequency analysis on the received Doppler radar echo signal, the instantaneous frequencies (IF) of the targets overlap or are too close in the spectrum, making it impossible to distinguish the IFs of different targets. This has a significant impact on the estimation of the target trajectory. The above problem is called "frequency ambiguity". Therefore, it is very important to find a technology that can effectively enhance time-frequency analysis, reduce frequency ambiguity and improve positioning stability.

[0004] Currently, mainstream time-frequency analysis methods include STFT, SST, SET, and MSST, but all of them have the following problems: 1. Due to the limitations of the Heisenberg uncertainty criterion, a balance cannot be achieved in terms of time and frequency resolution, resulting in low frequency or time resolution; 2. When targets within a wall have the same radial velocity, their instantaneous frequencies will also be the same. In this case, methods based on STFT will have frequency ambiguity problems, and instantaneous frequencies will be indistinguishable at frequency intersections in the time-frequency spectrum; 3. Strong signal noise can interfere with the time-frequency analysis process.

[0005] In summary, most current positioning methods based on time-frequency analysis suffer from low target positioning accuracy and significant offset during the positioning process. Summary of the Invention

[0006] The purpose of this invention is to provide a Doppler radar positioning method based on complex convolutional neural networks and time-frequency analysis, which improves accuracy, reduces offset, enhances robustness, and improves noise resistance.

[0007] The Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis provided by this invention includes the following steps:

[0008] S1. Construct the echo signal after Doppler demodulation of multiple mobile Doppler signals. Using the constructed echo signal, construct the corresponding ideal time-frequency spectrum and construct the training dataset.

[0009] S2. Using a deep learning network as the basic framework, construct a preliminary prediction model for the time spectrogram. At the same time, use the training dataset from step S1 to train the constructed preliminary prediction model to obtain the final time spectrogram prediction model.

[0010] S3. Using the time-spectrum prediction model obtained in step S2, predict the radar echo signal acquired in real time and obtain the predicted time-spectrum.

[0011] S4. Using the predicted time spectrum obtained in step S3, the trajectory of the target inside the wall is predicted to achieve radar positioning.

[0012] Step S1 involves constructing the multi-person moving Doppler through-wall demodulation echo signal. Using the constructed echo signal, a corresponding ideal time-spectrum diagram is built, and a training dataset is constructed. Specifically, this includes:

[0013] Construct the echo signal and instantaneous frequency of the multi-person mobile Doppler through-wall demodulation based on sine and cosine basis, and simultaneously construct the corresponding ideal time spectrum using the constructed echo signal;

[0014] The echo signal x1(n) constructed based on the sine and cosine basis is expressed by the following formula:

[0015]

[0016] Where N represents the total number of sub-signals, each signal consists of N sub-signals, and the value of N is uniformly distributed in the range (1, 6); A i The amplitude of the i-th signal component is represented by the following formula:

[0017] A i =0.01+5g i

[0018] Among them, g i This represents a standard normal distribution;

[0019] B i C represents the Doppler frequency shift, which is a random distribution with values ​​ranging from (-100, 100); i The amplitude of the i-th component is represented by a uniform distribution with values ​​ranging from [0.1, 6]; D i E represents the second frequency shift, which is a random distribution with values ​​ranging from [-4, 4]. i This represents a frequency shift, which is a random distribution with values ​​ranging from [-6, 6]; F i This represents a frequency offset constant, which is a random distribution with values ​​ranging from [0, 2π], where π represents the mathematical constant pi.

[0020] The instantaneous frequency is expressed by the following formula:

[0021] IF i (n)=B i -2πC i (2D i n+E i )×sin(2π(D i n 2 +E i n+F i ))

[0022] All signals in the training dataset contain Gaussian noise, with a signal-to-noise ratio ranging from 0dB to 30dB.

[0023] The following formula represents the signal received by the transmitter after echo modulation:

[0024]

[0025] Where, α m Let f(t) represent the scattering index of the m-th target; A represent the amplitude; f1 and f2 represent the carrier frequencies of the radar transmitter; φ1 and φ2 represent the initial phase of transmission; d(t) represent the noise signal generated by multipath effects; n(t) represent the ambient noise signal; τ im Indicates receiver R i The delay of the m-th target is calculated using the following formula:

[0026] τ im =(R 1k +R im ) / C

[0027] Among them, R im Indicates receiver R i Distance to the m-th target; R 1m T represents the distance between transmitter T1 and the m-th target; C represents the speed of light;

[0028] Step S2 describes constructing a preliminary time-spectrum prediction model using a deep learning network as the basic framework. Simultaneously, the training dataset from step S1 is used to train the constructed preliminary prediction model to obtain the final time-spectrum prediction model. Specifically, this includes:

[0029] (2-1) Constructing a preliminary prediction model for the time-spectrum graph:

[0030] The model includes an input complex signal, a one-dimensional complex convolutional layer Conv1, an energy compression structure, a skip residual structure, and an output structure;

[0031] The input complex signal is a one-dimensional complex signal; the one-dimensional complex signal includes a real part and a complex part;

[0032] The one-dimensional complex convolutional layer Conv1 consists of two one-dimensional convolutional layers; the two one-dimensional convolutional layers undergo a linear transformation to obtain a feature vector; the feature vector is then reshaped to obtain a matrix of (L,W) dimensions.

