Radar target recognition method based on cross-scale feature aggregation network

Abstract

The application discloses a radar target recognition method based on a cross-scale feature aggregation network, and comprises the following steps: collecting skin vibration echoes generated by breathing of a static target by using a radar to obtain a radar echo signal; processing the radar echo signal by using an SSST algorithm to generate a time-frequency image containing breathing features; dividing and aggregating the time-frequency image in different scale sizes by using a feature pyramid network, and extracting a micro-Doppler feature map from feature maps of each scale output by the feature pyramid by using a CSFA network; and inputting the aggregated micro-Doppler feature map into a SoftMax classifier to obtain a recognition result of the static target. The radar echo signal of the static target is processed by using the SSST algorithm, so that clutter such as multipath interference can be suppressed, and then a time-frequency image with high resolution and breathing features is generated. Meanwhile, the micro-Doppler feature map in the time-frequency image is extracted by using the cross-scale feature aggregation network, so that the recognition accuracy of the target is improved, and the efficiency of radar target detection is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal processing, and particularly relates to a radar target recognition method based on a cross-scale feature aggregation network. BACKGROUND

[0002] The ultra-wideband continuous wave radar is an emerging sensing technology, and electromagnetic wave signals emitted by the ultra-wideband continuous wave radar have the characteristics of strong penetration ability and high resolution, and can reflect the scattering point energy distribution and other characteristics of a target, and can be used for detection of vital sign signals.

[0003] In order to avoid the radar echo signal from causing multipath clutter interference due to the environment where the stationary target is located, and to improve the resolution of the target image, it is crucial to improve the accuracy of target recognition. At present, in the related art, a MSRA (Multiscale Residual Attention) network is used to recognize stationary targets, and the method includes three parts of radar signal processing, a multiscale learning architecture, and a residual attention learning mechanism. However, the method uses a two-dimensional pseudo-color image to recognize stationary targets, and the effect of suppressing clutter interference is poor, so that it is difficult to generate a high-resolution image, resulting in that more invalid features are extracted from the image, and the accuracy of target recognition is reduced. SUMMARY

[0004] In order to solve the above problems in the prior art, the application provides a radar target recognition method based on a cross-scale feature aggregation network. The technical problem to be solved by the application is solved by the following technical scheme:

[0005] The application provides a radar target recognition method based on a cross-scale feature aggregation network, comprising:

[0006] The radar is used to collect skin vibration echoes generated by breathing of a stationary target, and radar echo signals are obtained;

[0007] A synchronous squeezing S transform time-frequency analysis algorithm is used to process the radar echo signals, and a time-frequency image containing breathing characteristics is generated;

[0008] A feature pyramid network is used to divide and aggregate the time-frequency image in different scale sizes, and a micro-Doppler feature map is extracted from each scale feature map output by the feature pyramid network by using a CSFA network;

[0009] After an aggregated micro-Doppler feature map is generated based on the micro-Doppler feature map, the aggregated micro-Doppler feature map is input into a SoftMax classifier, and a recognition result of the radar target is obtained.

[0010] In an embodiment of the present application, before the step of generating an aggregated micro-Doppler feature map based on the micro-Doppler feature map, the method further comprises:

[0011] A pre-trained Resnet-18 model is obtained, and a time-frequency map data set model file is loaded;

[0012] After the aggregated micro-Doppler feature map is generated based on the micro-Doppler feature map, the method further comprises:

[0013] After the aggregated micro-Doppler feature map is generated based on the micro-Doppler feature map, the method further comprises:

[0014] In an embodiment of the present application, before the step of processing the radar echo signal by using a synchrosqueezing S transform time-frequency analysis algorithm to generate a time-frequency image containing breathing features, the method further comprises:

[0015] The radar echo signal is preprocessed to obtain a radar echo signal containing breathing features;

[0016] Based on the radar echo signal containing breathing features, a two-dimensional range image is generated by time accumulation.

[0017] In an embodiment of the present application, the step of processing the radar echo signal by using a synchrosqueezing S transform time-frequency analysis algorithm to generate a time-frequency image containing breathing features comprises:

[0018] The effective channel data in the two-dimensional range image is processed by using a synchrosqueezing S transform time-frequency analysis algorithm, and all the effective channel data is aggregated into a time-frequency image.

