SAR small target detection and track association method based on environmental perception AIS guidance

Abstract

The application discloses an AIS guidance-based SAR small target detection and track association method based on environment perception, and relates to the technical field of ship monitoring.The application constructs a 5-channel environment perception input by fusing a dual-polarization image, an AIS guidance heat map and a water depth wind field, fully utilizes multi-source prior knowledge to guide the model to pay attention to the target area, significantly reduces sea clutter interference, effectively solves the problem of low signal-to-noise ratio of SAR small targets and easy missed detection, embeds a CBAM attention module in a YOLOv8 model, adaptively focuses on key features through channel and spatial attention mechanisms, suppresses background noise, and greatly improves the ship identification accuracy in complex sea conditions, introduces a Kalman filtering model based on wind speed adaptive adjustment process noise, combines a multi-dimensional Hungarian matching algorithm of position, speed and heading, effectively deals with ship maneuvering and severe sea conditions, reduces ID switching and track fragmentation, and generates high-precision and long-time-sequential ship continuous tracks, thereby providing reliable data support for marine management.

Description

Technical Field

[0001] This invention relates to the field of ship monitoring technology, specifically to a method for SAR small target detection and trajectory association based on environmental perception AIS guidance. Background Technology

[0002] Ship monitoring is a core requirement for maritime security and resource protection. Current technologies primarily rely on Automatic Identification Systems (AIS) to provide ship identity, location, and track information. However, AIS suffers from defects such as signal loss, intentional shutdown, or deception, especially in open ocean waters where small vessels can easily become "hidden ships," leading to monitoring blind spots. Synthetic Aperture Radar (SAR), with its all-weather, all-time imaging capabilities, has become an important means to compensate for the shortcomings of AIS. Sentinel-1 SAR images offer high resolution and wide coverage, but small targets have few pixels in the images (typically <20-50 pixels), making them severely susceptible to interference from sea clutter, wind, waves, and speckle noise. Traditional CFAR detectors have high false alarm rates, and single deep learning models (such as YOLO) still have relatively high false detection rates in complex sea conditions.

[0003] Existing SAR-AIS fusion methods mostly employ simple nearest neighbor matching or velocity estimation for point-to-point association, neglecting the significant modulating effect of dynamic marine environmental factors (wind field, water depth) on clutter intensity. Furthermore, trajectory association often relies on greedy algorithms, which are prone to mismatches in scenarios with dense multi-target activity or AIS interruption, making long-term continuous tracking of obscure vessels impossible. While recent research has focused on SAR vessel detection and multi-target tracking based on attention mechanisms, a systematic fusion framework integrating environmental grids, AIS-guided heatmaps, lightweight attention, and Kalman-Hangarian global optimization has yet to be established, failing to meet the demands of real-time, edge-deployed marine monitoring. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a method for SAR small target detection and trajectory association based on environment-aware AIS guidance. This technical solution solves the problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for small target detection and trajectory association in SAR based on environment-aware AIS guidance includes: Acquire Sentinel-1 synthetic aperture radar dual-polarization images, AIS matching tags from the automatic identification system, and marine environmental grid data; Construct a multi-channel input that integrates dual-polarization images, AIS-guided heatmaps, and environmental grids; An enhanced YOLOv8 model with an embedded attention module is used to detect small targets from multi-channel input, resulting in ship detection boxes. Based on Kalman filter prediction and Hungarian global matching algorithm, trajectory association and tracking are performed on detection boxes in consecutive frames; The continuity of the tracked trajectory is verified to generate a continuous ship trajectory.

[0006] Preferably, the construction of the multi-channel input, which integrates the dual-polarization image, the AIS guided heatmap, and the environment raster, specifically includes the following steps: The dB values ​​of the VV polarization channel and VH polarization channel of the dual polarization image are normalized to obtain the normalized VV channel and normalized VH channel. Based on the latitude and longitude coordinates of the AIS matching tags, the corresponding pixel coordinates in the SAR image are determined. A Gaussian-distributed attention mask is generated with the pixel coordinates of each AIS matching tag as the center, and the AIS heatmap channel is obtained. Bilinear upsampling of the water depth grid, wind speed grid, and wind direction grid in the marine environment grid to the SAR image resolution was performed, and normalization and angle coding were performed respectively to obtain the water depth channel and wind field channel. The normalized VV channel, normalized VH channel, AIS heat map channel, water depth channel and wind field channel are stitched together according to the channel dimension to form a 5-channel SAR environmental perception input image; The 5-channel SAR environmental perception input image is divided into image blocks of size 640×640 pixels with a slicing step size of 400 pixels. Image blocks containing AIS matching labels and randomly selected negative sample image blocks are retained.

[0007] Preferably, the dB value normalization processing of the VV polarization channel and VH polarization channel of the dual polarization image specifically includes the following steps: For any pixel in the VV polarization channel image, its original dB value is Normalized pixel values satisfy: when ,but ; When 0 ,but ; when ,but ; in, This represents the original backscattering coefficient (dB) value of a pixel in the VV polarization channel image. The pixel value after normalization of the VV polarization channel, with a value range of [0, 1]; The normalization process for the VH polarization channel image is the same as that for the VV polarization channel image, resulting in normalized VH channel pixel values. ,in This represents the original dB value of the VH polarization channel. The pixel value after VH channel normalization, with a value range of [0, 1]; Furthermore, the step of generating a Gaussian-distributed attention mask centered on the pixel coordinates of each AIS matched tag to obtain the AIS heatmap channel specifically includes the following steps: For each AIS matching tag, obtain its pixel coordinates in the SAR image. ,in These are the SAR image column coordinates corresponding to the longitude of the AIS label. The SAR image row coordinates corresponding to the latitude of the AIS tag; by Centered on a Gaussian function, an attention mask is generated, and the standard deviation of the Gaussian function is... The value is between 25 and 35 pixels, and the Gaussian function formula is: ; in, In the AIS heatmap channel The pixel value of a pixel, ranging from [0, 1]. For any pixel coordinate in the SAR image, Match the pixel coordinates of the tags to AIS. Let $\mathbf{a}$ be the standard deviation of the Gaussian distribution. When the attention masks of multiple AIS matching tags overlap, the maximum value of the pixel value in the overlapping area is taken as the AIS heatmap channel value of that pixel.

