Radar target detection method based on multi-scale deep learning under strong maneuver
Through multi-scale deep learning model and feature fusion technology, the problem of limited radar target detection accuracy under strong maneuverability is solved, and efficient, accurate and robust target detection is achieved in a sea-clutter interference environment.
Patent Information
- Application Number
- CN202411001040.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the case of strong maneuvering, the accuracy of the radar target detection algorithm is limited, which is mainly due to insufficient target feature extraction caused by the spatial offset of radar electromagnetic wave transmission and reception and sea clutter interference, resulting in insufficient detection accuracy and robustness.
Using a radar target detection model based on multi-scale deep learning, the dependence between historical frames and current frame echo images is analyzed through attention mechanisms and feature fusion mechanisms, spatial variation characteristics, target orientation variation characteristics and global spatial and temporal variation characteristics are extracted, and spatial variation characteristics are filtered. To remove clutter information, use the Swin-Spatial Transformer network and the Conv-LSTM network to extract features and output multi-level and multi-scale features to improve detection accuracy and robustness.
Under strong maneuvering, the accuracy and robustness of radar target detection can be improved, and the target distance and direction information can be detected more accurately, reducing the error detection rate.
Smart Images

Figure CN118915007A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of radar target detection, and more specifically, relates to a method for processing navigation radar echo data to detect and extract target distance and direction information. Background Art
[0002] Radar target detection technology is the core function of the radar system. Its main task is to determine whether there is a target in the echo signal and obtain the distance and direction information of the target. Traditional radar target detection methods mainly rely on the processing and analysis of the original radar echo data stream, which is easily affected by lighting, occlusion, weather, clutter, target posture, etc., and have poor adaptability to complex environments, and their robustness and detection accuracy are limited.
[0003] With the improvement of radar image quality and the advancement of artificial intelligence technology, it has become a new trend to use deep learning methods to process radar image data to achieve radar target detection. Deep learning methods have powerful complex nonlinear mapping fitting capabilities and can extract useful information from data through hierarchical features. This method can automatically learn how to effectively use the information in the data according to different task objectives, avoid performance degradation caused by mismatch of the preset model by learning mapping from the data, and at the same time, due to the ease of parallel implementation, computational efficiency is also guaranteed.
[0004] Although some progress has been made in the application of deep learning in radar target detection, the accuracy of target detection algorithms is usually limited when the radar carrier (such as an unmanned boat) is in a rotational maneuver or the detected target is in a strong maneuvering state when moving at high speed. This is mainly caused by the following two factors:
[0005] (1) Strong maneuvers will affect the transmission and reception of radar electromagnetic waves, causing the echo image to shift in space. Existing radar target detection methods based on deep learning rely on a single echo image for training and detection, which cannot describe the spatial variation characteristics caused by the radar carrier and target maneuvers. Therefore, when the radar carrier rotates or the detected target moves at high speed, it is difficult to ensure the target detection accuracy.
[0006] (2) The interference of sea clutter will lead to insufficient extraction of target features, resulting in a large deviation in the target orientation output by the detection algorithm and a certain false detection rate.
[0007] Therefore, how to further improve the performance of deep learning-based radar target detection technology under strong maneuvering conditions remains an urgent problem to be solved. Summary of the invention
[0008] Aiming at the problem of limited radar target detection accuracy caused by spatial offset of radar echo images and sea clutter interference in strong maneuvers, a radar target detection model based on multi-scale deep learning was constructed. The dependency relationship between historical frame and current frame echo images was analyzed to extract spatial variation features, target azimuth variation features and global spatiotemporal variation features. The attention mechanism was used to extract effective target features while discarding invalid features and clutter information, thereby achieving efficient, accurate and robust detection of radar targets in strong maneuvers.
[0009] To achieve the above object, the present invention provides a radar target detection method based on multi-scale deep learning under strong maneuvers, comprising:
[0010] The original radar echo data is filtered and processed into an echo image as the current frame echo image, the clutter noise in the current frame echo image is filtered out, and the current frame echo image is normalized into an echo image of the same scale;
[0011] The feature map of the echo image of the current frame is extracted by using the Swin-Spatial Transformer network equipped with an attention mechanism. On the one hand, the feature map of the echo image of the current frame is transmitted to the spatiotemporal feature fusion module. On the other hand, the feature map of the echo image of the current frame is processed by the Faster R-CNN network to generate the spatial features of the region of interest after preliminary correction.
