Millimeter wave radar human body target detection method based on deep learning

Abstract

The invention proposes a millimeter-wave radar human body target detection method based on deep learning, and the method comprises the steps: carrying out the filtering of a signal through a finite impulse response filter in a manner of combining the deep learning with the echo signal of a millimeter-wave radar, and carrying out the windowing of the signal, performing time-phase and time-space expansion on the received radar signal to obtain phase difference information and time-space information, and performing Fourier transform on the radar signal to obtain distance information and distance-Doppler information; all information is input into a ConvRNN network for feature extraction, then features are input into an improved ResNet network, a completely pre-activated residual block is used in the network, average pooling is introduced into shortcut connection, and a Bottleneck structure is introduced; and finally, the test data are input into the network model, so that the accuracy problem of human body target detection can be solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of human body detection, and in particular relates to a millimeter wave radar human body target detection method based on deep learning. Background Art

[0002] As global environmental problems continue to intensify, people are paying more and more attention to energy conservation and emission reduction. Reducing carbon emissions has become a top priority for protecting the ecological balance of the earth. In this context, people are beginning to pay attention to how to achieve energy conservation and environmental protection through technical means, especially in the energy utilization of working areas such as offices, conference rooms, and production workshops. At the same time, they are also paying more and more attention to energy conservation in their home life.

[0003] With the development of semiconductor technology and the widespread application of deep learning technology, radar can now be used to measure more characteristic information of the human body. At the same time, with the rise of automated driving technology, radar has shifted from the military field to the civilian field, especially millimeter-wave radar. Compared with lidar, it not only has a price advantage, but also is not affected by heat or light, which can greatly reduce the impact of environmental factors on measurement.

[0004] The existing methods for human target detection mainly include: (1) Detection method based on traditional millimeter wave radar algorithm, after obtaining the intermediate frequency signal of the target returned by the millimeter wave radar, it is subjected to one-dimensional Fourier transform, then static clutter is filtered out, and then a constant false alarm probability detection algorithm is adopted. (2) Perception method based on machine learning, this method also uses the intermediate frequency echo signal of the millimeter wave radar, and then performs a two-dimensional Fourier transform, and uses the obtained RDM map, that is, the range-Doppler map, to classify the RDM map using methods such as SVM and naive Bayes, and then obtain the detection result. However, the above methods do not make full use of the target information in the millimeter wave radar echo signal, and it is easy to filter out the human target information during the filtering process, resulting in misjudgment in practical applications. Therefore, it is necessary to propose a method for detecting human targets to solve the above problems. Summary of the invention

[0005] The main problem to be solved by the present invention is to provide a millimeter wave radar human target detection method based on deep learning to improve the accuracy of human target detection.

[0006] To solve the above problems, the present invention provides a millimeter wave radar human target detection method based on deep learning, the steps of which include:

[0007] S1. The millimeter wave radar transmits a frequency modulated continuous wave to obtain an intermediate frequency signal; a finite impulse response filter is used to perform convolution processing on the intermediate frequency signal to filter out low-frequency signals; the intermediate frequency signal is then windowed to reduce spectrum leakage; after the windowing processing is completed, the intermediate frequency signal is time-phase and time-space-time unfolded to obtain a time-phase difference matrix and a time-space-time information matrix, and then Fourier transform is performed to obtain a distance information matrix and a distance-Doppler information matrix.

[0008] S2. The millimeter-wave radar has three receiving antennas. Within one frame, the intermediate frequency signals received by different antenna signals are processed by S1, and the results of time-phase and time-space-time expansion are data-associated. At the same time, the distance information matrix and the speed information matrix are fused together to form a three-dimensional phase difference matrix, a space-time information matrix, a distance information matrix and a distance-Doppler information matrix.

[0009] S3. Input the phase difference matrix, space-time information matrix, distance information matrix and distance-Doppler information matrix into the deep learning neural network model, and then adjust the feature extraction capability of the network model according to the recognition accuracy of the model.

[0010] Wherein S1 specifically comprises the following steps:

[0011] S11. The intermediate frequency signal is filtered along the signal at a fast time using a complex finite impulse response filter; the intermediate frequency signal is filtered along the slow time by convolution with a bandpass filter to filter out low-frequency signals, and the bandpass filter is a group of Sinc filters with a length of Nf=49, which reduces the complexity of subsequent calculations and memory overhead.