[0008] Preferably, the bilinear upsampling of the water depth grid, wind speed grid, and wind direction grid in the marine environment grid to SAR image resolution, followed by normalization and angle coding, specifically includes the following steps: Water depth raster processing: The water depth raster data is upsampled to the spatial resolution of the SAR image through bilinear interpolation to obtain an upsampled water depth raster; the maximum value in the upsampled water depth raster is then processed. Based on the water depth value of each pixel Normalization is performed; the normalization formula is: This yields the pixel values ​​of the water depth channel; where, This represents the water depth value of a pixel in the upsampled water depth raster. This represents the maximum water depth value in the sampled water depth grid. These are the pixel values ​​after normalization for the water depth channel; Wind speed raster processing: The wind speed raster data is upsampled to the spatial resolution of the SAR image using bilinear interpolation to obtain an upsampled wind speed raster; the maximum value in the upsampled wind speed raster is then used to process the data. Based on the wind speed value of each pixel Normalization is performed; the normalization formula is: The normalized wind speed value is obtained; among which, This is the wind speed value of the pixel in the upsampled wind speed grid. This represents the maximum wind speed value in the upsampled wind speed grid. This is the normalized value of the wind speed, which ranges from [0, 1]. Wind direction raster processing: The wind direction raster data is upsampled to the spatial resolution of the SAR image using bilinear interpolation to obtain the upsampled wind direction raster; for any pixel in the upsampled wind direction raster, its wind direction angle is... (Unit: degrees, range 0° to 360°), convert θ to radians. ,calculate and Where θ is the wind direction angle of the pixel in the upsampled wind direction grid. This represents the radian value corresponding to the wind direction angle. and These are the sine and cosine codes for the wind direction, respectively, with values ​​ranging from [-1, 1]. Wind field channel generation: Normalize the wind speed value and , The pixel values ​​are merged into the wind field channel, and the merging method is the same as the pixel values ​​of the wind field channel. ;in, The pixel value of the pixel in the wind field channel, with a value range of: [- , ].

[0009] Preferably, the step of dividing the 5-channel SAR environmental perception input image into image blocks of size 640×640 pixels with a slicing step size of 400 pixels, retaining image blocks containing AIS matching tags and randomly selected negative sample image blocks, specifically includes the following steps: The total pixel size of the SAR image is obtained as W×H (W is the number of pixels in the width of the SAR image, and H is the number of pixels in the height of the SAR image). The SAR image is divided into multiple overlapping image blocks, each with a size of 640×640 pixels. The horizontal and vertical step sizes of adjacent image blocks are both 400 pixels. For each image patch, determine whether it contains the pixel coordinates of at least one AIS matching tag: if it does, retain the image patch as a positive sample image patch; N image blocks are randomly selected from image blocks that do not contain AIS matching labels as negative sample image blocks, where N is 1 to 2 times the number of positive sample image blocks; All retained image patches are subjected to random horizontal flipping, random brightness adjustment, and random speckle noise addition. The speckle noise is added by multiplying each pixel value of the image patch by a random number that follows a normal distribution N(1, 0.12). Here, N(1, 0.12) represents a normal distribution with a mean of 1 and a variance of 0.01, which is used to simulate speckle noise in SAR images.

[0010] Preferably, the step of using an enhanced YOLOv8 model with an embedded attention module to perform small target detection on the multi-channel input to obtain ship detection boxes specifically includes the following steps: Modify the input layer of the YOLOv8n model, increasing the number of input channels from 3 to 5, corresponding to the 5 channels of the 5-channel SAR environment perception input image; A lightweight convolutional block attention module (CBAM) is inserted into the Neck layer of the YOLOv8n model. The CBAM module includes a channel attention submodule and a spatial attention submodule. The processing steps of the channel attention submodule are as follows: Process the feature map output by the Neck layer... C represents the number of feature map channels, H represents the feature map height in pixels, and W represents the feature map width in pixels. Global average pooling is performed to obtain the channel descriptors. ;Will Inputting two layers of a multilayer perceptron (MLP), the first MLP has C / 4 neurons and uses ReLU activation, while the second MLP has C neurons and uses Sigmoid activation, yields the channel attention weight vector. ;Will Multiplying the feature map F channel by channel yields the channel-weighted feature map. Where ⊙ represents multiplication by channel; The processing steps of the spatial attention submodule are as follows: weighted feature maps of the channels. Global average pooling and global max pooling are performed to obtain feature maps, respectively. and ;Will and By concatenating along the channel dimension, a concatenated feature map is obtained. ;right Perform a 7×7 convolution operation with a kernel of 1 and an activation function of sigmoid to obtain the spatial attention weight map. ;Will Channel-weighted feature map Pixel-by-pixel multiplication yields a spatially weighted feature map. Where ⊙ represents pixel-by-pixel multiplication; Spatial weighted feature map Input the head layer of the YOLOv8n model, use the CIoU loss function to calculate the bounding box regression loss, and use the Focal Loss function to calculate the class classification loss; The enhanced YOLOv8 model was trained using a SAR image dataset containing small target ships, employing data augmentation methods such as random rotation, random scaling, and wind field disturbance simulation during training.

