Non-cooperative spacecraft attitude estimation method based on contour and linear structure

Abstract

The invention discloses a non-cooperative spacecraft attitude estimation method based on a contour and a linear structure, and the method comprises the steps: S1, enabling a target three-dimensional model to rotate and project on a 2D plane with a mass center as a center, so as to build an attitude template library; s2, performing noise filtering on the input original ISAR image by using a filter, redistributing the image brightness, and enhancing the image edge to obtain an enhanced image Img1; s3, introducing an attention module (CBAM) into the image segmentation network to extract a contour image Img2 of the target from the Img1; s4, contour key points are extracted from the Img2, a contour image Img3 with the key points is obtained, and then a linear structure image Img4 of the target spacecraft is extracted; and S5, performing similarity comparison on the Img4 and the attitude template library to obtain the attitude of the target spacecraft. The invention provides a non-cooperative spacecraft attitude estimation method based on a contour and a linear structure. Compared with deep learning and traditional template matching, three attitude angles including a pitch angle, a rolling angle and a yaw angle extracted from an ISAR image have smaller errors.

Description

Technical Field

[0001] The present invention relates to the field of spacecraft navigation, and more specifically, to a non-cooperative spacecraft attitude estimation method based on contours and linear structures on ISAR images and target three-dimensional models. Background Art

[0002] Non-cooperative spacecraft refers to spacecraft that do not have the ability to cooperate in orbit or are not specially designed for docking, maintenance, capture and other operations. They may be abandoned satellites, failed spacecraft, space debris or other spacecraft that are not actively equipped with docking hardware and communication interfaces. Due to the unpredictable behavior and status, it brings additional difficulties to tasks such as on-orbit servicing and space exploration.

[0003] The signal processing technology of Inverse Synthetic Aperture Radar (ISAR) is used for imaging of moving targets in the range-Doppler (or range and azimuth) domain and is currently widely used in spacecraft pose estimation.

[0004] In practical applications, the position and attitude estimation of spacecraft is the basis and key technology of space on-orbit services, and is one of the important tasks of satellite navigation and space maintenance. It helps to simulate the attitude evolution of retired objects and realize the identification, detection and tracking of objects. Especially when the motion state of an object is unclear or irregular, its force condition, motion trajectory, etc. can be predicted with the help of attitude estimation. Relevant personnel can more effectively design and optimize satellite orbits and control systems, so as to better complete space missions and more effectively prevent potential safety risks.

[0005] However, due to the lack of effective information exchange between non-cooperative spacecraft, it is impossible to directly return its own status and operation information through sensors. Therefore, there are currently two main methods for estimating the attitude of non-cooperative spacecraft:

[0006] The first is pose measurement based on deep learning. However, the application of deep learning to realize non-cooperative target pose estimation requires a large amount of real data for training the model. Problems such as small number of samples and sparse features make it difficult to use deep learning for non-cooperative spacecraft pose estimation on ISAR images.

[0007] The second is the structural feature template matching method, which uses a pre-established template library to compare the similarity between the input image and the predefined template to estimate the object's posture. It is relatively direct and easy to implement in simple application scenarios, but speckle noise will appear in the imaging process of ISAR images and a large amount of details will be lost. Image distortion caused by motion will also cause blurred edges of the spacecraft. Traditional template matching methods are sensitive to feature sparseness and noise, and are usually used for rough estimation of posture. Summary of the invention

[0008] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0009] In order to achieve these objectives and other advantages of the present invention, a non-cooperative spacecraft attitude estimation method based on contour and linear structure is provided, comprising:

[0010] S1, rotating and projecting the target 3D model onto a 2D plane with the center of mass as the center to establish a posture template library;

[0011] S2, using a filter to filter out noise from the input original ISAR image, redistribute the image brightness, and enhance the image edge to obtain an enhanced image Img1;

[0012] S3, introduce the attention module CBAM into the image segmentation network to extract the target contour image Img2 from Img1;

[0013] S4, extracting contour key points from Img2 to obtain contour image Img3 with key points, and then extracting linear structure image Img4 of the target spacecraft;

[0014] S5. Perform a similarity comparison between Img4 and the attitude template library to obtain the attitude of the target spacecraft.