[0012] The time and space features of the echo images of the historical n frames are extracted through the Conv-LSTM network, and are fused with other extracted scale features. The deep features containing the time and space variation features are output through the subsequent modules.
[0013] The multi-dimensional deep features generated by fusion are used to obtain the multi-level and multi-scale features of the radar echo image, and the target direction information is output through a fully connected network.
[0014] In some optional implementation schemes, filtering the original radar echo data and processing it into an echo image as the current frame echo image, filtering out clutter noise in the current frame echo image, and normalizing the current frame echo image into an echo image of the same scale, includes:
[0015] For each complete original echo data, the redundant edge information is removed and converted into an original echo image of size (NH, NW), and the original echo image of size (NH, NW) is reduced and adjusted to (H, W), where (NH, NW) represents the height and width of the original echo image, and (H, W) represents the height and width of the adjusted image, respectively;
[0016] The image of (H, W) size is clutter-filtered to obtain the current frame echo image. Each time the current frame echo image is generated, it is stored in a buffer with a capacity of n and is called as a historical echo image.
[0017] In some optional implementation schemes, the feature map of the echo image of the current frame is extracted by using the Swin-SpatialTransformer network equipped with an attention mechanism, and the feature map of the echo image of the current frame is transmitted to the spatiotemporal feature fusion module on the one hand, and the feature map of the echo image of the current frame is processed by the Faster R-CNN network on the other hand to generate the spatial features of the region of interest after preliminary correction, including:
[0018] Keep the channel information of each current frame echo image received by the improved Faster R-CNN module as C=1;
[0019] The Swin-Spatial Transformer network is used to extract key features and spatial features from the current frame echo image to form a feature map F1;
[0020] The feature map F1 output by the Swin-Spatial Transformer network is passed to the spatiotemporal feature fusion module and the Faster R-CNN network at the same time. The Faster R-CNN network extracts deep overall features from the feature map F1 to form a feature map F3 with a size of (1, h).
[0021] In some optional implementation schemes, the method of extracting key features and spatial features from the current frame echo image using the Swin-Spatial Transformer network to form a feature map F1 includes:
[0022] The Swin-Spatial Transformer network is designed to replace the convolutional layer in the front-end of the Faster R-CNN network to extract features and form the feature map F1.
[0023] In some optional implementation schemes, the Swin-Spatial Transformer network includes: a multi-layer perceptron MLP, a normalization layer LP, a convolutional neural network CNN, a fully connected layer FC, a window multi-head self-attention W-MSA, a moving window multi-head self-attention SW-MSA and a spatial transformer network STN, wherein the STN acts once before and after the attention mechanism acts.
[0024] In some optional implementation schemes, the input and output of STN are both feature maps. Assuming the input shape is (C, H, W) and the output shape is (C, H′, W′), the output feature map V has a feature V in each channel C.i c Expressed as a function of the input feature map U on the neural network parameters: in, represents the feature of the input feature map U at position (m,n) in channel C, k(·) represents a linear interpolation method, Φ x ,Φ y represents the parameters of the linear interpolation method. To ensure that the network can perform back propagation, k(·) here uses bilinear interpolation. H′, W′ are the height and width of the output feature map. It is expressed as: Represents the affine transformation matrix, which is the output of the localization network (Localisation Net); are the coordinate points of the output feature map.
[0025] In some optional implementation schemes, the temporal and spatial features of the echo images of the historical n frames are extracted through the Conv-LSTM network, and are fused with other extracted scale features, and the deep features containing the temporal and spatial variation features are output through the subsequent modules, including:
[0026] The Conv-LSTM network is used to extract spatiotemporal features from the historical frame echo data with the shape of (T, C, W, H) and generate a shape of (1, C f ,H f ,W f ), wherein C=1, and T=n represents the number of input frames of the historical echo image;
[0027] The feature map F2 is concatenated and fused with the feature map F1 output by the Swin-Spatial Transformer network;
[0028] A residual block is used to reduce the number of channels of the concatenated fusion features to 1, and a maximum pooling layer is used to process the output features of the residual block to form a feature map F4.