[0011] Preferably, the trajectory association and tracking of detection boxes in consecutive frames based on Kalman filter prediction and Hungarian global matching algorithm specifically includes the following steps: For each target corresponding to a detection box, a Kalman filter state model is established. The state vector of the state model includes the target's pixel coordinates, velocity components, and heading angle. Based on the detection box of the current frame and the Kalman filter prediction state of the previous frame, a Hungarian matching cost matrix is ​​constructed. The cost matrix contains a weighted combination of position distance, velocity difference and heading difference. The Hungarian algorithm is used to perform global optimal matching on the cost matrix to obtain the association result between the detection box of the current frame and the trajectory of the previous frame; The target trajectory is updated based on the association results. When an AIS matching tag exists, the position, velocity, and heading information of the AIS tag are used to correct the Kalman filter state. When an AIS matching tag is lost, the Kalman filter prediction state is used to maintain the trajectory.

[0012] Preferably, the step of establishing a Kalman filter state model for each target corresponding to a detection box specifically includes the following steps: Define the state vector of the Kalman filter. ,in For the goal in the The x-coordinate of the frame's pixels. For the goal in the The pixel ordinate of the frame. For the goal in the The velocity in the horizontal direction of the frame. For the goal in the The velocity in the vertical direction of the frame. For the goal in the The heading angle of the frame; Define a state transition matrix Φ to describe the transition of the state vector from the 1st to the 2nd state. Frame to the Linear transition of frames: ; in, This is the inter-frame time interval, and its value is the revisit period of Sentinel-1 SAR; Define the process noise covariance matrix Q to describe the uncertainty in the state transition process. The diagonal elements of Q are... ,in and For location noise variance, and For the velocity noise variance, For heading noise variance; Based on the wind speed value of the current frame The actual wind speed is derived by back-calculating the normalized wind speed value from the wind field channel. The diagonal elements of the noise covariance matrix Q are then used for adaptive adjustment. when hour, , , , , ; when At that time, Q retains its initial value. ; when hour, , , , , ; in These are the initial diagonal elements of the process noise covariance matrix; According to the Frame state estimate Covariance Matrix Predicting the first Frame state vector Predicting the covariance matrix .

[0013] Preferably, the step of constructing the Hungarian matching cost matrix based on the detection box of the current frame and the Kalman filter prediction state of the previous frame specifically includes: For each detection box in the current frame Obtain the center coordinates of its bounding box. velocity components and heading angle ; For each trajectory in the previous frame Obtain the center coordinates predicted by its Kalman filter. Predicting velocity components and predicted heading angle ; Calculate the detection box With trajectory The distance between them ; Calculate the detection box With trajectory The speed difference between them ; Calculate the detection box With trajectory The difference in course between them ; Construct a cost matrix C, where elements It is a weighted sum of positional distance, speed difference, and heading difference, with the following weights: , , The calculation formula is: ; The cost matrix C is normalized by dividing each element by the maximum value of the cost matrix to obtain the normalized cost matrix. ,in ; Furthermore, the use of the Hungarian algorithm to perform global optimal matching on the cost matrix to obtain the association result between the current frame detection box and the trajectory of the previous frame specifically includes the following steps: The Hungarian algorithm is used to normalize the cost matrix. Perform global optimal matching to obtain the detection box. With trajectory Set of matching pairs ; For each matching pair Calculate the detection box bounding box and trajectory The intersection-union ratio (IoU) between the predicted bounding boxes is used to determine the match. If the IoU is greater than 0.3, the match is considered valid. For valid matching pairs Using a detection frame The center coordinates, velocity components, and heading angle-corrected Kalman filter state vector: ; in, For detection box The observation vector includes center coordinates, velocity, and heading. Let K be the observation matrix, and K be the Kalman gain, calculated using the following formula: , To measure the noise covariance matrix, the diagonal elements are... ; For unmatched detection boxes If the confidence level is greater than 0.7, a new trajectory is created, and the Kalman filter state vector is initialized to the detection box. The center coordinates, velocity components, initial velocity of 0 and heading angle, initial heading of 0, and covariance matrix P initialized to 10 times the identity matrix; For unmatched trajectories If the number of consecutive unmatched frames is less than 5, the trajectory remains in the prediction state; if the number of consecutive unmatched frames is greater than or equal to 5, the trajectory is marked as terminated.

[0014] Preferably, the step of verifying the continuity of the tracked trajectory to generate a continuous ship trajectory specifically includes: Maintain a list of active tracks, containing all tracks that have not yet terminated; For each active trajectory, count the number of consecutively successfully matched frames. :like If a frame is encountered, the trajectory is marked as a valid trajectory; if... If the frame is not cleared, then the trajectory is deleted. For a valid trajectory, the center coordinates, velocity, and heading state vector of each frame are converted into geographic coordinates, i.e., longitude and latitude. The conversion formula is as follows: ; ; in, and The geographic coordinates of the top-left pixel in the SAR image. and The pixel geographic resolution of the SAR image; The geographic coordinate sequence of the effective trajectory is arranged in chronological order to generate the final continuous ship trajectory. The trajectory data includes the timestamp, longitude, latitude, speed and heading information for each frame.

