Multi-wave-potential pulse radar target detection method and device based on deep learning, and medium

Abstract

The invention discloses a multi-wave-potential pulse radar target detection method and device based on deep learning and a medium, and the method comprises the steps: obtaining radar original echo data in different detection scenes, and forming a radar original echo data set; training the 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, a real part and an imaginary part of radar original echo data are separated, linear layer networks with different weight matrixes are used for carrying out feature mapping on the real part and the imaginary part of different wave position data, the features of the real part and the imaginary part are calculated, then the phase spectrum features are calculated, and a multi-dimensional feature map is formed; training a target detection network by using the echo data set; and obtaining echo data to be detected, and inputting the echo data to the target detection model for detection. According to the method, high-precision and low-calculation-overhead target detection can be realized in a complex scene, and meanwhile, the detection performance and robustness are remarkably improved.

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, existing deep learning-based multi-wavelength pulse radar target detection methods typically require converting complex signal processing into radar images to accommodate image-based convolutional neural network models. Complex signal processing operations, such as Fourier transforms and time-frequency analysis, are then performed on the radar echo data for each wavelength to extract information such as range and Doppler as input to the neural network. This process is not only computationally expensive and complex to implement, but can also lead to loss or distortion of the target's original features due to the fixed signal transformation method, thus affecting detection performance. Furthermore, the signal processing for each wavelength is separated from the neural network training process, making it impossible for the model to adaptively adjust the feature extraction strategy based on data characteristics. This makes the entire process non-end-to-end, severely restricting end-to-end learning capabilities, affecting the network's real-time performance, the integrity and accuracy of feature extraction, and the network's ability to extract the globally optimal features, thereby reducing the accuracy and robustness of the target detection network. Summary of the Invention

[0006] The technical problem to be solved by the present invention is: In response to the technical problems existing in the prior art, the present invention provides a multi-wavelength pulse radar target detection method, equipment and medium based on deep learning, which can achieve high-precision and low-computational overhead target detection in complex scenarios, while significantly improving detection performance and robustness.

[0007] In order to solve the above technical problems, the technical solution proposed by the present invention is: A multi-wavelength pulse radar target detection method based on deep learning, comprising the following steps: 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.

[0008] Furthermore, 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, respectively, including: Extract different wave position data according to the pulse number of different wave position data in the echo data; Different wave position 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 wave positions. The weights of the corresponding real weight block and imaginary weight block are configured according to the size and coordinates of each sub-matrix block.

[0009] Furthermore, in the weights of the real weight block and the imaginary weight block corresponding to the size and coordinate configuration of each sub-matrix block, 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.

[0010] Furthermore, in the input preprocessing layer, the real part feature of the Fourier transform result is obtained by complex operation based on the real part weight matrix and the imaginary part weight matrix. real_feature and imaginary part characteristics imag_feature , according to the extracted real part features 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.

[0011] Furthermore, 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, and semantic features are further extracted through an intermediate layer between the encoding layer and the decoding layer.

[0012] Furthermore, each layer of the encoder includes a maximum pooling layer and parallel multi-branch convolution units, at least one of the branch convolution units 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.

[0013] Furthermore, 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 spliced ​​together to achieve multi-scale features through a cross-layer feature splicing unit. Parallel multi-branch convolution is used to perform feature transformation and channel dimension scale compression on the spliced ​​multi-scale features.

[0014] Furthermore, 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.

[0015] An electronic device includes a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0016] A computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention sets an input preprocessing layer in the target detection network, and the input preprocessing layer directly processes the radar raw echo data, and uses two linear layer networks with different weight matrices to perform feature mapping on the real and imaginary parts of different wave position data. It can not only avoid the information loss caused by traditional Fourier transform, but also directly use the original echo data of multi-wave position pulse radar to achieve end-to-end optimization from data to detection results, and solve the problems of non-adaptive feature extraction and information loss and non-end-to-end process of traditional deep learning methods. Therefore, combined with the adaptive learning ability of deep neural networks, it can achieve high-precision and low-computational overhead target detection in complex scenarios, and at the same time, it can fully capture the multi-dimensional features of the target such as amplitude and phase, significantly improve the detection performance, and at the same time reduce the sensitivity to complex scenarios, and improve the efficiency and robustness of radar target detection.