[0029] In some optional implementation schemes, the multi-dimensional deep features generated by the fusion obtain multi-level and multi-scale features of the radar echo image, and output the target position information through a fully connected network, including:
[0030] The feature maps F3 and F4 are fused to generate multi-scale and multi-level deep features. The feature shape is (1, H f W f / 4+h);
[0031] The feature map F4 is processed using a fully connected layer network to obtain the spatial position of the target proposal box, where each proposal box is represented as a four-tuple (X i ,Yi ,h i ,w i ), where (X i ,Y i ) represents the coordinates of the center point of the suggestion box, h i and w i Respectively represent the height and width of the suggestion box;
[0032] According to the target's azimuth coordinates (X i ,Y i ),Depend on Calculate the distance D between the target and the radar carrier i ,Depend on Calculate the direction
[0033] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:
[0034] The present invention can improve the radar target detection accuracy and detection robustness.
[0035] (1) Improve detection accuracy
[0036] For high-maneuver situations, the accuracy of radar target detection can be improved by automatically learning complex patterns in radar echo data through deep learning models.
[0037] (2) Improve detection robustness
[0038] Through the attention mechanism and feature fusion mechanism, the robustness of radar target detection facing sea clutter and noise can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 This is a general framework diagram of a radar target detection model provided by an embodiment of the present invention: the functional module composition of the model and its data flow;
[0040] Figure 2 It is a data preprocessing process provided by an embodiment of the present invention;
[0041] Figure 3 It is an internal structure of a Swin-Spatial Transformer network provided by an embodiment of the present invention;
[0042] Figure 4 It is a feature change diagram of a spatiotemporal fusion module provided by an embodiment of the present invention: multi-scale feature fusion in two dimensions of time and space;
[0043] Figure 5 is a radar scanning schematic diagram provided by an embodiment of the present invention;
[0044] Figure 6 It is an example of an original echo image provided by an embodiment of the present invention: an example of an original echo image containing clutter generated by scanning;
[0045] Figure 7 This is an example of a radar target detection result provided by an embodiment of the present invention: the detection result is represented by a rectangular frame, which includes target distance and azimuth information. DETAILED DESCRIPTION
[0046] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0047] The present invention comprises four functional modules, namely, data preprocessing module, improved Faster R-CNN module, spatiotemporal feature fusion module, and target orientation output module. Figure 1 shown.
[0048] (1) Data preprocessing module
[0049] The original radar echo data is simply filtered and processed into an echo image (called the current frame echo image) to filter out obvious clutter noise in the echo data to prevent excessive interference with the subsequent deep neural network feature extraction. This process also ensures that the echo image is normalized to the same scale.
[0050] (2) Improve the Faster R-CNN module
[0051] Receive the current frame echo image, use the Swin-Spatial Transformer network equipped with an attention mechanism to extract the feature map of the echo image, and then on the one hand transmit the feature map of the current frame echo image to the spatiotemporal feature fusion module, and on the other hand, process it through the Faster R-CNN network to generate the preliminary corrected region of interest features (ROI features).
[0052] (3) Spatiotemporal feature fusion module
[0053] Input historical n frames of echo images, extract the temporal and spatial features of the historical echo images through the Conv-LSTM network, and fuse them with other extracted scale features. The subsequent modules output deep features containing temporal and spatial variation characteristics.
[0054] (4) Target direction output module
[0055] By integrating the multi-dimensional features generated by the above modules, the multi-level and multi-scale features of the radar echo image are obtained, and the target orientation information is output through the fully connected layer (Full Connection) network.
[0056] Furthermore, the specific implementation process of the data preprocessing module includes:
[0057] like Figure 2 As shown, the collected raw echo data is first processed into a two-dimensional echo image as the current frame echo image, and its image size is adjusted to a standard size to unify the model input.
[0058] For each complete raw echo data, its size may be inconsistent. Its redundant edge information is removed to convert it into a raw echo image of size (NH, NW). Next, in order to ensure the uniformity of the input size of each model and to reduce the computational complexity of the model, the image size is reduced to (H, W). Clutter filtering is performed after the image size is adjusted, so that clutter filtering can be performed on a smaller image size to reduce the amount of filtering calculations. Each echo image generated is stored in a buffer with a capacity of n at the same time so that it can be called as a historical echo image. For the first echo image generated, its historical echo image is itself.