[0019] In an embodiment of the present application, the radar echo signal is processed by using a synchrosqueezing S transform time-frequency analysis algorithm according to the following formula:

[0020]

[0021] In the formula, f k , f c and Δf c respectively represent the discrete frequency of the S transform, the center frequency of the squeezing interval, and the bandwidth of the squeezing interval, b represents the time axis displacement parameter, and Δf k = f k -f k-1 , Δf c = f c-f c-1 , ST(f k ,b) represents the S transform of the radar echo signal, f c (f k ,b) represents the instantaneous frequency of the radar echo signal.

[0022] In an embodiment of the present application, the CSFA network comprises a channel attention model and a spatial attention model;

[0023] The step of dividing and aggregating the time-frequency image in different scale sizes by using a feature pyramid network and extracting a micro-Doppler feature map from the feature maps of each scale output by the feature pyramid network by using a CSFA network comprises:

[0024] Dividing the time-frequency image in different scale sizes by using a feature pyramid network;

[0025] Inputting the input feature map F i (i = 2, 3, 4, 5) of each scale into the channel attention model to obtain a one-dimensional channel attention feature map M c1 After that, multiplying the one-dimensional channel attention feature map M c1 by the input feature map F i to obtain a channel feature F c1 Inputting the channel feature F c1 into the spatial attention model to obtain a two-dimensional spatial attention feature map M s1 ;

[0026] Multiplying the channel feature F c1 by the two-dimensional spatial attention feature map M s1 to obtain a spatial feature F s ;

[0027] Determining a feedback variable F i by multiplying the input feature map F s and the spatial feature F FB , and inputting the feedback variable F FB into the channel attention model to obtain a one-dimensional channel attention feature map M c2 After that, multiplying the one-dimensional channel attention feature map M c2 by the feedback variable F FB to obtain a channel feature F c2 ;

[0028] Inputting the channel feature F c2 into the spatial attention model to obtain a two-dimensional spatial attention feature map M s2 ;

[0029] Multiplying the channel feature F c2The two-dimensional spatial attention feature map M s2 The input feature map F i The micro-Doppler feature map

[0030] In an embodiment of the present application, the channel attention model comprises a first max-pooling layer, a first average-pooling layer, and a multi-layer fully connected neural network.

[0031] Each scale feature map F i of the feature pyramid network is input into the channel attention model to obtain a channel feature .

[0032] The input feature map F i is transmitted to the first max-pooling layer and the first average-pooling layer respectively to obtain a first max-pooling feature map and a first average-pooling feature map

[0033] The first max-pooling feature map and the first average-pooling feature map are input into the multi-layer fully connected neural network with the activation function ReLU in a forward input manner, and the two output features of the multi-layer fully connected neural network are added to obtain a one-dimensional channel attention feature map M c1 . The one-dimensional channel attention feature map M c1 is multiplied by the current input feature map F i to obtain a channel feature F c1 .

[0034] In an embodiment of the present application, the spatial attention model comprises a second max-pooling layer, a second average-pooling layer, and a convolution kernel of 5*5.

[0035] The channel feature F c1 is input into the spatial attention model to obtain a two-dimensional spatial attention feature map M s1 .

[0036] The channel feature F c1 is input into the second max-pooling layer and the second average-pooling layer respectively, the obtained two-dimensional max-pooling feature map and the two-dimensional average-pooling feature map are combined into a matrix, and after dimension reduction by the convolution layer with the convolution kernel of 5*5, the two-dimensional spatial attention feature map M s1 is generated.

[0037] In an embodiment of the present application, the feedback variable represents a dot product operation. Compared with the prior art, the present application has the beneficial effects that:

[0038] The application provides a radar target recognition method based on a cross-scale feature aggregation network.

[0039] The application will be further described in detail below in combination with the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is a flowchart of the radar target recognition method based on the cross-scale feature aggregation network provided by the embodiment of the application;

[0041] Figure 2 is an image acquisition diagram of the time-frequency image provided by the embodiment of the application;

[0042] Figure 3 is a contrast diagram of different time-frequency images provided by the embodiment of the application;

[0043] Figure 4 is a schematic diagram of the overall architecture of the radar target recognition method based on the cross-scale feature aggregation network provided by the embodiment of the application;

[0044] Figure 5 is a structural schematic diagram of the cross-scale feature aggregation network provided by the embodiment of the application;

[0045] Figure 6 is a contrast diagram of the micro-Doppler feature map of the time-frequency image provided by the embodiment of the application;

[0046] Figure 7a is a schematic diagram of an experimental scene provided by the embodiment of the application;

[0047] Figure 7b is a schematic diagram of another experimental scene provided by the embodiment of the application;

[0048] Figure 8 is a contrast diagram of the convergence curves of the time-frequency analysis algorithm provided by the embodiment of the application.