[0018] 2. The present invention further inputs different wave position data into two independent linear layer networks for feature mapping to process the real and imaginary parts of the radar echo data at different wave positions, and initializes the weight matrices of the two linear layer networks to discrete approximate forms of the Fourier transform basis functions, which can simulate the process of extracting frequency domain information by Fourier transform, thereby enabling the optimal radar target frequency domain features to be adaptively extracted during the subsequent network training process.

[0019] 3. The present invention further uses a multi-branch convolution structure to match the large amount of data, small targets and no obvious shape, color, texture and other features of multi-wavelength pulse radar echo data, thereby improving the real-time nature of network reasoning and the accuracy of target feature characterization; at the same time, combined with the efficient reasoning ability of the parallel multi-branch convolution unit during reasoning and the lightweight channel attention mechanism, it can achieve high-precision and low-computational overhead target detection in complex scenarios, and at the same time, through adaptive learning, comprehensively capture the target's amplitude, phase and other multi-dimensional features, significantly improving detection performance and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of the implementation flow of the multi-wavelength pulse radar target detection method based on deep learning in this embodiment.

[0021] Figure 2 This is a complete flowchart of implementing multi-wavelength pulse radar target detection in this embodiment.

[0022] Figure 3 Schematic diagram of the implementation flow of signal processing in this embodiment.

[0023] Figure 4 Schematic diagram of the structural principle of the target detection network constructed in this embodiment.

[0024] Figure 5 Schematic diagram of the data processing flow of the input preprocessing layer in this embodiment.

[0025] Figure 6 It is a schematic diagram of the structural principle of a single encoder used in a specific application embodiment of the present invention.

[0026] Figure 7 It is a schematic diagram of the structural principle of a single decoder used in a specific application embodiment of the present invention.

[0027] Figure 8 Schematic diagram of the structural principle of the output layer used in a specific application embodiment of the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the scope of protection of the present invention is not limited thereby.

[0029] As shown in the disclosure of the present invention, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not specifically refer to the singular, but also include the plural. The words "first", "second" and similar words used in the disclosure of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0030] For ease of understanding, the relevant technical background of the present invention is first introduced by way of example.

[0031] Radar sensors offer advantages such as long detection range, all-weather operation, high accuracy, and strong penetration, making them suitable for a variety of complex scenarios. Radar target detection, as the core component of radar signal processing, effectively separates the target's motion echo from complex background clutter and noise. Multi-wavelength pulse radar, based on pulse radar, utilizes multiple different wavelengths for detection to improve radar target detection performance.

[0032] Traditional multi-wavelength pulse radar target detection methods based on statistical signal processing typically assume a uniform distribution of background noise or clutter. However, in real-world scenarios, clutter often exhibits significant non-uniformity (such as terrain clutter, sea clutter, and weather clutter). This non-uniformity can cause traditional methods to generate false alarms in areas with strong clutter and miss detections in areas with weak clutter, significantly degrading detection performance. Furthermore, in multi-target environments, traditional detection methods based on statistical signal processing are susceptible to interference from neighboring targets, leading to inaccurate threshold estimation and, in turn, target masking, where strong targets mask weaker targets or the threshold is set too high, leading to missed detection of weaker targets.

[0033] Furthermore, traditional detection methods based on statistical signal processing rely on manually designed background clutter statistical models and threshold strategies (based on thresholds to determine whether a template exists). These methods require pre-set parameters for different scenarios, making them difficult to fully adapt to complex and changing dynamic environments. Frequent parameter adjustments also limit the system's generalization and real-time performance. The manual feature extraction process is susceptible to noise interference, and deviations in statistical characteristics can further exacerbate target obscuration and increased false alarm rates. Furthermore, frequent parameter adjustments not only consume significant resources but also limit the system's generalization capabilities, making it difficult to achieve efficient and stable target detection in practical applications.