[0015] Preferably, in S1, the process of establishing the posture template library includes:

[0016] S10, reading the data of the target three-dimensional model, and using a data structure to store all vertex coordinates (v, x, y, z), where v represents a vertex, and (x, y, z) represents the three-dimensional coordinates of the vertex;

[0017] S11. Set the center of mass V c use The main axis vector and the secondary axis vector of the model are represented by V1=(v 1x ,v 1y ,v 1z )、V2=(v 2x ,v 2y ,v 2z ) is used to indicate;

[0018] S12, the model, the main axis vector, and the secondary axis vector are combined and rotated around the center of mass on the XYZ axis, and projected onto a two-dimensional plane, and the contour, main axis direction, and secondary axis direction of the target three-dimensional model are recorded to establish a posture template library.

[0019] Preferably, in S2, the process of obtaining Img1 includes:

[0020] S20, for each pixel in the original ISAR image, use an exponential weighting function related to the distance between pixels and local statistical characteristics in the filter to perform filtering processing, and the filter output ξ(x, y) is expressed by the following formula;

[0021]

[0022] In the above formula, N(x,y) is the neighborhood window centered at the pixel point (x,y), I(i,j) is the pixel value of the input image at position (i,j), d(i,j) is the Euclidean distance between the input image at position (i,j) and the center image (x,y), α is the adaptive damping coefficient, λ is the weight parameter, and the expression of λ is τ is a constant used to control the sensitivity of the filter, σ(x,y) is the local standard deviation of the neighborhood window centered on the pixel point (x,y), and V mean is the global mean variance of the image;

[0023] S21, redistributing the brightness of the filtered image by contrast-limited adaptive histogram equalization, and the pixel output I(x, y) of the redistributed brightness is expressed as:

[0024]

[0025] In the above formula, is the histogram density function in the local area of ​​the pixel, and δ is the contrast limit threshold;

[0026] The Img1 is obtained after two processes of filtering and redistributing the brightness.

[0027] Preferably, in S3, the process of obtaining Img2 includes:

[0028] In the original UNet3+ segmentation network model, the attention module CBAM is introduced after each downsampling and upsampling 3x3 convolution operation;

[0029] Img1 is input into the UNet3+ segmentation network model integrated with CBAM, and the feature map is passed to the attention module. The feature map adjusted by the attention mechanism is then passed to the next layer or fused through jump connections to enhance the feature expression to obtain Img2.

[0030] Preferably, in S4, the process of obtaining Img4 includes:

[0031] S40, mark the coordinates of the N contour key points as (x i ,y i ), perform a one-dimensional discrete Fourier transform on each contour key point to obtain the Fourier coefficient Z(k) of the kth frequency component to obtain Img3:

[0032]

[0033] In the above formula, (x i +jy i ) is the complex form of each contour key point, j is the imaginary unit

[0034] S41, spindle direction data θ' main By calculating the phase angle of the lowest frequency component Z(1) of the Fourier transform, we get:

[0035] θ′ main =arg(Z(1))

[0036] Secondary axis direction data θ' sub By calculating the phase angle of the second lowest frequency component Z(2) of the Fourier transform, we get:

[0037] θ′ sub =arg(Z(2))

[0038] S42, record the data of the major and minor axis directions (θ' main ,θ' sub ), and the line vectors are marked with lines on the image to obtain Img4, where the line vectors of the major axis and the minor axis in the linear structure are the farthest distances between the intersection of the major axis direction and the contour edge and the intersection of the minor axis direction and the contour edge, respectively.

[0039] Preferably, in S5, the similarity comparison is achieved by adding a local direction information to constrain the local features based on the shape comparison, then the total similarity measurement matrix C A for:

[0040] C A =(1-γ)D+γC

[0041] In the above formula, D is the shape similarity matrix, C is the direction similarity matrix, γ is the adaptive weight parameter, and the expression of γ is μ is a tuning parameter used to adjust the steepness of the function, and t is the mean of the shape similarity matrix D.

[0042] Preferably, in S5, after the similarity comparison, the images are arranged in descending order according to the total similarity, and the top N images are used as the image sequence to be selected;

[0043] The major and minor axis direction data (θ main ,θ sub ) and the major and minor axis direction data (θ' main ,θ' sub) to compare and calculate the angle difference Diff all The smallest image is used as the target pose estimation result:

[0044] Diff all =Diff main +Diff sub

[0045] Diff main =min(|θ main -θ′ main |,360-|θ mai n-θ′ main |)

[0046] Diff sub =min(|θ sub -θ′ sub |,360-|θ sub -θ′ sub |)

[0047] In the above formula, Diff main is the main axis angle difference, Diff sub The angle difference of the secondary axis is returned. all The smallest image can obtain the optimal spacecraft attitude estimation result.

