ISAR image satellite target attitude estimation method based on geometric feature constraint

Abstract

The invention discloses an ISAR (Inverse Synthetic Aperture Radar) image satellite target attitude estimation method based on geometric feature constraint, which comprises the following steps of: segmenting typical satellite components in a sequence ISAR image based on a Conv-LSTM network, and extracting component scattering centers by adopting an attribute scattering center model; estimating the direction and length of the main axis of the satellite main body by using principal component analysis (PCA); extracting axis features of the sailboard through DBSCAN clustering and an L-shaped rectangular search algorithm; and finally, a projection geometric model of a target three-dimensional to two-dimensional ISAR imaging plane is constructed in combination with radar sight line information, an optimization target function is established, and a satellite attitude angle is solved through a particle swarm optimization algorithm. The spatial-temporal information of the sequence image and the geometric structure constraint of the target are fully fused, the problem of projection change caused by target motion in ISAR imaging can be effectively solved, high-precision estimation of the satellite attitude is achieved, and the method has high anti-interference capacity and practical application value.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints. Background Technology

[0002] With the development of aerospace technology, the monitoring and identification of satellite targets has become one of the important tasks for ensuring space security. Satellite target attitude estimation is a key issue in this process. It is directly related to orbit prediction, communication stability, and collision early warning, and is an indispensable critical link in this field.

[0003] Traditional satellite attitude estimation methods, such as ground-based remote sensing and optical imaging technologies, while achieving high accuracy under certain conditions, are limited by environmental factors such as weather, lighting, and observation angle, making it difficult to meet the needs of stable monitoring in all weather and at all times. Against this backdrop, Inverse Synthetic Aperture Radar (ISAR), with its ability to acquire high-resolution images under diverse climatic conditions, has gradually become an important technique for satellite target attitude estimation.

[0004] However, ISAR technology has a problem in the process of acquiring images: it is difficult to eliminate the influence of changes in target attitude on the image projection geometry. Due to the complex motion of satellite targets in orbit, their attitude includes changes in multiple directions such as pitch, yaw, and roll, which causes the projection relationship of the target in the ISAR image to change constantly. This highly dynamic feature makes it extremely difficult to accurately extract the target's attitude information from a single image or a few frames.

[0005] To address the aforementioned issues, the ISAR image satellite target attitude estimation method based on geometric feature constraints has significant advantages.

[0006] This method utilizes temporal series information from sequential ISAR images to accurately estimate the attitude of satellite targets through image projection geometry. Compared to traditional ISAR methods for acquiring target attitude information, this method offers several advantages: First, it employs a Convolutional Long Short-Term Memory (Conv-LSTM) network to process the spatiotemporal dependencies in sequential ISAR images, enabling effective segmentation and identification of different satellite target components within dynamically changing images, thus providing a reliable foundation for subsequent attitude estimation. Second, accurate extraction of the target's scattering centers is crucial for accurate attitude estimation. A scattering center extraction method based on an attribute scattering center model, combined with the target's geometric features, further enhances the accuracy of attitude estimation. Finally, typical satellite structural components, such as solar panels and antennas, provide important geometric information during attitude estimation. Parameter estimation based on these components not only constrains attitude estimation but also effectively improves estimation accuracy. By optimizing the geometric relationships across multiple ISAR frames, the attitude of the satellite target at different time points can be accurately estimated.

[0007] This method fully considers the complexity of satellite targets in orbital motion and the distribution characteristics of scattering points in ISAR images, thus effectively solving several problems in target attitude estimation. Summary of the Invention

[0008] Technical problem: To overcome the shortcomings of the existing technology, the present invention provides a satellite target attitude estimation method based on geometric feature constraints of ISAR images, which realizes the attitude estimation of satellite targets.

[0009] Technical solution: The present invention provides a method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints, comprising the following steps: Step 1: The typical component segmentation method of sequential ISAR images based on the Convolution Long Short-Term Memory (Conv-LSTM) network is used to segment typical satellite components; Step 2: Establish a simplified attribute scattering center model, introduce sparse representation theory to improve the traditional relaxation algorithm RELAX, and transform the parameter estimation problem into a sparse coefficient vector under a dictionary set. The norm optimization problem involves alternating updates to the position and amplitude parameters of the scattering center, and through iteration, the estimation results gradually converge to the maximum likelihood solution. Step 3: Extract the principal axis features of the satellite body, and estimate the orientation and length of the satellite body based on principal component analysis (PCA); Step 4: Perform clustering on the extracted satellite solar panel scattering centers to find two point sets, and then use the L-shaped rectangle search algorithm to extract the axis features of the satellite solar panel. Step 5: Based on the elevation and azimuth angles corresponding to the radar line of sight, construct the projection matrix from the three-dimensional coordinates of the spatial target scattering center to the ISAR two-dimensional imaging plane. On this basis, establish the constraint relationship between the projection of two-dimensional components in the ISAR image and the three-dimensional spatial attitude. Accurate estimation of the target attitude is achieved by solving the joint optimization equation of geometric features.

[0010] in: In step 1, the typical component segmentation method for sequential ISAR images based on the Convolutional Long Short-Term Memory (Conv-LSTM) network is used. The input is a sequential ISAR image, and features are extracted through the backbone network model EfficientNet-B4. Low-level features are directly fed into the decoder, while high-level features are processed by the Atrous Spatial Pyramid Pooling (ASPP) module and then input together with the low-level features into the Conv-LSTM module to fuse spatiotemporal information. Finally, the component segmentation results of the satellite body and the satellite solar panels are obtained.

