A Radar Multi-Angle and Multi-Feature Fusion Target Classification Method

Abstract

The present application discloses a radar multi-angle and multi-feature fusion target classification method, which relates to the technical field of radar signal processing. The method includes obtaining a multi-angle and multi-feature fusion network target classification model; the multi-angle and multi-feature fusion network target classification model includes: a temporal feature extraction module, an attention fusion module, and a feature fusion module; obtaining radar echo data of a target in real time; processing the radar echo data of the target to obtain processed radar echo data; the processed radar echo data includes: a time-domain signal component diagram of the target and a time-frequency diagram of the target; inputting the processed radar echo data into the multi-angle and multi-feature fusion network target classification model to obtain a target classification result. The present application can improve the accuracy of target classification.

Description

Technical Field

[0001] The present application relates to the technical field of radar signal processing, and particularly to a radar multi-angle and multi-feature fusion target classification method. Background Art

[0002] In recent years, due to advantages such as low price and strong practicability, the usage rate of civilian small unmanned aerial vehicles has increased rapidly. However, the widespread increase in drones has brought many potential threats. Many airports have successively experienced drone intrusion incidents, seriously threatening public safety. Since the threats posed by different types of drones and birds are different, it is imperative to identify drones and birds. As radar systems are not easily affected by scene changes and weather changes, radar has become an effective means of target surveillance, but there is still a lack of an effective method for identifying drones and birds through radar.

[0003] Micro-Doppler is a Doppler effect that reflects the micro-motion characteristics of a target and is a tiny additional modulation in the spectral changes caused by the target's motion. It can capture the weak frequency shifts caused by the non-rigid motion of the target (such as vibration, swing, or rotation) and is an important feature in modern radar signal processing and target recognition. Under the assumption of the ideal scattering center model, the time-frequency distribution is introduced to extract the micro-motion characteristics of the rotary-wing UAV. Since the radar return signal consists of complex components, including the body Doppler signal caused by translation, vibration interference, and environmental clutter effects, the time-frequency representation of the micro-Doppler signal is always mixed with other signals and cannot be directly extracted. Empirical Mode Decomposition has advantages in solving the mode mixing problem. Empirical Mode Decomposition (EMD) is a completely data-driven method that can be used to analyze non-linear and non-stationary signals and can adaptively decompose any signal into a set of complete and finite local amplitude frequencies. However, the EMD method is prone to mode aliasing during the decomposition of complex signals and it is difficult to distinguish different types of UAVs. The Short-Time Fourier Transform can extract the micro-motion characteristics of rotary-wing UAVs and birds. Due to the aliasing phenomenon of the micro-Doppler of multi-rotor UAVs with the increase in the number of rotors, it is difficult to describe and classify them through mathematical models and parameters in complex motions and environments. In recent years, with the development of deep learning and machine learning technologies, feature fusion has become an important trend in the field of micro-motion feature extraction of UAVs. The combination of radar target classification and recognition tasks with deep learning has greatly improved the target classification and recognition ability. Deep learning methods extract high-dimensional features, avoiding the limitations of manual feature extraction. Convolutional Neural Networks (CNN), as a classic method in deep learning, has been widely used in the fields of image detection and classification. For example, in current technologies, the combination of Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD) improves the classification accuracy of the target by splicing the decomposed features and sending them into the deep learning network for classification. There is also the combination of EMD and Long Short-Term Memory (LSTM) to extract the temporal sequence features of the time-domain signal, which improves the classification accuracy of the target. In addition, there is also the use of FMCW radar to collect UAV and bird data, merge the ranging spectrogram and micro-Doppler features, and use multi-channel DCNN for target classification, which improves the accuracy of UAV and bird classification. Thus, it can be seen that radar and deep learning are of great significance for the classification task of low, slow, and small targets.In addition, by fusing the micro-motion features with traditional radar features (such as distance, Doppler velocity, etc.), constructing a multi-modal dataset, and using a deep learning model to jointly learn the features, the accuracy of UAV recognition can still be significantly improved. The multi-sensor fusion technology is also widely used in the extraction of UAV micro-motion features. By fusing multi-feature radar data, although theoretically it can effectively reduce the defects of a single feature, thereby enhancing the stability and accuracy of the micro-motion features, the UAV micro-motion feature extraction and fusion technology has only achieved certain results in the experimental environment. In practical applications, the classification task of low, slow, and small targets still faces some challenges and cannot truly reduce the defects of a single feature and enhance the stability and accuracy of the micro-motion features. Summary of the Invention

[0004] The object of this application is to provide a radar multi-angle and multi-feature fusion target classification method, which can enhance the stability and accuracy of the micro-motion features, and further improve the accuracy of target classification.

