Wall-penetrating radar feature enhancement method and system based on preseg-gan

Abstract

The application discloses a through-wall radar feature enhancement method and system based on preseg-GAN, which adds an unsupervised image segmentation network as an auxiliary branch in the discriminator of the DCGAN network, utilizes the physical prior information of human body motion shown by the image segmentation, enhances the ability of the discriminator, and prompts the generator to generate a human body micro-Doppler feature time-frequency graph which is closer to the actual situation. Finally, the traditional classifier and the CNN-based classifier are trained by using the data set expanded by the PreSeg-DCGAN and other several generation models, and the effectiveness verification result shows that the accuracy of the data set expanded based on the Segment DCGAN is 98.61% under the Resnet18 algorithm, reaching the highest; the method can expand the data set with high precision based on the existing action data, improve the subsequent through-wall human body action recognition effect, and the generated human body micro-Doppler feature time-frequency graph is closer to the actual scene, and the method has the ability of generating high-quality human body micro-Doppler feature graphs.

Description

Technical Field

[0001] This invention relates to the field of radar or radar-based motion recognition technology, and in particular to a method and system for augmenting human motion data using through-wall radar based on neural networks. Background Technology

[0002] Radar is a wireless detection and sensing technology, recognized internationally as a cutting-edge emerging technology. This technology uses special electromagnetic waves emitted by radar as a carrier. Due to the characteristics of these electromagnetic waves, they can penetrate obstacles such as clothing, rubble, and walls. When these electromagnetic waves strike a target surface, they are reflected. The movement and shape of a human target affect the reflection of these waves, allowing the radar to predict the target's location when it receives the reflected waves. Human movements cause varying degrees of micro-movements in different body parts; capturing these micro-movements allows for motion recognition and even target identification. Radar possesses the characteristics of being non-contact, having strong penetrating power, and enabling precise positioning, making it a promising technology with broad applications in fields such as biomedicine, national security, and emergency rescue.

[0003] Human motion recognition through walls plays a crucial role in counter-terrorism and human-computer interaction. While techniques such as wall removal processing and feature enhancement for through-wall radar echoes are largely mature, the problem of oversaturation in recognition models due to limited data volume remains to be addressed. This issue is particularly critical now that deep learning methods have become the mainstream approach in human motion detection and recognition research. Deep network models require large datasets for training to avoid overfitting and meet the demands of robustness and accuracy in practical applications. However, obtaining large datasets of motion data is often difficult in practice, leading to decreased accuracy in motion recognition and posing a significant challenge to human motion identification and classification. Therefore, how to expand existing motion data with high accuracy under through-wall conditions has become a hot and challenging issue in human motion recognition research. Summary of the Invention

[0004] This invention aims to provide a method for augmenting radar data in wall-penetrating scenarios. Based on existing motion data, the proposed neural network framework is used to augment the dataset with high precision, thereby solving the problems of overfitting and poor robustness of the recognition network model caused by the small amount of motion sample data.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A method for enhancing the features of through-wall radar based on preseg-GAN, comprising the following steps:

[0006] Perform an inverse Fourier transform on the radar echo data in SFCW mode to obtain the original range-time map.

[0007] The RPCA method is used to remove the DC component caused by walls and multipath in the original time-distance map, resulting in a clean distance-time map.

[0008] STFT is applied to a clean time-range map to obtain multiple time-frequency map slices, and a JRTFR radar data cube RDC is constructed. In the three-dimensional data volume, the maximum projection of each time-frequency map along the range dimension is extracted based on the R-max TFR method to obtain a high-quality time-frequency map.

[0009] The obtained high-quality time-frequency plot is used as the real image. The PreSeg-DCGAN neural network outputs a time-frequency plot of human movement through the wall that is closer to the actual scene.

[0010] Furthermore, the PreSeg-DCGAN neural network includes an improved DCGAN network structure and a segmentation network. The segmentation network is an auxiliary branch in the DCGAN discriminator, which is pre-trained and used to extract physical prior information about human motion and incorporate it into the DCGAN structure. The DCGAN is used to generate high-quality time-frequency maps. After training the DCGAN, the generator is retained and the discriminator is discarded.