[0048] The present invention has at least the following beneficial effects: First, in view of the problem of edge blur caused by speckle noise, detail loss and image distortion in ISAR images, the present invention proposes a filtering method that uses weighted averaging of local image regions to adaptively adjust filtering parameters, and reduces the weight of edge pixels to reduce the smoothness of edge regions. The clarity of edge details is maintained in heterogeneous regions, and smoother output is achieved in homogeneous regions. This method can effectively suppress image noise and retain spacecraft components and edge details as much as possible.

[0049] Secondly, in order to solve the problem of sparse spacecraft features on ISAR images, the present invention uses UNet3+ with CBAM attention module to extract spacecraft contour features. In order to improve the segmentation accuracy of image segmentation network UNet3+ applied to ISAR images, the present invention introduces CBAM attention module after each encoding layer and before the decoding layer of the original segmentation network to enhance features and improve segmentation accuracy. And through key point detection and primary and secondary axis recognition, the key points and linear structure features of the spacecraft are extracted.

[0050] Thirdly, in order to solve the problem that the difference between segmented contour and projected contour in template matching will affect the accuracy of similarity comparison results, the present invention proposes a method of adding local directional information on the basis of shape histogram matrix to constrain local features and optimizing the comparison results by linear structure, and performing weighted summation of directional similarity measurement matrix and shape similarity measurement matrix to obtain total similarity measurement matrix, and finally using linear structure extracted from contour to further reduce similarity comparison error. Traditional template matching is usually used for rough estimation of attitude, and the present method can achieve high-precision attitude estimation on ISAR images by combining contour and linear structure for template matching.

[0051] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of a complete process of spacecraft attitude estimation in one embodiment of the present invention;

[0053] Figure 2 is a real ISAR image input in an embodiment of the present invention;

[0054] Figure 3 For Figure 2 The enhanced output image;

[0055] Figure 4 For Figure 3 Schematic diagram of the result after segmentation processing of the image in;

[0056] Figure 5 For Figure 4 Schematic diagram of the result after linear structure extraction of the image in ( Figure 5 The coordinate axes in represent the pixel coordinates of the image, where the x-axis is the column index and the y-axis is the row index);

[0057] Figure 6 The attitude estimation results obtained by using the method of the present invention to conduct experiments on multiple real ISAR images;

[0058] Figure 7 A flowchart of a non-cooperative spacecraft attitude estimation method using the present invention;

[0059] Figure 8 This is a schematic diagram of the network structure of the UNet3+ segmentation network integrated with the CBAM module of the present invention. DETAILED DESCRIPTION

[0060] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0061] A non-cooperative spacecraft attitude estimation method based on contour and linear structure is mainly performed by cutting the spacecraft in the ISAR image and rotating and projecting the target 3D model onto a 2D plane with the center of mass as the center to obtain the spacecraft contour and its linear structure, thereby obtaining the target spacecraft attitude through similarity comparison. Specifically, Figure 7 As shown, it mainly includes:

[0062] S1. Simulate the imaging principle of ISAR, rotate and project the 3D model and axis vector around the center of mass to obtain the target attitude template library.

[0063] Specifically, by reading the spacecraft three-dimensional model data, the data structure is used to store all vertex coordinates. The vertex coordinates are usually represented by (v, x, y, z), where (x, y, z) is the three-dimensional coordinate of the vertex and the centroid is The major and minor axis vectors of the model are denoted as V1=(v 1x ,v 1y ,v 1z ) and V2=(v 2x ,v 2y ,v 2z ). The model and the major and minor axis vectors are rotated around the center of mass on the XYZ axis and projected onto a two-dimensional plane. The rotation matrix expression on each axis is as follows:

[0064]

[0065] The rotation matrix R applied to each vertex and axis vector is a combination of the rotation matrices on different axes, resulting in new point coordinates and axis vectors.

[0066]

[0067] V′1=R·V1

[0068] V′2=R·V2

[0069] Project the rotated model and axis vector onto a two-dimensional plane and record the contour, major and minor axis directions (θ main ,θ sub ), establish a posture template library.