[0005] To achieve the above object, this application provides the following solutions:

[0006] In the first aspect, this application provides a radar multi-angle and multi-feature fusion target classification method, including:

[0007] Obtain a multi-angle and multi-feature fusion network target classification model; the multi-angle and multi-feature fusion network target classification model includes: a time-series feature extraction module, an attention fusion module, and a feature fusion module;

[0008] Real-time obtain the radar echo data of the target;

[0009] Process the radar echo data of the target to obtain the processed radar echo data; the processed radar echo data includes: the time-domain signal component map of the target and the time-frequency map of the target;

[0010] Input the processed radar echo data into the multi-angle and multi-feature fusion network target classification model to obtain the target classification result.

[0011] Optionally, the construction process of the multi-angle and multi-feature fusion network target classification model includes:

[0012] Obtain the radar echo data of the sample target;

[0013] Process the radar echo data of the sample target to obtain the processed radar echo data of the sample target; the processed radar echo data of the sample target includes: the time-domain signal component map of the sample target and the time-frequency map of the sample target;

[0014] Construct a training dataset based on the processed radar echo data of the sample target;

[0015] Construct an initial multi - angle and multi - feature fusion network target classification model;

[0016] Train the initial multi - angle and multi - feature fusion network target classification model based on the training data set until the sample target classification result output by the initial multi - angle and multi - feature fusion network target classification model reaches the set result, and obtain the trained initial multi - angle and multi - feature fusion network target classification model; Use the trained initial multi - angle and multi - feature fusion network target classification model as the multi - angle and multi - feature fusion network target classification model.

[0017] Optionally, the process of obtaining the radar echo data of the sample target includes:

[0018] Use radars with different bands to collect multi - angle radar echo data for each target in the set area, and obtain the multi - angle radar echo data of all targets in the set area;

[0019] Obtain the radar echo data of the sample target based on the multi - angle radar echo data of all targets in the set area.

[0020] Optionally, processing the radar echo data of the sample target to obtain the processed radar echo data of the sample target includes:

[0021] Use moving target indication to perform negative frequency removal and clutter filtering on the radar echo data of the sample target to obtain the pre - processed radar echo data of the sample target;

[0022] Obtain an extraction signal based on the pre - processed radar echo data of the sample target;

[0023] Perform empirical mode decomposition on the extraction signal to obtain the time - domain signal component diagram of the sample target;

[0024] Perform short - time Fourier transform on the extraction signal to obtain the time - frequency diagram of the sample target.

[0025] Optionally, constructing a training data set based on the processed radar echo data of the sample target includes:

[0026] Construct a first data set based on the time - domain signal component diagram of the sample target;

[0027] Construct a second data set based on the time - frequency diagram of the sample target;

[0028] Construct a training data set based on the first data set and the second data set.

[0029] Optionally, train the initial multi-angle multi-feature fusion network target classification model based on the training data set until the sample target classification result output by the initial multi-angle multi-feature fusion network target classification model reaches a set result, and obtain the trained initial multi-angle multi-feature fusion network target classification model, including:

[0030] Use the initial multi-angle multi-feature fusion network target classification model to obtain a sample target classification result based on the training data set;

[0031] Use the AdamW algorithm to train the initial multi-angle multi-feature fusion network target classification model based on the sample target classification result until the sample target classification result reaches the set result, and obtain the trained initial multi-angle multi-feature fusion network target classification model;

[0032] Use the trained initial multi-angle multi-feature fusion network target classification model as the multi-angle multi-feature fusion network target classification model.

[0033] Optionally, the attention fusion module includes: a CBAM attention feature extraction unit and a visual channel attention feature extraction unit; during the training process, the process of obtaining a sample target classification result using the initial multi-angle multi-feature fusion network target classification model includes:

[0034] Use the temporal feature extraction module and the CBAM attention feature extraction unit to obtain a first feature map based on the first data set;

[0035] Use the CBAM attention feature extraction unit and the visual channel attention feature extraction unit to obtain a second feature map based on the second data set;

[0036] Use the feature fusion module to obtain the sample target classification result based on the first feature map and the second feature map.

[0037] Optionally, it is characterized in that the short-time Fourier transform is performed on the extracted signal using the formula to obtain the time-frequency diagram of the sample target;

[0038] In the formula, represents a sliding window, and a Gaussian window function is used; represents an integration variable, represents the time dimension, represents the frequency dimension, represents the result after the short-time Fourier transform; represents the extracted signal; is the kernel of the Fourier transform, representing the basis functions of different frequency components.

[0039] Optionally, it is characterized in that the empirical mode decomposition is performed on the extracted signal by using the formula to obtain the time-domain signal component diagram of the sample target;

[0040] In the formula, represents the input signal, represents the th intrinsic mode function, represents the number of intrinsic mode functions obtained by decomposition, represents the process of extracting the residual intrinsic mode function.