[0012] S12, after filtering out the low-frequency data of the millimeter-wave radar intermediate frequency signal, firstly use the Blackman-Harris Window algorithm to process the intermediate frequency signal of a single transmission signal to reduce spectrum leakage.

[0013] S13, time-phase unfolding of the intermediate frequency signal. In one frame of the millimeter wave radar, multiple detection signals are transmitted and multiple intermediate frequency signals are received. The phase difference between two adjacent intermediate frequency signals on a single antenna is calculated to record the movement of the detected target. The phase φ of the Nth sample N The calculation formula is as follows:

[0014] φ N =Δψ=ψ t -ψ t-1 (1)

[0015] where ψ t and ψ t-1is the phase of two adjacent intermediate frequency signals on the same antenna, and t is the slow time index; by performing time-phase expansion on the intermediate frequency signal, the phase change information in a frame of signal can be obtained, and the time-phase difference matrix of the previous and next signals in a frame can be obtained.

[0016] S14, perform time-space expansion on the intermediate frequency signal, first calculate the spatial phase difference between different antennas in the slow time, and then create a synthetic signal to record the accumulation of the spatial phase difference in the slow time. The phase of the synthetic signal of the Nth sample The calculation formula is as follows:

[0017]

[0018] in, and is the phase of the first and second antennas, and t is the slow time index; each frame of the intermediate frequency signal is time-space expanded to obtain the reception situation of the same transmitted signal on different antennas, and the time-space information matrix is ​​obtained.

[0019] Wherein S2 specifically comprises the following steps:

[0020] S21. The millimeter-wave radar has three antennas. The same antenna will continuously receive multiple intermediate frequency signals within one frame. The phases of two consecutive intermediate frequency signals received on the same antenna are differentiated, and the default initial phase difference is set to zero to obtain three phase difference matrices. The phase difference matrices are then superimposed to form a three-dimensional phase difference matrix.

[0021] S22. The millimeter-wave radar has three antennas. At the same time in a frame, different antennas will receive intermediate frequency signals from the same transmission signal. The phases of the intermediate frequency signals of the different antennas are differentiated and then summed in slow time to obtain a space-time information matrix. There are three ways to combine the phase difference values ​​of the three antennas. The three obtained space-time information matrices are superimposed to obtain a three-dimensional space-time information matrix.

[0022] S23, superimposing the distance information matrix and the range-Doppler information matrix obtained by performing one-dimensional and two-dimensional Fourier transform on the intermediate frequency signals of the different antennas to obtain a three-dimensional distance information matrix and a range-Doppler information matrix.

[0023] Wherein S3 specifically comprises the following steps:

[0024] S31. The training model includes a ConvRNN network and an improved ResNet network. The three-dimensional phase difference matrix, the spatiotemporal information matrix, the distance information matrix, and the distance-Doppler information matrix are respectively transmitted to the ConvRNN network to extract the distance feature information of the human target, and then the distance feature information extracted from the human target is transmitted to the next layer of the ResNet network.

[0025] S32. The ConvRNN network includes a convolutional neural network and a recurrent neural network. The convolutional neural network extracts spatial feature information of the human target based on the spatiotemporal information matrix and the distance information matrix of the human target. The recurrent neural network extracts micro-motion feature information of the human target based on the speed information matrix.

[0026] S33. The improved ResNet network uses a fully pre-activated residual block, changes the connection order of the activation function in the original residual block, so that the gradient can flow unimpeded to any other layer through the shortcut connection; introduces average pooling in the shortcut connection to reduce data loss; and introduces a Bottleneck structure, using the Bottleneck structure to replace the original two-layer residual block structure, so that the number of convolutional layers in each residual block is increased by one layer.

[0027] S34, the improved ResNet network performs model training according to the spatial feature information and micro-motion feature information of the human target obtained by the ConvRNN network to complete human target detection.

[0028] The data set input into the network model includes human presence data and non-human presence data, wherein:

[0029] The human presence data includes data of normal human activities, data of a human body lying still on a chair, data of a human body lying still on a table, and data of a human body lying down and sleeping.

[0030] The no-person data includes data that there is no person in the environment, data that only fans exist in the environment, data that only curtains exist in the environment, and data that only mobile smart devices exist in the environment.