[0070] S2. Enhanced adaptive filtering is used to filter out speckle noise in ISAR images, and contrast-limited adaptive histogram equalization is introduced to redistribute image brightness, and a threshold is set to limit pixel over-enhancement.

[0071] Specifically, the original ISAR image is filtered to remove the image speckle noise, and each pixel in the image is processed. An exponential weighting function related to the distance between pixels and local statistical characteristics is used, so that the filter can be dynamically adjusted according to the local characteristics of the image, thereby achieving different filtering effects on noise in edge areas and non-edge areas, effectively suppressing noise while retaining image edge details and clarity. The image after filtering is expressed as:

[0072]

[0073] Where N(x,y) is the neighborhood window centered at (x,y), I(i,j) is the pixel value of the input image at position (i,j), d(i,j) is the Euclidean distance between the image (i,j) and the center image (x,y), and α is the adaptive damping coefficient.

[0074] λ is the weight parameter, and its expression is:

[0075]

[0076] τ is a constant used to control the sensitivity of the filter. The larger the τ value, the more obvious the smoothing effect of the edge area. The smaller the τ value, the more edge details are retained. σ(x,y) is the local standard deviation of the neighborhood window centered on the pixel (x,y). V mean is the global average variance of the image. For edge pixels, the local standard deviation is large, and λ will be adaptively increased to reduce the smoothness of the edge area and maintain the details and clarity of the edge. For non-edge pixels, the local standard deviation is small, and λ is adaptively reduced to achieve smoother output in homogeneous areas.

[0077] After filtering out the coherent speckle noise caused by multipath reflection and phase interference, in order to better preserve the image details, the image brightness is redistributed by contrast-limited adaptive histogram equalization. A threshold is set to limit the problem of noise amplification caused by excessive enhancement in areas with high contrast. Specifically, the brightness of the filtered image is redistributed by contrast-limited adaptive histogram equalization, making the difference between different brightness in the image more obvious and the distinction between objects and background clearer. The image after redistribution of brightness is expressed as:

[0078]

[0079] in, is the histogram density function in the local area of ​​the pixel, and δ is the contrast limit threshold.

[0080] After two processes of filtering and redistributing brightness, an enhanced image Img1 is obtained which can be used for contour segmentation network. The processed image can more easily extract accurate features in subsequent image segmentation tasks and improve segmentation accuracy.

[0081] S3. Integrate the CBAM attention module after each convolutional layer in the UNet3+ network model to improve the contour segmentation accuracy and detect and extract contour key points.

[0082] like Figure 8 As shown in Figure 1, the enhanced image Img1 obtained in S2 is input into the UNet3+ segmentation network integrated with the CBAM module. The CBAM module is added after each encoding layer and before the decoding layer of the original UNet3+ segmentation network, and the feature map is passed to the attention module. The feature map adjusted by the attention mechanism is then passed to the next layer or fused through jump connections to enhance feature expression.

[0083] The dataset consisting of multiple real ISAR images of spacecraft in different postures is randomly divided into training set and test set in a ratio of 8:2, the batch-size is set to 16, the learning rate is 0.00001, the Focal Loss is used as the network loss function, and the training cycle is 100. The trained segmentation network UNet3+ with integrated CBAM attention module is used to segment the input ISAR image and output the contour image Img2 of the target non-cooperative spacecraft.

[0084] In order to improve the segmentation accuracy of the segmentation network model on ISAR images, the CBAM attention module is introduced after each downsampling and upsampling 3x3 convolution operation of the original UNet3+ network model. The channel attention processes the channel descriptors of the maximum pooling and global average pooling through a multi-layer perceptron, followed by an activation function Sigmoid, and the spatial attention performs global pooling of features on the basis of the channel attention, and uses a 7x7 convolution layer to further extract spatial features, adjust the model's attention to different spatial and channel key information, enhance the model's extraction of image features, and finally generate a feature map to improve the model's segmentation accuracy IoU. Specifically, it is by extracting contour key points from the contour image Img2 in step 3. The pixel coordinates on each contour point are marked as (x c ,y c ), calculate the gradient information I of each pixel position of the contour in the horizontal and vertical directions x and I y . At each pixel position, I x and I y The pixels within the local window are summed to construct the gradient matrix M, which describes the local gradient features of the image.

[0085]

[0086] Solve the gradient matrix M at each pixel to obtain the minimum eigenvalues ​​λ1 and λ2, and convert the minimum eigenvalue λ min It is used as the corner point response value to determine whether the point is a contour key point.

