Deep learning-based multi-wave position pulse radar target detection method, device and medium

Abstract

The application discloses a multi-wave-position pulse radar target detection method and device based on deep learning, and a medium, and the method comprises the following steps: acquiring radar original echo data under different detection scenes to form a radar original echo data set; training a target detection network by using the radar original echo data set to obtain a target detection model; in an input preprocessing layer of the target detection network, the real part and the imaginary part of the radar original echo data are separated, the real part and the imaginary part of different wave-position data are respectively subjected to feature mapping by using linear layer networks with different weight matrices, the real part and the imaginary part features are calculated, the phase spectrum features are further calculated, and a multi-dimensional feature map is formed; the target detection network is trained by using the echo data set; and echo data to be detected is acquired and input into the target detection model for detection. The application can realize high-precision and low-computing-cost target detection in a complex scene, and significantly improves the detection performance and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-wavelength pulse radar, and in particular to a multi-wavelength pulse radar target detection method, device and medium based on deep learning. Background Art

[0002] Multi-wavelength pulse radar is based on pulse radar and uses multiple different wavelengths for detection. Each wavelength has corresponding frequency, phase and other parameters. By jointly processing the multi-wavelength echo signals, the detection probability, positioning accuracy and recognition ability of radar target detection can be improved.

[0003] For multi-wavelength pulse radar target detection, existing technologies typically use methods based on statistical signal processing. These methods extract the radar's raw multi-wavelength echo data by wavelength, analyze and process it separately, and then further distinguish the target from background noise or clutter on the resulting range Doppler maps for each wavelength. However, these methods rely on manual feature extraction and threshold setting strategies, and their robustness and anti-interference capabilities are poor in complex clutter environments or low signal-to-noise ratio conditions. In practice, it is difficult to effectively separate target signals from background noise. In particular, in multipath reflection or dense interference scenarios, target echoes are easily drowned out, resulting in low detection accuracy and high false alarm rates. Furthermore, their performance is highly dependent on prior knowledge of the statistical characteristics of background clutter. When multiple targets are close together, they are prone to target obscuration and mutual interference, resulting in a significant decrease in detection performance.

[0004] Deep learning automatically extracts features from input data through deep neural networks and uses these features to classify targets and backgrounds. It offers advantages such as automatic feature extraction, high classification accuracy, strong robustness, and good generalization. Applying deep learning methods to multi-wavelength pulse radar target detection can automatically learn deep features from radar echo signals, effectively distinguishing different types of targets, and improving the adaptability and accuracy of radar detection in complex environments.

[0005] However, in the prior art, in order to adapt to the image-based convolutional neural network model, the complex signal processing needs to be converted into a radar image first, and complex signal processing operations such as Fourier transform and time-frequency analysis are performed on the radar echo data of each wave position to extract distance, Doppler and other information as the input of the neural network. This process not only has large computational overhead and is complex to implement, but also may cause loss or distortion of the original characteristics of the target due to the fixed signal transformation method, thereby affecting the detection performance. In addition, the signal processing process for each wave position and the neural network training process are mutually separated, which causes the model to be unable to adaptively adjust the feature extraction strategy according to the data characteristics, so that the entire process is non-end-to-end, which seriously restricts the end-to-end learning ability, affects the real-time performance, feature extraction integrity and accuracy of the network, and the global optimal feature extraction of the network, thereby reducing the precision and robustness of the target detection network. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a multi-wave position pulse radar target detection method, device and medium based on deep learning, which can realize high-precision and low-computational-overhead target detection in complex scenes, while significantly improving the detection performance and robustness.

[0007] To solve the above technical problems, the technical solution provided by the present application is:

[0008] A multi-wave position pulse radar target detection method based on deep learning, comprising the following steps:

[0009] Obtaining radar raw echo data of multi-wave position pulses in different detection scenes to form a radar raw echo data set;

[0010] Training a target detection network based on a neural network using the radar raw echo data set, and obtaining a target detection model after training; the target detection network comprises an input preprocessing layer, an encoding layer, a decoding layer and an output layer connected in sequence, the input preprocessing layer is used for multi-wave position and multi-dimensional feature extraction of the input radar raw echo data, and the extracted features are output after the encoding layer, the decoding layer and the output layer; in the input preprocessing layer, the real part and the imaginary part of the radar raw echo data are separated, two linear layer networks with different weight matrices are used to map the real part and the imaginary part of the different wave position data respectively, the real part feature and the imaginary part feature are calculated according to the mapping results, and the phase spectrum feature is calculated according to the real part feature and the imaginary part feature, and the real part feature, the imaginary part feature and the phase spectrum feature are used to construct a multi-dimensional feature map output;

[0011] The radar raw echo data of the multi-wave position pulse to be detected is input into the target detection model obtained after training for detection to obtain a target detection result output.

[0012] Further, in the input preprocessing layer, the real part and the imaginary part of the different wave position data are respectively mapped by two linear layer networks with different weight matrices.

[0013] Different wave position data are extracted according to the number of pulses of different wave position data in the echo data.

[0014] The different wave position data are respectively input into two independent linear layer networks for feature mapping. The weight matrix of each linear layer network is composed of a plurality of fixed-dimension sub-matrix blocks to correspond to process radar echo data under different wave positions, and the weights of the real part weight block and the imaginary part weight block are configured according to the size and coordinates of each sub-matrix block.