[0049] Figure 9 is a contrast diagram of the accuracy rate curves of the human recognition results of different training models provided by the embodiment of the application. DETAILED DESCRIPTION

[0050] The application will be further described in detail below in combination with the drawings and embodiments.

[0051] Figure 1 is a flow chart of a radar target recognition method based on a cross-scale feature aggregation network provided by an embodiment of the present application. As shown in Figure 1 , the present application provides a radar target recognition method based on a cross-scale feature aggregation network, comprising:

[0052] S1, collecting the echo of skin micro-vibration caused by target respiration by using a radar to obtain a radar echo signal;

[0053] S2, processing the radar echo signal by using a synchronous squeezing S transform time-frequency analysis algorithm to generate a time-frequency image containing respiration features;

[0054] S3, dividing and aggregating the time-frequency image in different scale sizes by using a feature pyramid network, and extracting a micro-Doppler feature map from the feature maps of each scale output by the feature pyramid network by using a CSFA (cross-scale feature aggregation) network;

[0055] S4, after generating an aggregated micro-Doppler feature map based on the micro-Doppler feature map, inputting the aggregated micro-Doppler feature map into a SoftMax classifier to obtain a recognition result of a radar detection target.

[0056] It should be understood that since the target's heartbeat vibrations are weak, the feature error extracted from the heartbeat signal is large, which will affect the accuracy of target identification. Therefore, this embodiment selects the respiratory signal caused by the target's chest rise and fall as the target radar echo signal. Specifically, the radar obtains the radar echo signal by collecting the skin micro-vibration echo caused by the target's breathing, and performs pre-processing operations such as local oscillator mixing, Hamming window filtering, inverse fast Fourier transform, moving target display, and constant false alarm detection on the radar echo signal. The radar echo accumulates over time to generate a two-dimensional range image. This image can show the change of the target position over time, but cannot represent the target's Doppler characteristic information and has relatively weak noise resistance. In the process of processing radar echo signals, the STFT (Short-time Fourier Transform) algorithm is usually used to obtain micro-Doppler characteristics. However, the resolution problem of this algorithm can lead to spectrum leakage, which will cause the time-frequency image to have a blurring effect. In view of this, an embodiment of the present invention utilizes the SSST (Synchrosqueezing S-Transform) algorithm to generate high-resolution time-frequency images containing respiratory features. The SSST algorithm is a continuous and reversible process for identifying and extracting oscillatory components such as time-varying frequency and amplitude from a uniformly sampled signal. This algorithm can overcome the time and frequency expansion problems in traditional time-frequency analysis methods, and is beneficial for improving spectral resolution, thereby facilitating the subsequent extraction of higher-resolution micro-Doppler feature maps. At the same time, the time-frequency images generated by SSST have significantly improved frequency focusing compared to those generated by STFT. The time-frequency images of respiratory features have a clear envelope and no frequency mutation terms.

[0057] Optionally, the radar echo signal is processed using the SSST time-frequency analysis algorithm according to the following formula:

[0058]

[0059] Where, f k 、f c and Δf c They represent the discrete frequency of S transform, the center frequency of the squeeze interval, and the bandwidth of the squeeze interval, respectively. b represents the time axis displacement parameter, and Δf k =f k -f k-1 , Δf c =f c -f c-1 ,ST(f k ,b) represents the S transform of radar echo signal, f c (f k ,b) represents the instantaneous frequency of the radar echo signal. This formula represents the center frequency f on the S transform spectrum c Surrounding frequency range [fc -0.5Δf c c +0.5Δf c c

[0060] In particular, the stepped frequency waveform s(t) transmitted by the radar can be written as follows:

[0061]

[0062] where N represents the total number of frequency points of the stepped frequency signal, T represents the duration of each frequency point, f0represents the starting carrier frequency, Δf represents the frequency bandwidth, and the function rect(·) is defined as follows: L

[0063]

[0064] The radar echo signal r(t) reflected from the target can be represented as follows:

[0065]

[0066] where K represents the number of target breathing frequency points, τi represents the round-trip time of the ith scattering point, and r(t) and r(t) represent the reflected echo of the wall and other environmental interference and noise, respectively. i ω n

[0067] First, the S transform of the radar echo signal r(t) can be represented as follows:

[0068]

[0069] where t represents time, f represents frequency, and b represents the time axis displacement parameter. Considering that the time-frequency spectrum energy distribution of the signal is at f = f0, but the actual obtained time-frequency spectrum has a false band near f0, the instantaneous frequency expression of the radar echo signal r(t) is as follows:

[0070]

[0071] Finally, the synchronous squeezing S transform of the radar echo signal r(t) can be represented as follows:

[0072]

[0073] According to the above formula, the SSST(f c ​​​​​​​b) a corresponding time-frequency image with breathing characteristics is obtained.