[0011] Step 2 of the improved traditional relaxation algorithm RELAX includes: transforming the parameter estimation problem into solving a sparse coefficient vector norm optimization problem, and then alternately updating the scattering center position and amplitude parameters, so that the estimation results gradually converge to the maximum likelihood solution; including the following steps: Step 2.1: Establish a simplified property scattering center model, targeting the first... A scattering center, its scattering response model The expression is: , in, It is a scattering center model. It's the azimuth. It's the radar frequency. It's the speed of light. This represents additive white Gaussian noise. Indicates the first The index of each point, The imaginary unit is represented for each scattering center. This represents the set of its corresponding attribute scattering parameters, where, Represents the range, Indicates the length of the scattering center. It is the scattering center direction angle. , Represents the position coordinates of the scattering center in the azimuth and range directions; By using sparse signal decomposition theory to extract the target's attribute scattering centers, the model can be represented as: Equation 2, in This is the result of vectorizing the scattering center. It is the dictionary set used throughout the parameter extraction process. It is a sparse coefficient vector. It's noise; Step 2.2: Transform the parameter estimation problem into solving for a sparse coefficient vector. Norm optimization is a problem of solving nondeterministic polynomial problems. Formula 3, in express Norm, express Norm, This indicates the parameter used to find the minimum value of the above expression within a user-defined domain. This represents the sparse vector of coefficients obtained from the solution. Indicates noise level; Step 2.3: Let Initialization is performed, zero-padding is applied to the signal in the frequency and angle domains, and then an inverse fast Fourier transform (IFFT) is used to obtain the position coordinate parameters of the first scattering center in the azimuth and range directions. A rough estimate, substituted into Equation 3, yields the length and orientation angle parameters of the first scattering center. A rough estimate, reconstructed from Equation 1, of the signal ; Step 2.4: Let Subtract the reconstructed signal from the original signal. After obtaining the remaining signal, the length, direction angle, and position coordinates of the second scattering center in the azimuth and range directions are obtained according to the method in step 3). Substitute and recalculate Then, the parameters are re-estimated using Equation 2. Repeat these two steps until convergence. Step 2.5: Let Similarly, subtract the reconstructed signal from the original signal. The remaining signal is obtained, and the length, direction angle, and position coordinates of the third scattering center in the azimuth and range directions are determined. The estimate, using and Recalculate At the same time, the parameters are re-estimated using formula (2). Similarly, based on the re-estimated parameters and Recalculate Re-estimate parameters Iterate these three steps together until convergence. Steps 2.3 to 2.5 above constitute a cyclic iterative process based on the increment of the index variable i. The iteration terminates when the energy difference between two adjacent remaining signals accounts for no more than 1% of the energy of the original signal, or when the preset maximum number of iterations is reached.

[0012] Step 3 includes: estimating the orientation and length of the satellite body using Principal Component Analysis (PCA); estimating its principal axis vector using the PCA algorithm based on the spatial distribution coordinates of the satellite body's scattering center in the ISAR image; including the following steps: Step 3.1: Extracting the main body of the satellite Coordinates of each scattering center The position matrix formed is , , Center all samples: , in, , Step 3.2: Calculate the sample covariance matrix: , Step 3.3: Eigenvalue decomposition of the covariance matrix: , in, It is a matrix. It's another matrix, with superscript... For transpose; Step 3.4: Based on the matrix sum matrix The first principal component is determined, and the unit vector corresponding to the projection direction of the satellite body on the ISAR two-dimensional imaging plane is obtained. ; Step 3.5: Obtain the projection length and project all scattering centers onto the unit vector. The projection value of each point is obtained, thus yielding the projected length of the satellite body. , .

[0013] In step 4, the density-based spatial clustering (DBSCAN) algorithm is used to cluster the extracted satellite solar panel scattering centers. The algorithm implementation process is as follows: Input parameters: sample point set neighborhood radius Sample points are the core points Points within the neighborhood Core clustering Number of clusters Non-core points The core point Points in the neighborhood Distance between , the minimum number of points ; Output: Set of clusters and set of noise points The cluster set includes core points and boundary points; Initialization: All points are marked as unprocessed; Step 4.1: Calculate the neighborhood radius , representing the sample point set Center and core point The distance is not greater than the neighborhood radius. For a subset of samples, Euclidean distance is used as the distance metric, expressed as: , Step 4.2: Determine the core point. When the core point... neighborhood radius At least contains When there are 1 sample: , This sample point is the core point; Step 4.3: Construct core point clusters. Based on density reachability, traverse all core points and process each unprocessed current core point. : a) Create a new cluster The core point Add a new cluster And mark it as processed; b) Density expansion: For new clusters Each core point Check its neighborhood radius All points within: If the points in the neighborhood If the core point is not processed, then the points in the neighborhood will be visited. Add a new cluster And mark it as processed to achieve indirect density reach; right Repeat the expansion process in step 2 until the current new cluster is reached. There are no unprocessed core points within the neighborhood of any of the core points in the process; c) The new cluster Add to cluster set Let the number of clusters be The count increments by 1, that is... , Step 4.4: Identify boundary points and noise points, and traverse all non-core points: For non-core points If any core point exists Make Then it is not the core point As a boundary point, it is classified as a core point. The cluster in which it is located; for non-core points If there is no core point Make Then it is not the core point Noise point; Step 4.5: Terminate the algorithm. After the neighborhood relationships of all core objects have been traversed, the algorithm terminates and outputs the clustering results.

