Unmanned aerial vehicle radar detection and identification method and system based on background contrast attention and double-flow micro-Doppler fusion
By combining background contrast attention with dual-stream micro-Doppler fusion, the problem of radar detection and identification of low-altitude, slow, and small UAVs in low-altitude environments was solved, achieving efficient and stable target detection and identification, meeting the requirements of real-time performance and high precision.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-19
AI Technical Summary
Existing radar technology has difficulty effectively detecting and identifying low-altitude, slow, and small UAVs in low-altitude environments. In particular, it suffers from low detection rate and high false alarm rate under conditions of low signal-to-noise ratio and complex clutter. Furthermore, the micro-Doppler feature is sensitive to the observation angle, resulting in poor robustness and overfitting due to scarce samples.
A background contrast attention and dual-stream micro-Doppler fusion method is adopted. The target and background are separated by a fully convolutional detection network BCA-FCN and a random mask multi-scale background learner RMBL to generate a time-Doppler map (TDM) and a time-cepstrum map (TCM). The Transformer network is combined to perform feature fusion and few-shot learning to achieve rapid target detection and robust recognition.
It significantly reduces the false alarm rate, improves the detection rate, enhances recognition stability and real-time performance in complex low-altitude environments, meets the real-time requirements of low-altitude monitoring, and achieves high-precision classification under conditions of few samples.
Smart Images
Figure CN122063584A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar signal processing and intelligent target recognition technology, and relates to a method and system for UAV radar detection and recognition based on background contrast attention and dual-stream micro-Doppler fusion. Background Technology
[0002] With the rapid development of the low-altitude economy, unmanned aerial vehicles (UAVs) are increasingly widely used in logistics, agricultural protection, and aerial surveying. However, non-cooperative UAVs, characterized by their low-altitude, slow, and small (LSS) characteristics, also pose a serious challenge to low-altitude airspace security. Radar, as an all-weather, all-day monitoring tool, is the core sensor of counter-UAV systems. However, in practical applications, UAVs have small radar cross-sections (RCS), weak echoes, and often fly in complex low-altitude environments such as urban buildings, trees, and sea clutter, making targets easily obscured by strong clutter and difficult to detect.
[0003] Existing radar target detection methods are mainly divided into the traditional Constant False Alarm Rate (CFAR) method and deep learning methods. The CFAR method relies on statistical modeling of background clutter, and its performance degrades significantly in non-uniform clutter environments. While deep learning-based methods improve feature extraction capabilities, existing high-performance networks such as Faster R-CNN typically employ a sliding window mechanism to transform the detection problem into a classification problem, resulting in high computational redundancy and difficulty in meeting real-time requirements. Furthermore, to address the missed detection problem under low signal-to-noise ratio (LowSCR) conditions, although some studies have proposed using attention mechanisms, most of these ignore the fundamental differences in feature distribution between the target and the background, failing to effectively suppress background interference.