[0074] Before the step of processing the radar echo signal by using the SSST time-frequency analysis algorithm to generate a time-frequency image containing breathing characteristics, the method further comprises:

[0075] The radar echo signal is preprocessed to obtain a radar echo signal containing breathing characteristics.

[0076] Based on the radar echo signal containing breathing characteristics, a two-dimensional range image is generated by time accumulation.

[0077] Figure 2 is a time-frequency image acquisition diagram provided by an embodiment of the present application. Specifically, as shown in Figure 2 The radar echo signal containing breathing characteristics is obtained by performing local oscillator mixing, Hamming window filtering, inverse fast Fourier transform, moving target display, constant false alarm detection and other preprocessing operations on the radar echo signal. The radar echo signal containing breathing characteristics is time-accumulated to generate a two-dimensional range image, which can display the change of target position over time, but cannot represent the Doppler feature information of the target, and the anti-noise capability is relatively weak. By selecting the data of the effective channels in the two-dimensional range image and processing it by using the SSST time-frequency analysis algorithm, the time-frequency image of each channel is obtained, and then all the data of the effective channels are aggregated to generate the final time-frequency image.

[0078] Figure 3 is a different time-frequency image comparison diagram provided by an embodiment of the present application, wherein the first row is a time-frequency image generated by using the SSST time-frequency analysis algorithm, the second row is a time-frequency image generated by using the STFT time-frequency analysis algorithm, the (a) column corresponds to a stationary person, the (b) column corresponds to a stationary pig, the (c) column corresponds to a stationary dog, and the (d) column corresponds to a stationary cat. Obviously, as shown in Figure 3 The image envelope of the time-frequency image generated by SSST is more clear than that of the time-frequency image generated by STFT, and the frequency focusing is significantly improved. At the same time, the SSST time-frequency analysis algorithm relieves the blurring effect of the time-frequency image, and each oscillation component of the radar echo signal can be well concentrated in the time-frequency image.

[0079] Although different targets can be distinguished according to different breathing characteristics in the time-frequency image, there is an overlapping interval between the breathing characteristics of each target, which makes it difficult to distinguish different targets. Therefore, in order to make the breathing characteristics contained in the generated time-frequency image meet the requirements of target recognition, the neural network is needed to further extract the micro-Doppler feature map in the time-frequency image.

[0080] Figure 4: is a schematic diagram of the overall architecture of a radar target recognition method based on a cross-scale feature aggregation network provided by an embodiment of the present invention, Figure 5 This is a schematic diagram of the structure of the cross-scale feature aggregation network provided by an embodiment of the present invention. Figures 4-5 ,CSFA network includes channel attention model and spatial attention model;

[0081] In step S3 above, the steps of using the feature pyramid network to divide and aggregate the time-frequency image into different scales, and using the CSFA network to extract the micro-Doppler feature map from the output feature map of each scale processed by the feature pyramid network include:

[0082] S301, using a feature pyramid network to divide the time-frequency image into different scales;

[0083] S302, the input feature map F of each scale obtained by division i Input channel attention model to obtain one-dimensional channel attention feature map M c1 After that, the one-dimensional channel attention feature map M c1 With the input feature map F i Multiply to get the channel feature F c1 , the channel feature F c1 Input into the spatial attention model to obtain a two-dimensional spatial attention feature map M s1 ;

[0084] S303, the channel feature F c1 and the two-dimensional spatial attention feature map M s1 Multiply to obtain the spatial feature F s ;

[0085] S304, according to the input feature map F i and spatial characteristics F s Multiply to determine the feedback variable F FB , and the feedback variable F FB Input into the channel attention model to obtain a one-dimensional channel attention feature map M c2 After that, the one-dimensional channel attention feature map M c2 and feedback variable F FB Multiply to get the channel feature F c2 ;

[0086] S305, the channel feature F c2 Input into the spatial attention model to obtain a two-dimensional spatial attention feature map M s2 ;

[0087] S306, channel feature F c2 and the two-dimensional spatial attention feature map M s2The input feature map F of the current i-th layer scale is obtained by multiplication i The micro-Doppler feature map of the current i-th layer scale is obtained by multiplication

[0088] It should be understood that the respiratory characteristics contained in the time-frequency image change over time, and the limitation of distance resolution will cause the problem of overlapping intervals in the time-frequency image, so the key to solving the target recognition problem is to extract the differentiated respiratory characteristics of different targets from the time-frequency image.