[0011] Furthermore, the improved DCGAN network structure is based on the cGAN network structure and the DCGAN network structure. The generator G and discriminator D of DCGAN are improved: First, the label is input into the Embedding layer for encoding, then concatenated with random noise space, input into the fully connected layer for information integration, and then reshaped into a two-dimensional image. The number of neurons in the fully connected layer is increased, and the number of convolutional modules is also increased.

[0012] The one-dimensional label vector is input into the Embedding layer for encoding, then into the fully connected layer for information integration, and finally reshaped into a two-dimensional image with the same resolution as the original image. The two-dimensional image generated by the reshaped one-dimensional label vector is input into the convolution module for feature extraction to eliminate the data differences between the two. The stitched multi-channel image is then input into the discriminator D to distinguish between the fake image and the real image.

[0013] Furthermore, the training process of the PreSeg-DCGAN neural network includes the following steps:

[0014] The distance-time map of the removed wall is obtained based on the RPCA method. The time-distance map of the removed wall is processed based on the STFT and R-max TFR methods to construct a high-quality time-frequency map dataset, which is used as the training label for the DCGAN network generator.

[0015] Digital labels representing the micro-Doppler features of human movement are input into the Embedding layer of the DCGAN network for encoding, then concatenated with random noise space, and input into the fully connected layer for information integration. After reshaping, the image is converted into a low-resolution 2D image of 64×64. The batch size is set, and the output of DCGAN, the label, and the loss function of the discriminator are calculated.

[0016] The Segment image segmentation network is pre-trained. The two-dimensional image generated by the DCGAN generator is input into the Segment network for image segmentation and information extraction. The output and the output after the one-dimensional label vector is upsized are concatenated and input into the discriminator for discrimination.

[0017] The generator or discriminator is fixed based on the loss function, and the parameters of the discriminator or generator are optimized using the Adam method based on the loss function.

[0018] The entire action dataset is traversed multiple times to train the PreSeg-DCGAN network. The trained parameters are saved to obtain the trained PreSeg-DCGAN network.

[0019] Furthermore, the Segment image segmentation network, as an auxiliary branch in the DCGAN discriminator, is used to extract features from image segmentation information, incorporating prior physical information about human motion into the network structure. During the alternating training of the generator and discriminator, the weight parameters of the Segment image segmentation network remain constant. The training of the Segment image segmentation network includes the following steps:

[0020] The input image is coarsely segmented to obtain the coarsely segmented image;

[0021] Adjust the algorithm parameters to assign the same image label to regions with the same information in the coarsely segmented image;

[0022] Using a convolutional neural network, features are extracted and classified from the input image, so that the output image segmentation result is close to the coarsely segmented image. The training continues until convergence.

[0023] Furthermore, the unsupervised Felzenszwalb algorithm is used for coarse segmentation of the input image. The Felzenszwalb algorithm determines the boundary between two regions based on the image representation of the graph, selects the image edges, and each pixel corresponds to a node on the undirected graph. The weight of each edge describes the degree of dissimilarity between two pixels of the node. The segmentation criteria are adjusted according to the degree of change of adjacent regions.

[0024] Furthermore, in the radar echo data, the center frequency of the transmitted signal is 2.39 GHz, and the system operating frequency is 380 MHz-4.4 GHz.

[0025] Based on the technical concept of the method, the present invention provides a through-wall radar feature enhancement system based on GAN and Attention U-NET structure, including a distance-time map acquisition module, a distance-time map processing module, a time-frequency map acquisition module, and a time-frequency map processing module;

[0026] The range-time map acquisition module is used to perform inverse Fourier transform on radar echo data in SFCW mode to obtain the original range-time map.

[0027] The distance-time map processing module is used to remove the DC component caused by walls and multipath in the original time-distance map according to the RPCA method, so as to obtain a clean distance-time map.