[0015] Compared with existing technologies, this invention provides a SAR small target detection and trajectory association method based on environment-aware AIS guidance, which has the following beneficial effects: This invention constructs a 5-channel environmental perception input that integrates dual-polarization images, AIS-guided heatmaps, and water depth and wind fields. It fully utilizes multi-source prior knowledge to guide the model to focus on the target area, significantly reducing sea clutter interference and effectively addressing the problems of low signal-to-noise ratio and easy missed detection of small SAR targets. A CBAM attention module is embedded in the YOLOv8 model, adaptively focusing on key features through channel and spatial attention mechanisms to suppress background noise and significantly improve ship identification accuracy in complex sea conditions. A Kalman filtering model based on wind speed adaptive adjustment of process noise is introduced, combined with a multi-dimensional Hungarian matching algorithm for position, speed, and heading, effectively addressing ship maneuvering and adverse sea conditions, reducing ID switching and trajectory fragmentation. Furthermore, a trajectory continuity verification mechanism eliminates short-term false alarms and accurately converts pixel coordinates to geographic coordinates, generating high-precision, long-term continuous ship trajectories, providing reliable data support for marine management. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow for S101-S105 in this invention; Figure 2 This is a schematic diagram of the method flow for S201-S205 in this invention; Figure 3 This is a schematic diagram of the method flow for S301-S304 in this invention; Figure 4 This is a schematic diagram of the method flow for S401-S404 in this invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Example 1 Please refer to Figure 1 As shown, the SAR small target detection and trajectory association method based on environment-aware AIS guidance includes: S101. Acquire Sentinel-1 synthetic aperture radar dual-polarization images, AIS matching tags, and marine environmental grid data; S102. Construct a multi-channel input that integrates dual-polarization images, AIS-guided heatmaps, and environmental grids; S103. An enhanced YOLOv8 model with an embedded attention module is used to detect small targets in the multi-channel input to obtain ship detection boxes; S104. Based on Kalman filter prediction and Hungarian global matching algorithm, trajectory association and tracking are performed on the detection boxes of consecutive frames; S105. Verify the continuity of the tracked trajectory and generate a continuous ship trajectory.

[0019] As will be understood by those skilled in the art, this invention constructs a 5-channel environmental perception input that integrates dual-polarization images, AIS-guided heatmaps, and water depth and wind fields. It fully utilizes multi-source prior knowledge to guide the model to focus on the target area, significantly reducing sea clutter interference and effectively solving the problems of low signal-to-noise ratio and easy missed detection of small SAR targets. By embedding a CBAM attention module into the YOLOv8 model, it adaptively focuses on key features through channel and spatial attention mechanisms, suppressing background noise and significantly improving the accuracy of ship identification in complex sea conditions. It introduces a Kalman filtering model based on wind speed adaptive adjustment of process noise, combined with a multi-dimensional Hungarian matching algorithm for position, speed, and heading, effectively addressing ship maneuvering and adverse sea conditions, reducing ID switching and trajectory fragmentation. Furthermore, it eliminates short-term false alarms through a trajectory continuity verification mechanism and accurately converts pixel coordinates into geographic coordinates, generating high-precision, long-term continuous ship trajectories, providing reliable data support for marine management.

[0020] Please refer to Figure 2 As shown, constructing a multi-channel input that integrates dual-polarization images, AIS-guided heatmaps, and environmental rasters includes the following steps: S201. Normalize the dB values ​​of the VV polarization channel and VH polarization channel of the dual polarization image to obtain the normalized VV channel and normalized VH channel. S202. Based on the latitude and longitude coordinates of the AIS matching tags, determine the corresponding pixel coordinates in the SAR image, and generate a Gaussian-distributed attention mask centered on the pixel coordinates of each AIS matching tag to obtain the AIS heatmap channel. S203. Bilinear upsampling of the water depth grid, wind speed grid and wind direction grid in the marine environment grid to the SAR image resolution is performed, and normalization and angle coding are performed respectively to obtain the water depth channel and wind field channel. S204. The normalized VV channel, normalized VH channel, AIS heat map channel, water depth channel and wind field channel are stitched together according to the channel dimension to form a 5-channel SAR environmental perception input image. S205. Divide the 5-channel SAR environmental perception input image into image blocks of size 640×640 pixels with a slicing step size of 400 pixels, and retain the image blocks containing AIS matching labels and randomly selected negative sample image blocks.

[0021] The dB values ​​of the VV and VH polarization channels of the dual-polarization image are normalized separately, specifically including the following steps: For any pixel in the VV polarization channel image, its original dB value is Normalized pixel values satisfy: when ,but ; When 0 ,but ; when ,but ; in, This represents the original backscattering coefficient (dB) value of a pixel in the VV polarization channel image. The pixel value after normalization of the VV polarization channel, with a value range of [0, 1]; The normalization process for the VH polarization channel image is the same as that for the VV polarization channel image, resulting in normalized VH channel pixel values. ,in This represents the original dB value of the VH polarization channel. The pixel value after VH channel normalization, with a value range of [0, 1]; Furthermore, a Gaussian-distributed attention mask is generated centered on the pixel coordinates of each AIS matched tag to obtain the AIS heatmap channel. The specific steps include the following: For each AIS matching tag, obtain its pixel coordinates in the SAR image. ,in These are the SAR image column coordinates corresponding to the longitude of the AIS label. The SAR image row coordinates corresponding to the latitude of the AIS tag; by Centered on a Gaussian function, an attention mask is generated, and the standard deviation of the Gaussian function is... The value is between 25 and 35 pixels, and the Gaussian function formula is: ; in, In the AIS heatmap channel The pixel value of a pixel, ranging from [0, 1]. For any pixel coordinate in the SAR image, Match the pixel coordinates of the tags to AIS. Let $\mathbf{a}$ be the standard deviation of the Gaussian distribution. When the attention masks of multiple AIS matching tags overlap, the maximum value of the pixel value in the overlapping area is taken as the AIS heatmap channel value of that pixel.