[0034] With the rapid development of radar technology, deep learning-based methods have gradually been introduced into the task of target detection using multi-wavelength pulse radar echo signals. Deep learning can automatically extract features from input data through deep neural networks, and these features can be used to effectively classify targets and backgrounds. It has the advantages of automatic feature extraction, high classification accuracy, strong robustness, and good generalization capabilities. In existing multi-wavelength pulse radar target detection methods based on deep learning, it is necessary to convert the multi-wavelength pulse radar echo data into radar images through complex signal processing to adapt to the image-based convolutional neural network model. For example, the raw echo data of the multi-wavelength pulse radar is usually converted into a range Doppler image using signal processing methods such as Fourier transform to adapt to the processing requirements of the convolutional neural network model.

[0035] However, when processing multi-wavelength radar echo data, the complex signal processing process of converting radar echo data into an image using Fourier transforms leads to complex network input preprocessing, information loss in the original echo data, and a non-end-to-end detection network. This affects the real-time performance of the network, the integrity and accuracy of feature extraction, and the network's ability to extract the globally optimal features, thereby reducing the accuracy and robustness of the target detection network. Specifically, the Fourier transform is a fixed mathematical transformation, rather than a data-driven feature extraction method, which limits the model's ability to automatically learn the optimal features from the data. Secondly, after the raw echo data is Fourier transformed and imaged, some information is inevitably lost compared to the original radar echo data. This weakens the model's ability to fully understand target features, leading to information loss and non-adaptive feature extraction. Finally, the Fourier transform, as a preprocessing step, splits the radar target detection task into multiple independent stages, resulting in a non-end-to-end process. This non-end-to-end process limits the model's global optimization capabilities, 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-wavelength pulse radar target detection tasks.

[0036] The present invention constructs a target detection network based on deep learning. The neural network constructed based on deep learning can automatically learn target features, effectively solve the problems of false alarms and missed detections in non-uniform clutter environments, and avoid target obstruction in multi-target scenarios. It avoids the defects of traditional methods that rely on manually designed statistical models and fixed threshold strategies, avoids the defects of traditional methods that rely on manually designed statistical models and fixed threshold strategies, adopts data-driven feature extraction to automatically learn optimal features, reduces dependence on preset parameters, and can significantly improve the generalization ability and adaptability of the system.

[0037] At the same time, according to the data characteristics of the multi-wavelength pulse radar echo data, the target detection network of the present invention is composed of an input preprocessing layer, a coding layer, a decoding layer and an output layer. The input preprocessing layer is used to extract multi-wavelength and multi-dimensional features from the input multi-wavelength radar original echo data. In combination with the coding layer, the decoding layer and the output layer, an efficient, adaptive and end-to-end multi-wavelength pulse radar target detection framework can be constructed. By setting an 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, and uses two linear layer networks with different weight matrices to perform feature mapping on the real part and the imaginary part of different wavelength data. According to the feature mapping results, the real part features and imaginary part features are calculated, and the phase spectrum features are calculated according to the real part features and the imaginary part features. The multidimensional feature map is obtained by the real part features, the imaginary part features and the phase spectrum features. It can not only avoid the information loss caused by the traditional Fourier transform, but also directly utilize the original data of the multi-wavelength pulse radar. Echo data enables end-to-end optimization from data to detection results, solving the problem that traditional deep learning methods are difficult to directly use the original echo data of multi-wavelength pulse radar. They need to use Fourier transform for complex signal processing, which leads to non-adaptive feature extraction, information loss and non-end-to-end processes. Combined with the adaptive learning ability of deep neural networks, it can achieve high-precision, low-computational overhead target detection in complex scenarios, and at the same time can fully capture the target's multi-dimensional features such as amplitude and phase, significantly improving detection performance, while reducing sensitivity to complex scenarios and improving the efficiency and robustness of radar target detection.

[0038] The present invention will be further described below with reference to specific embodiments.