[0004] After target detection, accurately identifying the type of UAV, such as quadcopters, fixed-wing drones, and bionic birds, is crucial. The micro-Doppler (mD) effect is a key characteristic for UAV identification. However, existing methods based on Time-Doppler Maps (TDM) have two drawbacks: first, TDM features are sensitive to the observation angle; when the radar viewpoint changes, the micro-Doppler features deform, reducing robustness; second, radar data acquisition is difficult, leading to a scarcity of labeled samples (the few-sample problem), making traditional deep learning models prone to overfitting. Therefore, how to achieve rapid detection and robust identification of UAVs under complex conditions of high interference, few samples, and multiple viewpoints remains to be studied. Summary of the Invention
[0005] In view of this, the purpose of the present invention is to provide a method and system for detecting and identifying unmanned aerial vehicles (UAVs) using radar based on background contrast attention and dual-stream micro-Doppler fusion.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A method for UAV radar detection and identification based on background contrast attention and dual-stream micro-Doppler fusion includes the following steps: S1: Perform pulse compression and coherent accumulation on the pulse Doppler signal received by the radar to construct a range-Doppler diagram (RD diagram) as the input to the detection network; S2: Construct a fully convolutional detection network (BCA-FCN) integrating background contrast attention mechanism. The network includes a random mask multi-scale background learner (RML), which learns the background clutter distribution features in the RD map through random masks, and uses the background contrast attention (BCA) module to calculate the difference between the target and the background at the feature level, outputting a classification heatmap containing the probability of the target's presence and a regression map with refined coordinate parameters. S3: Based on the detection and localization results of S2, extract the fast-time and slow-time data segments where the target is located, and perform short-time fourier transform (STFT) and short-time cepstral transform (STCT) respectively to generate a time-Doppler map (TDM) containing the scintillation features of the rotating blades and a time-cepstrum map (TCM) containing the rotation frequency features. S4: Construct a Transformer-based dual-stream micro-Doppler feature fusion recognition network, map TDM and TCM into feature vectors, extract complementary features using a tensor outer product fusion module, enhance long-range dependencies through a multi-head self-attention mechanism, and finally calculate the distance relationship score with the support set samples under the Few-Shot Learning (FSL) framework to determine the UAV target type.
[0008] Furthermore, in S2, the Random Mask Multi-Scale Background Learner (RMBL) divides the input RD map into blocks and applies random masks, forcing the network to focus on the background texture of non-target regions; the loss function of the fully convolutional detection network introduces Jensen-Shannon divergence (JS divergence) to maximize the difference between the target feature distribution and the background feature distribution. The JS divergence loss function is defined as:
[0009] in, This represents the target feature vector after processing by the BCA module. This represents the background feature vector extracted by RMBL. For normalization function, It is a non-negative constant.
[0010] Furthermore, in S2, the feature enhancement process of the background contrast attention (BCA) module is as follows: First, the downsampled target features are fused through a splicing operation. and expanded background features Obtain combined features :
[0011] Pixel-level max pooling and convolution are used to generate intermediate features after background suppression. :
[0012] The final generated target attention-weighted features for:
[0013] in, For the original target features, It is the Sigmoid activation function. This represents element-wise dot product.
[0014] Furthermore, in S2, the fully convolutional detection network uses a regression branch to predict the fine location of the target within the grid cell, and the regression loss function... Defined as:
[0015] in, A set of grids containing the target. This is the offset predicted by the network. This is the actual offset; actual offset The calculation formula is:
[0016] in, For the target's true coordinates, The coordinates of the top left corner of the grid cell. The width and height of the grid cell.
[0017] Furthermore, in S3, the generation process of the time-cepstrum (TCM) includes: For discrete echo signals Short-time Fourier transform is performed to obtain Then a short-time cepstral transform is performed. The formula is:
[0018] in, Indicates the inverse Fourier transform. For time indexing, This is an inverted frequency index.
[0019] Furthermore, in S4, the dual-stream feature fusion employs a fusion module based on tensor outer product, assuming... and These are the TDM and TCM feature vectors after convolutional embedding, respectively, after passing through a linear mapping layer. After dimensionality reduction, we get and fusion feature tensor The calculation is as follows:
[0020] in, This represents the tensor outer product operation, and the fused features. Input into the Transformer module for relation learning.
[0021] Furthermore, in S4, the recognition network employs a few-shot learning framework, which involves constructing a support set. and query set The mean squared error is used as the objective function for training.
[0022] in, The similarity score output by the relationship module. and These are the real labels for the support set and query set samples, respectively. For indicator functions, and These are the network parameters for the feature extraction module and the relation module, respectively.
[0023] Furthermore, the post-processing stage of S2 employs a minimum deviation localization algorithm to calculate the regression deviation magnitude of each unit within the candidate connected components selected by the initial probability threshold. The unit with the smallest deviation is selected as the final positioning center of the target.