[0028] The time-frequency map acquisition module is used to apply STFT to a clean time-range map, obtain multiple time-frequency map slices, and construct the JRTFR radar data cube RDC; in the three-dimensional data volume, the maximum projection of each time-frequency map along the range dimension is extracted based on the R-max TFR method to obtain a high-quality time-frequency map;

[0029] The time-frequency plot processing module is used to input the obtained high-quality time-frequency plot results as real image inputs to the PreSeg-DCGAN neural network, which outputs a time-frequency plot of human motion through walls that is closer to the actual scene.

[0030] The present invention also provides a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and the processor can realize the preseg-GAN-based through-wall radar feature enhancement method described in the present invention when executing part or all of the computer executable program.

[0031] A computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, it can implement the preseg-GAN-based through-wall radar feature enhancement method described in this invention.

[0032] Compared with existing technologies, this invention has at least the following advantages: Firstly, this invention proposes using a multi-branch GAN neural network to augment through-wall radar motion data. Based on a segment network structure, this invention extracts features from image segmentation information and integrates prior physical information about human motion into an improved DCGAN network structure, thereby improving the realism of the synthesized human motion micro-Doppler features. Compared to the improved DCGAN network, PreSeg-DCGAN, after training, can generate time-frequency maps that are closer to real-world scenarios, further reducing background interference components. This invention can accurately augment existing motion data, improving the subsequent through-wall human motion recognition effect. The generated human micro-Doppler feature time-frequency map is closer to the real-world scenario, demonstrating the ability to generate high-quality human micro-Doppler feature maps. Attached Figure Description

[0033] Figure 1 This is a block diagram illustrating the working principle of radar.

[0034] Figure 2 This is a block diagram of the Preseg-DCGAN architecture.

[0035] Figure 3 This is a diagram of the improved DCGAN structure.

[0036] Figure 4 This is a diagram of the Segment network structure.

[0037] Figure 5 The radar system used and the experimental scenario are described.

[0038] Figure 6 This is a comparison chart of the results from the method. Detailed Implementation

[0039] The present invention will be described in detail below with reference to specific embodiments.

[0040] This embodiment provides a brief explanation of the working principle and system parameters of UWB-based through-wall radar: Figure 1 The system principle block diagram of through-wall radar is as follows: Figure 1As shown, the system includes a transmitter subsystem, a receiver subsystem, an antenna subsystem, an A / D subsystem, and a control and processing subsystem. After the system software issues a working command, the waveform generator generates a baseband signal and mixes it with the local oscillator signal to achieve up-conversion. The mixed signal is then amplified by the transmitter power divider and sent to the transmitting antenna. The receiver subsystem is responsible for mixing and filtering the echo signal received by the receiving antenna with the internal local oscillator signal to convert it into an intermediate frequency (IF) signal. The IF signal is then quadrature-demodulated to obtain the baseband signal. After being digitized and sampled by the A / D subsystem, the baseband signal is sent to the control and processing subsystem. The processing subsystem and the antenna subsystem employ planar logarithmic spiral antennas with opposite polarizations (left-handed transmitting and right-handed receiving). The transmitting antenna radiates electromagnetic wave signals into space, while the receiving antenna receives echo signals and feeds them into the receiver. The antenna subsystem is flexibly connected to the rest of the radar system via fiber optic and coaxial cables, allowing for flexible arrangement of the transmitting and receiving antennas to form the required array configuration based on actual detection needs. The control and processing subsystem runs multi-information sensing system software and is responsible for parameter setting, command transmission, echo processing, and data storage to provide a human-machine interface for the bio-radar system.

[0041] The main technical specifications of the system are shown in Table 1. The array antenna is configured with an operating bandwidth of approximately 4 GHz, ensuring both improved penetration and sensitivity while meeting the requirements for multi-information perception of the human body. The radar involved in this invention adopts a stepped frequency system. The center frequency of the transmitted signal is 2.39 GHz, and the system operating frequency is 380 MHz-4.4 GHz. The large bandwidth ensures high range resolution of the radar, while the lower frequency band ensures that it can penetrate walls to obtain the movement information of people behind the walls.