[0041] Optionally, it is characterized in that the attention fusion module includes: a CBAM attention feature extraction unit and a visual channel attention feature extraction unit; inputting the processed radar echo data into the multi-angle multi-feature fusion network target classification model to obtain a target classification result, including:

[0042] Using the time series feature extraction module and the CBAM attention feature extraction unit to obtain a third feature map based on the processed radar echo data;

[0043] Using the CBAM attention feature extraction unit and the visual channel attention feature extraction unit to obtain a fourth feature map based on the processed radar echo data;

[0044] Using the feature fusion module to obtain the target classification result based on the third feature map and the fourth feature map.

[0045] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0046] The present application provides a radar multi-angle multi-feature fusion target classification method. By using a multi-angle multi-feature fusion network target classification model to obtain a target classification result based on the processed radar echo data, it can overcome the differences in single neural network feature extraction, enhance the stability and accuracy of micro-motion features, and thus improve the accuracy of target classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0048] Figure 1 is a flowchart of a radar multi-angle multi-feature fusion target classification method in an embodiment of the present application;

[0049] Figure 2 Schematic diagram of the target classification process of a radar multi - angle and multi - feature fusion target classification method provided by an embodiment of the present application;

[0050] Figure 3 Schematic diagram of the multi - angle and multi - feature fusion network target classification model provided by an embodiment of the present application;

[0051] Figure 4 Schematic diagram of the structure of the time - series feature extraction module and the CBAM attention feature extraction unit provided by an embodiment of the present application;

[0052] Figure 5 Schematic diagram of the structure of the CBAM attention feature extraction unit and the visual channel attention feature extraction unit provided by an embodiment of the present application;

[0053] Figure 6 Schematic diagram of the structure of the feature fusion module provided by an embodiment of the present application;

[0054] Figure 7 Schematic diagram of the structure of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0055] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0056] To make the above - mentioned objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0057] In an exemplary embodiment, as Figure 1 shown, a radar multi - angle and multi - feature fusion target classification method is provided. This method is executed by a computer device, specifically, it can be executed alone by a computer device such as a terminal or a server, or jointly executed by a terminal and a server. In the embodiments of the present application, taking the target classification of this method as an example for illustration, it includes:

[0058] Step 100: Obtain a multi - angle and multi - feature fusion network target classification model. The multi - angle and multi - feature fusion network target classification model includes: a time - series feature extraction module, an attention fusion module, and a feature fusion module.

[0059] Step 200: Real - time obtain the radar echo data of the target.

[0060] Step 300: Process the radar echo data of the target to obtain the processed radar echo data. The processed radar echo data includes: the time-domain signal component diagram of the target and the time-frequency diagram of the target.

[0061] Step 400: Input the processed radar echo data into the multi-angle multi-feature fusion network target classification model to obtain the target classification result.

[0062] As an optional implementation, to improve the accuracy of the radar multi-angle multi-feature fusion target classification method, the construction process of the multi-angle multi-feature fusion network target classification model in Step 100 includes:

[0063] 101. Obtain the radar echo data of the sample target. The process of obtaining the radar echo data of the sample target includes: Using radars with different frequency bands to collect the radar echo data of each target in the set area from multiple angles, obtaining the multi-angle radar echo data of all targets in the set area. Based on the multi-angle radar echo data of all targets in the set area, obtain the radar echo data of the sample target.

[0064] 102. Process the radar echo data of the sample target to obtain the processed radar echo data of the sample target. The processed radar echo data of the sample target includes: the time-domain signal component diagram of the sample target and the time-frequency diagram of the sample target. Among them, the process of processing the radar echo data of the sample target to obtain the processed radar echo data of the sample target includes: Using moving target indication to perform negative frequency removal and clutter filtering on the radar echo data of the sample target to obtain the preprocessed radar echo data of the sample target. Based on the preprocessed radar echo data of the sample target, obtain the extracted signal. Perform empirical mode decomposition on the extracted signal to obtain the time-domain signal component diagram of the sample target. Perform short-time Fourier transform on the extracted signal to obtain the time-frequency diagram of the sample target.

[0065] For example, use the formula to perform short-time Fourier transform on the extracted signal to obtain the time-frequency diagram of the sample target. In the formula, represents the sliding window, using a Gaussian window function; represents the integration variable, represents the time dimension, represents the frequency dimension, represents the result after short-time Fourier transform; represents the extracted signal, is the kernel of the Fourier transform, representing the basis functions of different frequency components.

[0066] Use the formula to perform empirical mode decomposition on the extracted signal to obtain the time-domain signal component diagram of the sample target. In the formula, represents the input signal, represents the nth intrinsic mode function, represents the number of decomposed intrinsic mode functions, represents the process of extracting the residual intrinsic mode function.