[0059] Furthermore, improving the Faster R-CNN module specifically includes the following steps:
[0060] ① Wave image preprocessing.
[0061] Improved Faster R-CNN Each current frame echo image received by the module usually has channel information, so the data shape input to the module each time is (C, H, W), where C represents the channel. For radar echo images, the image channel information is not important, so its channel always keeps C = 1 to reduce the computational complexity and improve the model operation efficiency.
[0062] ② Use the Swin-Spatial Transformer network to extract key features and spatial features from the current frame echo image to form a feature map F1.
[0063] The Faster R-CNN network has second-order accuracy in its two-stage target detection process, and can provide high detection accuracy for radar target detection tasks. When using the Faster R-CNN network for target detection, in addition to the target detection network itself, it usually also includes a pre-link and a post-link. The pre-link consists of several convolution layers (Convolution) to generate feature maps; the post-link consists of several pooling layers (Pooling) to reduce feature dimensions.
[0064] The computational complexity of the convolutional layer in the front-end of the Faster R-CNN network is high, which makes it difficult to meet the real-time requirements of radar target detection. Therefore, the Swin-Spatial Transformer network is designed to replace the convolutional layer for feature extraction. On the one hand, it aims to extract key feature information through the attention mechanism and filter out irrelevant sea clutter information, and on the other hand, it aims to extract spatial features in the echo image, thereby coping with the problem of spatial offset of the echo image caused by the rotation maneuver to a certain extent.
[0065] The conventional Swin-Transformer network consists of patch partitioning, linear embedding, and Swin-Transformer blocks. Although the Swin-Transformer network can extract certain spatial features, it focuses more on extracting the dependency information between global features and is insufficient in representing spatial features. The present invention refers to the Swin-Transformer network and designs a Swin-SpatialTransformer network that can explicitly extract spatial features and specifically handle spatial transformations. The internal structure of the Swin-SpatialTransformer network is shown in Figure 2. Figure 3 As shown in the figure, MLP (Multilayer Perceptron) represents multi-layer perceptron, LN (Layer Normalization) represents layer normalization, CNN (Convolutional Neural Networks) represents convolutional neural network, FC (Fully Connection) represents fully connected layer, W-MSA (Window Multi-head Self-Attention) represents window multi-head self-attention, SW-MSA (Shifted Window Multi-head Self-Attention) represents shifted window multi-head self-attention, and STN (Spatial Transformer Network) represents spatial transformer network. The Swin-Spatial Transformer network involves the use of attention mechanism for feature extraction, and STN acts once before and after the attention mechanism acts to extract multi-level, more abstract, and more essential spatial features.
[0066] In order to obtain the feature map after spatial transformation, the input feature map is affine transformed, so the parameter θ of the transformation matrix is 6-dimensional. The input and output of STN are both feature maps. Assume that the input shape is (C, H, W) and the output shape is (C, H′, W′). Each feature V in the output feature map V in channel C ic , which can be expressed as a function of the input feature map U on the neural network parameters:
[0067]
[0068] in, represents the feature of the input feature map U at position (m,n) in channel C, k(·) represents a linear interpolation method, Φ x ,Φ y represents the parameters of the linear interpolation method. To ensure that the network can perform back propagation, k(·) here uses bilinear interpolation. H′, W′ are the height and width of the output feature map. It is expressed as:
[0069]
[0070] in, Represents the affine transformation matrix, which is the output of the local network (LocalisationNet); are the coordinate points of the output feature map.
[0071] ③The feature map F1 output by the Swin-Spatial Transformer network is passed to the spatiotemporal feature fusion module and the Faster R-CNN network at the same time. The Faster R-CNN network extracts deep overall features from F1 to form a feature map F3.
[0072] For the feature map F1 entering the Faster R-CNN network, a deeper feature extraction will be performed on the echo image. The feature map F3 output by the ROI pooling layer in the later stage of the Faster R-CNN network is a one-dimensional vector flattened from the two-dimensional feature map, and its size is (1, h). The feature map F3 is only used as a type of overall feature extracted by the improved Faster R-CNN module for the input current frame echo image, and will be further fused with the global feature map F4 generated by the spatiotemporal feature fusion module, instead of providing separate suggestion box features and classification features.