[0089] In the embodiment, after obtaining the micro-Doppler feature map, the input feature map F of the current i-th layer scale is aggregated with the micro-Doppler feature map of the feature map of the previous layer scale i The micro-Doppler feature map of the current i-th layer scale is obtained by multiplication The micro-Doppler feature map of the current i-th layer scale is obtained by multiplication The aggregated micro-Doppler feature map Map is obtained by upsampling. i-1 Further, the aggregated micro-Doppler feature map is input into a SoftMax classifier to obtain the recognition result of the radar target.

[0090] The inventors found in the research process that the performance of deep CNN is better than that of shallow CNN, which can capture micro-Doppler features from time-frequency images in layers to obtain good classification performance. However, the accuracy of deep CNN is always in a saturated state, which hinders the process of feature optimization. In order to solve this problem, the deep residual learning network is used to extract the feature information of the image in the embodiment, and the convolution features of different levels can improve the complementary feature information, which helps to improve the recognition performance of humans and animals.

[0091] As shown in Figure 4 , Resnet-18 has better feature extraction capability on a small image dataset, and the performance overhead is relatively small, so the Resnet-18 model is used as Backbone to extract convolution features in the embodiment, and the micro-Doppler feature map is extracted from the time-frequency image by the feature pyramid network. Specifically, as shown in Figure 5 , the bottom-up flow line is the forward processing of the network, and the feature layer extracted each time is the output of the previous layer in the same order. The time-frequency image is the input of the feature pyramid network based on Resnet-18, and the output of the feature pyramid network is 4 layers of feature pyramids, i.e. P2, P3, P4, P5, obtained by applying the step size as {4, 8, 16, 32}. In the top-down flow line, the feature maps output by the CSFA network are aggregated by upsampling high-level semantic feature information. Among these feature maps, the high-level feature maps contain more information related to the respiratory characteristics, and the low-level feature maps contain more information of the details of the respiratory characteristics.

[0092] The feature pyramid network can realize multi-feature representation and improve the performance of image target recognition. However, the high-level feature map ignores the information of the respiratory feature details, and the low-level feature map ignores the information related to the respiratory feature. In order to solve this problem, the CSFA architecture is designed to extract the micro-Doppler feature map in the time-frequency image. The network is mainly used for feature aggregation and feature enhancement of each micro-Doppler feature, while paying attention to the depth features of interest and suppressing unnecessary features from the channel and spatial response, thereby effectively extracting the micro-Doppler feature map from the time-frequency image.

[0093] Exemplarily, the CSFA network includes a channel attention model and a spatial attention model, the channel attention model includes a first max-pooling layer, a first average-pooling layer and a multi-layer fully connected neural network, and the spatial attention model includes a second max-pooling layer, a second average-pooling layer and a convolution kernel of 5x5;

[0094] The feature pyramid of each scale feature map F i is input into the channel attention model to obtain a channel feature F ic1 . The step includes:

[0095] The input feature map F i is transmitted to the first max-pooling layer and the first average-pooling layer respectively to obtain a first max-pooling feature map F and a first average-pooling feature map F .

[0096] The first max-pooling feature map F and the first average-pooling feature map F are input into the multi-layer fully connected neural network with the activation function of ReLU in the forward input mode, and the two output features of the multi-layer fully connected neural network are added to obtain a one-dimensional channel attention feature map M c1 . The one-dimensional channel attention feature map M c1 is multiplied by the current input feature map F i to obtain a channel feature F c1 .

[0097] Further, the channel feature F c1 is input into the spatial attention model to obtain a two-dimensional spatial attention feature map M s1 . The step includes:

[0098] The channel feature F c1 is input into the second max-pooling layer and the second average-pooling layer respectively, the obtained two-dimensional max-pooling feature map F and the two-dimensional average-pooling feature map F are combined into a matrix, and after dimension reduction by the convolution layer with the convolution kernel of 5x5, the two-dimensional spatial attention feature map M s1.