[0015] Further, in the input preprocessing layer, the real part and the imaginary part of the different wave position data are respectively mapped by two linear layer networks with different weight matrices. cos -2πij / k ), and the weight of the imaginary part weight block is sin -2πij / k , wherein k is the block size, i , j are the horizontal and vertical coordinate values of the block, respectively.

[0016] Further, in the input preprocessing layer, the real part feature real_feature and the imaginary part feature imag_feature of the Fourier transform result are obtained by complex operation according to the real part weight matrix and the imaginary part weight matrix. real_feature imag_feature The phase spectrum feature is obtained by an inverse tangent function according to the extracted real part feature and imaginary part feature, and the extracted real part feature, imaginary part feature and phase spectrum feature are spliced to form the multi-dimensional feature map.

[0017] Further, the encoding layer in the target detection network includes a plurality of encoders for feature extraction and spatial dimension compression and channel dimension expansion, the decoding layer includes a plurality of decoders for feature fusion and gradual recovery of spatial dimensions, and the output layer is used for fusing feature maps of different scales and outputting the final target detection result. Each layer of the encoder and the decoder is connected in a step-by-step stacking manner, each layer of the encoder is connected to a layer of the decoder, and the intermediate layer is further used to extract semantic features between the encoding layer and the decoding layer.

[0018] ​​​Further, each layer of the encoder comprises a max-pooling layer and parallel multi-branch convolution units, at least one branch convolution unit in each branch convolution unit is used to extract local detail features, at least one branch convolution unit is used to realize cross-channel information interaction and dimension transformation, and at least one branch convolution unit is used to retain the identity mapping of the original input feature map.

[0019] Further, each layer of the decoder realizes up-sampling through a bilinear interpolation unit, and the feature map obtained after up-sampling is spliced with the feature map output by the previous layer of the decoder through a cross-layer feature splicing unit to realize multi-scale feature splicing, and the spliced multi-scale features are subjected to feature transformation and channel dimension compression using parallel multi-branch convolution.

[0020] Further, the output layer of the target detection network comprises a convolution prediction head, a multi-scale up-sampling unit, a channel attention mechanism and a convolution fusion unit, the convolution prediction head is used to predict the prediction results of a plurality of different scale feature layers output by the decoding layer and adjust the channel number, the multi-scale up-sampling unit is used to up-sample the prediction results of each different scale feature layer, the channel attention mechanism forms a multi-scale feature map by weighting the feature map after multi-scale up-sampling based on the attention mechanism, and the convolution fusion unit is used to fuse the weighted multi-scale feature map and generate a final target detection result.

[0021] An electronic device comprising a processor and a memory, the memory being configured to store a computer program, and the processor being configured to execute the computer program to perform the method as described above.

[0022] A computer-readable storage medium storing a computer program, the computer program being executed by a processor to implement the method as described above.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] 1、The present application directly processes the radar raw echo data by the input preprocessing layer in the target detection network, respectively uses two linear layer networks with different weight matrices to perform feature mapping on the real part and the imaginary part of different wave position data, can not only avoid the information loss caused by the traditional Fourier transform, but also directly use the original echo data of the multi-wave position pulse radar to realize end-to-end optimization from data to detection result, solve the problems of non-adaptive feature extraction and information loss and non-end-to-end process of the traditional deep learning method, thereby combining the adaptive learning ability of the deep neural network, can realize high-precision and low-computing-cost target detection in a complex scene, can comprehensively capture the multi-dimensional features such as amplitude and phase of the target, significantly improve the detection performance, reduce the sensitivity to complex scenes, improve the efficiency and robustness of radar target detection.

[0025] 2. The application further maps the features by inputting different wave position data into two independent linear layer networks respectively to correspond to process the real part and the imaginary part of the radar echo data under different wave positions, and simultaneously initializes the weight matrix of the two linear layer networks as a discrete approximation form of the Fourier transform basis function, which can simulate the process of extracting frequency domain information by Fourier transform, and further enables adaptive extraction of optimal radar target frequency domain features in the subsequent network training process.

[0026] 3. The application further uses a multi-branch convolution structure, which can match the large data volume, small target and no obvious shape, color, texture and other characteristics of the multi-wave position pulse radar echo data, improve the real-time performance of network reasoning and the accuracy of target feature representation; at the same time, combining the high efficiency reasoning ability of parallel multi-branch convolution unit in reasoning and the lightweight channel attention mechanism, it can realize high-precision and low-computing-overhead target detection in complex scenes, and through adaptive learning, it can comprehensively capture the multi-dimensional features such as amplitude and phase of the target, and significantly improve the detection performance and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is the implementation flowchart of the multi-wave position pulse radar target detection method based on deep learning of the embodiment.

[0028] Figure 2 is the complete flowchart of the multi-wave position pulse radar target detection in the embodiment.

[0029] Figure 3 is the implementation flowchart of signal processing in the embodiment.

[0030] Figure 4 is the structural principle diagram of the target detection network constructed in the embodiment.

[0031] Figure 5 is the data processing flowchart of the input preprocessing layer in the embodiment.

[0032] Figure 6 is the structural principle diagram of a single encoder used in the specific application embodiment of the application.

[0033] Figure 7 is the structural principle diagram of a single decoder used in the specific application embodiment of the application.

[0034] Figure 8 is the structural principle diagram of the output layer used in the specific application embodiment of the application. DETAILED DESCRIPTION

[0035] The present application will be further described by way of illustration with reference to the accompanying drawings and specific preferred embodiments described hereinbelow.