[0014] In step 4, an L-shaped rectangle search algorithm is used to extract rectangular features from the scattering center of the sail, and the axis direction and length are calculated. The process is as follows: Input: The set of points with the minimum distance between the longer and shorter sides of a rectangle. , The long and short sides of the rectangle and , Output: The minimum sum of distances from all points to the long and short sides of the rectangle to be fitted. Step 4.6: Based on the L-shaped rectangle model assumption, for each set of points, use the least squares method to find the rectangle that best fits these points, ensuring that the sum of the distances from all points to the long and short sides of the rectangle to be fitted is as small as possible. This involves the following optimization problem: , in Indicates the coordinates of the input data points. Indicates the first The index of each point, The principal orientation angle representing the L-shaped structure is in the first term. It is the normal vector of the first straight line, i.e., the longer side of the rectangle, in the second term. It is the normal vector of the second straight line, i.e., the shorter side of the rectangle; these two vectors are orthogonal, ensuring that the fitted shape is a right-angled L-shape; These represent the projected distances from the origin of the coordinate system to the boundaries of the two lines mentioned above, respectively, and their values ​​range from the set of real numbers. ; Indicates will Each observation point is assigned to an index set of either the long side or the short side of a rectangle, satisfying the following condition: ; Step 4.7: Traverse all possible directions of the rectangle and find a rectangle that contains all the points in each direction; Step 4.8: Calculate the distance from all points to the four sides of the rectangle, and divide the points into sets based on the distances. and And calculate the corresponding squared error as the objective function of the formula in step 4.7; Step 4.9: Repeat steps 4.7 and 4.8 to obtain the squared errors of rectangles in all directions, find the optimal direction that yields the minimum squared error, and fit a rectangle based on the optimal direction.

[0015] In step 5, a satellite target orbital plane coordinate system O-XYZ is established, where OZ points towards the Earth's center. The plane formed by OZ and the satellite target's direction of motion is called the orbital plane. OX lies within the orbital plane and points towards the satellite target's direction of motion. The OY direction is determined using the right-hand rule. The instantaneous radar line of sight is then obtained. ,in , Let be the radar line-of-sight azimuth and elevation angles in the orbital plane coordinate system, respectively. Let the unit direction vector of any linear structure in the target in three-dimensional space be: ,in The pitch angle of this linear structure; According to the projection theory in ISAR imaging, the geometric representation of a target's three-dimensional structure in the two-dimensional ISAR imaging plane can be considered as its projection result in the radar radial direction and the equivalent Doppler direction; whereby the projection component of the structure on the i-axis in the range dimension is expressed as: , In terms of orientation The projection components on the axis are approximately expressed as: , in, , Instantaneous radar line of sight Azimuth and elevation angles , They are respectively , The derivative with respect to time; The unit direction vector representing the linear structure Projection in the radar equivalent Doppler direction; The projection components in the radar radial direction and the equivalent Doppler direction are described above. and After scale calibration, it is mapped to range and azimuth coordinates in the ISAR imaging plane, thereby establishing a linear relationship between the three-dimensional structure orientation vector and the two-dimensional ISAR imaging coordinates; let This represents the two-dimensional projection of the linear structure extracted from the satellite target onto the ISAR image, where... and Corresponding to the distance and azimuth coordinates respectively, we have: , Where the projection matrix Defined as: , in, For distance resolution, For azimuth resolution, It's the speed of light. It is the radar bandwidth. It's the wavelength. It is the rotation angle of the radar ray during the observation period; Let the azimuth and elevation angles of the radar line of sight at the start of the imaging accumulation time be respectively... and At the end of the time, the azimuth and elevation angles of the radar line of sight are respectively and Therefore, at the initial moment, the radar ray vector is: , At the end time, the radar ray vector is: , Then the rotation angle of the radar ray at the start and end times for: , Let the absolute length of any linear structure vector be L, and its corresponding unit direction vector in three-dimensional space be... Then the projection vector in a two-dimensional ISAR image is represented as: , Extracting the projection features of this structure from ISAR images ,definition and The distance between them is the optimization function: , in This represents the number of ISAR image frames involved in the optimization, and the function is minimized using the particle swarm optimization algorithm. The estimated values ​​of the attitude angles are obtained, thereby achieving attitude estimation of the satellite target.

[0016] Beneficial Effects: This invention provides a satellite target attitude estimation method based on geometric feature constraints in ISAR images, achieving attitude estimation of satellite targets. First, a Convolutional Long Short-Term Memory (Conv-LSTM) network is used to accurately segment satellite target components, effectively improving the extraction accuracy of these components. Then, an attribute scattering center model is used to extract the scattering centers of key components, and Principal Component Analysis (PCA) is used to estimate the direction and length of the satellite's main axis. DBSCAN clustering and an L-shaped rectangle search algorithm are used to extract the axis features of the solar panels. Finally, a projection geometric model of the target from three-dimensional to two-dimensional ISAR imaging plane is constructed by combining radar line-of-sight information, and an optimized objective function is established to solve for the satellite attitude angles. This invention fully integrates the spatiotemporal information of sequential images with the geometric constraints of the target, effectively overcoming the projection changes caused by target motion in ISAR imaging, achieving high-precision satellite attitude estimation, and possessing strong anti-interference capabilities and practical application value.