[0099] Specifically, the input feature map Among them, C, H, and W represent the input feature map F i The number of channels, image height and image width; for the channel attention model, the input feature map F i After the first maximum pooling layer and the first average pooling layer respectively, the first maximum pooling feature map is obtained And the first average pooling feature map Then, the two feature maps are input into the multi-layer fully connected neural network with ReLU activation function in the forward input mode, and the two output features of the multi-layer fully connected neural network are added to obtain a C×1×1 one-dimensional channel attention feature map M c1 , and finally the one-dimensional channel attention feature map M c1 With the current input feature map F i Multiply to get the channel feature F c1 .

[0100] For the spatial attention model, the channel feature F c1 As its input, it passes through the second maximum pooling layer and the second average pooling layer respectively to obtain the two-dimensional maximum pooling feature map and two-dimensional average pooling feature map Then merge these two feature maps into a matrix, and then reduce the dimension through a convolution layer with a convolution kernel of 5×5 to generate a two-dimensional spatial attention feature map M of H×W×1 s1 , and finally M s1 With F c1 Multiply to obtain spatial feature F s , the processing process can be written as follows:

[0101]

[0102]

[0103] in, represents the input feature map, Indicates multiplication by array elements. The complete process of the attention model can be written as follows:

[0104]

[0105]

[0106] in, is the concatenation operator, σ represents the Sigmoid function, and denotes the weights of the MLP network, ReLU activation function followed by W0, r is the decay ratio, f 5×5 denotes a convolution operation with a kernel size of 5x5.

[0107] In this embodiment, the spatial feature F s is multiplied by the current feature map F i to obtain the feedback variable F FB , and the feedback variable F FB is determined. FB As the new input feature map of the channel attention and spatial attention model, the final output F out is the output feature map processed by the above model. In this way, the feature information of the current layer can use feedback connection to obtain more micro-Doppler feature maps, so that the new features are more different. FB It can be written as follows:

[0108]

[0109] The feedback variable F FB is used as a bidirectional information flow feature, which greatly improves the connection of each feature in the spatial and channel dimensions, so that high-level features and low-level features are both used to achieve high-quality target recognition. Alternatively, before the step of inputting the aggregated micro-Doppler feature map into the SoftMax classifier to obtain the recognition result of the target, the method further comprises:

[0110] obtain the pre-trained Resnet-18 model and load the time-frequency graph dataset model file;

[0111] In step S4, the step of inputting the aggregated micro-Doppler feature map into the SoftMax classifier to obtain the recognition result of the target comprises:

[0112] The aggregated micro-Doppler feature map is input into the SoftMax classifier, so that the SoftMax classifier compares the time-frequency graph dataset model file with the aggregated micro-Doppler feature map. The result after comparison is the recognition result of the target.

[0113] In this embodiment, the pre-trained Resnet-18 model can be obtained before target recognition, and the time-frequency graph dataset model file is loaded. The diversity of the dataset is improved by adding 5-10 dB of Gaussian white noise to the radar echo; further, the aggregated micro-Doppler feature map is input into the SoftMax classifier, and the SoftMax classifier compares and evaluates in the time-frequency graph dataset, and finally outputs the recognition result of the target, completing the recognition of humans and animals.

[0114] Figure 6is a contrastive diagram of micro-Doppler feature maps of time-frequency images generated by different model methods provided by the embodiment of the present application. Figure 6 As shown in the figure, when the CSFA network is introduced, more effective micro-Doppler features can be obtained from the time-frequency image, which can fully reflect the breathing characteristics of different targets.

[0115] Further, the embodiment performs four sets of comparative experiments respectively taking human and three different animals as targets. Specifically, in this experiment, the radar works in a frequency range of 1.0-2.0 GHz, the duration of each frequency is 100 s, the pulse repetition interval (PRI) is 70 ms, and a transmit-receive array antenna is used to transmit and receive signals.

[0116] Figure 7a 、 7b is an experimental scene schematic diagram provided by the embodiment of the present application, in which the thickness of the brick wall is about 0.20 m. Please refer to Figures 7a-7b The human, dog, cat and pig are divided into four groups of experimental objects, and each group performs 10 experiments. In each experiment, the human and animal as experimental objects need to be kept in a stationary state. In order to ensure that the experimental objects are in a stationary state in each experiment, the experimental detection time is set to 90 s. It should be noted that since it cannot be ensured that the cat is always in a stationary state, the cat is placed in a plastic box during the experimental test in this embodiment to avoid interference of metal on radar echo collection. In addition, according to the size of the experimental scene, the sensing range of the through-wall radar is set to 0-5 m.