[0036] As shown in the present disclosure, unless the context clearly indicates otherwise, the words "one", "an", "a", and / or "the" do not mean "only one", but can include a plurality or "one or more" than one. The words "first", "second", and similar words do not necessarily indicate any order, quantity, or importance, but are only used to distinguish different components. Similarly, the words "comprise", "comprises" and "comprising", or words of similar import do not exclude other elements or articles from occurring in the process, method, object or apparatus. The words "connected" or "coupled" or similar words do not necessarily mean physically or mechanically connected, but can include electrical connection, whether direct or indirect.

[0037] For the sake of understanding, first exemplary related technical background involved in the present application is introduced.

[0038] Radar sensors have the advantages of long detection distance, all-weather operation, high precision, and strong penetration, and are suitable for various complex scenes. Radar target detection, as the core step of radar signal processing, its core task is to effectively separate the target's motion echo from complex background clutter and noise. Multi-wave position pulse radar is based on pulse radar, and uses multiple different wave positions for detection to improve the performance of radar target detection.

[0039] Traditional multi-wave position pulse radar target detection methods based on statistical signal processing usually assume that the background noise or clutter is uniformly distributed. However, in actual scenarios, the clutter often shows obvious non-uniformity (such as terrain clutter, sea clutter, and weather clutter), which can cause false alarms in strong clutter areas and missed detection in weak clutter areas, thereby significantly reducing the detection performance. In addition, in a multi-target environment, traditional detection methods based on statistical signal processing are easily disturbed by adjacent targets, leading to inaccurate threshold estimation and thus causing target shadowing phenomenon, i.e. strong targets covering weak targets, or setting the threshold too high, resulting in missed detection of weak targets.

[0040] In addition, the traditional detection method based on statistical signal processing needs to rely on the manually designed background clutter statistical model and threshold strategy (judging whether the template exists based on the threshold), needs to preset parameters for different scenes, is difficult to fully adapt to complex and changeable dynamic environment, frequent parameter adjustment also limits the generalization ability and real-time performance of the system, and the artificial feature extraction process is easily disturbed by noise, and the deviation of statistical characteristics will further exacerbate the problems of target masking and false alarm rate. At the same time, frequent parameter adjustment not only consumes a lot of resources, but also limits the generalization ability of the system, making it difficult to achieve efficient and stable target detection in practical application.

[0041] With the rapid development of radar technology, methods based on deep learning are gradually introduced into the target detection task of multi-wave position pulse radar echo signals. Deep learning can automatically extract the features of input data through deep neural networks, and these features can effectively classify targets and backgrounds, which has the advantages of automatic feature extraction, high classification accuracy, strong robustness and good generalization ability. In the existing technology, the multi-wave position pulse radar target detection method based on deep learning needs to convert the multi-wave position echo data into radar images through complex signal processing to adapt to the convolutional neural network model based on images. For example, the original echo data of the multi-wave position pulse radar is usually converted into a range-doppler image form through signal processing methods such as Fourier transform, in order to adapt to the processing needs of the convolutional neural network model.

[0042] However, when processing multi-wave position radar echo data, the complex signal processing process of relying on Fourier transform to convert radar echo data into image form will cause problems such as complex network input preprocessing, loss of original echo data information, and non-end-to-end detection network, affecting the real-time performance of the network, the completeness and accuracy of feature extraction, and the network's ability to extract globally optimal features, thereby reducing the precision and robustness of the target detection network. Specifically, first, Fourier transform is a fixed mathematical transformation, not a data-driven feature extraction method, which limits the model's ability to automatically learn optimal features from data; second, after the original echo data is Fourier transformed and imaged, compared to the original radar echo data, some information is inevitably lost, which weakens the model's ability to fully understand target features, leading to information loss and non-adaptive feature extraction problems; finally, Fourier transform as a preprocessing step divides the radar target detection task into multiple independent stages, resulting in a non-end-to-end process, which limits the model's global optimization ability, making it difficult to achieve global optimization. The above problems affect the performance of radar detection and limit the application potential of deep learning in multi-wave position pulse radar target detection tasks.

[0043] The application can automatically learn target features based on the neural network constructed based on deep learning, can effectively solve the false alarm and missed detection problems in a non-uniform clutter environment, can avoid the target shielding phenomenon in a multi-target scene, avoids the defects that the traditional method needs to rely on artificial design of a statistical model and a fixed threshold strategy, adopts a data-driven feature extraction mode to automatically learn optimal features, reduces the dependence on preset parameters, and can significantly improve the generalization ability and adaptability of the system.

[0044] Meanwhile, according to the data characteristics of the multi-wave position pulse radar echo data, the target detection network of the application is composed of an input preprocessing layer, an encoding layer, a decoding layer and an output layer, the input preprocessing layer is used for multi-wave position and multi-dimensional feature extraction on the input multi-wave position radar original echo data, and the encoding layer, the decoding layer and the output layer can be combined to form an efficient, adaptive and end-to-end multi-wave position pulse radar target detection framework. By setting the input preprocessing layer in the target detection network, the input preprocessing layer directly processes the radar original echo data, separates the real part and the imaginary part of the radar original echo data, respectively uses two linear layer networks with different weight matrices to perform feature mapping on the real part and the imaginary part of the different wave position data, calculates the real part feature and the imaginary part feature according to the feature mapping result, calculates the phase spectrum feature according to the real part feature and the imaginary part feature, and obtains the multi-dimensional feature map from the real part feature, the imaginary part feature and the phase spectrum feature. Not only can the information loss caused by the traditional Fourier transform be avoided, but also the end-to-end optimization from data to detection result can be realized directly using the original echo data of the multi-wave position pulse radar, the problems of non-adaptive feature extraction and information loss and non-end-to-end process caused by the traditional deep learning method which needs to use Fourier transform for complex signal processing are solved, and the adaptive learning ability of the deep neural network is combined to realize high-precision and low-computing-cost target detection in a complex scene, to comprehensively capture the amplitude, phase and other multi-dimensional features of the target, to significantly improve the detection performance, to reduce the sensitivity to complex scenes, and to improve the efficiency and robustness of radar target detection.