[0033] The energy compression structure consists of L one-dimensional complex convolutional layers; the (L,W) dimension matrix is ​​obtained through the one-dimensional convolutional layers and is separated into L (1,W) dimension matrices in the energy compression structure; one-dimensional complex convolution is performed on each (1,W) dimension matrix, and the processed results are concatenated to obtain a (L,W) complex matrix.

[0034] The skip residual structure comprises 20 encoders and decoders; each encoder includes a 2D convolution and a ReLU activation function; each decoder includes a 2D deconvolution and a ReLU activation function; the encoders are used to extract time-frequency features from the spectrogram; the decoders are used to reconstruct the high-resolution time-spectrum; the first encoder is selected as the input to the second encoder and also as the input to the twentieth decoder; the second encoder is selected as the input to the third encoder and also as the input to the nineteenth decoder; the third encoder is selected as the input to the fourth encoder and also as the input to the eighteenth decoder; the fourth encoder is selected as the input to the fifth encoder and also as the input to the seventeenth decoder; the fifth encoder is selected as the input to the sixth encoder and also as the input to the sixteenth decoder; the sixth encoder is selected as the input to the seventh encoder and also as the input to the fifteenth decoder; the seventh encoder is selected as the input to the eighth encoder and also as the input to the fourteenth decoder; the eighth encoder is selected as the input to the ninth encoder and also as the input to the thirteenth decoder; the ninth encoder... The encoder serves as the input to the tenth encoder and also as the input to the twelfth decoder; the tenth encoder serves as the input to the eleventh encoder; the eleventh encoder serves as the input to the twelfth encoder and also as the input to the tenth decoder; the twelfth encoder serves as the input to the thirteenth encoder and also as the input to the ninth decoder; the thirteenth encoder serves as the input to the fourteenth encoder and also as the input to the eighth decoder; the fourteenth encoder serves as the input to the fifteenth encoder and also as the input to the seventh decoder; the fifteenth encoder serves as the input to the sixteenth encoder and also as the input to the sixth decoder; the sixteenth encoder serves as the input to the seventeenth encoder and also as the input to the fifth decoder; the seventeenth encoder serves as the input to the eighteenth encoder and also as the input to the fourth decoder; the eighteenth encoder serves as the input to the nineteenth encoder and also as the input to the third decoder; the nineteenth encoder serves as the input to the twentieth encoder and also as the input to the second decoder; the twentieth encoder serves as the input to the first encoder.

[0035] The output structure includes a two-dimensional convolution with a size of (1,1) and an output channel of 1, ultimately yielding an optimized spectrogram (L,W) with the same dimensions as the input.

[0036] The input complex signal is passed through a one-dimensional complex convolutional layer to obtain feature matrix A; the obtained matrix A is passed through an energy compression structure to obtain feature matrix B; the obtained matrix B is passed through a skip residual structure to obtain feature matrix C; the obtained matrix C is passed through convolutional dimensionality reduction to obtain the final optimized spectrum.

[0037] (2-2) Obtain the final time-spectrum prediction model:

[0038] Using the training dataset from step S1, perform basic testing on the preliminary model built in step (2-1);

[0039] During training, the network performs backpropagation to obtain a preliminary prediction model of the time-spectral graph that converges during training.

[0040] Convergence means that the model's loss function remains unchanged within a defined interval;

[0041] The enhanced time-frequency spectrum of the model output is observed by the input signal. If the time-frequency resolution of the time-frequency spectrum is within the set range, the training ends and the corresponding model is selected as the final time-frequency spectrum prediction model. If the time-frequency resolution of the time-frequency spectrum is not within the set range, the training is carried out in a intensive training session until the set conditions are met, the testing stops, and the final time-frequency spectrum prediction model is obtained.

[0042] Step S4, which uses the predicted time-spectrum map obtained in step S3 to predict the trajectory of the target inside the wall and achieve radar positioning, specifically includes:

[0043] Using the predicted time-frequency spectrum obtained in step S3, the instantaneous frequency is extracted;

[0044] Radar positioning is achieved by predicting the trajectory of targets inside the wall using radial distance estimation algorithms and angle of arrival estimation algorithms.

[0045] The distance between the target and the receiver is determined using a radial distance estimation algorithm.

[0046] The angle of arrival between the target and the transmitter is estimated using an angle of arrival estimation algorithm.

[0047] (4-1) Radial distance estimation algorithm:

[0048] The instantaneous frequency of the target is estimated based on the predicted time-frequency spectrum, thereby allowing for real-time estimation of the radial distance or XY axis between the target and the transmitter;

[0049] The radar signal T transmitted by the radar is expressed by the following formula. x (t):

[0050]

[0051] Among them, f c1 f c2 These represent the two carrier frequencies that the transmitter can transmit; θ1 and θ2 represent the initial phase of the signal.

[0052] The signal R received by the radar receiver is expressed by the following formula. x(t):

[0053]

[0054] Where R represents the radial distance between the transmitter and the target; R1 represents the radial distance between the target and receiver 1; c represents the speed of light; and t represents the time variable.