[0042] Based on unsupervised image segmentation networks in neural networks, a Preseg-DCGAN network is proposed that can efficiently generate human micro-Doppler feature maps. The proposed network further optimizes the structure of the improved DCGAN by adding a segmentation network branch to the discriminator of the improved DCGAN. This branch is used to extract physical prior auxiliary information from the feature map and incorporate it into the DCGAN network, thereby enhancing the realism of the synthesized human action micro-Doppler features.

[0043] The principle behind generating high-quality through-wall radar feature maps is explained as follows: The final generated radar feature map in this invention is a time-frequency map. The time-frequency map contains rich information about human actions, and using it as a feature representation of human actions behind a wall can improve the accuracy of subsequent human action classification. Unlike free space, the presence of a wall causes the feature map to contain more noise, which manifests as a DC component in the time-frequency map. This can mask the motion features of the target human behind the wall. Traditional GAN ​​structures generate image feature components that are discretized and lack continuity. Therefore, this invention incorporates a Segment image segmentation network into the DCGAN discriminator to provide the network with physical prior auxiliary information, making the generated wall-free feature map more realistic.

[0044] Table 1

[0045]

[0046] Figure 2 This invention presents the overall structure of the Preseg-DCGAN neural network proposed in this paper. The neural network structure proposed in this invention generates results that more closely resemble real-world scenarios, further reducing background interference. The Preseg-DCGAN neural network framework includes an improved DCGAN network structure and a segmentation network. The segmentation network is an auxiliary branch of the DCGAN discriminator, pre-trained, and incorporates prior physical information about human motion into the DCGAN network structure. Furthermore, the weight parameters of the segmentation network remain constant throughout the alternating training of the generator and discriminator. The DCGAN is used to generate high-quality time-frequency maps. After training the DCGAN, the generator is retained, while the discriminator is discarded. The two parts of the proposed framework will be discussed in detail below.

[0047] The improved DCGAN network structure specifically involves combining the characteristics of cGAN and DCGAN network structures, and improving the DCGAN generator G as follows: Figure 3 As shown on the left. Digital labels representing the micro-Doppler features of human movement are input into the DCGAN network in the same manner as the cGAN network input. First, the labels are encoded in the embedding layer, then concatenated with random noise, and input into the fully connected layer for information integration. Then, reshaping is performed to generate a low-resolution 64×64 2D image. The number of neurons in the fully connected layer is increased to improve the initial resolution of the image. Furthermore, the number of convolutional modules is increased for better feature extraction, improving the performance of the generator G in generating indistinguishable images, while reducing the number of upsampling layers to prevent information loss as the image resolution increases. Improvements to the discriminator D of DCGAN are as follows: Figure 3 As shown on the right.

[0048] To simultaneously input labels and either fake or real images into the discriminator D, where the labels are one-dimensional vectors and the images are two-dimensional, directly concatenating them is not feasible. This invention designs a dimensionality transformation network that inputs the one-dimensional label vector into the output two-dimensional image for concatenation with the image. First, the one-dimensional label vector is input into the embedding layer for encoding, then into a fully connected layer for information integration, and finally reshaped into a two-dimensional image with the same resolution as the image. To prevent excessive data differences between the two images to be concatenated, the reshaped two-dimensional image from the one-dimensional label vector is input into a convolutional module for feature extraction to eliminate these differences. The concatenated multi-channel image is then input into the discriminator D to distinguish between fake and real images. The addition of a convolutional module to the discriminator D improves its performance in distinguishing between fake and real images.

[0049] Segment image segmentation network structure:

[0050] Due to the lack of labeled data after segmentation, this invention employs an unsupervised image segmentation network. In unsupervised scenarios, no labels are provided beforehand for training images or pixels. Therefore, the target image is input, pixel labels and feature representations are jointly optimized, and the network's weight parameters are updated using gradient descent. Unsupervised image segmentation algorithms are shown in Table 2, where PreSeg() is a classic unsupervised image segmentation algorithm. This invention uses the Felzenszwalb algorithm, a classic unsupervised image segmentation algorithm. The fully convolutional neural network structure used by Net() is as follows... Figure 4 As shown, when the target image is input, the following two sub-problems are addressed alternately: predicting cluster labels with fixed network parameters and training network parameters with fixed predicted cluster labels. The former corresponds to the network's forward convolution process and superpixel thinning. The latter corresponds to the backpropagation process of the gradient descent-based network. Similar to supervised learning, the cross-entropy loss between the network output xn and the improved cluster label cn′ is calculated. Then, through backpropagation, the network weight parameters are updated using stochastic gradient descent.