[0045] The application will be further described below in combination with specific embodiments.

[0046] As shown in Figure 1 , the steps of the multi-wave position pulse radar target detection method based on deep learning of the embodiment include:

[0047] Step S01. Data preparation: obtaining multi-wave position pulse radar original echo data in different detection scenes to form a radar original echo data set;

[0048] Step S02. Model training: Using the radar original echo data set to train the target detection network constructed based on the neural network, and obtaining the target detection model after the training is completed; the target detection network includes an input preprocessing layer, an encoding layer, a decoding layer and an output layer connected in sequence, the input preprocessing layer is used to extract multi-wavelength and multi-dimensional features of the input radar original echo data, and the extracted features are output as target detection results after passing through the encoding layer, the decoding layer and the output layer; in the input preprocessing layer, by separating the real and imaginary parts of the radar original echo data, two linear layer networks with different weight matrices are used to perform feature mapping on the real and imaginary parts of the different wavelength data, respectively, and the real and imaginary features and imaginary features are calculated according to the feature mapping results, and the phase spectrum features are calculated according to the real and imaginary features, and the real features, the imaginary features and the phase spectrum features are used to construct a multidimensional feature map output;

[0049] Step S03. Target detection: Obtain the radar original echo data of the multi-wavelength pulse to be detected, input it into the trained target detection model for detection, and obtain the target detection result output.

[0050] Through the above steps, this embodiment implements end-to-end multi-wavelength pulse radar target detection based on deep learning. The neural network structure can directly process the original echo data of the multi-wavelength pulse radar to extract features, avoiding the information loss caused by Fourier transform. It can also be trained end-to-end with the downstream network to build an end-to-end detection framework, driving the network to adaptively extract key features of the target signal from the original radar data, achieving efficient optimization from data to results. Combined with the adaptive learning ability of the deep neural network, it can achieve high-precision and low-computational-cost target detection in complex scenarios. At the same time, through adaptive learning, it comprehensively captures the multi-dimensional features of the target such as amplitude and phase, significantly improving detection performance and robustness, and realizing a more efficient and intelligent multi-wavelength pulse radar target detection system.

[0051] In a specific application embodiment, the present invention is divided into three stages: data set construction, model construction and training, and model deployment, which correspond to steps S01 to S03 respectively. Figure 2As shown, in the data set construction phase, first, the multi-wave position pulse radar original echo data under multiple detection scenes is collected, the collected echo data is cleaned to improve the data quality, and after the cleaned data is processed, the distance Doppler graph needed for subsequent data labeling is obtained, the target information is labeled on the distance Doppler graph after signal processing, the labeled information required for network training is obtained, the original echo data after data cleaning is converted into a unified format and saved as training data and label information are input into the network for training. In the model construction and training phase, according to the data characteristics of the multi-wave position pulse radar echo data, a target detection network is constructed, the input of the network is the neural network of the original radar echo data, the label data and training data obtained in the data preparation phase are input into the constructed target detection network for network model training, and finally, the model parameters obtained by neural network training are saved to obtain the target detection model. In the model deployment phase, the real-time original echo data collected by the radar is preprocessed (such as normalization) and then input into the saved model for inference prediction to obtain real-time target detection results.

[0052] In the data preparation phase of step S01, the original echo data of the multi-wave position pulse radar needs to be collected to construct the data set required for model training and testing. In a specific application embodiment, the following steps are included:

[0053] Step S101. Data collection: Collecting original echo data of multi-wave position pulse radar under different detection scenes.

[0054] Specifically, data collection work can be carried out within the radar detection range, and multi-wave original echo data of radar targets at different heights and distances in 360 degrees can be collected. The collection process can cover the echo conditions of targets in stationary, uniform motion, accelerated / decelerated motion, and direction-changing motion to ensure data diversity and representativeness. At the same time, radar equipment can be arranged in different typical complex scenes such as suburban fields, mountain forests, and urban high-rise buildings, and multiple sets of original echo data can be collected in each scene to enrich the background environment data and make the data set fully reflect the clutter characteristics and target features in different environments. Further, the position information of the target relative to the radar during data collection can be recorded to facilitate the differentiation between the target and the background in subsequent data labeling.

[0055] Step S102. Data cleaning: cleaning the collected multi-wave position pulse radar original echo data.

[0056] Specifically, by analyzing the original data, invalid data such as outliers and noise interference can be removed to ensure data quality.

[0057] Step S103. Signal processing: the original echo data of the cleaned multi-wave position pulse radar is processed to obtain a complete range-Doppler map.