[0031] Beneficial Effects

[0032] The present invention proposes a millimeter wave radar human target detection method based on deep learning. By combining deep learning with the echo signal of the millimeter wave radar, a finite impulse response filter is first used to filter the signal, and then a windowing process is performed on the signal. The received radar signal is then expanded in time-phase and time-space to obtain phase difference information and time-space information, and the radar signal is Fourier transformed to obtain distance information and distance-Doppler information; all the information is input into a ConvRNN network for feature extraction, and then the feature is input into an improved ResNet network, in which a fully pre-activated residual block is used, average pooling is introduced in a shortcut connection, and a Bottleneck structure is introduced; and test data is input into a network model, so that the problem of accuracy of human target detection can be solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure (1) is an overall flow chart of a method for detecting human targets by a millimeter-wave radar based on deep learning according to an embodiment of the present invention;

[0034] Figure (2) is a schematic flow chart of a method for detecting human targets by a millimeter-wave radar based on deep learning according to an embodiment of the present invention;

[0035] Figure (3) is a neural network framework diagram used in the millimeter wave radar human target detection method based on deep learning according to an embodiment of the present invention;

[0036] Figure (4) is a schematic diagram of radar installation of the millimeter wave radar human target detection method based on deep learning according to an embodiment of the present invention; Specific implementation plan

[0037] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, in the following description, known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0038] Referring to FIG. (1), the present invention proposes a method for detecting human targets using a millimeter wave radar based on deep learning, which includes but is not limited to:

[0039] S1. The millimeter wave radar transmits a frequency modulated continuous wave to obtain an intermediate frequency signal; a finite impulse response filter is used to perform convolution processing on the intermediate frequency signal to filter out low-frequency signals; the intermediate frequency signal is then windowed to reduce spectrum leakage; after the windowing processing is completed, the intermediate frequency signal is time-phase and time-space-time unfolded to obtain a time-phase difference matrix and a time-space-time information matrix, and then Fourier transform is performed to obtain a distance information matrix and a distance-Doppler information matrix.

[0040] In one example, S1 includes the following steps:

[0041] S11. The millimeter-wave radar transmits a frequency-modulated continuous wave signal through a transmitting antenna, and returns an echo signal after encountering the human body. The echo signal data is processed and analyzed. The method used for data processing is to obtain the analog signal of the echo on all receiving antennas of the millimeter-wave radar, and obtain the digital data of the echo signal after AD conversion. The echo signal is superimposed with the frequency and phase of the frequency-converting signal emitted by the radar to obtain an intermediate frequency signal; the intermediate frequency signal is filtered along the signal at a fast time using a complex finite impulse response filter; the intermediate frequency signal is filtered along the slow time by convolution with a bandpass filter to filter out low-frequency signals. The bandpass filter is a group of Sinc filters with a length of Nf=49, which reduces the complexity of subsequent calculations and memory overhead.

[0042] S12, after filtering out the low-frequency data of the intermediate frequency signal of the millimeter-wave radar, firstly use the Blackman-Harris Window algorithm to perform windowing processing on the intermediate frequency signal of a single transmission signal, and multiply the intermediate frequency signal by the Blackman-Harris Window function point by point to reduce spectrum leakage.

[0043] S13, time-phase unfolding of the intermediate frequency signal. In one frame of the millimeter wave radar, multiple detection signals are transmitted and multiple intermediate frequency signals are received. The phase difference between two adjacent intermediate frequency signals on a single antenna is calculated to record the movement of the detected target. The phase φ of the Nth sample N The calculation formula is as follows:

[0044] φ N =Δψ=ψ t -ψ t-1 (3)

[0045] where ψ t and ψ t-1 It is the phase of two adjacent intermediate frequency signals on the same antenna, and t is the slow time index. After performing time-phase expansion on the intermediate frequency signal, the phase change information in a frame of signal can be obtained. Since the millimeter-wave radar will transmit multiple detection signals within one frame, the received intermediate frequency signal is sampled to obtain a row of discrete data of the intermediate frequency signal. All the intermediate frequency signals in one frame are arranged together to obtain a two-dimensional matrix, and each row represents the discrete value of the intermediate frequency signal at different times. According to the time-phase expansion formula, the phase of the latter row is subtracted from the phase of the previous row to obtain the phase change of the adjacent intermediate frequency signal in one frame, and finally the time-phase differential information is obtained.