[0117] All models are configured with NVIDIA GTX 1080 GPU and Intel i5-8400, and the optimizer selects the SGD mode with a learning rate of 0.01. The main evaluation index adopted by the embodiment is the average precision (AP), and the definition of the accuracy parameter can be written in the following form:

[0118]

[0119]

[0120] The definition of AP can be written in the following form:

[0121]

[0122] In the formula, TP represents true positive, FP represents false positive, p represents precision, r represents recall, and p is a function of r, which is equal to the total area under the curve.

[0123] Figure 8 is a convergence curve contrastive diagram of time-frequency analysis algorithms provided by the embodiment of the present application. As shown in the figure, Figure 8As shown in the figure, the data set of the STFT algorithm converges after 60 iterations, and the data set of the SST algorithm converges after 20 iterations. The comparison results show that the data set of the SST algorithm converges faster than the data set of the STFT algorithm. The trained network includes Resnet-18, Resnet-18+CBAM, MSRA network and CSFA network architecture, and the accuracy curves of the four identification models are as shown in the figure. Figure 9 Compared with other networks, the network model of the application has higher identification accuracy, which proves the effectiveness of the CSFA network, and the training results are as shown in Table 1.

[0124] Table 1

[0125]

[0126]

[0127] As shown in Table 1, the arrangement order of the identification performance of different targets of the four methods is: Resnet-18, Resnet-18+CBAM, MSRA network and the method for target identification based on the through-wall radar provided by the application. Among them, the MSRA network is based on the Resnet-18 network and adds a multi-scale attention mechanism, so the performance of identifying targets is improved. The experimental results show that the CSFA network improves the accuracy by 2.57% compared with CBAM, and compared with the MSRA network, the classification accuracy of the application for human, dog, cat and pig is improved by 1.23%, 0.37%, 2.02% and 1.7% respectively. Therefore, it can be concluded that the CSFA network can effectively extract the micro-Doppler feature from the time-frequency graph, and improve the accuracy of identifying different target time-frequency graphs.

[0128] From the above embodiments, the application has the following beneficial effects:

[0129] The application provides a radar target identification method based on a cross-scale feature aggregation network. The radar echo signal of the target is processed by a synchronous squeezing S transform time-frequency analysis algorithm, which can suppress multipath interference and other clutter, and further generate a high-resolution time-frequency graph with breathing characteristics. At the same time, the application extracts the micro-Doppler feature map in the time-frequency graph through the cross-scale feature aggregation network, which improves the identification accuracy of the target. In the description of the application, the terms "first" and "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features limited by "first" and "second" can explicitly or implicitly include one or more features. In the description of the application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0130] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate different embodiments or examples described in the specification.

[0131] Although the present application is described herein in conjunction with various embodiments, it is understood that other variations of the disclosed embodiments can be understood and realized by those skilled in the art upon viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and the indefinite article "a" or "an" does not exclude a plurality. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0132] The above is a further detailed description of the present application in conjunction with specific preferred embodiments, and cannot be regarded as a limitation of the specific implementation of the present application. For those skilled in the art, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be regarded as falling within the scope of protection of the present application.

Claims

1. A radar target recognition method based on a cross-scale feature aggregation network, characterized in that, The method comprises the steps of: Collecting echo of skin vibration generated by respiration of a static target by radar to obtain radar echo signal; Processing the radar echo signal by using a synchronous squeezing S transform time-frequency analysis algorithm to generate a time-frequency image containing respiration characteristics; Dividing and aggregating the time-frequency image in different scale sizes by using a feature pyramid network, and extracting a micro-Doppler feature map from feature maps of each scale output by the feature pyramid network by using a cross-scale feature aggregation CSFA network; After generating an aggregated micro-Doppler feature map based on the micro-Doppler feature map, inputting the aggregated micro-Doppler feature map into a SoftMax classifier to obtain a recognition result of the radar target; The cross-scale feature aggregation CSFA network comprises a channel attention model and a spatial attention model; The step of dividing and aggregating the time-frequency image in different scale sizes by using a feature pyramid network, and extracting a micro-Doppler feature map from feature maps of each scale output by the feature pyramid network by using a CSFA network, comprises the steps of: Dividing the time-frequency image in different scale sizes by using a feature pyramid network; input feature map of each scale obtained by the division input the channel attention model to obtain a one-dimensional channel attention feature map then input the one-dimensional channel attention feature map and the input feature map to obtain channel features input the channel features to the spatial attention model to obtain a two-dimensional spatial attention feature map ; ; obtaining channel features multiplying the two-dimensional spatial attention feature map obtaining spatial features ; According to the input feature map And the spatial features Multiply to determine the feedback variable And input the feedback variable Into the channel attention model to obtain a one-dimensional channel attention feature map Then, the one-dimensional channel attention feature map And the feedback variable Multiply to obtain the channel features ; channel features are input into the spatial attention model to obtain a two-dimensional spatial attention feature map ; Channel features are obtained by multiplying the two-dimensional spatial attention feature map with the current input feature map of the layer scale to obtain a micro-Doppler feature map .