[0017] Low Earth Orbit (LEO) satellite simulation was performed using STK with a target satellite orbital altitude of 300 km. ISAR satellite images were generated through simulation using a satellite CAD 3D model. Electromagnetic simulation was performed using satellite orbit data and a Jason-3 satellite model to obtain echo data, followed by high-resolution ISAR imaging using the RD algorithm. A typical component segmentation method based on Conv-LSTM and an improved RELAX algorithm were used to estimate attribute scattering center parameters, obtaining the scattering center positions after component segmentation. Principal axis features of the satellite body were extracted, and the direction and length of the principal axis were estimated using PCA technology. Clustering of the extracted scattering centers from the solar panel was performed, followed by L-shaped shape fitting to extract the solar panel's axis features. Parameter estimation was performed using an optimization function constructed based on the projection matrix. Attitude estimation results for the satellite body and solar panel were obtained by constructing multi-frame projection geometric constraint equations. Attached Figure Description

[0018] Figure 1 This is a flowchart of the satellite target attitude estimation process for sequential ISAR images based on projection geometry optimization proposed in this invention.

[0019] Figure 2 These are the first and second frames of ISAR images from the Jason-3 satellite.

[0020] Figure 3 Location of the attribute scattering center after ISAR image segmentation; Figure 3 (a) in the image is the location of the attribute scattering center after the segmentation of the first frame of the ISAR image; Figure 3 (b) in the image is the location of the attribute scattering center after the segmentation of the second frame ISAR image.

[0021] Figure 4The results of feature extraction for typical components of the Jason-3 satellite; Figure 4 (a) in the image represents the feature extraction result of the first frame. Figure 4 (b) in the image is the feature extraction result of the component in the second frame.

[0022] Figure 5 This is a schematic diagram of the preset satellite coordinates.

[0023] Figure 6 This is the satellite attitude estimation result of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: This invention can be implemented in many different forms and should not be considered as limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully express the scope of the invention to those skilled in the art.

[0025] like Figure 1 As shown, Figure 1 This is a schematic diagram of the overall process of the ISAR image satellite target attitude estimation method based on geometric feature constraints proposed in this invention, which specifically includes the following steps: Step 1: The typical component segmentation method of sequential ISAR image based on convolutional long short-term memory (Conv-LSTM) network is used to segment typical satellite components; Specifically, the method includes the following steps: a typical component segmentation method for sequential ISAR images based on Convolutional Long Short-Term Memory (Conv-LSTM) networks. The input is a sequential ISAR image, and features are extracted through the backbone network model EfficientNet-B4. Low-level features are directly fed into the decoder, while high-level features are processed by the dilated spatial convolution pooling (ASPP) module and then input together with the low-level features into the Conv-LSTM module to fuse spatiotemporal information. Finally, the component segmentation results of the satellite body and the satellite solar panels are obtained.

[0026] Step 2: Establish a simplified attribute scattering center model, introduce sparse representation theory to improve the traditional relaxation algorithm RELAX, and transform the parameter estimation problem into a sparse coefficient vector under a dictionary set. The norm optimization problem involves alternating updates to the position and amplitude parameters of the scattering center, and through iteration, the estimation results gradually converge to the maximum likelihood solution. The method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints is characterized by the following steps: Step 2, which improves the traditional relaxation algorithm RELAX, involves transforming the parameter estimation problem into solving a sparse coefficient vector norm optimization problem, and then alternately updating the scattering center position and amplitude parameters, causing the estimation results to gradually converge to the maximum likelihood solution. The method includes the following steps: Step 2.1: Establish a simplified property scattering center model, targeting the first... A scattering center, its scattering response model The expression is: , in, It is a scattering center model. It's the azimuth. It's the radar frequency. It's the speed of light. This represents additive white Gaussian noise. Indicates the first The index of each point, The imaginary unit is represented for each scattering center. This represents the set of its corresponding attribute scattering parameters, where, Represents the range, Indicates the length of the scattering center. It is the scattering center direction angle. , Represents the position coordinates of the scattering center in the azimuth and range directions; By using sparse signal decomposition theory to extract the target's attribute scattering centers, the model can be represented as: Equation 2, in This is the result of vectorizing the scattering center. It is the dictionary set used throughout the parameter extraction process. It is a sparse coefficient vector. It's noise; Step 2.2: Transform the parameter estimation problem into solving for a sparse coefficient vector. Norm optimization is a problem of solving nondeterministic polynomial problems. Formula 3, in express Norm, express Norm, This indicates the parameter used to find the minimum value of the above expression within a user-defined domain. This represents the sparse vector of coefficients obtained from the solution. Indicates noise level; Step 2.3: Let Initialization is performed, zero-padding is applied to the signal in the frequency and angle domains, and then an inverse fast Fourier transform (IFFT) is used to obtain the position coordinate parameters of the first scattering center in the azimuth and range directions. A rough estimate, substituted into Equation 3, yields the length and orientation angle parameters of the first scattering center. A rough estimate, reconstructed from Equation 1, of the signal ; Step 2.4: Let Subtract the reconstructed signal from the original signal. After obtaining the remaining signal, the length, direction angle, and position coordinates of the second scattering center in the azimuth and range directions are obtained according to the method in step 3). Substitute and recalculate Then, the parameters are re-estimated using Equation 2. Repeat these two steps until convergence. Step 2.5: Let Similarly, subtract the reconstructed signal from the original signal. The remaining signal is obtained, and the length, direction angle, and position coordinates of the third scattering center in the azimuth and range directions are determined. The estimate, using and Recalculate At the same time, the parameters are re-estimated using formula (2). Similarly, based on the re-estimated parameters and Recalculate Re-estimate parameters Iterate these three steps together until convergence. Steps 2.3 to 2.5 above constitute a cyclic iterative process based on the increment of the index variable i. The iteration terminates when the energy difference between two adjacent remaining signals accounts for no more than 1% of the energy of the original signal, or when the preset maximum number of iterations is reached.