[0067] 103. Construct a training data set based on the radar echo data of the processed sample target. For example, construct a first data set based on the time-domain signal component diagram of the sample target. Construct a second data set based on the time-frequency diagram of the sample target. Construct a training data set based on the first data set and the second data set.

[0068] 104. Construct an initial multi-angle and multi-feature fusion network target classification model.

[0069] 105. Train the initial multi-angle and multi-feature fusion network target classification model based on the training data set until the sample target classification result output by the initial multi-angle and multi-feature fusion network target classification model reaches the set result, and obtain the trained initial multi-angle and multi-feature fusion network target classification model. Use the trained initial multi-angle and multi-feature fusion network target classification model as the multi-angle and multi-feature fusion network target classification model.

[0070] Among them, the process of training the initial multi-angle and multi-feature fusion network target classification model includes: using the initial multi-angle and multi-feature fusion network target classification model to obtain the sample target classification result based on the training data set. Using the AdamW algorithm to train the initial multi-angle and multi-feature fusion network target classification model based on the sample target classification result until the sample target classification result reaches the set result, and obtain the trained initial multi-angle and multi-feature fusion network target classification model. Use the trained initial multi-angle and multi-feature fusion network target classification model as the multi-angle and multi-feature fusion network target classification model.

[0071] The attention fusion module includes: a CBAM attention feature extraction unit and a visual channel attention feature extraction unit. The process of obtaining the sample target classification result by using the initial multi-angle and multi-feature fusion network target classification model during the training process includes: using the time series feature extraction module and the CBAM attention feature extraction unit to obtain a first feature map based on the first data set. Using the CBAM attention feature extraction unit and the visual channel attention feature extraction unit to obtain a second feature map based on the second data set. Using the feature fusion module to obtain the sample target classification result based on the first feature map and the second feature map.

[0072] The process of inputting the processed radar echo data into the multi-angle and multi-feature fusion network target classification model in step 400 to obtain the target classification result includes: using the CBAM attention feature extraction unit and the visual channel attention feature extraction unit to obtain the third feature map based on the processed radar echo data. Using the CBAM attention feature extraction unit and the visual channel attention feature extraction unit to obtain the fourth feature map based on the processed radar echo data. Using the feature fusion module to obtain the target classification result based on the third feature map and the fourth feature map.

[0073] The output vector of the multi-angle and multi-feature fusion network target classification model is , where is the score of the model on the th class, and the probability of the th class is calculated using the Softmax function , which is expressed as: . Among them, is the score of the th class, and is the predicted probability of the th class.

[0074] In an exemplary embodiment, taking the process of training the initial multi-angle and multi-feature fusion network target classification model by combining steps 101 to 105 of the above embodiment to obtain the multi-angle and multi-feature fusion network target classification model as an example, as Figure 2 shown.

[0075] S1. Obtain the radar echo data of the sample target, including the K-band radar echo data and the L-band radar echo data. Process the radar echo data of the sample target to obtain the processed radar echo data of the sample target.

[0076] First, in order to eliminate the negative frequency signal in the radar echo data, the data is preliminarily processed. A high-pass filter is used to effectively remove the DC component in the echo signal (i.e., the radar echo data). This step aims to enhance the dynamic range of the signal and ensure the accurate extraction of high-frequency components. The moving target indication (MTI) technology is used to perform negative frequency removal and clutter filtering on the radar echo data, filtering out the stationary clutter on the ground and in the air, thereby improving the target detection ability. MTI can effectively suppress the stationary or slow-moving background noise by comparing the phase changes between adjacent pulses, making the target signal more prominent, including negative frequency demodulation, demodulation, and pulse compression processing. The preprocessed radar echo data of the sample target is obtained through the MTI technology.

[0077] After the above-mentioned preprocessing steps, the radar echo data of the preprocessed sample target is obtained. Based on the radar echo data of the preprocessed sample target, a range periodogram of the target is drawn, and the echo signal of the range cell where the target is located is extracted to obtain the extracted signal (including the K-band extracted signal and the L-band extracted signal). The L-band extracted signal is subjected to empirical mode decomposition (EMD) to obtain the time-domain signal component diagram of IMF1 (i.e., the time-domain signal component diagram of the sample target):

[0078] (1)

[0079] In the formula, represents the input signal, represents the th intrinsic mode function, represents the process of extracting IMF of the residual, represents the number of intrinsic mode functions decomposed.