[0022] The water depth, wind speed, and wind direction grids in the marine environment grid are bilinearly upsampled to SAR image resolution, and then normalized and angle-coded. The specific steps include the following: Water depth raster processing: The water depth raster data is upsampled to the spatial resolution of the SAR image through bilinear interpolation to obtain an upsampled water depth raster; the maximum value in the upsampled water depth raster is then processed. Based on the water depth value of each pixel Normalization is performed; the normalization formula is: This yields the pixel values ​​of the water depth channel; where, This represents the water depth value of a pixel in the upsampled water depth raster. This represents the maximum water depth value in the sampled water depth grid. These are the pixel values ​​after normalization for the water depth channel; Wind speed raster processing: The wind speed raster data is upsampled to the spatial resolution of the SAR image using bilinear interpolation to obtain an upsampled wind speed raster; the maximum value in the upsampled wind speed raster is then used to process the data. Based on the wind speed value of each pixel Normalization is performed; the normalization formula is: The normalized wind speed value is obtained; among which, This is the wind speed value of the pixel in the upsampled wind speed grid. This represents the maximum wind speed value in the upsampled wind speed grid. This is the normalized value of the wind speed, which ranges from [0, 1]. Wind direction raster processing: The wind direction raster data is upsampled to the spatial resolution of the SAR image using bilinear interpolation to obtain the upsampled wind direction raster; for any pixel in the upsampled wind direction raster, its wind direction angle is... (Unit: degrees, range 0° to 360°), convert θ to radians. ,calculate and Where θ is the wind direction angle of the pixel in the upsampled wind direction grid. This represents the radian value corresponding to the wind direction angle. and These are the sine and cosine codes for the wind direction, respectively, with values ​​ranging from [-1, 1]. Wind field channel generation: Normalizing wind speed values and , The pixel values ​​are merged into the wind field channel, and the merging method is the same as the pixel values ​​of the wind field channel. ;in, The pixel value of the pixel in the wind field channel, with a value range of: [- , ].

[0023] Please refer to Figure 3As shown, the 5-channel SAR environmental perception input image is divided into image blocks of size 640×640 pixels with a slicing step size of 400 pixels. Image blocks containing AIS matching tags and randomly selected negative sample image blocks are retained. The specific steps include the following: S301. Obtain the total pixel size of the SAR image as W×H (W is the number of pixels in the width of the SAR image, and H is the number of pixels in the height of the SAR image). Divide the SAR image into multiple overlapping image blocks. The size of each image block is 640×640 pixels, and the horizontal and vertical step sizes of adjacent image blocks are both 400 pixels. S302. For each image block, determine whether it contains the pixel coordinates of at least one AIS matching tag: if it does, retain the image block as a positive sample image block. S303. Randomly select N image blocks from image blocks that do not contain AIS matching labels as negative sample image blocks, where N is 1 to 2 times the number of positive sample image blocks; S304. Perform random horizontal flipping, random brightness adjustment, and random speckle noise addition on all retained image blocks. The speckle noise is added by multiplying each pixel value of the image block by a random number that follows a normal distribution N(1, 0.12). Here, N(1, 0.12) represents a normal distribution with a mean of 1 and a variance of 0.01, used to simulate speckle noise in SAR images.

[0024] An enhanced YOLOv8 model with an embedded attention module is used to detect small targets from multi-channel input to obtain ship detection boxes. The specific steps include the following: Modify the input layer of the YOLOv8n model, increasing the number of input channels from 3 to 5, corresponding to the 5 channels of the 5-channel SAR environment perception input image; Lightweight convolutional block attention module (CBAM) is inserted into the Neck layer of the YOLOv8n model. The CBAM module includes channel attention submodule and spatial attention submodule. The processing steps of the channel attention submodule are as follows: Process the feature map output by the Neck layer... C represents the number of feature map channels, H represents the feature map height in pixels, and W represents the feature map width in pixels. Global average pooling is performed to obtain the channel descriptors. ;Will Inputting two layers of a multilayer perceptron (MLP), the first MLP has C / 4 neurons and uses ReLU activation, while the second MLP has C neurons and uses Sigmoid activation, yields the channel attention weight vector. ;Will Multiplying the feature map F channel by channel yields the channel-weighted feature map. Where ⊙ represents multiplication by channel; The processing steps of the spatial attention submodule are as follows: weighted feature maps of the channels. Global average pooling and global max pooling are performed to obtain feature maps, respectively. and ;Will and By concatenating along the channel dimension, a concatenated feature map is obtained. ;right Perform a 7×7 convolution operation with a kernel of 1 and an activation function of sigmoid to obtain the spatial attention weight map. ;Will Channel-weighted feature map Pixel-by-pixel multiplication yields a spatially weighted feature map. Where ⊙ represents pixel-by-pixel multiplication; Spatial weighted feature map Input the head layer of the YOLOv8n model, use the CIoU loss function to calculate the bounding box regression loss, and use the Focal Loss function to calculate the class classification loss; The enhanced YOLOv8 model was trained using a SAR image dataset containing small target ships, employing data augmentation methods such as random rotation, random scaling, and wind field disturbance simulation during training.

[0025] Please refer to Figure 4 As shown, based on Kalman filtering prediction and the Hungarian global matching algorithm, trajectory association and tracking of detection boxes in consecutive frames are performed, specifically including the following steps: S401. For each target corresponding to a detection box, establish a Kalman filter state model. The state vector of the state model includes the target's pixel coordinates, velocity components, and heading angle. S402. Based on the detection box of the current frame and the Kalman filter prediction state of the previous frame, construct the Hungarian matching cost matrix. The cost matrix contains a weighted combination of position distance, velocity difference and heading difference. S403. The Hungarian algorithm is used to perform global optimal matching on the cost matrix to obtain the association result between the detection box of the current frame and the trajectory of the previous frame. S404. Update the target trajectory based on the association results. When an AIS matching tag exists, use the position, speed, and heading information of the AIS tag to correct the Kalman filter state. When an AIS matching tag is lost, use the Kalman filter to predict the state and maintain the trajectory.