[0039] like Figure 1 As shown, the steps of the multi-wavelength pulse radar target detection method based on deep learning in this embodiment include: Step S01. Data preparation: obtaining radar raw echo data of multi-wavelength pulses under different detection scenarios to form a radar raw echo data set; 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; 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.

[0040] 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.

[0041] 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 figure, during the dataset construction phase, raw echo data from multi-wavelength pulse radars is collected from various detection scenarios. After data cleaning, such as denoising, to improve data quality, the cleaned data undergoes signal processing to obtain the range-Doppler maps required for subsequent data annotation. Target information is then annotated on the processed range-Doppler maps to obtain the annotation information required for network training. The cleaned raw echo data is then converted to a unified format and saved. This data is then fed into the network along with the label information as training data for training. During the model construction and training phase, a target detection network is constructed based on the data characteristics of the multi-wavelength pulse radar echo data. The input to this network is a neural network based on the raw radar echo data. The labeled data and training data obtained during the data preparation phase are fed into the constructed target detection network for network model training. Finally, the model parameters obtained from the trained neural network are saved to obtain the target detection model. During the model deployment phase, the real-time raw echo data collected by the radar undergoes preprocessing (such as normalization) and is then fed into the saved model for inference and prediction, resulting in real-time target detection results.

[0042] In the data preparation stage of step S01, it is necessary to collect raw echo data of the multi-wavelength pulse radar to construct a data set required for model training and testing. In a specific application embodiment, the following steps are included: Step S101. Data acquisition: collecting raw echo data of the multi-wavelength pulse radar under different detection scenarios.

[0043] Specifically, data collection can be carried out within the radar's detection range, collecting multi-wavelength raw echo data of radar targets in all directions, at different heights and distances. The collection process can cover the echo conditions of the target in various motion states, such as static, uniform motion, acceleration / deceleration, and directional movement, to ensure data diversity and representativeness. At the same time, radar equipment can be deployed in different typical complex scenes, such as suburban fields, mountain forests, and urban high-rise buildings. Multiple sets of raw echo data can be collected in each scene to enrich the background environmental data, so that the data set can fully reflect the clutter characteristics and target characteristics in different environments. Furthermore, the target's position information relative to the radar during data collection can be recorded to facilitate the distinction between target and background in subsequent data annotation.

[0044] Step S102. Data cleaning: performing data cleaning on the collected original echo data of the multi-wavelength pulse radar.

[0045] Specifically, the original data can be analyzed to eliminate invalid data such as outliers and noise interference to ensure data quality.

[0046] Step S103: Signal processing: Processing the cleaned original echo data of the multi-wavelength pulse radar to obtain a complete range Doppler map.

[0047] Specifically, if Figure 3 As shown, first, read the cleaned radar raw data to obtain the complex data matrix of the total number of pulse units H and the total number of range units W ( H, W ), which come from different wave positions ( r =1,2,3,..., n , n The radar data of the total number of wave positions are arranged in the order of wave positions. r The number of pulse units in the echo data Different, the number of distance units is W ,in ; Then for different wave positions r The echo data is processed by pulse compression and Fourier transform to generate range Doppler maps at different wave positions. , to reflect the target's distance and Doppler frequency information; 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 ).

[0048] Step S104. Data annotation: perform data annotation on the obtained complete range-Doppler map.

[0049] Specifically, the targets are first marked on the complete range Doppler map obtained by signal processing, and the category label and normalized bounding box information of each target are obtained, including the center point coordinates ( x_center, y_center ) and width and height dimensions ( width, height ) etc. Based on the annotation information, a multi-bit (e.g., 8-bit) single-channel annotated image of the same size as the original range-Doppler spectrum is generated. In this annotated image, background pixel values ​​are set to 0, while each target region is assigned a different pixel value based on its class label (e.g., class 1 can be set to 1, class 2 to 2, and so on). The resulting annotated image, along with the original radar data, will form the training dataset for subsequent deep learning model training.