[0024] A UAV radar detection and identification system based on background contrast attention and dual-stream micro-Doppler fusion includes: The signal preprocessing module is used to perform pulse compression and coherent accumulation on the pulse Doppler signal received by the radar, and construct the range-Doppler map (RD map) as the input of the detection network. The detection network module is used to construct a fully convolutional detection network BCA-FCN with an integrated background contrast attention mechanism. The network includes a random mask multi-scale background learner RMBL and a background contrast attention BCA module, and outputs a classification heatmap and a regression map. The feature generation module is used to extract target data segments based on the detection and localization results, perform short-time Fourier transform (STFT) and short-time cepstral transform (STCT), and generate time-Doppler graph (TDM) and time-cepstral graph (TCM). The identification network module is used to construct a Transformer-based dual-stream micro-Doppler feature fusion identification network, which determines the type of UAV target through tensor outer product fusion and few-shot learning framework.
[0025] Furthermore, the fully convolutional detection network of the detection network module uses regression branches to predict the fine position of the target within the grid cells, and employs a minimum deviation localization algorithm for post-processing.
[0026] The beneficial effects of this invention are as follows: (1) For low signal-to-noise ratio and strong clutter environments, this invention achieves effective separation of the target and background at the feature level through a random mask multi-scale background learner (RMBL) and a background contrast attention (BCA) module. The RMBL forces the network to learn the background clutter distribution, while the BCA module calculates the difference between the target and the background through feature enhancement, thereby significantly reducing the false alarm rate and improving the detection rate. For example, the introduction of the JS divergence loss function maximizes the distribution difference between the two types of features, ensuring stable detection of weak targets in complex low-altitude environments.
[0027] (2) By generating a time-cepstrum (TCM) and combining it with a time-Doppler image (TDM), this invention overcomes the sensitivity of micro-Doppler features to the observation angle by using dual-stream feature fusion. The rotational frequency feature provided by the TCM has angle invariance and complements the blade scintillation feature of the TDM. Furthermore, by learning the relationship through tensor outer product and Transformer network, the stability of the recognition model under the change of viewing angle is significantly improved.
[0028] (3) The fully convolutional detection network BCA-FCN adopts full-image inference and grid regression mechanism, which avoids the computational redundancy of traditional sliding window. Combined with the minimum deviation localization MDP algorithm for post-processing, the detection speed is increased by tens of times compared with traditional methods, which can meet the real-time requirements of low-altitude monitoring scenarios.
[0029] (4) Identify the network embedding of a few-shot learning framework. By calculating the relationship between the support set and the query set, the model can still achieve high-precision classification under the condition of scarce labeled samples. This effectively solves the overfitting problem caused by the difficulty of radar data acquisition and improves the practicality and generalization ability of the method.
[0030] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the overall process of the method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the BCA-FCN detection network in an embodiment of the present invention; Figure 3 This is a schematic diagram of the background contrast attention (BCA) module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dual-stream feature fusion recognition network structure based on Transformer in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the principle of grid regression and minimum deviation localization method (MDP) in an embodiment of the present invention. Detailed Implementation
[0032] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0033] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0034] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0035] This invention proposes a method and system for UAV radar detection and recognition based on background contrast attention and dual-stream micro-Doppler fusion. It can solve the problems of low detection rate and high false alarm rate of low, slow, small (LSS) targets in low signal-to-noise ratio environments, as well as poor target recognition robustness under conditions of large changes in radar observation angle and scarce samples.
[0036] like Figure 1 As shown in the figure, this embodiment provides a method for UAV radar detection and identification based on background contrast attention and dual-stream micro-Doppler fusion. The specific steps are as follows: S1: Radar signal preprocessing and range-Doppler image construction First, the pulse Doppler signal received by the radar is preprocessed. A linear frequency modulated signal is used as the transmitted waveform. After mixing and low-pass filtering, the received signal undergoes pulse compression to improve range resolution. Matched filtering is used for pulse compression. The compressed signal is represented as follows: Next, regarding Coherent accumulation is performed, and the Doppler frequency shift is calculated along the slow time dimension using Fast Fourier Transform to generate a range-Doppler map (RD map). The horizontal axis of the RD map represents the range gate, and the vertical axis represents the Doppler frequency; each pixel value corresponds to the signal energy intensity. This RD map serves as the input to the detection network, and its size can be set to 256×256 pixels to balance computational efficiency and resolution. The preprocessing stage must ensure optimized signal-to-noise ratio to provide a clean data source for subsequent detection.