[0027] Step 3: Extract the principal axis features of the satellite body, and estimate the orientation and length of the satellite body based on principal component analysis (PCA); The method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints is characterized by the following steps: Step 3 includes estimating the orientation and length of the satellite body using Principal Component Analysis (PCA); estimating its principal axis vector using the spatial distribution coordinates of the scattering center of the satellite body in the ISAR image and the PCA algorithm; including the following steps: Step 3.1: Extracting the main body of the satellite Coordinates of each scattering center The position matrix formed is , , Center all samples: , in, , Step 3.2: Calculate the sample covariance matrix: , Step 3.3: Eigenvalue decomposition of the covariance matrix: , in, It is a matrix. It's another matrix, with superscript... For transpose; Step 3.4: Based on the matrix sum matrix The first principal component is determined, and the unit vector corresponding to the projection direction of the satellite body on the ISAR two-dimensional imaging plane is obtained. ; Step 3.5: Obtain the projection length and project all scattering centers onto the unit vector. The projection value of each point is obtained, thus yielding the projected length of the satellite body. , .

[0028] Step 4: Perform clustering on the extracted satellite solar panel scattering centers to find two point sets, and then use the L-shaped rectangle search algorithm to extract the axis features of the satellite solar panel. The method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints is characterized by the following steps: Step 4, which uses the density clustering (DBSCAN) algorithm to cluster the extracted satellite solar panel scattering centers, includes the following steps: Input parameters: sample point set neighborhood radius Sample points are the core points Points within the neighborhood Core clustering Number of clusters Non-core points The core point Points in the neighborhood Distance between , the minimum number of points ; Output: Set of clusters and set of noise points The cluster set includes core points and boundary points; Initialization: All points are marked as unprocessed; Step 4.1: Calculate the neighborhood radius , representing the sample point set Center and core point The distance is not greater than the neighborhood radius. For a subset of samples, Euclidean distance is used as the distance metric, expressed as: , Step 4.2: Determine the core point. When the core point... neighborhood radius At least contains When there are 1 sample: , This sample point is the core point; Step 4.3: Construct core point clusters. Based on density reachability, traverse all core points and process each unprocessed current core point. : a) Create a new cluster Add the core point P to the new cluster And mark it as processed; b) Density expansion: For new clusters Each core point Check its neighborhood radius All points within: If the points in the neighborhood If the core point is not processed, then the points in the neighborhood will be visited. Add a new cluster And mark it as processed to achieve indirect density reach; right Repeat the expansion process in step 2 until the current new cluster is reached. There are no unprocessed core points within the neighborhood of any of the core points in the process; c) The new cluster Add to cluster set Let the number of clusters be The count increments by 1, that is... , Step 4.4: Identify boundary points and noise points, and traverse all non-core points: For non-core points If any core point exists Make Then it is not the core point As a boundary point, it is classified as a core point. The cluster in which it is located; for non-core points If there is no core point Make Then it is not the core point Noise point; Step 4.5: Terminate the algorithm. After the neighborhood relationships of all core objects have been traversed, the algorithm terminates and outputs the clustering results.

[0029] The method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints is characterized by the following steps: In step 4, an L-shaped rectangle search algorithm is used to extract rectangular features from the scattering center of the solar panel, and the axial direction and length are determined. Input: The set of points with the minimum distance between the longer and shorter sides of a rectangle. , The long and short sides of the rectangle and , Output: The minimum sum of distances from all points to the long and short sides of the rectangle to be fitted. Step 4.6: Based on the L-shaped rectangle model assumption, for each set of points, use the least squares method to find the rectangle that best fits these points, ensuring that the sum of the distances from all points to the long and short sides of the rectangle to be fitted is as small as possible. This involves the following optimization problem: , in Indicates the coordinates of the input data points. Indicates the first The index of each point, The principal orientation angle representing the L-shaped structure is in the first term. It is the normal vector of the first straight line, i.e., the longer side of the rectangle, in the second term. It is the normal vector of the second straight line, i.e., the shorter side of the rectangle; these two vectors are orthogonal, ensuring that the fitted shape is a right-angled L-shape; These represent the projected distances from the origin of the coordinate system to the boundaries of the two lines mentioned above, respectively, and their values ​​range from the set of real numbers. ; Indicates will Each observation point is assigned to an index set of either the long side or the short side of a rectangle, satisfying the following condition: ; Step 4.7: Traverse all possible directions of the rectangle and find a rectangle that contains all the points in each direction; Step 4.8: Calculate the distance from all points to the four sides of the rectangle, and divide the points into sets based on the distances. and And calculate the corresponding squared error as the objective function of the formula in step 4.7; Step 4.9: Repeat steps 4.7 and 4.8 to obtain the squared errors of rectangles in all directions, find the optimal direction that yields the minimum squared error, and fit a rectangle based on the optimal direction.

[0030] Step 5: Based on the elevation and azimuth angles corresponding to the radar line of sight, construct the projection matrix from the three-dimensional coordinates of the spatial target scattering center to the ISAR two-dimensional imaging plane. On this basis, establish the constraint relationship between the projection of two-dimensional components in the ISAR image and the three-dimensional spatial attitude. Accurate estimation of the target attitude is achieved by solving the joint optimization equation of geometric features.