[0050] In a specific application embodiment, an ellipse-based annotation method can be used. The center of the annotation information is used as the ellipse center, and half the width and height of the target are used as the major and minor axes of the ellipse, respectively. A solid ellipse is drawn at the corresponding location in the annotated image, and the interior of the ellipse is filled with the pixel values ​​corresponding to the target category. This ellipse-based annotation method better matches the diffusion characteristics of radar targets in the range-Doppler domain and more accurately reflects the target's actual energy distribution range than rectangular box annotation.

[0051] Step S105. Data conversion: converting the format of the cleaned raw radar echo data.

[0052] Specifically, by performing unified conversion processing on the original radar echo data, such as converting the cleaned radar echo data into a unified binary or structured data format and storing it using standardized naming conventions, it can support the standardized data processing process of subsequent offline training models.

[0053] In this embodiment, the input preprocessing layer is used to extract multi-wavelength and multi-dimensional features of the radar echo signal by separating the real and imaginary parts of the original radar echo data, and using two linear layer networks with different weight matrices to perform feature mapping on the real and imaginary parts of different wavelength data. The real and imaginary features are calculated according to the feature mapping results, and the phase spectrum features are calculated based on the real and imaginary features. The real features, the imaginary features and the phase spectrum features are used to construct a multi-dimensional feature map output.

[0054] In this embodiment, the input preprocessing layer uses two linear layer networks with different weight matrices 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 Fourier transform basis functions.

[0055] This embodiment uses two linear layer networks to respectively process the real and imaginary parts of different wave position data, and initializes the two linear layer networks in the above manner, which can simulate the process of extracting frequency domain information by Fourier transform, thereby enabling the optimal radar target frequency domain features to be adaptively extracted during the subsequent network training process.

[0056] Specifically, the weight matrices of the two linear layer networks are initialized as discrete approximations of the Fourier transform basis function, including: configuring the weights of the corresponding real weight block and imaginary weight block according to the size and coordinates of each sub-matrix block, where 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 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.

[0057] In the input preprocessing layer of this embodiment, the real part feature of the Fourier transform result can be obtained through complex number operation based on the real part weight matrix and the imaginary part weight matrix. real_feature and imaginary part characteristics imag_feature , the real weight matrix is ​​the weight matrix of the first linear layer network corresponding to the real part data processing, and the imaginary weight matrix is ​​the weight matrix of the second linear layer network corresponding to the imaginary part data processing. The calculation expression is shown as follows:

[0058]

[0059] in, represents the real part weight matrix, represents the imaginary weight matrix, represents the real part signal, Represents the imaginary part signal.

[0060] In this embodiment, according to the extracted real part features real_feature and imaginary part characteristics imag_feature The phase spectrum features can be obtained through the inverse tangent function, and the extracted real part features, imaginary part features and phase spectrum features can be spliced ​​together to construct a multi-dimensional feature map.

[0061] In a specific application embodiment, Figure 4 As shown in the figure, an input preprocessing layer is set at the input end of the target detection network. The input preprocessing layer first extracts multi-wavelength and multi-dimensional features from the input multi-wavelength radar raw echo data, and then passes through the encoding layer, decoding layer, and output layer in sequence to obtain the final detection result. Figure 5 As shown in FIG, the detailed processing flow of the input preprocessing layer includes: first, the radar echo signal usually contains two parts, the real part and the 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 number of pulses of different wave position data in the echo data, and sends the different wave position data to two independent linear layer networks for feature mapping. The weight matrices of the two networks are initialized to the discrete approximation form of the Fourier transform basis function, where the weight matrix is ​​composed of multiple sub-matrix blocks of fixed dimensions, corresponding to the processing of radar echo data at different wave positions, and the element value in each sub-block follows the transformation principle of Fourier transform: for the block size k and coordinates ( i , j ), the real weight block is taken cos ( - 2π ij / k ), the imaginary weight block is sin ( - 2π ij / k) so that the linear layer network can simulate the process of Fourier transform to extract frequency domain information, and adaptively extract the best radar target frequency domain features in the subsequent network training process; after the real and imaginary signals are projected by the customized weight matrix respectively, the real part features of the Fourier transform result are obtained through the complex operation formula real_feature and imaginary part characteristics imag_ feature Finally, based on the extracted real and imaginary features, the phase spectrum features are obtained through the inverse tangent function, and the three feature channels of real part, imaginary part and phase are spliced ​​into a multidimensional feature map to provide further processing for subsequent networks.