[0037] S2: Construction and Training of the Fully Convolutional Detection Network BCA-FCN Figure 2This is a schematic diagram of the overall architecture of the BCA-FCN detection network. This step enables rapid target localization. The network structure is based on a fully convolutional architecture, avoiding computational redundancy associated with sliding windows. The network input is the RD map generated by S1, and the output includes a classification heatmap (target presence probability) and a regression map (coordinate offset parameters).
[0038] Random Masked Multi-Scale Background Learner (RMBL): The input RD image is divided into blocks of 32×32 pixels, and a random mask (30% masking rate) is applied to force the network to learn background clutter texture. RMBL extracts background features through multi-scale convolutional layers. The scales include 3×3, 5×5 and 7×7 convolution kernels to capture different ranges of background statistical properties.
[0039] Figure 3 This is a schematic diagram of the Background Contrast Attention (BCA) module. The BCA module enhances target features and suppresses background interference. Specifically, it downsamples target features. With extended background features spliced as Intermediate features are generated through convolution and pixel-level max pooling. Finally, the original target features are weighted by upsampling and Sigmoid activation. The formula is:
[0040]
[0041]
[0042] This process effectively maximizes the JS divergence between the target and the background, and the loss function... constants in C Set it to 1.0 for stable training.
[0043] Regression branch and loss function: The regression branch predicts the fine offset of the target within the grid cell, and the loss function... Using Smooth L1 loss, the true offset is calculated as follows: The network was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 16.
[0044] S3: Micro-Doppler Feature Generation – Construction of TDM and TCM Based on the detection results of S2, a fast-time-slow-time data segment containing the target is extracted. The segment length is set to 256 pulses with an overlap rate of 50% to balance temporal resolution and continuity.
[0045] Time-Doppler Map (TDM) Generation: Short-Time Fourier Transform (STFT) was applied to the segment data, with a Hamming window function and a window length of 64 points to generate the TDM. The TDM reflects the micro-Doppler characteristics of blade scintillation, but it is significantly affected by the observation angle.
[0046] Time-Cepstral Transform (TCM) Generation: To enhance robustness, a short-time cepstral transform (STCT) is performed on the STFT results. The formula is as follows: ,in This is an inverse Fourier transform. The TCM extracts rotational frequency features independent of angle, improving recognition stability. The generated TDM and TCM are both 128×128 pixels in size and used as input to the recognition network.
[0047] S4: Execution of the Two-Stream Feature Fusion Recognition Network The recognition network is based on the Transformer architecture and is designed to handle few-shot learning tasks.
[0048] Feature extraction and fusion: TDM and TCM are input into the convolutional embedding layer respectively, and output feature vectors. and After dimensionality reduction via a linear mapping layer, the fusion is achieved through tensor outer product. The outer product operation captures complementary information between modes, and the dimension of the fused tensor Z is 256.
[0049] Transformer relation learning: Z is input to a Transformer encoder with 8 heads and 512 hidden layer dimensions. Long-range dependencies are modeled using a multi-head self-attention mechanism. The output features are processed by the relation module to calculate similarity scores.
[0050] Few-shot learning framework: The support set S contains 5 classes of drones, with 5 samples from each class, and the query set Q contains the samples to be identified. The training objective function is the mean squared error.
[0051] in, The similarity score output by the relationship module. and These are the real labels for the support set and query set samples, respectively. For indicator functions, and These are the network parameters for the feature extraction module and the relationship module, respectively. During inference, the category with the highest score is selected as the recognition result.