[0031] The method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints is characterized by the following steps: establishing a satellite target orbital plane coordinate system O-XYZ, where OZ points towards the Earth's center, the plane formed by OZ and the satellite target's motion direction is called the orbital plane, OX lies within the orbital plane and points towards the satellite target's motion direction, and the OY direction is determined using the right-hand rule; obtaining the instantaneous radar line of sight: ,in , Let be the radar line-of-sight azimuth and elevation angles in the orbital plane coordinate system, respectively. Let the unit direction vector of any linear structure in the target in three-dimensional space be: ,in The pitch angle of this linear structure; According to the projection theory in ISAR imaging, the geometric representation of a target's three-dimensional structure in the two-dimensional ISAR imaging plane can be considered as its projection result in the radar radial direction and the equivalent Doppler direction; whereby the projection component of the structure on the i-axis in the range dimension is expressed as: , In terms of orientation The projection components on the axis are approximately expressed as: , in, , Instantaneous radar line of sight Azimuth and elevation angles , They are respectively , The derivative with respect to time; The unit direction vector representing the linear structure Projection in the radar equivalent Doppler direction; The projection components in the radar radial direction and the equivalent Doppler direction are described above. and After scale calibration, it is mapped to range and azimuth coordinates in the ISAR imaging plane, thereby establishing a linear relationship between the three-dimensional structure orientation vector and the two-dimensional ISAR imaging coordinates; let This represents the two-dimensional projection of the linear structure extracted from the satellite target onto the ISAR image, where... and Corresponding to the distance and azimuth coordinates respectively, we have: , Where the projection matrix Defined as: , in, For distance resolution, For azimuth resolution, It's the speed of light. It is the radar bandwidth. It's the wavelength. It is the rotation angle of the radar ray during the observation period; Let the azimuth and elevation angles of the radar line of sight at the start of the imaging accumulation time be respectively... and At the end of the time, the azimuth and elevation angles of the radar line of sight are respectively and Therefore, at the initial moment, the radar ray vector is: , At the end time, the radar ray vector is: , Then the rotation angle of the radar ray at the start and end times for: , Let the absolute length of any linear structure vector be L, and its corresponding unit direction vector in three-dimensional space be... Then the projection vector in a two-dimensional ISAR image is represented as: , Extracting the projection features of this structure from ISAR images ,definition and The distance between them is the optimization function: , in This represents the number of ISAR image frames involved in the optimization, and the function is minimized using the particle swarm optimization algorithm. The estimated values ​​of the attitude angles are obtained, thereby achieving attitude estimation of the satellite target.

[0032] To verify the beneficial effects of the present invention, the following experiments were conducted: Low Earth Orbit (LEO) satellite simulation was performed using STK with a target satellite orbital altitude of 300 km. ISAR satellite images were generated through simulation using a satellite CAD 3D model. Electromagnetic simulation was performed using satellite orbit data and a Jason-3 satellite model to obtain echo data, followed by high-resolution ISAR imaging using the RD algorithm. A typical component segmentation method based on Conv-LSTM and an improved RELAX algorithm were used to estimate attribute scattering center parameters, obtaining the scattering center positions after component segmentation. Principal axis features of the satellite body were extracted, and the direction and length of the principal axis were estimated using PCA. Clustering of the extracted scattering centers from the solar panel was performed, followed by L-shaped shape fitting to extract the solar panel's axis features. Parameter estimation was performed using an optimization function constructed based on the projection matrix. Attitude estimation results for the satellite body and solar panel were obtained by constructing multi-frame projection geometric constraint equations.

[0033] The experimental results are as follows: ISAR imaging results, refer to... Figure 2 The location of the scattering center after component segmentation is referenced. Figure 3 The estimated direction and length of the main shaft, and the axial characteristics of the sail, are referenced. Figure 4 The attitude estimation results for the main body and the sail are referenced. Figure 6 The estimation results are shown in Table 1 below. The estimation errors for the azimuth and pitch angles of the main body are both around 2°, while the attitude estimation error for the solar panel is only around 1°. This indicates that the method can effectively recover the spatial attitude information of the satellite target even under complex imaging conditions, verifying its robustness and accuracy.

[0034] Table 1. Estimation results of attitude parameters for typical components of the Jason-3 satellite. .

[0035] It should be noted that the above-described embodiments only illustrate some implementation methods of the present invention, and their description should not be construed as limiting the scope of the present invention. It should be pointed out that those skilled in the art can make several improvements without departing from the concept of the present invention, and these improvements should all fall within the protection scope of the present invention.

Claims

1. A method for estimating the attitude of satellite targets in ISAR images based on geometric feature constraints, characterized in that: Includes the following steps, Step 1: The typical component segmentation method of sequential ISAR image based on convolutional long short-term memory (Conv-LSTM) network is used to segment typical satellite components; Step 2: Establish a simplified attribute scattering center model, introduce sparse representation theory to improve the traditional relaxation algorithm RELAX, and transform the parameter estimation problem into a sparse coefficient vector under a dictionary set. The norm optimization problem involves alternating updates to the position and amplitude parameters of the scattering center, and through iteration, the estimation results gradually converge to the maximum likelihood solution. Step 3: Extract the principal axis features of the satellite body, and estimate the orientation and length of the satellite body based on principal component analysis (PCA); Step 4: Perform clustering on the extracted satellite solar panel scattering centers to find two point sets, and then use the L-shaped rectangle search algorithm to extract the axis features of the satellite solar panel. Step 5: Based on the elevation and azimuth angles corresponding to the radar line of sight, construct the projection matrix from the three-dimensional coordinates of the spatial target scattering center to the ISAR two-dimensional imaging plane. On this basis, establish the constraint relationship between the projection of two-dimensional components in the ISAR image and the three-dimensional spatial attitude. Accurate estimation of the target attitude is achieved by solving the joint optimization equation of geometric features.