2. The radar target recognition method based on the cross-scale feature aggregation network according to claim 1, characterized in that, Before the step of generating an aggregated micro-Doppler feature map based on the micro-Doppler feature map, the method further comprises the steps of: Obtaining a pre-trained Resnet-18 model, and loading a time-frequency image data set model file; The step of inputting the aggregated micro-Doppler feature map into a SoftMax classifier to obtain a recognition result of the radar target after generating an aggregated micro-Doppler feature map based on the micro-Doppler feature map, comprises the steps of: After generating an aggregated micro-Doppler feature map based on the micro-Doppler feature map, inputting the aggregated micro-Doppler feature map into a SoftMax classifier to enable the SoftMax classifier to compare the aggregated micro-Doppler feature map by using the time-frequency image data set model file, and obtaining a recognition result of the radar target.

3. The radar target recognition method based on the cross-scale feature aggregation network according to claim 2, characterized in that, Before the step of processing the radar echo signal by using a synchronous squeezing S transform time-frequency analysis algorithm to generate a time-frequency image containing respiration characteristics, the method further comprises the steps of: Pretreating the radar echo signal to obtain radar echo signal containing respiration characteristics; Generating a two-dimensional range image by time accumulation based on the radar echo signal containing respiration characteristics.

4. The radar target recognition method based on the cross-scale feature aggregation network according to claim 3, characterized in that, The step of processing the radar echo signal by using a synchronous squeezing S transform time-frequency analysis algorithm to generate a time-frequency image containing respiration characteristics, comprises the steps of: Processing effective channel data in the two-dimensional range image by using a synchronous squeezing S transform time-frequency analysis algorithm, and aggregating all the effective channel data into a time-frequency image.

5. The radar target recognition method based on the cross-scale feature aggregation network according to claim 4, characterized in that, The radar echo signal is processed by using a synchronous squeezing S transform time-frequency analysis algorithm according to the following formula: wherein , and denote the discrete frequency of the S-transform, the center frequency of the squeezing interval and the bandwidth of the squeezing interval, respectively, denotes the time axis shift parameter, and , , denotes the S-transform of the radar echo signal, denotes the instantaneous frequency of the radar echo signal.

6. The radar target recognition method based on the cross-scale feature aggregation network according to claim 1, characterized in that, The channel attention model comprises a first maximum pooling layer, a first average pooling layer and a multi-layer fully connected neural network; each scale feature map of the feature pyramid network is input into a channel attention model to obtain channel features inputting the channel attention model to obtain channel features the step of The input feature map is transmitted to the first max-pooling layer and the first average-pooling layer respectively to obtain a first max-pooling feature map and a first average-pooling feature map ;​​ The first maximum pooling feature map and the first average pooling feature map is input to a multi-layer fully connected neural network with an activation function of ReLU in a forward input manner, and two output features of the multi-layer fully connected neural network are added to obtain a one-dimensional channel attention feature map The one-dimensional channel attention feature map is multiplied by the current input feature map to obtain a channel feature .

7. The radar target recognition method based on the cross-scale feature aggregation network according to claim 6, characterized in that, The spatial attention model comprises a second max-pooling layer, a second average-pooling layer and a convolution kernel of channel features are input into the spatial attention model to obtain a two-dimensional spatial attention feature map comprises: The channel features are respectively input to a second max-pooling layer and a second average-pooling layer, the obtained two-dimensional max-pooling feature map and the two-dimensional average-pooling feature map are merged into a matrix, and after dimension reduction by a convolution kernel , a two-dimensional spatial attention feature map is generated.

8. The radar target recognition method based on the cross-scale feature aggregation network according to claim 7, characterized in that, The feedback variable , denotes a point multiplication operation.

Images