[0051] First, a classic unsupervised image segmentation algorithm is used to coarsely segment the input image. The algorithm parameters are then adjusted to assign the same image label to regions with similar image information. Next, a convolutional neural network is used to extract features and classify the input image, making the output image segmentation result close to the coarse segmentation result. Training continues until convergence.

[0052] Table 2 - Unsupervised Image Segmentation Algorithms

[0053]

[0054] Based on the technical concept of the method, the present invention provides a through-wall radar feature enhancement system based on GAN and Attention U-NET structure, including a distance-time map acquisition module, a distance-time map processing module, a time-frequency map acquisition module, and a time-frequency map processing module;

[0055] The range-time map acquisition module is used to perform inverse Fourier transform on radar echo data in SFCW mode to obtain the original range-time map.

[0056] The distance-time map processing module is used to remove the DC component caused by walls and multipath in the original time-distance map according to the RPCA method, so as to obtain a clean distance-time map.

[0057] The time-frequency map acquisition module is used to apply STFT to a clean time-range map, obtain multiple time-frequency map slices, and construct the JRTFR radar data cube RDC; in the three-dimensional data volume, the maximum projection of each time-frequency map along the range dimension is extracted based on the R-max TFR method to obtain a high-quality time-frequency map;

[0058] The time-frequency plot processing module is used to input the obtained high-quality time-frequency plot results as real image inputs to the PreSeg-DCGAN neural network, which outputs a time-frequency plot of human motion through walls that is closer to the actual scene.

[0059] Optionally, the present invention also provides a computer device, including a processor and a memory, the memory for storing a computer executable program, the processor reading part or all of the computer executable program from the memory and executing it, and the processor executing part or all of the compute executable program can implement the steps of the through-wall radar feature enhancement method based on GAN and Attention U-NET structure described in the present invention.

[0060] And a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps of the through-wall radar feature enhancement method based on GAN and Attention U-NET structure described in this invention.

[0061] A program that can be written in a computer programming language to perform the methods described in this application can be used. The computer program can be in the form of source code, object code, executable file or some intermediate form. The computer programming language can be C++, Java, Fortran, C# or Python.

[0062] The device for compressing the deep neural network model based on combined pruning can be a laptop, tablet, desktop computer, mobile phone, or workstation.

[0063] The processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or an off-the-shelf programmable gate array (FPGA).

[0064] The memory described in this invention can be an internal storage unit of a laptop, tablet, desktop computer, mobile phone, or workstation, such as memory or hard disk; or it can be an external storage unit, such as a portable hard disk or flash memory card.

[0065] Computer-readable storage media can include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media can include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. Random access memory can include resistive random access memory (ReRAM).

[0066] Design experiments to evaluate the effectiveness of the above methods:

[0067] Experimental setup: The data acquisition scenario in this invention Figure 5 The antenna rests against the outer surface of a single-layer reinforced concrete wall. The wall is 30 centimeters thick. A human subject stands at the radar's front edge and performs different types of designated actions. This invention employs... Figure 6 The 2-transmitter, 4-receiver radar system shown is used to collect data on the echoes of human movement targets behind a wall. Using a 2-transmitter, 4-receiver MIMO radar system can reduce data acquisition time and obtain a variety of datasets, reducing the uniformity caused by single-transmitter, single-receiver radar systems collecting data at the same location.

[0068] There are nine common actions, including walking forward (type a), walking backward (type b), sitting in a chair (type c), standing up from a seat (type d), picking up an object (type e), crawling forward (type f), crawling backward (type g), falling forward (type h), and falling backward (type i). The high-quality time-frequency maps obtained by processing the time-distance maps after wall removal using Principal Component Analysis (RPCA) with STFT and R-max TFR methods are considered as ground truth labels for training the generator.