2. The ISAR image satellite target attitude estimation method based on geometric feature constraints according to claim 1, characterized in that: In step 1, the typical component segmentation method for sequential ISAR images based on the Convolutional Long Short-Term Memory (Conv-LSTM) network is used. The input is a sequential ISAR image, and features are extracted through the backbone network model EfficientNet-B4. Low-level features are directly fed into the decoder, while high-level features are processed by the dilated spatial convolution pooling (ASPP) module and then input together with the low-level features into the Conv-LSTM module to fuse spatiotemporal information. Finally, the component segmentation results of the satellite body and the satellite solar panels are obtained.

3. The ISAR image satellite target attitude estimation method based on geometric feature constraints according to claim 1, characterized in that: Step 2 of the improved traditional relaxation algorithm RELAX includes: transforming the parameter estimation problem into solving a sparse coefficient vector norm optimization problem, and then alternately updating the scattering center position and amplitude parameters, so that the estimation results gradually converge to the maximum likelihood solution; including the following steps: Step 2.1: Establish a simplified property scattering center model, targeting the first... A scattering center, its scattering response model The expression is: ， in, It is a scattering center model. It's the azimuth. It's the radar frequency. It's the speed of light. This represents additive white Gaussian noise. Indicates the first The index of each point, The imaginary unit is represented for each scattering center. This represents the set of its corresponding attribute scattering parameters, where, Represents the range, Indicates the length of the scattering center. It is the scattering center direction angle. , Represents the position coordinates of the scattering center in the azimuth and range directions; By using sparse signal decomposition theory to extract the target's attribute scattering centers, the model can be represented as: Equation 2, in This is the result of vectorizing the scattering center. It is the dictionary set used throughout the parameter extraction process. It is a sparse coefficient vector. It's noise; Step 2.2: Transform the parameter estimation problem into solving for a sparse coefficient vector. Norm optimization is a problem of solving nondeterministic polynomial problems. Formula 3, in express Norm, express Norm, This indicates the parameter used to find the minimum value of the above expression within a user-defined domain. This represents the sparse vector of coefficients obtained from the solution. Indicates noise level; Step 2.3: Let Initialization is performed, zero-padding is applied to the signal in the frequency and angle domains, and then an inverse fast Fourier transform (IFFT) is used to obtain the position coordinate parameters of the first scattering center in the azimuth and range directions. A rough estimate, substituted into Equation 3, yields the length and orientation angle parameters of the first scattering center. A rough estimate, reconstructed from Equation 1, of the signal ; Step 2.4: Let Subtract the reconstructed signal from the original signal. After obtaining the remaining signal, the length, direction angle, and position coordinates of the second scattering center in the azimuth and range directions are obtained according to the method in step 3). Substitute and recalculate Then, the parameters are re-estimated using Equation 2. Repeat these two steps until convergence. Step 2.5: Let Similarly, subtract the reconstructed signal from the original signal. The remaining signal is obtained, and the length, direction angle, and position coordinates of the third scattering center in the azimuth and range directions are determined. The estimate, using and Recalculate At the same time, the parameters are re-estimated using formula (2). Similarly, based on the re-estimated parameters and Recalculate Re-estimate parameters Iterate these three steps together until convergence. Steps 2.3 to 2.5 above constitute a cyclic iterative process based on the increment of the index variable i. The iteration terminates when the energy difference between two adjacent remaining signals accounts for no more than 1% of the energy of the original signal, or when the preset maximum number of iterations is reached.

4. The ISAR image satellite target attitude estimation method based on geometric feature constraints according to claim 1, characterized in that: Step 3 includes: estimating the orientation and length of the satellite body using Principal Component Analysis (PCA); estimating its principal axis vector using the PCA algorithm based on the spatial distribution coordinates of the satellite body's scattering center in the ISAR image; including the following steps: Step 3.1: Extracting the main body of the satellite Coordinates of each scattering center The position matrix formed is , ， Center all samples: ， in, ， Step 3.2: Calculate the sample covariance matrix: ， Step 3.3: Eigenvalue decomposition of the covariance matrix: ， in, It is a matrix. It's another matrix, with superscript... For transpose; Step 3.4: Based on the matrix sum matrix The first principal component is determined, and the unit vector corresponding to the projection direction of the satellite body on the ISAR two-dimensional imaging plane is obtained. ; Step 3.5: Obtain the projection length and project all scattering centers onto the unit vector. The projection value of each point is obtained, thus yielding the projected length of the satellite body. , 。 5. The ISAR image satellite target attitude estimation method based on geometric feature constraints according to claim 1, characterized in that: In step 4, the density clustering (DBSCAN) algorithm is used to cluster the extracted satellite solar panel scattering centers. The algorithm implementation process is as follows: Input parameters: Sample point set D, neighborhood radius Sample points are the core points Points within the neighborhood Core clustering Number of clusters Non-core points The core point Points in the neighborhood Distance between , the minimum number of points ; Output: Set of clusters and set of noise points The cluster set includes core points and boundary points; Initialization: All points are marked as unprocessed; Step 4.1: Calculate the neighborhood radius , representing the sample point set Center and core point The distance is not greater than the neighborhood radius. For a subset of samples, Euclidean distance is used as the distance metric, expressed as: ， Step 4.2: Determine the core point. When the core point... neighborhood radius When the sample contains at least MinPts samples: ， This sample point is the core point; Step 4.3: Construct core point clusters. Based on density reachability, traverse all core points and process each unprocessed current core point. : a) Create a new cluster The core point Add a new cluster And mark it as processed; b) Density expansion: For new clusters Each core point Check its neighborhood radius All points within: If the points in the neighborhood If the core point is not processed, then the points in the neighborhood will be visited. Add a new cluster And mark it as processed to achieve indirect density reach; right Repeat the expansion process in step 2 until the current new cluster is reached. There are no unprocessed core points within the neighborhood of any of the core points in the process; c) The new cluster Add to cluster set Let the number of clusters be The count increments by 1, that is... , Step 4.4: Identify boundary points and noise points, and traverse all non-core points: For non-core points If any core point exists Make Then it is not the core point As a boundary point, it is classified as a core point. The cluster in which it is located; for non-core points If there is no core point Make Then it is not the core point Noise point; Step 4.5: Terminate the algorithm. After the neighborhood relationships of all core objects have been traversed, the algorithm terminates and outputs the clustering results.