[0080] In order to further extract the micro-motion characteristics of the target, the short-time Fourier transform (STFT) is performed on the K-band extracted signal to obtain the time-frequency diagram of the sample target. Assuming that the K-band echo signal of the target is expressed as , the fast Fourier transform (FFT) is performed on the echo signal within the time window by means of a sliding window to obtain the time-frequency diagram of the sample target:

[0081] (2)

[0082] Among them, represents the sliding window, and a Gaussian window function is taken in this embodiment; represents the integration variable, represents the time dimension, represents the frequency dimension, represents the result after short-time Fourier transform. By observing the range Doppler of the UAV, the range cell where the target is located is selected, and the signal of the range cell where the target is located is extracted for STFT to obtain the time-frequency diagrams of the target signals at multiple angles.

[0083] S2. Based on the processed radar echo data of the sample target, a training data set is constructed. The training data set includes a first data set and a second data set. After processing the echo data of the sample target through step S1, the time-domain signal component diagram of the sample target and the time-frequency diagram of the sample target are obtained. The first data set is constructed based on the time-domain signal component diagram of the sample target. The second data set is constructed based on the time-frequency diagram of the sample target.

[0084] S3. Construct an initial multi - angle and multi - feature fusion network target classification model. The constructed target classification model includes a temporal feature extraction module (ConvLSTM), an attention fusion module, and a feature fusion module (AFF). The attention fusion module includes a CBAM attention feature extraction unit and a visual channel attention feature extraction unit (SEAttention). As Figure 3 shown, the temporal feature extraction module and the CBAM attention feature extraction unit are used to obtain a first feature map based on the first data set. The CBAM attention feature extraction unit is used to obtain CBAM - extracted features based on the second data set; the SEAttention unit is used to obtain SEAttention - extracted features based on the second data set; the CBAM - extracted features and the SEAttention - extracted features are weighted and fused to obtain a second feature map.

[0085] 1) Temporal feature extraction module (ConvLSTM).

[0086] Different types of rotor unmanned aerial vehicles exhibit different dynamic characteristics in the time - domain signal. Therefore, obtaining and analyzing the temporal information of the signal is crucial for target recognition. To effectively extract the features of continuous time - domain signals, a temporal feature extraction method based on ConvLSTM is adopted. ConvLSTM is a deep model that combines a convolutional neural network (CNN) and a long short - term memory network (LSTM), which is specifically designed to process data with spatio - temporal dependencies. The LSTM network has significant advantages in processing sequence data, especially in capturing long - distance time dependencies. Its special gating mechanism enables it to effectively retain important time information. CNN is good at extracting local spatial features from data and obtaining high - dimensional spatial dependency information through convolution operations. The ConvLSTM model integrates convolution operations into the gating units of the LSTM model, enabling it to simultaneously learn the time - series features and spatial features of the data, thus adapting to the processing requirements of spatio - temporal data. The ConvLSTM model can not only effectively capture the dynamic changes of the rotor unmanned aerial vehicle signal in the time dimension but also extract the temporal features of the time - domain signal in the spatial dimension. In all embodiments provided in this application, the ConvLSTM model is used to process the continuous time - domain data of radar echo signals, extract the dependencies between signals at different times and the temporal features that change over time, enabling the target classification model to better distinguish different types of rotor unmanned aerial vehicles. As Figure 4 shown, the internal structure of the ConvLSTM model is presented, including the convolution operation processes of the input gate , forget gate , and output gate , which can strengthen the capture of temporal dependencies while retaining spatial features. The implementation of temporal feature extraction based on ConvLSTM is as follows:

[0087] (3)

[0088] (4)

[0089] (5)

[0090] (6)

[0091] (7)

[0092] (8)

[0093] (9)

[0094] Among them, represents the convolution operation, , , , are all convolution kernels, represents element-wise multiplication, represents the input at the current time step, Conv() represents the convolution operation on the input, and CBAM() represents the operation of extracting attention features, represents the hidden state at the previous time step, represents the input gate, represents the cell state at the previous time step, represents the convolution operation, represents the Sigmoid activation function, which restricts the output to be between [0, 1], represents the forget gate (controlling which information in the cell state is retained), represents the hyperbolic tangent function, , , , represent the bias terms of the forget gate, input gate, cell state, and output gate respectively, represents the candidate cell state, represents the cell state at the current cell step, represents the bias of the input gate, represents the output gate (controlling the output of the final hidden state ). The input of the ConvLSTM model is a five-dimensional tensor. By taking five consecutive frames as the input of the ConvLSTM model, the temporal features between each frame are extracted, and at the same time, the spatial features of the feature map are extracted during the convolution process.

[0095] 2) Attention fusion module.

[0096] Echo data of different rotor UAVs (i.e., radar echo data of sample targets) are collected by radar. After performing STFT on the obtained data, features are extracted through convolution operations. The obtained target echo data from multiple angles are subjected to STFT to obtain the micro-motion features of the rotor UAV. The micro-motion features obtained from two radars (K-band and L-band) are weighted and fused through convolution operations to obtain features of higher dimensions.