[0069] Segmentation results: Results of augmenting human action Doppler feature data from three data models: cGAN, improved DCGAN, and PreSeg-DCGAN networks are as follows: Figure 6 As shown, images generated using cGAN exhibit a noticeable "graininess," with discretized feature components lacking continuity. Furthermore, images augmented with cGAN data suffer from significant background noise. While images generated by the improved DCGAN are smoother than those generated by cGAN, they still exhibit considerable background noise. The PreSeg-DCGAN network, by incorporating information extracted from the time-frequency map segmentation of human motion micro-Doppler features, provides data augmentation results that more closely resemble real-world scenarios, further reducing background noise.

[0070] Method Effectiveness Analysis: The effectiveness of the proposed structure was verified by training traditional machine learning classifiers PCA+SVM, PCA+KNN, and a CNN-based classifier, ResNet18. Table 3 shows the human classification accuracy of the three data models using PCA+SVM, PCA+KNN, and ResNet18, respectively. The dataset augmented by PreSeg-DCGAN achieved the highest accuracy of 98.61% using the ResNet18 algorithm, demonstrating that this invention has the ability to generate high-quality human micro-Doppler feature maps and effectively expand the dataset.

[0071] Table 3 Accuracy of Human Movement Classification

[0072]

[0073] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for enhancing the features of through-wall radar based on preseg-GAN, characterized in that, Includes the following steps: Perform an inverse Fourier transform on the radar echo data in SFCW mode to obtain the original range-time map. The RPCA method is used to remove the DC component caused by walls and multipath in the original time-distance map, resulting in a clean distance-time map. STFT is applied to a clean time-range map to obtain multiple time-frequency map slices, and a JRTFR radar data cube RDC is constructed. In the three-dimensional data volume, the maximum projection of each time-frequency map along the range dimension is extracted based on the R-max TFR method to obtain a high-quality time-frequency map. The obtained high-quality time-frequency plot results are used as real images. The PreSeg-DCGAN neural network outputs a time-frequency plot of human movement through the wall that is closer to the actual scene. The PreSeg-DCGAN neural network includes an improved DCGAN network structure and a segmentation network. The segmentation network is an auxiliary branch in the DCGAN discriminator, pre-trained, and used to extract physical prior information about human motion and incorporate it into the DCGAN structure. The DCGAN is used to generate high-quality time-frequency maps. After training the DCGAN, the generator is retained and the discriminator is discarded. The improved DCGAN network structure is based on the cGAN and DCGAN network structures, and improves the DCGAN generator. Improvements to the discriminator D: First, the label is input into the Embedding layer for encoding, then concatenated with random noise space, input into the fully connected layer for information integration, and then reshaped into a two-dimensional image. The number of neurons in the fully connected layer is increased, and the number of convolutional modules is also increased. The one-dimensional label vector is input into the Embedding layer for encoding, then into the fully connected layer for information integration, and finally reshaped into a two-dimensional image with the same resolution as the original image. The two-dimensional image generated by the reshaped one-dimensional label vector is input into the convolution module for feature extraction to eliminate the data differences between the two. The stitched multi-channel image is then input into the discriminator D to distinguish between the fake image and the real image.

2. The through-wall radar feature enhancement method based on preseg-GAN according to claim 1, characterized in that, The training process of the PreSeg-DCGAN neural network includes the following steps: The distance-time map of the removed wall is obtained based on the RPCA method. The time-distance map of the removed wall is processed based on the STFT and R-max TFR methods to construct a high-quality time-frequency map dataset, which is used as the training label for the DCGAN network generator. Digital labels representing the micro-Doppler features of human motion are input into the Embedding layer of the DCGAN network for encoding, then concatenated with random noise space, and input into the fully connected layer for information integration. After reshaping, the image is converted into a low-resolution 64×64 two-dimensional image. The batch size is set, and the output of DCGAN, the label, and the loss function of the discriminator are calculated. The Segment image segmentation network is pre-trained. The two-dimensional image generated by the DCGAN generator is input into the Segment network for image segmentation and information extraction. The output and the output after the one-dimensional label vector is upsized are concatenated and input into the discriminator for discrimination. The generator or discriminator is fixed based on the loss function, and the parameters of the discriminator or generator are optimized using the Adam method based on the loss function. The entire action dataset is traversed multiple times to train the PreSeg-DCGAN network. The trained parameters are saved to obtain the trained PreSeg-DCGAN network.