6. The method for estimating the attitude of ISAR image satellite targets based on geometric feature constraints according to claim 1, characterized in that: In step 4, an L-shaped rectangle search algorithm is used to extract rectangular features from the scattering center of the sail, and the axis direction and length are calculated. The process is as follows: Input: The set of points with the minimum distance between the longer and shorter sides of a rectangle. , The long and short sides of the rectangle and , Output: The minimum sum of distances from all points to the long and short sides of the rectangle to be fitted. Step 4.6: Based on the L-shaped rectangle model assumption, for each set of points, use the least squares method to find the rectangle that best fits these points, ensuring that the sum of the distances from all points to the long and short sides of the rectangle to be fitted is as small as possible. This involves the following optimization problem: ， in Indicates the coordinates of the input data points. Indicates the first The index of each point, The principal orientation angle representing the L-shaped structure is in the first term. It is the normal vector of the first straight line, i.e., the longer side of the rectangle, in the second term. It is the normal vector of the second straight line, i.e., the shorter side of the rectangle; these two vectors are orthogonal, ensuring that the fitted shape is a right-angled L-shape; These represent the projected distances from the origin of the coordinate system to the boundaries of the two lines mentioned above, respectively, and their values ​​range from the set of real numbers. ; Indicates will Each observation point is assigned to an index set of either the long side or the short side of a rectangle, satisfying the following condition: ; Step 4.7: Traverse all possible directions of the rectangle and find a rectangle that contains all the points in each direction; Step 4.8: Calculate the distance from all points to the four sides of the rectangle, and divide the points into sets based on the distances. and And calculate the corresponding squared error as the objective function of the formula in step 4.7; Step 4.9: Repeat steps 4.7 and 4.8 to obtain the squared errors of rectangles in all directions, find the optimal direction that yields the minimum squared error, and fit a rectangle based on the optimal direction.

7. The method for estimating the attitude of ISAR image satellite targets based on geometric feature constraints according to claim 1, characterized in that: In step 5, a satellite target orbital plane coordinate system O-XYZ is established, where OZ points towards the Earth's center. The plane formed by OZ and the satellite target's direction of motion is called the orbital plane. OX lies within the orbital plane and points towards the satellite target's direction of motion. The OY direction is determined using the right-hand rule. The instantaneous radar line of sight is then obtained. ,in , Let be the radar line-of-sight azimuth and elevation angles in the orbital plane coordinate system, respectively. Let the unit direction vector of any linear structure in the target in three-dimensional space be: ,in The pitch angle of this linear structure; According to the projection theory in ISAR imaging, the geometric representation of a target's three-dimensional structure in the two-dimensional ISAR imaging plane can be considered as its projection result in the radar radial direction and the equivalent Doppler direction; where the projection component of the structure on the i-axis in the range dimension is expressed as: ， In terms of orientation The projection components on the axis are approximately expressed as: ， in, , Instantaneous radar line of sight Azimuth and elevation angles , They are respectively , The derivative with respect to time; The unit direction vector representing the linear structure Projection in the radar equivalent Doppler direction; The projection components in the radar radial direction and the equivalent Doppler direction are described above. and After scale calibration, it is mapped to range and azimuth coordinates in the ISAR imaging plane, thereby establishing a linear relationship between the three-dimensional structure orientation vector and the two-dimensional ISAR imaging coordinates; let This represents the two-dimensional projection of the linear structure extracted from the satellite target onto the ISAR image, where... and Corresponding to the distance and azimuth coordinates respectively, we have: ， Where the projection matrix Defined as: ， in, For distance resolution, For azimuth resolution, It's the speed of light. It is the radar bandwidth. It's the wavelength. It is the rotation angle of the radar ray during the observation period; Let the azimuth and elevation angles of the radar line of sight at the start of the imaging accumulation time be respectively... and At the end of the time, the azimuth and elevation angles of the radar line of sight are respectively and Therefore, at the initial moment, the radar ray vector is: ， At the end time, the radar ray vector is: ， Then the rotation angle of the radar ray at the start and end times for: ， Let the absolute length of any linear structure vector be L, and its corresponding unit direction vector in three-dimensional space be... Then the projection vector in a two-dimensional ISAR image is represented as: ， Extracting the projection features of this structure from ISAR images ,definition and The distance between them is the optimization function: ， in This represents the number of ISAR image frames involved in the optimization, and the function is minimized using the particle swarm optimization algorithm. The estimated values ​​of the attitude angles are obtained, thereby achieving attitude estimation of the satellite target.

Images