[0026] For each target corresponding to a detection box, a Kalman filter state model is established, which includes the following steps: Define the state vector of the Kalman filter. ,in For the goal in the The x-coordinate of the frame's pixels. For the goal in the The pixel ordinate of the frame. For the goal in the The velocity in the horizontal direction of the frame. For the goal in the The velocity in the vertical direction of the frame. For the goal in the The heading angle of the frame; Define a state transition matrix Φ to describe the transition of the state vector from the 1st to the 2nd state. Frame to the Linear transition of frames: ; in, This is the inter-frame time interval, and its value is the revisit period of Sentinel-1 SAR; Define the process noise covariance matrix Q to describe the uncertainty in the state transition process. The diagonal elements of Q are... ,in and For location noise variance, and For the velocity noise variance, For heading noise variance; Based on the wind speed value of the current frame The actual wind speed is derived by back-calculating the normalized wind speed value from the wind field channel. The diagonal elements of the noise covariance matrix Q are then used for adaptive adjustment. when hour, , , , , ; when At that time, Q retains its initial value. ; when hour, , , , , ; in These are the initial diagonal elements of the process noise covariance matrix; According to the Frame state estimate Covariance Matrix Predicting the first Frame state vector Predicting the covariance matrix .

[0027] Based on the detection bounding box of the current frame and the Kalman filter prediction state of the previous frame, a Hungarian matching cost matrix is ​​constructed, which specifically includes: For each detection box in the current frame Obtain the center coordinates of its bounding box. velocity components and heading angle ; For each trajectory in the previous frame Obtain the center coordinates predicted by its Kalman filter. Predicting velocity components and predicted heading angle ; Calculate the detection box With trajectory The distance between them ; Calculate the detection box With trajectory The speed difference between them ; Calculate the detection box With trajectory The difference in course between them ; Construct a cost matrix C, where elements It is a weighted sum of positional distance, speed difference, and heading difference, with the following weights: , , The calculation formula is: ; The cost matrix C is normalized by dividing each element by the maximum value of the cost matrix to obtain the normalized cost matrix. ,in ; Furthermore, the Hungarian algorithm is used to perform global optimal matching on the cost matrix to obtain the association result between the detection box of the current frame and the trajectory of the previous frame. The specific steps include the following: The Hungarian algorithm is used to normalize the cost matrix. Perform global optimal matching to obtain the detection box. With trajectory Set of matching pairs ; For each matching pair Calculate the detection box bounding box and trajectory The intersection-union ratio (IoU) between the predicted bounding boxes is used to determine the match. If the IoU is greater than 0.3, the match is considered valid. For valid matching pairs Using a detection frame The center coordinates, velocity components, and heading angle-corrected Kalman filter state vector: ; in, For detection box The observation vector includes center coordinates, velocity, and heading. Let K be the observation matrix, and K be the Kalman gain, calculated using the following formula: , To measure the noise covariance matrix, the diagonal elements are... ; For unmatched detection boxes If the confidence level is greater than 0.7, a new trajectory is created, and the Kalman filter state vector is initialized to the detection box. The center coordinates, velocity components, initial velocity of 0 and heading angle, initial heading of 0, and covariance matrix P initialized to 10 times the identity matrix; For unmatched trajectories If the number of consecutive unmatched frames is less than 5, the trajectory remains in the prediction state; if the number of consecutive unmatched frames is greater than or equal to 5, the trajectory is marked as terminated.

[0028] The tracked trajectory is then validated for continuity to generate a continuous ship trajectory, specifically including: Maintain a list of active tracks, containing all tracks that have not yet terminated; For each active trajectory, count the number of consecutively successfully matched frames. :like If a frame is encountered, the trajectory is marked as a valid trajectory; if... If the frame is not cleared, then the trajectory is deleted. For a valid trajectory, the center coordinates, velocity, and heading state vector of each frame are converted into geographic coordinates, i.e., longitude and latitude. The conversion formula is as follows: ; ; in, and The geographic coordinates of the top-left pixel in the SAR image. and The pixel geographic resolution of the SAR image; The geographic coordinate sequence of the effective trajectory is arranged in chronological order to generate the final continuous ship trajectory. The trajectory data includes the timestamp, longitude, latitude, speed and heading information for each frame.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A SAR small target detection and track association method based on environmental perception AIS guidance, characterized in that, include: Acquire Sentinel-1 synthetic aperture radar dual-polarization images, AIS matching tags from the automatic identification system, and marine environmental grid data; Construct a multi-channel input that integrates dual-polarization images, AIS-guided heatmaps, and environmental grids; An enhanced YOLOv8 model with an embedded attention module is used to detect small targets from multi-channel input, resulting in ship detection boxes. Based on Kalman filter prediction and Hungarian global matching algorithm, trajectory association and tracking are performed on detection boxes in consecutive frames; The continuity of the tracked trajectory is verified to generate a continuous ship trajectory.

2. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 1, characterized in that, The construction of a multi-channel input that integrates dual-polarization images, AIS-guided heatmaps, and environmental grids specifically includes the following steps: The dB values ​​of the VV polarization channel and VH polarization channel of the dual polarization image are normalized to obtain the normalized VV channel and normalized VH channel. Based on the latitude and longitude coordinates of the AIS matching tags, the corresponding pixel coordinates in the SAR image are determined. A Gaussian-distributed attention mask is generated with the pixel coordinates of each AIS matching tag as the center, and the AIS heatmap channel is obtained. Bilinear upsampling of the water depth grid, wind speed grid, and wind direction grid in the marine environment grid to SAR image resolution is performed, and normalization and angle coding are performed respectively to obtain the water depth channel and wind field channel. The normalized VV channel, normalized VH channel, AIS heat map channel, water depth channel and wind field channel are stitched together according to the channel dimension to form a 5-channel SAR environmental perception input image; The 5-channel SAR environmental perception input image is divided into image blocks of size 640×640 pixels with a slicing step size of 400 pixels. Image blocks containing AIS matching labels and randomly selected negative sample image blocks are retained.

3. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 2, characterized in that, The dB value normalization processing of the VV and VH polarization channels of the dual-polarization image includes the following steps: For any pixel in the VV polarization channel image, its original dB value is Normalized pixel values satisfy: when ,but ; When 0 ,but ; when ,but ; in, This represents the original backscattering coefficient (dB) value of a pixel in the VV polarization channel image. The pixel value after normalization of the VV polarization channel, with a value range of [0, 1]; The normalization process for the VH polarization channel image is the same as that for the VV polarization channel image, resulting in normalized VH channel pixel values. ,in This represents the original dB value of the VH polarization channel. The pixel value after VH channel normalization, with a value range of [0, 1]; Furthermore, the step of generating a Gaussian-distributed attention mask centered on the pixel coordinates of each AIS matched tag to obtain the AIS heatmap channel specifically includes the following steps: For each AIS matching tag, obtain its pixel coordinates in the SAR image. ,in These are the SAR image column coordinates corresponding to the longitude of the AIS label. These are the SAR image row coordinates corresponding to the latitude of the AIS tag; by Centered on a Gaussian function, an attention mask is generated, and the standard deviation of the Gaussian function is... The value is between 25 and 35 pixels, and the Gaussian function formula is: ； in, In the AIS heatmap channel The pixel value of a pixel, ranging from [0, 1]. For any pixel coordinate in the SAR image, Match the pixel coordinates of the tags to AIS. The standard deviation of the Gaussian distribution; When the attention masks of multiple AIS matching tags overlap, the maximum value of the pixel value in the overlapping area is taken as the AIS heatmap channel value of that pixel.

4. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 3, characterized in that, The process of bilinearly upsampling the water depth grid, wind speed grid, and wind direction grid in the marine environment grid to SAR image resolution, and then performing normalization and angle coding processing, specifically includes the following steps: Water depth raster processing: The water depth raster data is upsampled to the spatial resolution of the SAR image through bilinear interpolation to obtain an upsampled water depth raster; the maximum value in the upsampled water depth raster is then processed. Based on the water depth value of each pixel Normalization is performed; the normalization formula is: This yields the pixel values ​​of the water depth channel; where, This represents the water depth value of a pixel in the upsampled water depth raster. This represents the maximum water depth value in the sampled water depth grid. These are the pixel values ​​after normalization for the water depth channel; Wind speed raster processing: The wind speed raster data is upsampled to the spatial resolution of the SAR image using bilinear interpolation to obtain an upsampled wind speed raster; the maximum value in the upsampled wind speed raster is then used to process the data. Based on the wind speed value of each pixel Normalization is performed; the normalization formula is: The normalized wind speed value is obtained; among which, This is the wind speed value of the pixel in the upsampled wind speed grid. This represents the maximum wind speed value in the upsampled wind speed grid. This is the normalized value of the wind speed, which ranges from [0, 1]. Wind direction raster processing: The wind direction raster data is upsampled to the spatial resolution of the SAR image using bilinear interpolation to obtain the upsampled wind direction raster; for any pixel in the upsampled wind direction raster, its wind direction angle is... (Unit: degrees, range 0° to 360°), convert θ to radians. ,calculate and Where θ is the wind direction angle of the pixel in the upsampled wind direction grid. This represents the radian value corresponding to the wind direction angle. and These are the sine and cosine codes for the wind direction, respectively, with a range of [-1, 1]. Wind field channel generation: Normalize the wind speed value and , The pixel values ​​are merged into the wind field channel, and the merging method is the same as the pixel values ​​of the wind field channel. ;in, The pixel value of the pixel in the wind field channel, with a value range of: [- , ].

5. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 4, characterized in that, The step of dividing the 5-channel SAR environmental perception input image into image blocks of size 640×640 pixels with a slicing step size of 400 pixels, retaining image blocks containing AIS matching tags and randomly selected negative sample image blocks, specifically includes the following steps: The total pixel size of the SAR image is obtained as W×H (W is the number of pixels in the width of the SAR image, and H is the number of pixels in the height of the SAR image). The SAR image is divided into multiple overlapping image blocks, each with a size of 640×640 pixels. The horizontal and vertical step sizes of adjacent image blocks are both 400 pixels. For each image patch, determine whether it contains the pixel coordinates of at least one AIS matching tag: if it does, retain the image patch as a positive sample image patch; N image blocks are randomly selected from image blocks that do not contain AIS matching labels as negative sample image blocks, where N is 1 to 2 times the number of positive sample image blocks; All retained image patches are subjected to random horizontal flipping, random brightness adjustment, and random speckle noise addition. The speckle noise is added by multiplying each pixel value of the image patch by a random number that follows a normal distribution N(1, 0.12). Here, N(1, 0.12) represents a normal distribution with a mean of 1 and a variance of 0.01, which is used to simulate speckle noise in SAR images.

6. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 5, characterized in that, The enhanced YOLOv8 model with embedded attention modules is used to detect small targets from multi-channel input to obtain ship detection boxes. This process includes the following steps: Modify the input layer of the YOLOv8n model, increasing the number of input channels from 3 to 5, corresponding to the 5 channels of the 5-channel SAR environment perception input image; A lightweight convolutional block attention module (CBAM) is inserted into the Neck layer of the YOLOv8n model. The CBAM module includes a channel attention submodule and a spatial attention submodule. The processing steps of the channel attention submodule are as follows: Process the feature map output by the Neck layer... C represents the number of feature map channels, H represents the feature map height in pixels, and W represents the feature map width in pixels. Global average pooling is performed to obtain the channel descriptors. ;Will Inputting two layers of a multilayer perceptron (MLP), the first MLP has C / 4 neurons and uses ReLU activation, while the second MLP has C neurons and uses Sigmoid activation, yields the channel attention weight vector. ;Will Multiplying the feature map F channel by channel yields a channel-weighted feature map; where ⊙ denotes channel-by-channel multiplication. The processing steps of the spatial attention submodule are as follows: weighted feature maps of the channels. Global average pooling and global max pooling are performed to obtain feature maps, respectively. and ;Will and By concatenating along the channel dimension, a concatenated feature map is obtained. ;right Perform a 7×7 convolution operation with a kernel of 1 and an activation function of sigmoid to obtain the spatial attention weight map. ;Will Channel-weighted feature map Pixel-by-pixel multiplication yields a spatially weighted feature map. Where ⊙ represents pixel-by-pixel multiplication; Spatial weighted feature map Input the head layer of the YOLOv8n model, use the CIoU loss function to calculate the bounding box regression loss, and use the Focal Loss function to calculate the class classification loss; The enhanced YOLOv8 model was trained using a SAR image dataset containing small target ships, employing data augmentation methods such as random rotation, random scaling, and wind field disturbance simulation during training.

7. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 6, characterized in that, The method based on Kalman filtering prediction and Hungarian global matching algorithm for trajectory association and tracking of detection boxes in consecutive frames specifically includes the following steps: For each target corresponding to a detection box, a Kalman filter state model is established. The state vector of the state model includes the target's pixel coordinates, velocity components, and heading angle. Based on the detection box of the current frame and the Kalman filter prediction state of the previous frame, a Hungarian matching cost matrix is ​​constructed. The cost matrix contains a weighted combination of position distance, velocity difference and heading difference. The Hungarian algorithm is used to perform global optimal matching on the cost matrix to obtain the association result between the detection box of the current frame and the trajectory of the previous frame; Update the target trajectory based on the association results. When an AIS matching tag exists, use the position, velocity and heading information of the AIS tag to correct the Kalman filter state. When AIS matching tags are lost, Kalman filtering is used to predict the state-maintaining trajectory.

8. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 7, characterized in that, The step of establishing a Kalman filter state model for each target corresponding to a detection box includes the following steps: Define the state vector of the Kalman filter. ,in For the goal in the The x-coordinate of the frame's pixels. For the goal in the The pixel ordinate of the frame. For the goal in the The velocity in the horizontal direction of the frame. For the goal in the The velocity in the vertical direction of the frame. For the goal in the The heading angle of the frame; Define a state transition matrix Φ to describe the transition of the state vector from the 1st to the 2nd state. Frame to the Linear transition of frames: ； in, This is the inter-frame time interval, and its value is the revisit period of Sentinel-1 SAR; Define the process noise covariance matrix Q to describe the uncertainty in the state transition process. The diagonal elements of Q are... ,in and For location noise variance, and For the velocity noise variance, For heading noise variance; Based on the wind speed value of the current frame The actual wind speed is derived by back-calculating the normalized wind speed value from the wind field channel. The diagonal elements of the noise covariance matrix Q are then used for adaptive adjustment. when hour, , , , , ; when At that time, Q retains its initial value. ; when hour, , , , , ; in These are the initial diagonal elements of the process noise covariance matrix; According to the Frame state estimate Covariance Matrix Predicting the first Frame state vector Predicting the covariance matrix .

9. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 8, characterized in that, The step of constructing the Hungarian matching cost matrix based on the detection box of the current frame and the Kalman filter prediction state of the previous frame specifically includes: For each detection box in the current frame Obtain the center coordinates of its bounding box. velocity components and heading angle ; For each trajectory in the previous frame Obtain the center coordinates predicted by its Kalman filter. Predicting velocity components and predicted heading angle ; Calculate the detection box With trajectory The distance between them ; Calculate the detection box With trajectory The speed difference between them ; Calculate the detection box With trajectory The difference in course between them ; Construct a cost matrix C, where elements It is a weighted sum of positional distance, speed difference, and heading difference, with the following weights: , , The calculation formula is: ； The cost matrix C is normalized by dividing each element by the maximum value of the cost matrix to obtain the normalized cost matrix. ,in ; Furthermore, the use of the Hungarian algorithm to perform global optimal matching on the cost matrix to obtain the association result between the current frame detection box and the trajectory of the previous frame specifically includes the following steps: The Hungarian algorithm is used to normalize the cost matrix. Perform global optimal matching to obtain the detection box. With trajectory Set of matching pairs ; For each matching pair Calculate the detection box bounding box and trajectory The intersection-union ratio (IoU) between the predicted bounding boxes is used to determine the match. If the IoU is greater than 0.3, the match is considered valid. For valid matching pairs Using a detection frame The center coordinates, velocity components, and heading angle-corrected Kalman filter state vector: ； in, For detection box The observation vector includes center coordinates, velocity, and heading. Let K be the observation matrix, and K be the Kalman gain, calculated using the following formula: , To measure the noise covariance matrix, the diagonal elements are... ; For unmatched detection boxes If the confidence level is greater than 0.7, a new trajectory is created, and the Kalman filter state vector is initialized to the detection box. The center coordinates, velocity components, initial velocity of 0 and heading angle, initial heading of 0, and covariance matrix P initialized to 10 times the identity matrix; For unmatched trajectories If the number of consecutive unmatched frames is less than 5, the trajectory remains in the prediction state; if the number of consecutive unmatched frames is greater than or equal to 5, the trajectory is marked as terminated.

10. The SAR small target detection and trajectory association method based on environment-aware AIS guidance according to claim 9, characterized in that, The step of verifying the continuity of the tracked trajectory to generate a continuous ship trajectory specifically includes: Maintain a list of active tracks, containing all tracks that have not yet terminated; For each active trajectory, count the number of consecutively successfully matched frames. :like If a frame is encountered, the trajectory is marked as a valid trajectory; if... If the frame is not cleared, then the trajectory is deleted. For a valid trajectory, the center coordinates, velocity, and heading state vector of each frame are converted into geographic coordinates, i.e., longitude and latitude. The conversion formula is as follows: ； ； in, and The geographic coordinates of the top-left pixel in the SAR image. and The pixel geographic resolution of the SAR image; The geographic coordinate sequence of the effective trajectory is arranged in chronological order to generate the final continuous trajectory of the ship. The trajectory data includes the timestamp, longitude, latitude, speed and heading information of each frame.

Images