[0062] In this embodiment, the target detection network adopts a U-shaped structure, such as Figure 4 As shown in the figure, the network can be structurally divided into five parts: input preprocessing layer, encoding layer, intermediate layer, decoding layer, and output layer. The encoding layer 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 at 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. Each layer of encoders and decoders is connected in a step-by-step stacking manner. Each layer of encoders is connected to a corresponding decoder. Semantic features are further extracted between the encoding layer and the decoding layer through the intermediate layer. By adopting the target detection network with the above structure, pixel-level target detection can be achieved.

[0063] In this embodiment, the encoder layer implements feature extraction, spatial dimension compression, and channel dimension expansion through a multi-layer encoder. Each encoder layer includes a maximum pooling layer and parallel multi-branch convolutional units. Each of these branch convolutional units includes at least one branch convolutional unit for extracting local detail features, at least one branch convolutional unit for implementing cross-channel information exchange and dimensionality transformation, and at least one branch convolutional unit for preserving the identity mapping of the original input feature map. By utilizing a multi-branch convolutional structure, the encoding layer can match the large data volume of multi-wavelength pulse radar echo data, small targets, and the lack of obvious shape, color, texture, and other features, thereby improving the real-time network inference and the accuracy of target feature representation.

[0064] In a specific application embodiment, Figure 6As shown, a single encoder can adopt a four-layer architecture, achieving spatial dimension compression and channel dimension expansion through stacked maximum pooling layers and convolution operations. Each encoder layer consists of a maximum pooling layer and a parallel three-branch convolution unit. The first branch uses 3×3 standard convolution to extract local detail features, the second branch uses 1×1 convolution to achieve cross-channel information interaction and dimensionality transformation, and the third branch retains the identity mapping of the original input feature map. The first encoder layer does not have a maximum pooling layer for feature downsampling, but instead uses 1×1 convolution for channel adjustment of the feature map.

[0065] In a specific application embodiment, the intermediate layer may be composed of a maximum pooling layer and a 3×3 convolutional layer to further extract deeper semantic features between the encoder and the decoder.

[0066] In this embodiment, the decoding layer adopts a multi-layer encoder structure to realize feature fusion of different layers and gradually restore the spatial dimension, while reducing the channel dimension. Each layer of the decoder is upsampling through a bilinear interpolation unit. The feature map obtained after upsampling and the feature map output by the decoder of the previous layer are spliced ​​together through a cross-layer feature splicing unit to realize multi-scale feature splicing. Parallel multi-branch convolution is used to perform feature transformation and channel dimension scale compression on the spliced ​​multi-scale features.

[0067] In a specific application embodiment, Figure 7 As shown, a single decoder adopts a four-layer structure, including a four-layer upsampling structure. c, h, w Represents the number of channels, height, and width of the feature map input to the current decoder. Each decoder layer contains a bilinear interpolation upsampling unit, a cross-layer feature splicing unit, and parallel multi-branch convolution. Upsampling is achieved through the bilinear interpolation unit, restoring the size of the feature map layer by layer. The feature map obtained after upsampling is combined with the feature map output by the decoder encoder of the previous layer through the cross-layer feature splicing unit to achieve multi-scale feature splicing. Parallel multi-branch convolution is also used in each decoder layer to perform feature transformation and channel dimension scale compression on the spliced ​​multi-scale features, achieving step-by-step feature decoding and multi-scale feature fusion.

[0068] In this embodiment, the output layer in the target detection network is used to fuse feature maps of different scales and output the final network prediction result, including 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 feature layers of different scales 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 feature layer of different scales. 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.