[0055] By adjusting the received signal, the following formula is obtained:

[0056]

[0057] Among them, R x1 (t) indicates that when the transmitter carrier is f c1 Receiver 1 receives the regulated radar echo signal;

[0058]

[0059] Among them, R x2 (t) indicates that when the transmitter carrier is f c2 Receiver 1 receives the regulated radar echo signal;

[0060] The phase angle of the signal acquired by the receiver is expressed by the following formula:

[0061]

[0062] Where ψ1(t) represents the transmitter carrier when f c1 Receiver 1 receives the phase of the regulated radar echo signal;

[0063]

[0064] Where ψ2(t) represents the transmitter carrier when f c2 Receiver 1 receives the phase of the regulated radar echo signal;

[0065] The phase difference is expressed by the following formula:

[0066]

[0067] Assuming R is equal to R1, the distance R between the target and the receiver is expressed by the following formula:

[0068]

[0069] (4-2) Angle of Arrival Estimation Algorithm:

[0070] The instantaneous frequency of the target is estimated based on the predicted time-frequency spectrum, thereby estimating the angle of arrival between the target and the transmitter;

[0071] Transmitter T x Capable of transmitting radar echo signals with two different carrier frequencies, the default is f. c1 By default, the radar's transmitted signal is represented by the following formula:

[0072]

[0073] Where φ1 represents the initial phase of the signal transmitted by the radar transmitting antenna;

[0074] The echo signal from the receiver is expressed by the following formula:

[0075] R x (t)=T x (t-τ)

[0076] Where τ represents the time delay between the radar transmitting signal and the received signal;

[0077]

[0078] R represents the radial distance between the transmitter and the target; R1 represents the distance between receiver 1 and the target; c represents the speed of light per second.

[0079] For two different receivers receiving signals from radar carrier frequency f c1 The signal R received by the two receivers is expressed by the following formula. x1 With R x2 :

[0080]

[0081]

[0082] Where θ0 represents the initial phase generated by the transmitter at carrier frequency f1; R1 represents the distance between the target and receiver 1; and R2 represents the distance between the target and receiver 2.

[0083] The phase received by receiver 1 is represented by the following formula:

[0084]

[0085] The phase received by receiver 2 is represented by the following formula:

[0086]

[0087] The phase difference Δθ between receiver 1 and receiver 2 is expressed by the following formula:

[0088]

[0089] Assuming the distances between receiver 1 and transmitter, and between receiver 1 and receiver 2, are both d, the path difference between the target radar echo signal and receiver 1 is expressed by the following formula:

[0090] x = d × sin(μ)

[0091] Where μ represents the angle between the target and the radar transmitter;

[0092] The phase difference formula is further obtained as follows:

[0093]

[0094] μ can be further expressed using the following formula:

[0095]

[0096] Based on the above calculations, the following formula is obtained:

[0097]

[0098] Among them, f 1Rx1 This indicates that receiver 1 is operating at a transmitter carrier frequency of f. c1 The instantaneous frequency of the received echo signal; f 1Rx2 This indicates that receiver 2 is operating at a transmitter carrier frequency of f. c1 The instantaneous frequency of the received echo signal; θ 1,2 λ represents the initial phase difference between the two signals; λ1 represents the carrier frequency f. c1 The corresponding wavelength;

[0099] (4-3) Trajectory Prediction:

[0100] Based on the radial distance R and arrival angle μ of the target calculated in steps (4-1) and (4-2), the trajectory of the target inside the wall is predicted using the following formula:

[0101] x i =R i sinμ i

[0102] y i =R i cosμ i

[0103] Where, x i The x-coordinate represents the x-coordinate of the i-th target; the y-coordinate represents the x-coordinate of i R represents the y-coordinate of the i-th target; i μ represents the radial distance to the i-th target. i Indicates the angle of the i-th target;

[0104] Based on the above calculation results, target localization in a two-dimensional Cartesian coordinate system is achieved, thereby enabling human body tracking.

[0105] The Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis provided by this invention suppresses the frequency ambiguity problem by combining deep learning with time-frequency analysis; it introduces an energy aggregation module, which performs a one-dimensional complex convolution on each column of the time-frequency spectrum to achieve energy aggregation without being restricted by the Heisenberg uncertainty criterion; by introducing a network with a skip residual structure, it performs secondary extraction on the time-frequency spectrum after the energy aggregation module, thereby further improving the time-frequency resolution; the method of this invention improves accuracy, reduces offset, enhances robustness, and improves noise resistance. Attached Figure Description

[0106] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0107] Figure 2 This is a schematic diagram illustrating the implementation of the method of the present invention.

[0108] Figure 3 This is a schematic diagram illustrating the construction of the end-to-end complex convolutional neural network and time-frequency analysis network in the method of this invention.

[0109] Figure 4 This is a schematic diagram of the Doppler through-wall radar positioning method based on end-to-end complex convolutional neural networks and time-frequency analysis in the present invention.