[0046] S14, perform time-space expansion on the intermediate frequency signal, first calculate the spatial phase difference between different antennas in the slow time, and then create a synthetic signal to record the accumulation of the spatial phase difference in the slow time. The phase of the synthetic signal of the Nth sample The calculation formula is as follows:

[0047]

[0048] in, and is the phase of the first and second antennas, and t is the slow time index; by performing time-phase expansion on the intermediate frequency signal, the spatiotemporal characteristic information of the human target in a frame of signal can be obtained. Since the millimeter wave radar has multiple antennas, three two-dimensional matrices can be obtained according to the above description, and each row represents each discrete value of the intermediate frequency signal. Therefore, within one frame, the three receiving antennas will obtain the intermediate frequency signal from the same transmitting signal, and perform differential processing on the phases of different antennas, and accumulate the phase difference in slow time to obtain time-space information.

[0049] S2. The millimeter-wave radar has three receiving antennas. Within one frame, the intermediate frequency signals received by different antenna signals are processed by S1, and the results of time-phase and time-space-time expansion are data-associated. At the same time, the distance information matrix and the speed information matrix are fused together to form a three-dimensional phase difference matrix, a space-time information matrix, a distance information matrix and a distance-Doppler information matrix.

[0050] In one example, S2 includes the following steps:

[0051] S21. The millimeter-wave radar has three antennas. Within one frame, the same antenna will continuously receive multiple intermediate frequency signals. The phases of two consecutive intermediate frequency signals received on the same antenna are differentiated, and the default initial phase difference is set to zero. Three two-dimensional matrices of the same size as in S13 are obtained. Each element in the matrix is ​​the phase difference between two adjacent intermediate frequency signals, and three phase difference matrices are obtained. The size of these three phase matrices is the same as that of the original intermediate frequency signal matrix. The three phase difference matrices are then superimposed to form a three-dimensional phase difference matrix.

[0052] S22. The millimeter-wave radar has three antennas. At the same time in one frame, different antennas will receive intermediate frequency signals from the same transmitted signal. The receiving antennas are paired in pairs for phase differentiation and then summed in slow time. Since the millimeter-wave radar has three antennas, three space-time information matrices will be obtained by pairing them in pairs. Therefore, the three space-time information matrices in one frame are superimposed to obtain a three-dimensional space-time information matrix.

[0053] S23. The distance information matrix and the distance-Doppler information matrix obtained by performing one-dimensional and two-dimensional Fourier transform on the intermediate frequency signals of the different antennas are also processed. Since the millimeter wave radar has three receiving antennas, the distance information matrices and the distance-Doppler information matrices on the three different antennas can be obtained. The three matrices are superimposed to obtain the three-dimensional distance information matrix and the distance-Doppler information matrix.

[0054] S3. Input the phase difference matrix, space-time information matrix, distance information matrix and distance-Doppler information matrix into the deep learning neural network model, and then adjust the feature extraction capability of the network model according to the recognition accuracy of the model.

[0055] In one example, S3 includes the following steps:

[0056] S31, the training model includes a ConvRNN network and an improved ResNet network, and the three-dimensional phase difference matrix, the spatiotemporal information matrix, the distance information matrix, and the distance-Doppler information matrix are respectively transmitted to the ConvRNN network to extract human target feature information, and then the extracted human target feature information is transmitted to the next layer of improved ResNet network.

[0057] S32. The ConvRNN network includes a convolutional neural network and a recurrent neural network. The convolutional neural network extracts the spatial feature information of the human target based on the spatiotemporal information matrix and the distance information matrix of the human target. The recurrent neural network extracts the micro-motion feature information of the human target based on the phase difference matrix and the distance-Doppler information matrix.

[0058] S33. The improved ResNet network uses a fully pre-activated residual block, changes the connection order of the activation function in the original residual block, so that the gradient can flow unimpeded to any other layer through the shortcut connection; introduces average pooling in the shortcut connection to reduce data loss; and introduces a Bottleneck structure, using the Bottleneck structure to replace the original two-layer residual block structure, so that the number of convolutional layers in each residual block is increased by one layer.