[0052] Post-processing and positioning optimization In the post-processing stage of S2, a minimum deviation localization algorithm is used to improve localization accuracy. For candidate connected components obtained by thresholding the classification heatmap (threshold 0.5), the regression deviation magnitude of each grid cell is calculated. ,choose dev The smallest unit is used as the final center of the target. This process reduces positioning errors, and is especially suitable for low signal-to-noise ratio scenarios.
[0053] In implementing this invention, GPU acceleration for training and inference is required, such as an NVIDIA Tesla V100 graphics card. In actual testing, the entire system achieved a detection frame rate exceeding 30fps, and the recognition accuracy reached over 95% under limited sample conditions, verifying its practicality. Those skilled in the art can adjust the above details according to specific radar parameters, without departing from the core concept of this invention.
[0054] Furthermore, in S4, the dual-stream feature fusion employs a fusion module based on tensor outer product. Let... and These are the TDM and TCM feature vectors after convolutional embedding, respectively, after passing through a linear mapping layer. After dimensionality reduction, we get and fusion feature tensor The calculation is as follows:
[0055] in, This represents the tensor outer product operation, and the fused features. Input into the Transformer module for relation learning. Figure 4 This is a schematic diagram of a Transformer-based dual-stream feature fusion recognition network structure.
[0056] Furthermore, in S4, the recognition network employs a few-shot learning framework by constructing a support set. and query set The mean squared error (MSE) is used as the objective function for training.
[0057] in, The similarity score output by the relationship module. and These are the real labels for the support set and query set samples, respectively. For indicator functions, and These are the network parameters for the feature extraction module and the relation module, respectively.
[0058] Furthermore, in the post-processing stage of S2, a grid-based regression prediction and minimum deviation localization (MDP) algorithm is employed. The network regression branch outputs the offset of the target relative to the grid cells, and the regression deviation magnitude of each cell within the candidate connected domain is calculated. The unit with the smallest deviation is selected as the final positioning center of the target to improve positioning accuracy. Figure 5 This is a schematic diagram illustrating the principle of MDP based on grid regression and minimum deviation localization.
[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for detecting and identifying unmanned aerial vehicles (UAVs) using radar based on background contrast attention and dual-stream micro-Doppler fusion, characterized in that: Includes the following steps: S1: Perform pulse compression and coherent accumulation on the pulse Doppler signal received by the radar to construct a range-Doppler RD map as the input to the detection network; S2: Construct a fully convolutional detection network BCA-FCN with integrated background contrast attention mechanism. The network includes a random mask multi-scale background learner RMBL, which learns the background clutter distribution features in the RD map through random masking, and uses the background contrast attention BCA module to calculate the difference between the target and the background at the feature level, outputting a classification heatmap containing the probability of the target's presence and a regression map with refined coordinate parameters. S3: Based on the detection and localization results of S2, the fast-time-slow-time data segment where the target is located is extracted, and short-time Fourier transform (STFT) and short-time cepstral transform (STCT) are performed respectively to generate time-Doppler map (TDM) containing the scintillation features of the rotating blade and time-cepstral map (TCM) containing the rotation frequency features. S4: Construct a Transformer-based dual-stream micro-Doppler feature fusion recognition network, map TDM and TCM into feature vectors, extract complementary features using a tensor outer product fusion module, enhance long-range dependencies through a multi-head self-attention mechanism, and finally calculate the distance relationship score with the support set samples under the few-shot learning FSL framework to determine the UAV target type.
2. The UAV radar detection and identification method based on background contrast attention and dual-stream micro-Doppler fusion as described in claim 1, characterized in that: In S2, the Random Mask Multi-Scale Background Learner (RMBL) divides the input RD map into blocks and applies random masks, forcing the network to focus on the background texture of non-target regions. The loss function of the fully convolutional detection network introduces JS divergence to maximize the difference between the target feature distribution and the background feature distribution. The JS divergence loss function is defined as follows: in, This represents the target feature vector after processing by the BCA module. This represents the background feature vector extracted by RMBL. For normalization function, It is a non-negative constant.