[0110] Figure 5 Here are schematic diagrams showing the results of using the method of the present invention and other methods for localization: Figure 5 (a) is a schematic diagram of the IF result obtained by STFT calculation; Figure 5 (b) is a schematic diagram of the STFT positioning results; Figure 5 (c) is a schematic diagram of the IF result obtained by SST calculation; Figure 5 (d) is a schematic diagram of the SST positioning results; Figure 5 (e) is a schematic diagram of the IF result obtained by Hof-Bessel calculation; Figure 5 (f) is a schematic diagram of the localization results of the Hof-Bessel tracking method; Figure 5 (g) is a schematic diagram of the IF results calculated by the energy-free accumulation module proposed by the method of the present invention; Figure 5 (h) is a schematic diagram of the positioning results of the energy-free accumulation module proposed by the method of the present invention; Figure 5 (i) is a schematic diagram of the IF result calculated by the method of the present invention; Figure 5 (j) is a schematic diagram of the positioning results obtained by the method of the present invention. Detailed Implementation

[0111] like Figure 1 The diagram shown is a flowchart of the method of the present invention. Figure 2 The diagram shown illustrates the implementation flow of the method of the present invention: The Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis provided by the present invention includes the following steps:

[0112] S1. Construct the Doppler through-wall demodulated echo signal of multiple mobile devices. Using the constructed echo signal, construct the corresponding ideal time-spectrum diagram and simultaneously construct the training dataset; specifically including:

[0113] Construct the echo signal and instantaneous frequency of the multi-person mobile Doppler through-wall demodulation based on sine and cosine basis, and simultaneously construct the corresponding ideal time spectrum using the constructed echo signal;

[0114] The echo signal x1(n) constructed based on the sine and cosine basis is expressed by the following formula:

[0115]

[0116] Where N represents the total number of sub-signals, each signal consists of N sub-signals, and the value of N is uniformly distributed in the range (1, 6); A i The amplitude of the i-th signal component is represented by the following formula:

[0117] A i =0.01+5g i

[0118] Among them, g i This represents a standard normal distribution;

[0119] B i C represents the Doppler frequency shift, which is a random distribution with values ​​ranging from (-100, 100); i The amplitude of the i-th component is represented by a uniform distribution with values ​​ranging from [0.1, 6]; D i E represents the second frequency shift, which is a random distribution with values ​​ranging from [-4, 4]. i This represents a frequency shift, which is a random distribution with values ​​ranging from [-6, 6]; F i This represents a frequency offset constant, which is a random distribution with values ​​ranging from [0, 2π], where π represents the mathematical constant pi.

[0120] The instantaneous frequency is expressed by the following formula:

[0121] iF i (n)=B i -2πC i (2D i n+Ei )×sin(2π(D i n 2 +E i n+F i ))

[0122] All signals in the training dataset contain Gaussian noise, with a signal-to-noise ratio ranging from 0dB to 30dB.

[0123] In the method of this invention, the training dataset includes 30,000 data points;

[0124] The following formula represents the signal received by the transmitter after echo modulation:

[0125]

[0126] Where, α m Let f(t) represent the scattering index of the m-th target; A represent the amplitude; f1 and f2 represent the carrier frequencies of the radar transmitter; φ1 and φ2 represent the initial phase of transmission; d(t) represent the noise signal generated by multipath effects; n(t) represent the ambient noise signal; τ im Indicates receiver R i The delay of the m-th target is calculated using the following formula:

[0127] τ im =(R 1k +R im ) / C

[0128] Among them, R im Indicates receiver R i Distance to the m-th target; R 1m T represents the distance between transmitter T1 and the m-th target; C represents the speed of light;

[0129] S2. Using a deep learning network as the basic framework, construct a preliminary prediction model for the time-spectrum graph. Simultaneously, use the training dataset from step S1 to train the constructed preliminary prediction model, obtaining the final time-spectrum graph prediction model; specifically including:

[0130] like Figure 3 The diagram shown illustrates the construction of the end-to-end complex convolutional neural network and time-frequency analysis network in the method of this invention:

[0131] (2-1) Constructing a preliminary prediction model for the time-spectrum graph:

[0132] The model includes an input complex signal, a one-dimensional complex convolutional layer Conv1, an energy compression structure, a skip residual structure, and an output structure;

[0133] The input complex signal is a one-dimensional complex signal; the one-dimensional complex signal includes a real part and a complex part;

[0134] The one-dimensional complex convolutional layer Conv1 consists of two one-dimensional convolutional layers; the two one-dimensional convolutional layers undergo a linear transformation to obtain a feature vector; the feature vector is then reshaped to obtain a matrix of (L,W) dimensions.

[0135] The energy compression structure consists of L one-dimensional complex convolutional layers; the (L,W) dimension matrix is ​​obtained through the one-dimensional convolutional layers and is separated into L (1,W) dimension matrices in the energy compression structure; one-dimensional complex convolution is performed on each (1,W) dimension matrix, and the processed results are concatenated to obtain a (L,W) complex matrix.