[0087] λ min =min(λ1,λ2)

[0088] Calculate the response value of each contour point, select the points whose response value is greater than the specified threshold as contour key points, and mark them with points in the contour image. The coordinates of each contour key point are recorded as (x i ,y i ) to obtain the contour image Img3 with key point information.

[0089] S4. Use Fourier transform to determine the directions of the major and minor axes and extract the contour linear structure.

[0090] The linear structure of the contour is extracted from the contour image Img3 with key point information obtained from S3 by Fourier transform to obtain the corresponding linear structure image Img4.

[0091] The coordinates of the N contour key points are marked as (x i ,y i ), perform one-dimensional discrete Fourier transform on the contour key points to obtain the Fourier coefficient Z(k) of the kth frequency component:

[0092]

[0093] Where (x i +jy i ) is the complex form of each contour key point, j is the imaginary unit

[0094] The directions of the major and minor axes are usually related to the low-frequency components of the Fourier transform, which represent the overall direction of the contour, so the direction of the major axis θ' main It can be obtained by calculating the phase angle of the lowest frequency component Z(1) of the Fourier transform:

[0095] θ′ main =arg(Z(1))

[0096] Similarly, the secondary axis direction θ' sub It can be obtained by calculating the phase angle of the second lowest frequency component Z(2) of the Fourier transform:

[0097] θ′ sub =arg(Z(2))

[0098] Record the data of the major and minor axis directions (θ'main ,θ' sub ), for easy observation, the line vector can be marked with a line on the image. The line vector of the major and minor axes in the linear structure is the farthest distance between the major and minor axis directions and the intersection of the contour edge.

[0099] S5. Perform template matching. On the basis of the shape histogram matrix, add local direction information to constrain local features, use linear structure to optimize similarity comparison results, and reduce posture errors. That is, perform template matching on the linear structure image Img4 in the posture template library.

[0100] In the template matching method, the distance between matching descriptors is used to calculate the matching cost. The smaller the matching cost, the more similar the images are. The global structural information in each key point on the shape contour is encoded into a feature vector that can describe the relative position relationship between point pairs, which mainly includes two dimensions: distance d and angle θ. d and θ are mapped to the bin filled histogram, and the distance between the two shape histograms is calculated to obtain the shape similarity matrix D. Specifically, several points on the segmentation contour are uniformly sampled and the distance between each point p is calculated. i and all other points p j The relative position relationship between (j≠i) is usually expressed in polar coordinates and encoded into a two-dimensional histogram. The similarity is compared by calculating the chi-square distance between the histograms of the corresponding points of the two shapes. The shape similarity measurement matrix D is:

[0101]

[0102] Among them, h i (k) and h j (k) are the values ​​in the kth bin in the descriptor histogram, and K is the dimension of the descriptor.

[0103] At the same time, a local direction information is added to constrain the local features based on the shape comparison, the gradient of the grayscale image of the input image and the template image is calculated, the tangent angles θ1 and θ2 corresponding to the contour points are obtained, and the tangent angle dissimilarity function is used as the basis for measuring the local direction. The expression of the direction similarity measurement matrix C is:

[0104]

[0105] Due to the loss of details in the real ISAR image, there is a difference in the accuracy of the segmented contour and the projected contour on the main shape of the spacecraft. In order to compare the shape and direction similarity comprehensively, the total similarity measurement matrix C A An adaptive weight parameter γ is introduced into the following calculation formula:

[0106] C A =(1-γ)D+γC

[0107] in, μ is a tuning parameter used to adjust the steepness of the function. t is the mean of the shape similarity matrix D. When the partial shape differences are large, the parameter γ is large, which reduces the influence of the shape similarity matrix on the overall similarity matrix. Then the total similarity measurement matrix C A The calculation formula can be transformed into:

[0108]

[0109] After similarity comparison in the template library, multiple images with the highest total similarity are obtained as the candidate image sequence, and the major and minor axis data in the linear structure of the input image and the candidate image sequence are further compared.

[0110] (θ main ,θ sub ) and (θ′ main ,θ′ sub ):

[0111] Diff main =min(|θ main -θ′ main |,360-|θ main -θ′ main |)

[0112] Diff sub =min(|θ sub -θ′ sub |,360-|θ sub -θ′ sub |)

[0113] Diff all =Diff main +Diff sub

[0114] Diff main is the main axis angle difference, Diff sub The angle difference of the secondary axis is returned. all The smallest image can obtain the optimal spacecraft attitude estimation result.