[0059] S34, the improved ResNet network performs model training according to the spatial feature information and micro-motion feature information of the human target obtained by the ConvRNN network to complete human target detection.

[0060] In this embodiment, the data set input into the network model includes human presence data and non-human presence data, wherein:

[0061] The human presence data includes data on normal human activities, data on a human body lying motionless on a chair, data on a human body lying motionless on a table, and data on a human body lying down and sleeping.

[0062] The unmanned data includes data about no one in the environment, data about only fans in the environment, data about only curtains in the environment, and data about only movable smart furniture in the environment.

[0063] For ease of understanding, this embodiment is described below using a detection process as an example, see Figure (2). As shown in Figure (2), after the millimeter wave radar receives the echo signal of the frequency modulated continuous wave, the millimeter wave radar echo signal is superimposed with the frequency and phase of the frequency conversion signal emitted by the radar to obtain an intermediate frequency signal. After the intermediate frequency signal is filtered out by a bandpass filter to remove the low frequency signal, all signals in a frame are subjected to time-phase expansion, time-space expansion, one-dimensional Fourier transform, and two-dimensional Fourier transform to obtain a phase difference matrix, a space-time information matrix, a distance information matrix, and a distance-Doppler information matrix. The information data in the obtained frame are then superimposed; the phase difference matrix, the space-time information matrix, the distance information matrix, and the distance-Doppler information matrix are then input into the ConvRNN network for feature extraction, and then the obtained features are input into the improved ResNet network for human target detection.

[0064] In one frame of signal, 128 transmission signals are transmitted to detect the environment, and 128 intermediate frequency signals are obtained. The echo signal is sampled to obtain 128 sampling points. The discrete form of the intermediate frequency signal in one frame of signal is a 128*128 matrix. Each row represents the echo data received from different transmission signals in the sent frame of signal. Filtering is performed row by row, and a set of Sinc filters with a length of Nf=49 is used as a bandpass filter to filter out low-frequency noise. Then, the Blackman-Harris Window is used to perform windowing on the data. The data in each row is multiplied point by point by the window function to obtain the windowed data to reduce spectrum leakage.

[0065] The size of the data after windowing remains unchanged, which is also a 128*128 matrix. The time phase is unfolded to calculate the phase difference between the previous moment and the next moment, which is reflected in the matrix as the phase difference between the next row and the previous row. The default initial phase difference is zero, so the size of the phase information matrix obtained after time-phase unfolding is also 128*128. At the same time, the millimeter-wave radar has three antennas, and three phase information matrices of the same size will be obtained. Superimposing them together, a 128*128*3 phase information matrix is ​​obtained.

[0066] The windowed data is then expanded in time and space, that is, the spatial phase difference between different antennas in slow time is calculated, and then the accumulation of spatial phase difference is recorded. The windowed data is a 128*128 matrix. The millimeter-wave radar has three antennas, so three 128*128 matrices are obtained, representing the data received by different antennas. The phases of different antennas are subtracted and summed to obtain the target's time-space information. Since there are three radar antennas, there are three arrangements for subtracting two antennas, so three time-space information matrices of the same size are obtained. Superimposing them together, a 128*128*3 time-space information matrix is ​​obtained.

[0067] The windowed data is subjected to one-dimensional and two-dimensional Fourier transform row by row to obtain three distance information matrices and range-Doppler information matrices of size 128*128, which are superimposed together to obtain two distance information matrices and range-Doppler information matrices of size 128*128*3.

[0068] ConvRNN refers to a hybrid network structure that combines convolutional neural network (CNN) and recurrent neural network (RNN). It combines CNN's ability to process spatial information with RNN's ability to model time series data. The basic idea is to introduce recurrent connections based on CNN, allowing the network to process time series data and giving the network a certain memory capacity. In ConvRNN, the convolution layer is used to extract spatial feature information from time series data, while the recurrent layer processes the changes in the time series and is used to extract the speed features of the input radar data, that is, the micro-motion feature information of the human body.

[0069] In this example, the convolution kernel size used by ConvRNN is 3*3, and 128 feature maps are generated. Then the recurrent layer of RNN uses 128 recurrent neurons, and the spatial information features obtained after convolution are input into RNN again. RNN finally outputs the matrix of hidden state, which is the micro-motion information features of the human body. Finally, the spatial information features and micro-motion information features of the human body are input into the improved ResNet network.