[0058] Specifically, as shown in Figure 3 , first, the cleaned radar original data is read to obtain a complex data matrix of the total number of pulse units H and the total number of distance units W ( H, W ), wherein the radar data from different wave positions ( r =1, 2, 3,..., n , n are arranged in wave position order, and the echo data of different wave positions r have different numbers of pulse units and the same number of distance units W , wherein ; then the echo data of different wave positions r are subjected to signal processing operations such as pulse compression and Fourier transform to generate range-Doppler maps of different wave positions to reflect the range and Doppler frequency information of the target; finally, the range-Doppler maps of different wave positions are spliced together according to the pulse dimension to obtain a complete range-Doppler map ( H, W ).

[0059] Step S104. Data labeling: the obtained complete range-Doppler map is labeled.

[0060] Specifically, first, the target on the complete range-Doppler map obtained by signal processing is labeled to obtain the class label and normalized bounding box information of each target, including the center point coordinate ( x_center, y_center ) and the width and height dimension ( width, height ). Based on the labeling information, a multi-bit (for example, 8 bits can be taken) single-channel labeled image with the same size as the original range-Doppler spectrum is generated. In the labeled image, the background region pixel value is set to 0, and each target region is assigned a different pixel value according to its class label (for example, class 1 can be set to 1, class 2 can be set to 2, etc.). The finally generated labeled image will together with the original radar data form a training data set for subsequent deep learning model training.

[0061] In a specific application embodiment, an elliptical labeling method can be used, taking the center point of the labeling information as the center of the ellipse, taking half of the target width and height as the major axis and minor axis of the ellipse respectively, drawing a solid ellipse region at the corresponding position of the labeled image, and filling the inside of the ellipse with the pixel value corresponding to the target class. By using the elliptical labeling method, the diffusion characteristics of the radar target in the range-Doppler domain can be better matched, and the actual energy distribution range of the target can be more accurately reflected compared to rectangular frame labeling.

[0062] Step S105. Data conversion: format conversion is performed on the original radar echo data after data cleaning.

[0063] Specifically, by uniformly converting the original radar echo data, for example, converting the cleaned radar echo data into a unified binary or structured data format, and storing it using standardized naming conventions, the standardized data processing flow of the subsequent offline training model can be supported.

[0064] In this embodiment, the input preprocessing layer is used for multi-wave position and multi-dimensional feature extraction of the radar echo signal. By separating the real part and the imaginary part of the radar original echo data, two linear layer networks with different weight matrices are used to map the real part and the imaginary part of the different wave position data. The real part feature and the imaginary part feature are calculated according to the feature mapping results, and the phase spectrum feature is calculated according to the real part feature and the imaginary part feature. The multi-dimensional feature map output is formed by the real part feature, the imaginary part feature and the phase spectrum feature.

[0065] In this embodiment, the input preprocessing layer uses two linear layer networks with different weight matrices to map the real part and the imaginary part of the different wave position data, which includes:

[0066] According to the number of pulses of different wave position data in the echo data, different wave position data is extracted;

[0067] The different wave position data is respectively input into two independent linear layer networks for feature mapping. The weight matrix of each linear layer network is composed of a plurality of fixed-dimension sub-matrix blocks to correspondingly process the radar echo data under different wave positions. The weight matrices of the two linear layer networks are initialized as a discrete approximation form of Fourier transform basis functions.

[0068] By using two linear layer networks to correspondingly process the real part and the imaginary part of different wave position data, and initializing the two linear layer networks in the above manner, the process of extracting frequency domain information by Fourier transform can be simulated, and the best radar target frequency domain feature can be adaptively extracted in the subsequent network training process.

[0069] Specifically, initializing the weight matrices of the two linear layer networks as a discrete approximation form of Fourier transform basis functions includes: according to the size and coordinate configuration of each sub-matrix block, the weights of the real part weight block and the imaginary part weight block are configured, wherein the weights of the real part weight block are taken as cos -2πij / k , and the weights of the imaginary part weight block are taken as sin -2πij / k , wherein k is the block size, i , j ​​respectively, the real part weight block is a submatrix block in a weight matrix of a first linear layer network corresponding to processing real part data, and the imaginary part weight block is a submatrix block in a weight matrix of a second linear layer network corresponding to processing imaginary part data.

[0070] In the input preprocessing layer, according to the real part weight matrix and the imaginary part weight matrix, the real part feature real_feature and the imaginary part feature imag_feature of the Fourier transform result can be obtained through complex operation, the real part weight matrix is a weight matrix of a first linear layer network corresponding to processing real part data, and the imaginary part weight matrix is a weight matrix of a second linear layer network corresponding to processing imaginary part data, and the calculation expression is as shown in the following formula:

[0071]

[0072]

[0073] wherein, represents the real part weight matrix, represents the imaginary part weight matrix, represents the real part signal, and the imaginary part signal.

[0074] In the embodiment, according to the extracted real part feature real_feature and the imaginary part feature imag_feature , the phase spectrum feature can be obtained through the inverse tangent function, and the multi-dimensional feature map can be constructed by splicing the extracted real part feature, imaginary part feature and phase spectrum feature.