3. The through-wall radar feature enhancement method based on preseg-GAN according to claim 2, characterized in that, The Segment image segmentation network, as an auxiliary branch in the DCGAN discriminator, is used to extract features from image segmentation information. It incorporates prior physical information about human motion into the network structure, and the weight parameters of the Segment image segmentation network remain constant throughout the alternating training of the generator and discriminator. Training the Segment image segmentation network includes the following steps: The input image is coarsely segmented to obtain the coarsely segmented image; Adjust the algorithm parameters to assign the same image label to regions with the same information in the coarsely segmented image; Using a convolutional neural network, features are extracted and classified from the input image, so that the output image segmentation result is close to the coarsely segmented image. The training continues until convergence.

4. The through-wall radar feature enhancement method based on preseg-GAN according to claim 3, characterized in that, The unsupervised Felzenszwalb algorithm is used for coarse segmentation of the input image. The Felzenszwalb algorithm determines the boundary between two regions based on the image representation of the graph, selects the image edges, and each pixel corresponds to a node on the undirected graph. The weight of each edge describes the degree of dissimilarity between two pixels of the node. The segmentation criteria are adjusted according to the degree of change of adjacent regions.

5. The through-wall radar feature enhancement method based on GAN and Attention U-NET structure according to claim 1, characterized in that, The center frequency of the transmitted signal in the radar echo data is 2.39 GHz, and the system operating frequency is 380 MHz-4.4 GHz.

6. A through-wall radar feature enhancement system based on GAN and Attention U-NET architecture, characterized in that, It includes a distance-time map acquisition module, a distance-time map processing module, a time-frequency map acquisition module, and a time-frequency map processing module; The range-time map acquisition module is used to perform inverse Fourier transform on radar echo data in SFCW mode to obtain the original range-time map. The distance-time map processing module is used to remove the DC component caused by walls and multipath in the original time-distance map according to the RPCA method, so as to obtain a clean distance-time map. The time-frequency map acquisition module is used to apply STFT to a clean time-range map, obtain multiple time-frequency map slices, and construct the JRTFR radar data cube RDC; in the three-dimensional data volume, the maximum projection of each time-frequency map along the range dimension is extracted based on the R-max TFR method to obtain a high-quality time-frequency map; The time-frequency map processing module uses the obtained high-quality time-frequency map results as input to the real image. The PreSeg-DCGAN neural network outputs a time-frequency map of human motion through walls that is closer to the actual scene. The PreSeg-DCGAN neural network includes an improved DCGAN network structure and a segmentation network. The segmentation network is an auxiliary branch in the DCGAN discriminator, which is pre-trained and used to extract the physical prior information of human motion and incorporate it into the DCGAN structure. The DCGAN is used to generate high-quality time-frequency maps. After training the DCGAN, the generator is retained and the discriminator is discarded. The improved DCGAN network structure is based on the cGAN and DCGAN network structures, and improves the DCGAN generator. Improvements to the discriminator D: First, the label is input into the Embedding layer for encoding, then concatenated with random noise space, input into the fully connected layer for information integration, and then reshaped into a two-dimensional image. The number of neurons in the fully connected layer is increased, and the number of convolutional modules is also increased. The one-dimensional label vector is input into the Embedding layer for encoding, then into the fully connected layer for information integration, and finally reshaped into a two-dimensional image with the same resolution as the original image. The two-dimensional image generated by the reshaped one-dimensional label vector is input into the convolution module for feature extraction to eliminate the data differences between the two. The stitched multi-channel image is then input into the discriminator D to distinguish between the fake image and the real image.

7. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer-executable program, the processor reading part or all of the computer-executable program from the memory and executing it, and the processor executing part or all of the computed executable program is able to implement the through-wall radar feature enhancement method based on any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the preseg-GAN-based through-wall radar feature enhancement method as described in claims 1 to 5.

Images