[0097] CBAM is a lightweight and efficient attention module that can adaptively assign different importance weights to the channel and spatial dimensions of feature maps, thereby improving network performance. As Figure 5 shown, CBAM takes the intermediate feature map (i.e., the time-frequency map in the second dataset) as input, and according to the channel attention module and the spatial attention module, sequentially derives the one-dimensional channel attention map and the two-dimensional spatial attention map , where represents the channel dimension, represents the height of the input feature, represents the width of the input feature. The entire attention feature extraction process can be summarized as:

[0098] (10)

[0099] (11)

[0100] Among them, represents element-wise multiplication, represents channel attention feature extraction, represents spatial attention feature extraction. represents the feature matrix after channel attention feature extraction, represents the feature matrix after spatial attention feature extraction. The channel attention module aims to assign different weights to each channel, use global pooling to gather global information, and then generate the importance weights of each channel through a multi-layer perceptron (MLP). The channel attention mechanism can be expressed as:

[0101] (12)

[0102] Among them, is the Sigmoid activation function, MLP is a two-layer perceptron with shared weights, AvgPool() represents the average pooling layer operation, and MaxPool() represents the max pooling layer operation. The spatial attention module aims to assign different weights to each spatial position to capture local regions that are more important for the target. The spatial attention mechanism can be expressed as:

[0103] (13)

[0104] Among them, represents the weight matrix generated after inputting features, represents concatenation by channel, represents the convolution operation.

[0105] Use the SEAttention unit to extract features from the multi-angle time-frequency map obtained by the radar, as Figure 5 shown. Assume that the extracted features are represented as , and its feature extraction can be expressed as:

[0106] (14)

[0107] Among them, represents a standard convolution operator, represents the extracted feature map:

[0108] (15)

[0109] Among them, is a c-dimensional vector representing the global feature of each channel; represents the height of the input feature map, represents the width of the input feature map, represents the longitudinal index in the feature map, represents the lateral index in the feature map, represents the channel index of the feature map; represents global pooling. The global feature of each channel is extracted through the operation of the global average pooling layer, and each channel of the feature map is compressed into a scalar:

[0110] (16)

[0111] (17)

[0112] (18)

[0113] Among them, represents the finally output feature, represents the global description of each channel, represents the output result, represents the ReLU function, represents the MLP to calculate the channel weights, and are learnable weight matrices, represents applying the channel weight to the original feature to complete channel recalibration, Indicates the final output feature, Indicates the index quantity, Indicates feature mapping.

[0114] In order to better extract the features of the time-frequency spectrograms obtained from two bands, two attention modules are combined. After extracting the features in the two spectrograms, they are weighted and fused to obtain two feature maps containing two features (i.e., the second feature map).

[0115] 3) Feature Fusion Module (AFF).

[0116] The feature fusion module can adapt to different levels of local and global features according to the input features. AFF is a network framework for multi-scale and multi-modal data fusion, mainly targeting the attention problem of different scale feature fusion in different network structures. The AFF feature fusion network realizes flexible and efficient multi-scale and multi-modal feature fusion through an adaptive weighting mechanism. Through the lightweight design of global average pooling and weight learning, it ensures the improvement of the accuracy and robustness of feature fusion while performing efficient calculations. Therefore, this module is used to obtain the features Figure 1 (i.e., the first feature map) obtained by the ConvLSTM model and the features Figure 2 (i.e., the second feature map) obtained from the time-frequency diagram of the target are fused to further extract the target signal features and improve the accuracy of target classification. The MS-CAM (Multi-Scale Channel Attention Module) used in the AFF module follows the idea of ParseNet (semantic segmentation network) in combining local and global features in CNN, as well as the spatial attention idea of aggregating multi-scale features within the attention module.

[0117] (19)

[0118] Among them, Is the fused feature, Represents the initial feature integration, X and Y represent the two input features. The overall framework of the AFF fusion module is as Figure 6 Shown.

[0119] MS-CAM combines multi-scale features and channel attention mechanisms to capture the details and global information of objects by processing different scale features of images. The channel attention mechanism in it calculates the weights of different channels, enabling the model to automatically focus on which features are more important. The calculation formula of channel attention can be expressed as :

[0120] (20)

[0121] Among them, B() represents the execution of the BatchNorm operation, Represents the ReLU activation function, represent Point convolution, and the calculated weight values are used for the input features to obtain the output after performing the attention operation , and the local channel information has the same size as the input features.

[0122] (21)

[0123] S4. The initial multi-angle multi-feature fusion network target classification model is adopted to obtain the sample target classification result based on the training dataset. The AdamW algorithm is used to train the initial multi-angle multi-feature fusion network target classification model based on the sample target classification result until the sample target classification result reaches the set result, and the trained initial multi-angle multi-feature fusion network target classification model is obtained. The trained initial multi-angle multi-feature fusion network target classification model is used as the multi-angle multi-feature fusion network target classification model.