[0073] Furthermore, the specific implementation process of the spatiotemporal feature fusion module includes:
[0074] This module aims to fuse the feature maps of the current frame echo image and the historical frame echo image to generate a global feature map containing time-varying features and spatial transformation features, improve the accuracy of the radar target detection network for the proposed box regression and orientation output, and alleviate the problem of reduced target detection accuracy caused by strong maneuvers. The shape changes of the feature map data flow in this module are as follows: Figure 4 The specific steps include:
[0075] ①Use the Conv-LSTM network to extract spatiotemporal features from the historical frame echo data with the shape of (T, C, W, H) and generate a shape of (1, C f ,H f ,W f ), where C=1, and T=n represents the number of input frames of the historical echo image.
[0076] ② Concatenate F2 and the feature map F1 output by the Swin-Spatial Transformer network to retain the information of each feature and facilitate multi-scale feature extraction.
[0077] ③ Using residual blocks to reduce the number of channels of fused features to 1 can improve the expressiveness of the target detection network model and alleviate the bottleneck problem of feature representation and the vanishing gradient problem of the loss function.
[0078] ④ Use the maximum pooling layer (MaxPool2d) to process the output features of the residual block to form the feature map F4, which can reduce the data dimension while retaining important feature information to reduce the amount of calculation and prevent overfitting.
[0079] Furthermore, the target position output module specifically includes the following steps:
[0080] ① The feature map F3 output by the improved Faster RCNN module and the feature map F4 output by the spatiotemporal feature fusion module are fused to generate multi-scale and multi-level deep features. The feature shape is (1, H f W f / 4+h), where h represents the dimension of the one-dimensional vector F3.
[0081] ②Use the full connection layer (Full Connection) network to process F4 and obtain the spatial position of the target suggestion box. Each suggestion box can be represented as a four-tuple (X i ,Y i ,h i ,w i ), where (X i ,Y i ) represents the coordinates of the center point of the suggestion box, h i and w i Represents the height and width of the suggestion box respectively. i ,Y i ), the distance D between the target and the radar carrier can be calculated i , and its direction
[0082]
[0083] Specific implementation process example
[0084] Consider a scenario that includes a highly maneuverable radar carrier, as well as multiple static targets, slow targets, and high-speed targets. The scenario is set in a sea environment, and radar scanning is susceptible to interference from sea clutter. The specific implementation process example of the radar target detection network includes the following steps:
[0085] ① The radar scans the surroundings to obtain echo images, such as Figure 5 After scanning, the data preprocessing module generates an echo image of uniform size. Due to the presence of interference such as sea clutter, there is often noise in the echo image, such as Figure 6 shown.
[0086] ② Input the current frame echo image into the improved Faster R-CNN module for noise suppression and feature extraction to obtain spatial variation features.
[0087] ③ Input the previously stored historical frame echo image (if it is the first frame, its historical frame is itself) into the spatiotemporal feature fusion module to obtain the global spatiotemporal change characteristics.
[0088] ④ Input the spatial variation characteristics and global spatiotemporal variation characteristics into the target orientation output module to obtain the target orientation information and display it in the Cartesian coordinate system, such as Figure 7 shown.
[0089] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.
[0090] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A radar target detection method based on multi-scale deep learning under strong maneuvers, characterized in that: include: The original radar echo data is filtered and processed into an echo image as the current frame echo image, the clutter noise in the current frame echo image is filtered out, and the current frame echo image is normalized into an echo image of the same scale; The feature map of the echo image of the current frame is extracted by using the Swin-Spatial Transformer network equipped with an attention mechanism. On the one hand, the feature map of the echo image of the current frame is transmitted to the spatiotemporal feature fusion module. On the other hand, the feature map of the echo image of the current frame is processed by the Faster R-CNN network to generate the spatial features of the region of interest after preliminary correction. The time and space features of the echo images of the historical n frames are extracted through the Conv-LSTM network, and are fused with other extracted scale features. The deep features containing the time and space variation features are output through the subsequent modules. The multi-dimensional deep features generated by fusion are used to obtain the multi-level and multi-scale features of the radar echo image, and the target direction information is output through a fully connected network.