[0075] In a specific application embodiment, as shown in Figure 4 , an input preprocessing layer is arranged at the input end of the target detection network, the input preprocessing layer first performs multi-wave position and multi-dimensional feature extraction on the input multi-wave position radar original echo data, and then sequentially passes through an encoding layer, a decoding layer and an output layer to obtain a final detection result. As shown in Figure 5 , the detailed processing procedure of the input preprocessing layer includes: first, the radar echo signal usually contains real part and imaginary part, the input preprocessing layer separates the real part and the imaginary part of the input data, extracts different wave position data according to the different pulse numbers of different wave position data in the echo data, and sends the different wave position data into two independent linear layer networks for feature mapping respectively, the weight matrices of the two networks are initialized as a discrete approximation form of the Fourier transform basis function, wherein the weight matrix is composed of a plurality of fixed-dimension submatrix blocks, corresponding to processing radar echo data under different wave positions, the element values in each sub-block follow the transformation principle of the Fourier transform: for the block size k and the coordinates i , j , the real part weight block takescos - 2π ij / k sin - 2π ij / k ), so that the linear layer network can simulate the process of Fourier transform to extract the frequency domain information, and adaptively extract the optimal radar target frequency domain features in the training process of the subsequent network; after the real part and the imaginary part signal pass through the projection of the customized weight matrix respectively, the real part feature real_feature and the imaginary part feature imag_ feature are obtained through the complex operation formula to obtain the real part feature of the Fourier transform result

[0076] The target detection network in the embodiment adopts a U-shaped structure, as shown in Figure 4 The network can be divided into five parts in structure, including an input preprocessing layer, an encoding layer, an intermediate layer, a decoding layer and an output layer. The encoding layer includes multiple layers of encoders for feature extraction and realization of spatial dimension compression and channel dimension expansion. The decoding layer includes multiple layers of decoders for feature fusion of different layers and gradual recovery of spatial dimension. The output layer is used for fusing feature maps of different scales and outputting the final target detection result. Each layer of encoder and each layer of decoder are connected in a step-by-step stacking manner, each layer of encoder is connected to one layer of decoder, and the encoding layer and the decoding layer further extract semantic features through the intermediate layer. By using the target detection network with the above structure, pixel-level target detection can be realized.

[0077] In the embodiment, the encoder layer realizes feature extraction, spatial dimension compression and channel dimension expansion through multiple layers of encoders. Each layer of encoder includes a max pooling layer and multiple parallel branch convolution units. At least one branch convolution unit in each branch convolution unit is used to extract local detail features, at least one branch convolution unit is used to realize cross-channel information interaction and dimension transformation, and at least one branch convolution unit is used to retain the identity mapping of the original input feature map. By using the multi-branch convolution structure, the encoding layer can match the large data volume, small target and no obvious shape, color, texture and other characteristics of the multi-wave bit pulse radar echo data, and improve the real-time performance of network reasoning and the accuracy of target feature representation.

[0078] In a specific application embodiment, as shown in Figure 6 ​​​As shown, a single encoder can employ a four-layer architecture to achieve spatial dimension compression and channel dimension expansion through progressively stacked max-pooling layers and convolution operations, and each layer of the encoder is composed of a max-pooling layer and a parallel three-branch convolution unit, where the first branch employs a 3x3 standard convolution to extract local detail features, the second branch implements cross-channel information interaction and dimension transformation through a 1x1 convolution, and the third branch preserves the identity mapping of the original input feature map. Among them, the first layer of the encoder does not have a max-pooling layer for feature down-sampling, but uses a 1x1 convolution for channel adjustment of the feature map.

[0079] In a specific application embodiment, the intermediate layer can be composed of a max-pooling layer and a 3x3 convolution layer for further extracting deeper semantic features between the encoder and the decoder.

[0080] In this embodiment, the decoding layer employs a multi-layer encoder structure to achieve feature fusion and gradual recovery of spatial dimensions, while reducing the channel dimension, where each layer of the decoder implements up-sampling through a bilinear interpolation unit, the feature map obtained after up-sampling is combined with the feature map output by the previous layer of the decoder through a cross-layer feature splicing unit to achieve multi-scale feature splicing, and a parallel multi-branch convolution is used to perform feature transformation and channel dimension compression on the spliced multi-scale features.

[0081] In a specific application embodiment, as shown in Figure 7 A single decoder employs a four-layer structure, including four up-sampling structures, where c, h, w represents the number of channels, height, and width of the feature map input to the current decoding layer. Each layer of the decoder includes a bilinear interpolation up-sampling unit, a cross-layer feature splicing unit, and a parallel multi-branch convolution, which implements up-sampling through a bilinear interpolation unit to restore the size of the feature map layer by layer; the feature map obtained after up-sampling is combined with the feature map output by the previous layer of the decoder through a cross-layer feature splicing unit to achieve multi-scale feature splicing; a parallel multi-branch convolution is also used in each layer of the decoder to perform feature transformation and channel dimension compression on the spliced multi-scale features, achieving gradual decoding and multi-scale feature fusion of the features.

[0082] In this embodiment, the output layer of the target detection network is used to fuse different scale feature maps and output the final network prediction result, including a convolution prediction head, a multi-scale up-sampling unit, a channel attention mechanism, and a convolution fusion unit. The convolution prediction head is used to predict the prediction results of multiple different scale feature layers output by the decoding layer and adjust the number of channels, the multi-scale up-sampling unit is used to up-sample the prediction results of each different scale feature layer, the channel attention mechanism forms a multi-scale feature map by weighting the multi-scale up-sampled feature map based on the attention mechanism, and the convolution fusion unit is used to fuse the weighted multi-scale feature map and generate the final target detection result.