[0069] In a specific application embodiment, Figure 8 As shown in the figure, the output layer first predicts the prediction results of the four different scale feature layers output by the output decoder through the convolutional prediction head and adjusts the number of channels. Then, the four different scale prediction results are upsampled through the multi-scale upsampling unit to make the feature size consistent with the original input size. Based on the lightweight channel attention mechanism, the channel features of the input feature map are analyzed and the multi-scale upsampled feature map is weighted to enhance the response of important feature channels and suppress unimportant feature channels. Finally, the weighted multi-scale feature map is fused through the convolutional fusion unit to generate the final prediction result.

[0070] This embodiment adopts the above structure and uses a multi-branch convolution structure and a lightweight channel attention module to improve the traditional Unet target detection model. A backbone feature extraction network is constructed based on the multi-branch convolutional neural network structure. The network can enhance the feature expression capability through a parallel multi-path structure in the model training stage, and can transform the multi-path structure into a single chain topology structure through parameter fusion in the inference stage, which not only ensures the model performance but also significantly improves the computational efficiency, so that the real-time performance of the pulse radar target detection task can be well met, and the real-time performance of the network inference and the accuracy of the target feature representation can be improved; at the same time, combined with the efficient reasoning capability of the parallel multi-branch convolution unit during inference and the lightweight channel attention mechanism, high-precision and low-computational-overhead target detection can be achieved in complex scenarios. At the same time, adaptive learning is used to fully capture the multi-dimensional features of the target such as amplitude and phase, significantly improving the detection performance and robustness.

[0071] In this embodiment, during the model training phase, a constructed training set—that is, the radar's raw echo data and corresponding label information—is fed into a constructed neural network for training. By fitting the model to actual data, a constructed validation sample set is fed into the neural network for verification after each training cycle. The accuracy of the network model is monitored in real time, and the weight parameters of the model with the best accuracy are saved. After training, a constructed test sample set is fed into the neural network for testing, and performance parameters such as the network model's detection rate, false positive rate, and missed detection rate are measured. If the network model meets the expected requirements, a deployable target detection model can be obtained through conversion.

[0072] In this embodiment, during the model deployment stage, the trained radar target detection model is imported into the radar target detection device, and the radar device with the imported detection model is used to monitor the target area, so that the trained model is used to process the multi-wavelength pulse radar raw echo data in real time. After obtaining the real-time raw radar echo data, the obtained radar raw echo data is normalized and preprocessed, and the preprocessed radar echo data is sent to the target detection model for prediction and reasoning to obtain the final target detection result.

[0073] Furthermore, the prediction results of the target detection model can be post-processed. For example, the network output can be binarized by setting a threshold to distinguish the target from the background area, and the connected domain detection algorithm can be used to extract the target connected area and determine its center coordinates. These pixel coordinates are converted into actual distance and speed physical quantities according to the radar's range unit and pulse unit parameters, and finally the distance-speed information of each target is output. The target's distance and speed information can be obtained, realizing efficient target detection in complex scenarios.

[0074] In view of the characteristics of multi-wavelength pulse radar data, the present invention adopts two independent linear layer networks to process the real and imaginary parts of different wavelength data respectively. It can directly extract frequency domain features of the original multi-wavelength pulse radar echo data and can be collaboratively optimized with the downstream network through end-to-end training.

[0075] This embodiment further provides an electronic device, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0076] It is understandable that the above method of this embodiment can be executed by a single device, such as a computer or server, etc., and can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps in the above method of this embodiment, and 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, etc., for executing relevant programs to implement the above method of this 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. When the above method of this embodiment is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0077] This embodiment further provides a computer-readable storage medium storing a computer program, which implements the above method when executed by a processor.

[0078] Those skilled in the art will appreciate that the above-described embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the functions described in the processes. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including the instruction device, which implements the function specified in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0079] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed above with reference to the preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent variations, and modifications to the above embodiment that do not depart from the technical solution of the present invention and are based on the technical essence of the present invention shall fall within the scope of protection of the technical solution of the present invention.

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