[0136] The skip residual structure comprises 20 encoders and decoders; each encoder includes a 2D convolution and a ReLU activation function; each decoder includes a 2D deconvolution and a ReLU activation function; the encoders are used to extract time-frequency features from the spectrogram; the decoders are used to reconstruct the high-resolution time-spectrum; the first encoder is selected as the input to the second encoder and also as the input to the twentieth decoder; the second encoder is selected as the input to the third encoder and also as the input to the nineteenth decoder; the third encoder is selected as the input to the fourth encoder and also as the input to the eighteenth decoder; the fourth encoder is selected as the input to the fifth encoder and also as the input to the seventeenth decoder; the fifth encoder is selected as the input to the sixth encoder and also as the input to the sixteenth decoder; the sixth encoder is selected as the input to the seventh encoder and also as the input to the fifteenth decoder; the seventh encoder is selected as the input to the eighth encoder and also as the input to the fourteenth decoder; the eighth encoder is selected as the input to the ninth encoder and also as the input to the thirteenth decoder; the ninth encoder... The encoder serves as the input to the tenth encoder and also as the input to the twelfth decoder; the tenth encoder serves as the input to the eleventh encoder; the eleventh encoder serves as the input to the twelfth encoder and also as the input to the tenth decoder; the twelfth encoder serves as the input to the thirteenth encoder and also as the input to the ninth decoder; the thirteenth encoder serves as the input to the fourteenth encoder and also as the input to the eighth decoder; the fourteenth encoder serves as the input to the fifteenth encoder and also as the input to the seventh decoder; the fifteenth encoder serves as the input to the sixteenth encoder and also as the input to the sixth decoder; the sixteenth encoder serves as the input to the seventeenth encoder and also as the input to the fifth decoder; the seventeenth encoder serves as the input to the eighteenth encoder and also as the input to the fourth decoder; the eighteenth encoder serves as the input to the nineteenth encoder and also as the input to the third decoder; the nineteenth encoder serves as the input to the twentieth encoder and also as the input to the second decoder; the twentieth encoder serves as the input to the first encoder.

[0137] The output structure includes a two-dimensional convolution with a size of (1,1) and an output channel of 1, ultimately yielding an optimized spectrogram (L,W) with the same dimensions as the input.

[0138] The input complex signal is passed through a one-dimensional complex convolutional layer to obtain feature matrix A; the obtained matrix A is passed through an energy compression structure to obtain feature matrix B; the obtained matrix B is passed through a skip residual structure to obtain feature matrix C; the obtained matrix C is passed through convolutional dimensionality reduction to obtain the final optimized spectrum.

[0139] (2-2) Obtain the final time-spectrum prediction model:

[0140] Using the training dataset from step S1, perform basic testing on the preliminary model built in step (2-1);

[0141] During training, the network performs backpropagation to obtain a preliminary prediction model of the time-spectral graph that converges during training.

[0142] Convergence means that the model's loss function remains unchanged within a defined interval;

[0143] The enhanced time-frequency spectrum of the model output is observed by the input signal. If the time-frequency resolution of the time-frequency spectrum is within the set range, the training ends and the corresponding model is selected as the final time-frequency spectrum prediction model. If the time-frequency resolution of the time-frequency spectrum is not within the set range, the training is carried out in a intensive training session until the set conditions are met, the testing stops, and the final time-frequency spectrum prediction model is obtained.

[0144] S3. Using the time-spectrum prediction model obtained in step S2, predict the radar echo signal acquired in real time and obtain the predicted time-spectrum.

[0145] S4. Using the predicted time-spectrum map obtained in step S3, trajectory prediction is performed on the target inside the wall to achieve radar positioning; specifically including:

[0146] Using the predicted time-frequency spectrum obtained in step S3, the instantaneous frequency is extracted;

[0147] Radar positioning is achieved by predicting the trajectory of targets inside the wall using radial distance estimation algorithms and angle of arrival estimation algorithms.

[0148] The distance between the target and the receiver is determined using a radial distance estimation algorithm.

[0149] The angle of arrival between the target and the transmitter is estimated using an angle of arrival estimation algorithm.

[0150] The parameters of the radar used in the method of this invention are shown in the table below:

[0151] parameter symbol value carrier frequency <![CDATA[f c1 / f c2 ]]> 2.4 / 2.39GHz sampling frequency f 200Hz Maximum power / Minimum power <![CDATA[P max / P min ]]> 30 / 15dBm Antenna bandwidth B 40MHz Antenna gain G 3.5dB

[0152] (4-1) Radial distance estimation algorithm:

[0153] The instantaneous frequency of the target is estimated based on the predicted time-frequency spectrum, thereby allowing for real-time estimation of the radial distance or XY axis between the target and the transmitter;

[0154] The radar signal T transmitted by the radar is expressed by the following formula. x (t):

[0155]

[0156] Among them, f c1 f c2 These represent the two carrier frequencies that the transmitter can transmit; θ1 and θ2 represent the initial phase of the signal.

[0157] The signal R received by the radar receiver is expressed by the following formula. x (t):

[0158]

[0159] Where R represents the radial distance between the transmitter and the target; R1 represents the radial distance between the target and receiver 1; c represents the speed of light; and t represents the time variable.