[0083] In a specific application embodiment, as shown in Figure 8 As shown in the figure, the output layer first predicts the prediction results of four different scale feature layers of the output decoder output through the convolution prediction head, adjusts the number of channels, then up-samples the four different scale prediction results through the multi-scale up-sampling unit to make the feature size consistent with the original input size; based on the lightweight channel attention mechanism, the multi-scale up-sampled feature map is weighted by analyzing the channel features of the input feature map, so as to enhance the response of important feature channels and suppress unimportant feature channels, and finally the weighted multi-scale feature map is fused through the convolution fusion unit to generate the final prediction result.

[0084] In this embodiment, the traditional Unet target detection model is improved by using the above structure, and the lightweight channel attention module and the multi-branch convolution structure are used to construct the backbone feature extraction network based on the multi-branch convolution neural network structure. The network can enhance the feature expression ability through the parallel multi-path structure in the model training stage, and can convert the multi-path structure into a single chain topology structure in the inference stage, which can not only ensure the model performance but also significantly improve the operation efficiency, so that the real-time performance of the pulse radar target detection task can be well met, the real-time performance of the network inference is improved, and the accuracy of the target feature representation is improved. At the same time, combined with the high inference ability of the parallel multi-branch convolution unit in inference and the lightweight channel attention mechanism, high-precision and low-computing-cost target detection can be realized in complex scenes, and the amplitude, phase and other multi-dimensional features of the target are comprehensively captured through adaptive learning, which significantly improves the detection performance and robustness.

[0085] In this embodiment, in the model training stage, the constructed training set, i.e. the original echo data of the radar and the corresponding label information, is sent into the constructed neural network for training. After each training period, the constructed verification sample set is input into the neural network for verification to detect the accuracy of the network model in real time, and the weight parameters of the best accuracy model are saved. After the training is completed, the constructed test sample set is input into the neural network for testing to detect the performance parameters such as detection rate, false detection rate and missed detection rate of the network model. If the network model meets the expected requirements, a deployable target detection model can be obtained through conversion.

[0086] In this embodiment, in the model deployment stage, the radar target detection model obtained by training is imported into the radar target detection device, and the radar device with the imported detection model is used to monitor the target area to process the multi-wave position pulse radar original echo data in real time using the trained model. After obtaining the real-time original radar echo data, the radar original echo data is normalized and preprocessed, the preprocessed radar echo data is input into the target detection model for prediction and inference, and the final target detection result is obtained.

[0087] Further, the prediction result of the target detection model can also be post-processed, for example, by setting a threshold to binarize the network output to distinguish target and background regions, using a connected component detection algorithm to extract target connected regions and determine their center coordinates, converting these pixel coordinates into actual distance and speed physical quantities according to the distance unit and pulse unit parameters of the radar, and finally outputting the distance-speed information of each target, so that the distance and speed information of the target can be obtained, and efficient target detection in complex scenes can be realized.

[0088] The present application is directed to the characteristics of multi-wave position pulse radar data, and two independent linear layer networks are used to correspond to process the real part and imaginary part data of different wave position data, which can directly extract the frequency domain features of the original multi-wave position pulse radar echo data, and can be optimized through end-to-end training and downstream network.

[0089] The embodiment further provides an electronic device, including a processor and a memory, the memory is used to store a computer program, and the processor is used to execute the computer program to perform the method as described above.

[0090] It can be understood that the above method of the embodiment can be executed by a single device, such as a computer or a server, and can also be applied to a distributed scenario to be completed by multiple devices in cooperation, in which case, one of the multiple devices can only execute one or more steps in the above method of the embodiment, and the multiple devices interact to complete the above method. The processor can be implemented in the form of a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, and is used to execute related programs to implement the above method of the embodiment. The memory can be implemented in the form of a read-only memory (ROM), a random access memory (RAM), a static storage device, and a dynamic storage device. The memory can store an operating system and other application programs, and when the above method of the embodiment is implemented by software or firmware, the related program codes are saved in the memory and executed by the processor.

[0091] The embodiment further provides a computer-readable storage medium having a computer program stored therein, and the computer program is executed by a processor to implement the method as described above.

[0092] Those skilled in the art will appreciate that the above-described embodiments of the present application can be embodied as a method, system, or computer program product. Accordingly, the present application can be embodied in hardware alone, software alone, or in combination with software and hardware. Moreover, the present application can be embodied in a computer program product on one or more computer readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and so forth) having computer usable program code embodied thereon. The present application is described in the general context of method, apparatus (system), and computer program product, according to one embodiment of the application. The computer program product can be implemented in a local memory of a computing device or in a memory of a computing device accessible via a network, such as the Internet or World Wide Web. Generally, computer program product comprises computer usable program code, which is a physical tangible entity that can be used directly, or indirectly, by a computing device to Figure 1 one or more functions specified in the flow or flows and / or blocks Figure 1 one or more functions specified in the flow or flows and / or blocks Figure 1 one or more functions specified in the flow or flows and / or blocks Figure 1 one or more functions specified in the flow or flows and / or blocks Figure 1 one or more functions specified in the flow or flows and / or blocks Figure 1 one or more functions specified in the flow or flows and / or blocks

[0093] The foregoing is considered as illustrative only of the principles of the application. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and practice described. Accordingly, all suitable modifications and equivalents can be resorted to falling within the scope of the application.