3. The UAV radar detection and identification method based on background contrast attention and dual-stream micro-Doppler fusion as described in claim 1, characterized in that: In S2, the feature enhancement process of the background contrast attention (BCA) module is as follows: First, the downsampled target features are fused through a splicing operation. and expanded background features Obtain combined features : Pixel-level max pooling and convolution are used to generate intermediate features after background suppression. : The final generated target attention-weighted features for: in, For the original target features, It is the Sigmoid activation function. This represents element-wise dot product.
4. The UAV radar detection and identification method based on background contrast attention and dual-stream micro-Doppler fusion as described in claim 1, characterized in that: In S2, the fully convolutional detection network uses a regression branch to predict the fine location of the target within the grid cell, and the regression loss function... Defined as: in, A set of grids containing the target. This is the offset predicted by the network. This is the actual offset; actual offset The calculation formula is: in, For the target's true coordinates, The coordinates of the top left corner of the grid cell. The width and height of the grid cell.
5. The UAV radar detection and identification method based on background contrast attention and dual-stream micro-Doppler fusion as described in claim 1, characterized in that: In S3, the generation process of the time-cepstrum (TCM) includes: For discrete echo signals Short-time Fourier transform is performed to obtain Then a short-time cepstral transform is performed. The formula is: in, Indicates the inverse Fourier transform. For time indexing, This is an inverted frequency index.
6. The UAV radar detection and identification method based on background contrast attention and dual-stream micro-Doppler fusion according to claim 1, characterized in that: In S4, the dual-stream feature fusion employs a fusion module based on tensor outer product, assuming... and These are the TDM and TCM feature vectors after convolutional embedding, respectively, after passing through a linear mapping layer. After dimensionality reduction, we get and fusion feature tensor The calculation is as follows: in, This represents the tensor outer product operation, and the fused features. Input into the Transformer module for relation learning.
7. The UAV radar detection and identification method based on background contrast attention and dual-stream micro-Doppler fusion according to claim 1, characterized in that: In S4, the recognition network employs a few-shot learning framework, which involves constructing a support set. and query set The mean squared error is used as the objective function for training. in, The similarity score output by the relationship module. and These are the real labels for the support set and query set samples, respectively. For indicator functions, and These are the network parameters for the feature extraction module and the relation module, respectively.
8. The UAV radar detection and identification method based on background contrast attention and dual-stream micro-Doppler fusion as described in claim 1, characterized in that: The post-processing stage of S2 employs a minimum deviation localization algorithm, calculating the regression deviation magnitude of each unit within the candidate connected components selected by the initial probability threshold. The unit with the smallest deviation is selected as the final positioning center of the target.
9. A UAV radar detection and identification system based on background contrast attention and dual-stream micro-Doppler fusion, characterized in that: include: The signal preprocessing module is used to perform pulse compression and coherent accumulation on the pulse Doppler signal received by the radar, and construct the range-Doppler map (RD map) as the input of the detection network. The detection network module is used to construct a fully convolutional detection network BCA-FCN with an integrated background contrast attention mechanism. The network includes a random mask multi-scale background learner RMBL and a background contrast attention BCA module, and outputs a classification heatmap and a regression map. The feature generation module is used to extract target data segments based on the detection and localization results, perform short-time Fourier transform (STFT) and short-time cepstral transform (STCT), and generate time-Doppler graph (TDM) and time-cepstral graph (TCM). The recognition network module is used to construct a Transformer-based dual-stream micro-Doppler feature fusion recognition network, which determines the type of UAV target through tensor outer product fusion and few-shot learning framework.
10. The UAV radar detection and identification system based on background contrast attention and dual-stream micro-Doppler fusion as described in claim 9, characterized in that: The fully convolutional detection network of the detection network module uses regression branches to predict the fine position of the target within the grid cells, and employs a minimum deviation localization algorithm for post-processing.