[0160] By adjusting the received signal, the following formula is obtained:

[0161]

[0162] Among them, R x1 (t) indicates that when the transmitter carrier is f c1 Receiver 1 receives the regulated radar echo signal;

[0163]

[0164] Among them, R x2 (t) indicates that when the transmitter carrier is f c2 Receiver 1 receives the regulated radar echo signal;

[0165] The phase angle of the signal acquired by the receiver is expressed by the following formula:

[0166]

[0167] Where ψ1(t) represents the transmitter carrier when f c1 Receiver 1 receives the phase of the regulated radar echo signal;

[0168]

[0169] Where ψ2(t) represents the transmitter carrier when f c2 Receiver 1 receives the phase of the regulated radar echo signal;

[0170] The phase difference is expressed by the following formula:

[0171]

[0172] Assuming R is equal to R1, the distance R between the target and the receiver is expressed by the following formula:

[0173]

[0174] (4-2) Angle of Arrival Estimation Algorithm:

[0175] The instantaneous frequency of the target is estimated based on the predicted time-frequency spectrum, thereby estimating the angle of arrival between the target and the transmitter;

[0176] Transmitter T x Capable of transmitting radar echo signals with two different carrier frequencies, the default is f. c1 By default, the radar's transmitted signal is represented by the following formula:

[0177]

[0178] Where φ1 represents the initial phase of the signal transmitted by the radar transmitting antenna;

[0179] The echo signal from the receiver is expressed by the following formula:

[0180] R x (t)=T x (t-τ)

[0181] Where τ represents the time delay between the radar transmitting signal and the received signal;

[0182]

[0183] R represents the radial distance between the transmitter and the target; R1 represents the distance between receiver 1 and the target; c represents the speed of light per second.

[0184] For two different receivers receiving signals from radar carrier frequency f c1 The signal R received by the two receivers is expressed by the following formula. x1 With R x2 :

[0185]

[0186]

[0187] Where θ0 represents the initial phase generated by the transmitter at carrier frequency f1; R1 represents the distance between the target and receiver 1; and R2 represents the distance between the target and receiver 2.

[0188] The phase received by receiver 1 is represented by the following formula:

[0189]

[0190] The phase received by receiver 2 is represented by the following formula:

[0191]

[0192] The phase difference Δθ between receiver 1 and receiver 2 is expressed by the following formula:

[0193]

[0194] Assuming the distances between receiver 1 and transmitter, and between receiver 1 and receiver 2, are both d, the path difference between the target radar echo signal and receiver 1 is expressed by the following formula:

[0195] x = d × sin(μ)

[0196] Where μ represents the angle between the target and the radar transmitter;

[0197] The phase difference formula is further obtained as follows:

[0198]

[0199] μ can be further expressed using the following formula:

[0200]

[0201] Based on the above calculations, the following formula is obtained:

[0202]

[0203] Among them, f 1Rx1 This indicates that receiver 1 is operating at a transmitter carrier frequency of f. c1 The instantaneous frequency of the received echo signal; f 1Rx2 This indicates that receiver 2 is operating at a transmitter carrier frequency of f. c1 The instantaneous frequency of the received echo signal; θ 1,2 λ represents the initial phase difference between the two signals; λ1 represents the carrier frequency f. c1 The corresponding wavelength;

[0204] (4-3) Trajectory Prediction:

[0205] like Figure 4 The diagram shown is a localization schematic of the Doppler through-wall radar localization method based on end-to-end complex convolutional neural networks and time-frequency analysis in this invention:

[0206] Based on the radial distance R and arrival angle μ of the target calculated in steps (4-1) and (4-2), the trajectory of the target inside the wall is predicted using the following formula:

[0207] xi =R i sinμ i

[0208] y i =R i cosμ i

[0209] Where, x i The x-coordinate represents the x-coordinate of the i-th target; the y-coordinate represents the x-coordinate of the i-th target. i R represents the y-coordinate of the i-th target; i μ represents the radial distance to the i-th target. i Indicates the angle of the i-th target;

[0210] Based on the above calculation results, target localization in a two-dimensional Cartesian coordinate system is achieved, thereby enabling human body tracking.

[0211] like Figure 5 The diagram shows the results of using the method of the present invention and other methods for localization in an example: where, Figure 5 (a) is a schematic diagram of the IF result obtained by STFT calculation; Figure 5 (b) is a schematic diagram of the STFT positioning results; Figure 5 (c) is a schematic diagram of the IF result obtained by SST calculation; Figure 5 (d) is a schematic diagram of the SST positioning results; Figure 5 (e) is a schematic diagram of the IF result obtained by Hof-Bessel calculation; Figure 5 (f) is a schematic diagram of the localization results of the Hof-Bessel tracking method; Figure 5 (g) is a schematic diagram of the IF results calculated by the energy-free accumulation module proposed by the method of the present invention; Figure 5 (h) is a schematic diagram of the positioning results of the energy-free accumulation module proposed by the method of the present invention; Figure 5 (i) is a schematic diagram of the IF result calculated by the method of the present invention; Figure 5 (j) is a schematic diagram of the positioning results obtained by the method of the present invention; Figure 5 This demonstrates that the proposed method can improve the effectiveness of extracting instantaneous frequencies from the time-spectrum map, thereby achieving better positioning results.