Claims

1. A multi-wavelength pulse radar target detection method based on deep learning, characterized in that the steps include: Obtain radar raw echo data of multi-wavelength pulses in different detection scenarios to form a radar raw echo data set; The radar original echo data set is used to train the target detection network constructed based on the neural network, and a target detection model is obtained after the training is completed; the target detection network includes an input preprocessing layer, an encoding layer, a decoding layer and an output layer connected in sequence, the input preprocessing layer is used to perform multi-wavelength and multi-dimensional feature extraction on the input radar original echo data, and the extracted features are output as target detection results after passing through the encoding layer, the decoding layer and the output layer; in the input preprocessing layer, the real part and the imaginary part of the radar original echo data are separated, and two linear layer networks with different weight matrices are used to perform feature mapping on the real part and the imaginary part of the different wavelength data, and the real part features and the imaginary part features are calculated according to the feature mapping results, and the phase spectrum features are calculated according to the real part features and the imaginary part features, and the real part features, the imaginary part features and the phase spectrum features are used to construct a multidimensional feature map output; The radar original echo data of the multi-wavelength pulse to be detected is obtained, and is input into the target detection model obtained after training for detection, and the target detection result is output.

2. The multi-wavelength pulse radar target detection method based on deep learning according to claim 1 is characterized in that: In the input preprocessing layer, two linear layer networks with different weight matrices are used to perform feature mapping on the real and imaginary parts of different wave position data, including: Extract different wave position data according to the pulse number of different wave position data in the echo data; Different waveposition data are input into two independent linear layer networks for feature mapping. The weight matrix of each linear layer network is composed of multiple sub-matrix blocks of fixed dimensions to process radar echo data at different wavepositions. The weight matrices of the two linear layer networks are initialized as discrete approximations of the Fourier transform basis function.

3. The multi-wavelength pulse radar target detection method based on deep learning according to claim 2 is characterized in that: Initializing the weight matrices of the two linear layer networks to the discrete approximate form of the Fourier transform basis function includes: configuring the weights of the corresponding real weight block and imaginary weight block according to the size and coordinates of each sub-matrix block, wherein the weight of the real weight block is cos ( -2πij / k ), the weight of the imaginary weight block is sin ( -2πij / k ),in k is the block size, i , j They are the horizontal and vertical coordinate values ​​of the blocks respectively. The real weight block is the sub-matrix block in the weight matrix of the first linear layer network corresponding to processing real data, and the imaginary weight block is the sub-matrix block in the weight matrix of the second linear layer network corresponding to processing imaginary data.

4. The multi-wavelength pulse radar target detection method based on deep learning according to claim 1 is characterized in that: In the input preprocessing layer, the real part features of the Fourier transform result are obtained by complex operations based on the real part weight matrix and the imaginary part weight matrix real_feature and imaginary part characteristics imag_feature The real part weight matrix is ​​the weight matrix of the first linear layer network corresponding to the real part data processing, and the imaginary part weight matrix is ​​the weight matrix of the second linear layer network corresponding to the imaginary part data processing. real_feature and imaginary part characteristics imag_feature The phase spectrum features are obtained by the inverse tangent function, and the extracted real part features, imaginary part features and phase spectrum features are spliced ​​together to form the multidimensional feature map.

5. The multi-wavelength pulse radar target detection method based on deep learning according to any one of claims 1 to 4, characterized in that: The encoding layer in the target detection network includes multiple layers of encoders for feature extraction and spatial dimension compression and channel dimension expansion. The decoding layer includes multiple layers of decoders for feature fusion of different layers and gradual restoration of spatial dimensions. The output layer is used to fuse feature maps of different scales and output the final target detection result. The encoders and decoding layers of each layer are connected in a step-by-step stacking manner. Each encoder layer is connected to a corresponding decoder layer. Semantic features are further extracted through an intermediate layer between the encoding layer and the decoding layer.

6. The multi-wavelength pulse radar target detection method based on deep learning according to claim 5 is characterized in that: Each layer of the encoder includes a maximum pooling layer and parallel multi-branch convolution units. Among the branch convolution units, at least one branch convolution unit is used to extract local detail features, at least one branch convolution unit is used to realize cross-channel information interaction and dimensionality transformation, and at least one branch convolution unit is used to preserve the identity mapping of the original input feature map.

7. The multi-wavelength pulse radar target detection method based on deep learning according to claim 5 is characterized in that: Each layer of the decoder achieves upsampling through a bilinear interpolation unit. The feature map obtained after upsampling and the feature map output by the previous layer of decoder are concatenated into multi-scale features through a cross-layer feature concatenation unit. Parallel multi-branch convolution is used to perform feature transformation and channel dimension scale compression on the concatenated multi-scale features.

8. The multi-wavelength pulse radar target detection method based on deep learning according to any one of claims 1 to 4, characterized in that: The output layer in the target detection network includes a convolutional prediction head, a multi-scale upsampling unit, a channel attention mechanism and a convolutional fusion unit. The convolutional prediction head is used to predict the prediction results of multiple different-scale feature layers output by the decoding layer and adjust the number of channels. The multi-scale upsampling unit is used to upsample the prediction results of each different-scale feature layer. The channel attention mechanism forms a multi-scale feature map by weighting the feature maps after multi-scale upsampling based on the attention mechanism. The convolutional fusion unit is used to fuse the weighted multi-scale feature maps and generate the final target detection result.

9. An electronic device comprising a processor and a memory, wherein the memory is used to store a computer program, wherein: The processor is configured to execute the computer program to perform the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images