[0124] The process of obtaining the sample target classification result by using the initial multi-angle multi-feature fusion network target classification model during the training process includes: using the temporal feature extraction module and the CBAM attention feature extraction unit to obtain the first feature map based on the first dataset. Using the CBAM attention feature extraction unit and the visual channel attention feature extraction unit to obtain the second feature map based on the second dataset. Using the feature fusion module to obtain the sample target classification result based on the first feature map and the second feature map.

[0125] When performing iterative optimization training, the AdamW gradient descent method is used, and the initial learning rate parameter (i.e., the scaling factor), the momentum factor , , the optimization parameter , the first instantaneous vector (i.e., the first-order momentum at the step), the second instantaneous vector (i.e., the second-order momentum at the step), the progress multiplier (i.e., the learning rate at the step).

[0126] (22)

[0127] (23)

[0128] (24)

[0129] (25)

[0130] (26)

[0131] (27)

[0132] Among them, represents the number of the current iteration, represents the learning rate, represents the first-order moment estimation of the gradient, represents the gradient at the current time step, represents the exponential decay rate of the first-order momentum, represents the second-order momentum at the step, represents the hyperparameter of the regularization strength, controlling the weight of the regularization term ; represents the objective function at the parameter at the previous moment, represents the model parameter at the step, represents the second-order momentum at the step, represents the corrected first-order momentum estimation at the step, represents the corrected second-order momentum estimation at the step; represents to the power. As increases, its value will approach 0; represents to the power.

[0133] When the sample target classification result obtained by the initial multi-angle and multi-feature fusion network target classification model reaches the set result, the trained initial multi-angle and multi-feature fusion network target classification model is obtained. For example, when the sample target classification result is close to the type of the set sample target, the training of the initial multi-angle and multi-feature fusion network target classification model is completed, and its network parameters are optimized, and the trained initial multi-angle and multi-feature fusion network target classification model is obtained.

[0134] According to all the above embodiments and combining the solution provided by the present application, the present application has the following advantages:

[0135] (1) Break through the limitations of traditional classification methods and be able to realize intelligent classification of radar targets in complex environments.

[0136] (2) It overcomes the deficiency of single neural network feature extraction from single radar data, and simultaneously obtains the time-frequency information and time-domain signal information in the echo signals collected from different bands of radar at multiple angles for feature extraction and fusion processing, thereby improving the target classification performance.

[0137] (3) By using the attention mechanism to enhance the feature extraction ability of ConvLSTM, the information of time-frequency features is extracted and fused through the attention mechanism. Finally, through the AFF fusion module, the spatio-temporal features and time-frequency features of time-domain information are fused, thus improving the accuracy of target classification and meeting the actual application requirements.

[0138] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 7 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data related to the radar multi-angle multi-feature fusion target classification method. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a radar multi-angle multi-feature fusion target classification method.

[0139] Those skilled in the art can understand that Figure 7 the structure shown in

[0140] is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0141] In an exemplary embodiment, a computer program product is provided, including a computer program which, when executed by a processor, implements the steps in the above method embodiments.

[0142] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0143] Those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0144] The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in this application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0145] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0146] In this article, specific examples are used to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation on the present application.

Claims

1. A method for classifying radar multi-angle and multi-feature fusion targets, characterized in that The radar multi-angle and multi-feature fusion target classification method includes: Obtain a multi-angle and multi-feature fusion network target classification model; the multi-angle and multi-feature fusion network target classification model includes: a temporal feature extraction module ConvLSTM, an attention fusion module, and a feature fusion module; the attention fusion module includes: a CBAM attention feature extraction unit and a visual channel attention feature extraction unit SEAttention; Obtain the multi-angle radar echo data of the target in real time; the multi-angle radar echo data of the target includes: K-band radar echo data and L-band radar echo data; Process the multi-angle radar echo data of the target to obtain the processed radar echo data, including: using moving target indication to perform negative frequency removal and clutter filtering on the multi-angle radar echo data of the target to obtain the preprocessed radar echo data of the target; obtaining an extraction signal based on the preprocessed radar echo data of the target; the extraction signal includes: a K-band extraction signal and an L-band extraction signal; performing empirical mode decomposition on the L-band extraction signal to obtain the time-domain signal component diagram of the target; performing short-time Fourier transform on the K-band extraction signal to obtain the time-frequency diagram of the target; the processed radar echo data includes: the time-domain signal component diagram of the target and the time-frequency diagram of the target; Input the processed radar echo data into the multi-angle and multi-feature fusion network target classification model to obtain a target classification result, including: Using the temporal feature extraction module and the CBAM attention feature extraction unit, obtain a third feature map based on the time-domain signal component diagram of the target; Using the CBAM attention feature extraction unit and the visual channel attention feature extraction unit, obtain a fourth feature map based on the time-frequency diagram of the target; Using the feature fusion module, obtain the target classification result based on the third feature map and the fourth feature map.