Claims

1. A Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis, comprising the following steps: S1. Construct the Doppler through-wall demodulated echo signal of multiple moving individuals. Using the constructed echo signal, construct the corresponding ideal time-spectrum diagram, and simultaneously construct the training dataset; specifically including: Construct the echo signal and instantaneous frequency of the multi-person mobile Doppler through-wall demodulation based on sine and cosine basis, and simultaneously construct the corresponding ideal time spectrum using the constructed echo signal; The echo signal constructed based on the sine and cosine basis is expressed by the following formula. : in, This represents the total number of sub-signals, each signal consisting of... Composed of individual signals, The range of values ​​for is Uniform distribution; Indicates the first The amplitude of each signal component is calculated using the following formula: in, This represents a standard normal distribution; This represents the Doppler frequency shift, and its value range is... The random distribution; Indicates the first The amplitude of the component is within a range of values. Uniform distribution; This represents the second frequency offset, and its value range is... The random distribution; This represents a frequency shift, and its value range is... The random distribution; This represents a frequency offset constant, with a value range of [value range missing]. The random distribution, where, Represents pi; The instantaneous frequency is expressed by the following formula: All signals in the training dataset contain Gaussian noise, and the corresponding signal-to-noise ratio is... ; The following formula represents the signal received by the transmitter after echo modulation: in, Indicates the first The scattering index of a target; Indicates amplitude; and Indicates the carrier frequency of the radar transmitter; and Indicates the initial phase of the emission; This represents the noise signal generated by the multipath effect; Indicates ambient noise signal; Indicates receiver The accepted first The time delay of each target is calculated using the following formula: in, Indicates receiver With the The distance between the targets; Indicates transmitter With the The distance between the targets; Represents the speed of light; S2. Using a deep learning network as the basic framework, construct a preliminary prediction model for the time spectrogram. At the same time, use the training dataset from step S1 to train the constructed preliminary prediction model to obtain the final time spectrogram prediction model. S3. Using the time-spectrum prediction model obtained in step S2, predict the radar echo signal acquired in real time and obtain the predicted time-spectrum. S4. Using the predicted time spectrum obtained in step S3, the trajectory of the target inside the wall is predicted to achieve radar positioning.

2. The Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis according to claim 1, characterized in that... Step S2 describes constructing a preliminary time-spectrum prediction model using a deep learning network as the basic framework. Simultaneously, the training dataset from step S1 is used to train the constructed preliminary prediction model to obtain the final time-spectrum prediction model. Specifically, this includes: (2-1) Constructing a preliminary prediction model for the time-frequency spectrum: The input complex signal is a one-dimensional complex signal; the one-dimensional complex signal includes a real part and a complex part; One-dimensional complex convolutional layer It includes two one-dimensional convolutional layers; the two one-dimensional convolutional layers undergo a linear transformation to obtain feature vectors; the feature vectors are then processed by... Processing, obtaining A matrix of dimensionality; Energy compression structure includes One-dimensional complex convolutional layers; obtained through one-dimensional convolutional layers The dimensionless matrix is ​​separated into the energy-compressed structure. indivual A matrix of dimensions; for each A matrix of dimension 1 is subjected to one-dimensional complex convolution, and the results are concatenated to obtain... Complex matrices; The skip residual structure consists of a 20-layer encoder and decoder; each encoder includes a 2D convolution and a... Activation function; each decoder consists of a two-dimensional deconvolution and an activation function. Activation function; encoder is used to extract time-frequency features from the spectrum; decoder is used to recover the high-resolution time-spectrum; the first encoder is selected as the input to the second encoder and also as the input to the twentieth decoder; the second encoder is selected as the input to the third encoder and also as the input to the nineteenth decoder; the third encoder is selected as the input to the fourth encoder and also as the input to the eighteenth decoder; the fourth encoder is selected as the input to the fifth encoder and also as the input to the seventeenth decoder; the fifth encoder is selected as the input to the sixth encoder and also as the input to the sixteenth decoder; the sixth encoder is selected as the input to the seventh encoder and also as the input to the fifteenth decoder; the seventh encoder is selected as the input to the eighth encoder and also as the input to the fourteenth decoder; the eighth encoder is selected as the input to the ninth encoder and also as the input to the thirteenth decoder; the ninth encoder is selected as the input to the tenth encoder and also as the input to the twelfth decoder; the tenth The encoder serves as the input to the eleventh encoder; the eleventh encoder serves as the input to the twelfth encoder and also as the input to the tenth decoder; the twelfth encoder serves as the input to the thirteenth encoder and also as the input to the ninth decoder; the thirteenth encoder serves as the input to the fourteenth encoder and also as the input to the eighth decoder; the fourteenth encoder serves as the input to the fifteenth encoder and also as the input to the seventh decoder; the fifteenth encoder serves as the input to the sixteenth encoder and also as the input to the sixth decoder; the sixteenth encoder serves as the input to the seventeenth encoder and also as the input to the fifth decoder; the seventeenth encoder serves as the input to the eighteenth encoder and also as the input to the fourth decoder; the eighteenth encoder serves as the input to the nineteenth encoder and also as the input to the third decoder; the nineteenth encoder serves as the input to the twentieth encoder and also as the input to the second decoder; the twentieth encoder serves as the input to the first encoder. The output structure consists of a two-dimensional convolution of size (1, 1) with one output channel, ultimately yielding an optimized spectral map of the same dimension as the input. ; The feature matrix is ​​obtained after the input complex signal is passed through a one-dimensional complex convolutional layer. The resulting matrix The characteristic matrix is ​​obtained after energy compression structure. The resulting matrix The feature matrix is ​​obtained after using the jump residual structure. The resulting matrix After dimensionality reduction through convolution, the final optimized spectrum is obtained; (2-2) Obtain the final time-spectrum prediction model: Using the training dataset from step S1, perform basic testing on the preliminary model built in step (2-1); During training, the network performs backpropagation to obtain a preliminary prediction model of the time-spectral graph that converges during training. Convergence means that the model's loss function remains unchanged within a defined interval; The enhanced time-frequency spectrum output by the model is observed by the input signal. If the time-frequency resolution of the time-frequency spectrum is within the set range, the training ends and the corresponding model is selected as the final time-frequency spectrum prediction model. If the time-frequency resolution of the time-frequency spectrum is not within the set range, the training is carried out in a intensive training session until the set conditions are met, the testing stops, and the final time-frequency spectrum prediction model is obtained.