[0070] The ResNet network is a convolutional neural network composed of multiple residual blocks. The residual blocks contain multiple convolutional layers and nonlinear activation functions. Through the residual blocks, the network can be made deeper, the classification effect is better, and the problem of gradient disappearance is solved.

[0071] The ResNet network in this example uses a fully pre-activated residual block. The specific network structure is shown in Figure (3). BN and ReLU activation functions are added before the convolution layer in the residual block. The input is first normalized and then the activation function is used to make it nonlinear. In this way, the gradient can flow unimpeded to any other layer through the shortcut connection.

[0072] In this example, average pooling is introduced in the shortcut connection of the ResNet network. Since the ResNet shortcut connection uses 1*1 downsampling with a step size of 2, it will cause 3 / 4 of the data loss. Average pooling is introduced before the downsampling operation. The pooling size is 2*2 and the step size is 2. The step size in the shortcut connection is changed to 1, so that the model can take more consideration of the overall characteristics of the data.

[0073] The ResNet network in this example is a modification of the original ResNet18 residual block. The Bottleneck structure is used to replace the original two-layer convolution residual structure. Bottleneck uses three convolution kernels with sizes of 1*1, 3*3, and 1*1, respectively, as residual blocks. It has one more convolution layer than the original ResNet18 residual block, so the total number of convolution layers becomes 26, which is called ResNet26.

[0074] In this example, the activation function of the ResNet26 network uses the ReLU activation function, and the pooling uses an Average pool of size 7*7. The output of the average pooling is connected to the fully connected layer, and then classified by Softmax.

[0075] After the model training is completed, the test data is input into the network model for human target detection.

[0076] In addition, the installation schematic diagram of a millimeter wave radar human target detection method based on deep learning of the present invention is shown in Figure (4). The millimeter wave radar is placed at a height of 3 meters. The millimeter wave radar is a 1-transmit 3-receive frequency modulated continuous wave millimeter wave radar with a center frequency of 60.75GHz, an intermediate frequency signal gain of 33, a maximum measurement data of 2.45m / s, a distance resolution of 0.15m, a speed resolution of 0.2m / s, and a sampling frequency of 1MHz for the echo signal. The emitted frequency modulated continuous wave hits the human body within a range of 4.8 meters and then bounces back to form an echo signal. The echo is then further processed and input into the model for prediction to obtain the result of human body detection.

[0077] The above description is only a description of the preferred embodiments of the present invention, and does not limit the scope of the present invention. Any modification, equivalent replacement or improvement made within the spirit and principle of the present invention based on the above disclosure should be included in the protection scope of the present invention.

Claims

1. A millimeter wave radar human target detection method based on deep learning, characterized in that the following step: S1. The millimeter wave radar transmits a frequency modulated continuous wave to obtain an intermediate frequency signal; a finite impulse response filter is used to perform convolution processing on the intermediate frequency signal to filter out low-frequency signals; the intermediate frequency signal is then windowed to reduce spectrum leakage; after the windowing processing is completed, the intermediate frequency signal is time-phase and time-space-time unfolded to obtain a time-phase difference matrix and a time-space-time information matrix, and then Fourier transform is performed to obtain a distance information matrix and a distance-Doppler information matrix. S2. The millimeter-wave radar has three receiving antennas. Within one frame, the intermediate frequency signals received by different antenna signals are processed by S1, and the results of time-phase and time-space-time expansion are data-associated. At the same time, the distance information matrix and the speed information matrix are fused together to form a three-dimensional phase difference matrix, a space-time information matrix, a distance information matrix and a distance-Doppler information matrix. S3. Input the phase difference matrix, space-time information matrix, distance information matrix and distance-Doppler information matrix into the deep learning neural network model, and then adjust the feature extraction capability of the network model according to the recognition accuracy of the model.