2. The radar multi-angle and multi-feature fusion target classification method according to claim 1, wherein The construction process of the multi-angle and multi-feature fusion network target classification model includes: Obtain the radar echo data of the sample target; Process the radar echo data of the sample target to obtain the processed radar echo data of the sample target; the processed radar echo data of the sample target includes: the time-domain signal component diagram of the sample target and the time-frequency diagram of the sample target; Construct a training data set based on the processed radar echo data of the sample target; Construct an initial multi-angle and multi-feature fusion network target classification model; Train the initial multi-angle and multi-feature fusion network target classification model based on the training data set until the sample target classification result output by the initial multi-angle and multi-feature fusion network target classification model reaches the set result, and obtain the trained initial multi-angle and multi-feature fusion network target classification model; use the trained initial multi-angle and multi-feature fusion network target classification model as the multi-angle and multi-feature fusion network target classification model.

3. The radar multi-angle and multi-feature fusion target classification method according to claim 2, wherein The acquisition process of the radar echo data of the sample target includes: Collect radar echo data of each target in a set area from multiple angles using radars of different bands, and obtain the multi-angle radar echo data of all targets in the set area; Obtain the radar echo data of the sample target based on the multi-angle radar echo data of all targets in the set area.

4. The radar multi-angle and multi-feature fusion target classification method according to claim 2, wherein, Process the radar echo data of the sample target to obtain the processed radar echo data of the sample target, including: Use moving target indication to perform negative frequency removal and clutter filtering on the radar echo data of the sample target to obtain the preprocessed radar echo data of the sample target; Obtain an extraction signal based on the preprocessed radar echo data of the sample target; Perform empirical mode decomposition on the extraction signal to obtain the time-domain signal component diagram of the sample target; Perform short-time Fourier transform on the extraction signal to obtain the time-frequency diagram of the sample target.

5. The radar multi-angle and multi-feature fusion target classification method according to claim 4, characterized in that Construct a training dataset based on the processed radar echo data of the sample target, including: Construct a first dataset based on the time-domain signal component diagram of the sample target; Construct a second dataset based on the time-frequency diagram of the sample target; Construct a training dataset based on the first dataset and the second dataset.

6. The radar multi-angle and multi-feature fusion target classification method according to claim 2, wherein Train the initial multi-angle multi-feature fusion network target classification model based on the training dataset until the sample target classification result output by the initial multi-angle multi-feature fusion network target classification model reaches the set result, and obtain the trained initial multi-angle multi-feature fusion network target classification model, including: Use the initial multi-angle multi-feature fusion network target classification model to obtain a sample target classification result based on the training dataset; Use the AdamW algorithm to train the initial multi-angle multi-feature fusion network target classification model based on the sample target classification result until the sample target classification result reaches the set result, and obtain the trained initial multi-angle multi-feature fusion network target classification model; Use the trained initial multi-angle multi-feature fusion network target classification model as the multi-angle multi-feature fusion network target classification model.

7. The radar multi-angle and multi-feature fusion target classification method according to claim 5, characterized in that The attention fusion module includes: a CBAM attention feature extraction unit and a visual channel attention feature extraction unit; during the training process, the process of obtaining a sample target classification result using the initial multi-angle multi-feature fusion network target classification model includes: Use the time series feature extraction module and the CBAM attention feature extraction unit to obtain a first feature map based on the first dataset; Use the CBAM attention feature extraction unit and the visual channel attention feature extraction unit to obtain a second feature map based on the second dataset; Use the feature fusion module to obtain the sample target classification result based on the first feature map and the second feature map.

8. The radar multi-angle and multi-feature fusion target classification method according to claim 4, wherein Using the formula Perform short-time Fourier transform on the extracted signal to obtain the time-frequency diagram of the sample target; In the formula, represents a sliding window, and a Gaussian window function is adopted; represents an integration variable, represents the time dimension, represents the frequency dimension, represents the result after short-time Fourier transform; represents the extracted signal; is the kernel of Fourier transform, representing the basis functions of different frequency components.

9. The radar multi-angle and multi-feature fusion target classification method according to claim 4, wherein Using the formula perform empirical mode decomposition on the extracted signal to obtain the time-domain signal component diagram of the sample target; In the formula, represents the input signal, represents the th eigenmode function, represents the number of decomposed eigenmode functions, represents the process of residual extraction of eigenmode functions.

Images