3. The Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis according to claim 2, characterized in that... Step S4, which uses the predicted time-spectrum map obtained in step S3 to predict the trajectory of the target inside the wall and achieve radar positioning, specifically includes: Using the predicted time-frequency spectrum obtained in step S3, the instantaneous frequency is extracted; Radar positioning is achieved by predicting the trajectory of targets inside the wall using radial distance estimation algorithms and angle of arrival estimation algorithms. The distance between the target and the receiver is determined using a radial distance estimation algorithm. The angle of arrival between the target and the transmitter is estimated using an angle of arrival estimation algorithm. Based on the calculated radial distance of the target and the angle of arrival The following formula is used to predict the trajectory of a target inside the wall: in, Indicates the first One goal coordinate; Indicates the first One goal coordinate; Indicates the first Radial distance of each target; Indicates the first From the perspective of a single goal; Based on the above calculation results, target localization in a two-dimensional Cartesian coordinate system is achieved, thereby enabling human body tracking.

4. The Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis according to claim 3, characterized in that... The included radial distance estimation algorithms specifically include: The instantaneous frequency of the target is estimated based on the predicted time-frequency spectrum, thereby allowing for real-time estimation of the radial distance or XY axis between the target and the transmitter; The radar signal transmitted by the radar is expressed by the following formula. : in, , This indicates the two carrier frequencies that the transmitter can transmit; , Indicates the initial phase of the signal; The signal received by the radar receiver is expressed by the following formula. : in, Indicates the radial distance between the transmitter and the target; This represents the radial distance between the target and receiver 1; Represents the speed of light; Represents a time variable; By adjusting the received signal, the following formula is obtained: in, Indicates that the transmitter carrier is At that time, receiver 1 received the mediated radar echo signal; in, Indicates that the transmitter carrier is At that time, receiver 1 received the mediated radar echo signal; The phase angle of the signal acquired by the receiver is expressed by the following formula: in, Indicates that the transmitter carrier is At that time, receiver 1 receives the phase of the regulated radar echo signal; in, Indicates that the transmitter carrier is At that time, receiver 1 receives the phase of the regulated radar echo signal; The phase difference is expressed by the following formula: Assumption and The distance between the target and the receiver is equal, expressed by the following formula. : 。 5. The Doppler radar localization method based on complex convolutional neural networks and time-frequency analysis according to claim 4, characterized in that... The included angle of arrival estimation algorithms specifically include: The instantaneous frequency of the target is estimated based on the predicted time-frequency spectrum, thereby estimating the angle of arrival between the target and the transmitter; transmitter Capable of transmitting radar echo signals with two different carrier frequencies, the default being... By default, the radar's transmitted signal is represented by the following formula: in, Indicates the initial phase of the signal transmitted by the radar transmitting antenna; The echo signal from the receiver is expressed by the following formula: in, This indicates the time delay between the radar transmitting a signal and receiving a signal; Indicates the radial distance between the transmitter and the target; This indicates the distance between receiver 1 and the target; This represents the speed at which light travels per second; For two different receivers receiving signals from radar carrier frequency... The signal received by the two receivers is expressed by the following formula. and : in, This indicates that the transmitter is at a carrier frequency of The resulting initial phase; This indicates the distance between the target and receiver 1; This indicates the distance between the target and receiver 2; The phase received by receiver 1 is represented by the following formula: The phase received by receiver 2 is represented by the following formula: The phase difference between receiver 1 and receiver 2 is expressed by the following formula. : Assume the distances between receiver 1 and transmitter, and between receiver 1 and receiver 2, are both... The path difference between the target radar echo signal transmitted to receiver 1 and receiver 2 is expressed by the following formula: in, Indicates the angle between the target and the radar transmitter; The phase difference formula is further obtained as follows: The following formula is used to further express this. : Based on the above calculations, the following formula is obtained: in, This indicates that receiver 1 is operating at the transmitter carrier frequency of... The instantaneous frequency of the received echo signal; This indicates that receiver 2 is operating at the transmitter carrier frequency of... The instantaneous frequency of the received echo signal; This represents the initial phase difference between two signals; Indicates carrier frequency The corresponding wavelength.

Images