2. The method for detecting human targets using a millimeter wave radar based on deep learning according to claim 1, characterized in that: The step S1 specifically comprises the following sub-steps: S11. The intermediate frequency signal is filtered along the signal at a fast time using a complex finite impulse response filter; the intermediate frequency signal is filtered along the slow time by convolution with a bandpass filter to filter out low-frequency signals, and the bandpass filter is a group of Sinc filters with a length of Nf=49, which reduces the complexity of subsequent calculations and memory overhead. S12: After filtering out the low-frequency data of the millimeter-wave radar intermediate frequency signal, a Blackman-Harris Window algorithm is used to process the intermediate frequency signal of a single transmission signal to reduce spectrum leakage. S13, time-phase unfolding of the intermediate frequency signal. In one frame of the millimeter wave radar, multiple detection signals are transmitted and multiple intermediate frequency signals are received. The phase difference between two adjacent intermediate frequency signals on a single antenna is calculated to record the movement of the detected target. The phase φ of the Nth sample N The calculation formula is as follows: f N =Dψ=ψ t -ψ t-1 (1) where ψ t and ψ t-1 is the phase of two adjacent intermediate frequency signals on the same antenna, and t is the slow time index; by performing time-phase expansion on the intermediate frequency signal, the phase change information in a frame of signal can be obtained, and the time-phase difference matrix of the previous and next signals in a frame can be obtained. S14, perform time-space expansion on the intermediate frequency signal, first calculate the spatial phase difference between different antennas in the slow time, and then create a synthetic signal to record the accumulation of the spatial phase difference in the slow time. The phase of the synthetic signal of the Nth sample The calculation formula is as follows: in, and is the phase of the first and second antennas, and t is the slow time index; each frame of the intermediate frequency signal is time-space expanded to obtain the reception situation of the same transmitted signal on different antennas, and the time-space information matrix is ​​obtained.

3. The method for detecting human targets using millimeter wave radar based on deep learning according to claim 1, characterized in that: The step S2 specifically comprises the following sub-steps: S21. The millimeter-wave radar has three antennas. The same antenna will continuously receive multiple intermediate frequency signals within one frame. The phases of two consecutive intermediate frequency signals received on the same antenna are differentiated, and the default initial phase difference is set to zero to obtain three phase difference matrices. The phase difference matrices are then superimposed to form a three-dimensional phase difference matrix. S22. The millimeter-wave radar has three antennas. At the same time in a frame, different antennas will receive intermediate frequency signals from the same transmission signal. The phases of the intermediate frequency signals of the different antennas are differentiated and then summed in slow time to obtain a space-time information matrix. The three obtained space-time information matrices are superimposed to obtain a three-dimensional space-time information matrix. S23, superimposing the distance information matrix and the range-Doppler information matrix obtained by performing one-dimensional and two-dimensional Fourier transform on the intermediate frequency signals of the different antennas to obtain a three-dimensional distance information matrix and a range-Doppler information matrix.

4. The method for detecting human targets using a millimeter wave radar based on deep learning according to claim 1, characterized in that: The step S3 specifically comprises the following sub-steps: S31. The training model includes a ConvRNN network and an improved ResNet network. The three-dimensional phase difference matrix, the spatiotemporal information matrix, the distance information matrix, and the distance-Doppler information matrix are respectively transmitted to the ConvRNN network to extract the distance feature information of the human target, and then the distance feature information extracted from the human target is transmitted to the next layer of the ResNet network. S32. The ConvRNN network includes a convolutional neural network and a recurrent neural network. The convolutional neural network extracts spatial feature information of the human target based on the spatiotemporal information matrix and the distance information matrix of the human target. The recurrent neural network extracts micro-motion feature information of the human target based on the speed information matrix. S33. The improved ResNet network uses a fully pre-activated residual block, changes the connection order of the activation function in the original residual block, so that the gradient can flow unimpeded to any other layer through the shortcut connection; introduces average pooling in the shortcut connection to reduce data loss; and introduces a Bottleneck structure, using the Bottleneck structure to replace the original two-layer residual block structure, so that the number of convolutional layers in each residual block is increased by one layer. S34, the improved ResNet network performs model training according to the spatial feature information and micro-motion feature information of the human target obtained by the ConvRNN network to complete human target detection.

5. A millimeter wave radar human target detection method based on deep learning, characterized in that: The data set input into the network model includes human presence data and non-human presence data, wherein: The human presence data includes data of a human body standing still, data of a human body lying still on a chair, data of a human body lying still on a table, and data of a human body lying down and sleeping. The no-person data includes data that there is no person in the environment, data that only fans exist in the environment, data that only curtains exist in the environment, and data that only mobile smart devices exist in the environment.

Images