2. The method according to claim 1, characterized in that The filtering of the original radar echo data and processing into an echo image as the current frame echo image, filtering out clutter noise in the current frame echo image, and normalizing the current frame echo image into an echo image of the same scale includes: For each complete original echo data, the redundant edge information is removed and converted into an original echo image of size (NH, NW), and the original echo image of size (NH, NW) is reduced and adjusted to (H, W), where (NH, NW) represents the height and width of the original echo image, and (H, W) represents the height and width of the adjusted image, respectively; The image of (H, W) size is clutter-filtered to obtain the current frame echo image. Each time the current frame echo image is generated, it is stored in a buffer with a capacity of n and is called as a historical echo image.
3. The method according to claim 2, characterized in that The Swin-Spatial Transformer network equipped with the attention mechanism is used to extract the feature map of the echo image of the current frame. On the one hand, the feature map of the echo image of the current frame is transmitted to the spatiotemporal feature fusion module. On the other hand, the feature map of the echo image of the current frame is processed by the Faster R-CNN network to generate the spatial features of the region of interest after preliminary correction, including: Keep the channel information of each current frame echo image received by the improved Faster R-CNN module as C=1; The Swin-Spatial Transformer network is used to extract key features and spatial features from the current frame echo image to form a feature map F1; The feature map F1 output by the Swin-Spatial Transformer network is passed to the spatiotemporal feature fusion module and the Faster R-CNN network at the same time. The Faster R-CNN network extracts deep overall features from the feature map F1 to form a feature map F3 with a size of (1, h).
4. The method according to claim 3, characterized in that The Swin-Spatial Transformer network is used to extract key features and spatial features from the current frame echo image to form a feature map F1, including: The Swin-Spatial Transformer network is designed to replace the convolutional layer in the front-end of the Faster R-CNN network to extract features and form the feature map F1.
5. The method according to claim 4, characterized in that The Swin-Spatial Transformer network includes: a multi-layer perceptron MLP, a normalization layer LP, a convolutional neural network CNN, a fully connected layer FC, a window multi-head self-attention W-MSA, a moving window multi-head self-attention SW-MSA and a spatial transformer network STN, wherein the STN acts once before and after the attention mechanism acts.
6. The method according to claim 5, characterized in that The input and output of STN are both feature maps. Assuming the input shape is (C, H, W) and the output shape is (C, H′, W′), the output feature map V is each feature V in channel C. i c Expressed as a function of the input feature map U on the neural network parameters: in, represents the feature of the input feature map U at position (m,n) in channel C, k(·) represents a linear interpolation method, Φ x ,Φ y represents the parameters of the linear interpolation method. To ensure that the network can perform back propagation, k(·) here uses bilinear interpolation. H′, W′ are the height and width of the output feature map. It is expressed as: Represents the affine transformation matrix, which is the output of the localization network (Localisation Net); are the coordinate points of the output feature map.
7. The method according to claim 6, characterized in that The Conv-LSTM network extracts the temporal and spatial features of the echo images of the historical n frames, and fuses them with the extracted other scale features. The subsequent modules output the deep features containing the temporal and spatial variation features, including: The Conv-LSTM network is used to extract spatiotemporal features from the historical frame echo data with the shape of (T, C, W, H) and generate a shape of (1, C f ,H f ,W f ), wherein C=1, and T=n represents the number of input frames of the historical echo image; The feature map F2 is concatenated and fused with the feature map F1 output by the Swin-Spatial Transformer network; A residual block is used to reduce the number of channels of the concatenated fusion features to 1, and a maximum pooling layer is used to process the output features of the residual block to form a feature map F4.
8. The method according to claim 7, characterized in that The multi-dimensional deep features generated by the fusion obtain the multi-level and multi-scale features of the radar echo image, and output the target position information through the fully connected network, including: The feature maps F3 and F4 are fused to generate multi-scale and multi-level deep features. The feature shape is (1, H f W f / 4+h); The feature map F4 is processed using a fully connected layer network to obtain the spatial position of the target proposal box, where each proposal box is represented as a four-tuple (X i ,Y i ,h i ,w i ), where (X i ,Y i ) represents the coordinates of the center point of the suggestion box, h i and w i Respectively represent the height and width of the suggestion box; According to the target's azimuth coordinates (X i ,Y i ),Depend on Calculate the distance D between the target and the radar carrier i ,Depend on Calculate the direction
Citation Information
Patent Citations
Radar echo extrapolation method and device based on deep space-time attention network
CN117233724A