[0115] Embodiment 1:

[0116] In this example, the three-dimensional model is first rotated and projected with its center of mass as the rotation center to establish a posture template library. A filter is used to filter out noise and redistribute image brightness from the input ISAR image, and image edges are enhanced. The CBAM module is introduced into the image segmentation network to improve the image segmentation quality. On this basis, the target linear structure is extracted. Finally, local direction information is added based on the shape histogram matrix to constrain local features and use the linear structure to improve the accuracy of template matching.

[0117] Taking Tiangong-1 and ISAR images as an example, the complete process of spacecraft attitude estimation is as follows: Figure 1 As shown:

[0118] (1) Read the 3D model data, calculate and record the model center of mass and the model's primary and secondary axis vectors, rotate and project the model and the primary and secondary axis vectors, and establish a posture template library.

[0119] (2) Filter out the input real ISAR image (such as Figure 2 The speckle noise in the image is removed and the image brightness is redistributed to enhance the image quality. The enhanced output image is as follows Figure 3 As shown:

[0120] (3) The dataset is randomly divided into training set and test set in a ratio of 8:2. The batch-size is set to 16, the learning rate is 0.00001, the Focal Loss is used as the network loss function, and the training cycle is 100. The trained segmentation network UNet3+ with integrated CBAM attention module is used to segment the input ISAR image. The segmentation results are shown in Figure 2. Figure 4 As shown:

[0121] (4) Figure 5 As shown, the linear structure extraction result is (63.60, -50.16);

[0122] (5) The segmented image in (3) is compared with the posture template library for similarity, and the linear structure information stored in the model and contour is used to optimize the posture information result finally returned.

[0123] On the same ISAR image, the accuracy of the segmentation network after integrating the CBAM attention module is improved by 3.45% compared with the segmentation network without the integrated attention module.

[0124] Figure 6 In, α i , β i , γ i is the attitude angle error corresponding to the i-th ISAR image. Figure 6It can be seen that compared with other traditional attitude estimation methods, the method of the present invention has a smaller error in extracting the three attitude angles of pitch angle, roll angle and yaw angle on the ISAR image.

[0125] From the above examples we can see that:

[0126] Compared with the original network model, the image segmentation network of the present invention improves the segmentation accuracy of spacecraft on ISAR images by 3.45% after the same parameters and rounds of training.

[0127] The pitch angle, roll angle and yaw angle extracted by the present invention on the ISAR image have smaller errors than those of deep learning and traditional template matching.

[0128] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacement and / or modification can be performed according to user needs.

[0129] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A non-cooperative spacecraft attitude estimation method based on contour and linear structure, characterized in that: include: S1, rotating and projecting the target 3D model onto a 2D plane with the center of mass as the center to establish a posture template library; S2, using a filter to filter out noise from the input original ISAR image, redistribute the image brightness, and enhance the image edge to obtain an enhanced image Img1; S3, introduce the attention module CBAM into the image segmentation network to extract the target contour image Img2 from Img1; S4, extracting contour key points from Img2 to obtain contour image Img3 with key points, and then extracting linear structure image Img4 of the target spacecraft; S5. Perform a similarity comparison between Img4 and the attitude template library to obtain the attitude of the target spacecraft.

2. The non-cooperative spacecraft attitude estimation method based on contour and linear structure as claimed in claim 1, characterized in that: In S1, the process of establishing the posture template library includes: S10, reading the data of the target three-dimensional model, and using a data structure to store all vertex coordinates (v, x, y, z), where v represents a vertex, and (x, y, z) represents the three-dimensional coordinates of the vertex; S11. Set the center of mass V c use The main axis vector and the secondary axis vector of the model are represented by V1=(v 1x ,v 1y ,v 1z )、V2=(v 2x ,v 2y ,v 2z ) is used to indicate; S12, the model, the main axis vector, and the secondary axis vector are combined and rotated around the center of mass on the XYZ axis, and projected onto a two-dimensional plane, and the contour, main axis direction, and secondary axis direction of the target three-dimensional model are recorded to establish a posture template library.

3. The non-cooperative spacecraft attitude estimation method based on contour and linear structure as claimed in claim 1, characterized in that: In S2, the acquisition process of Img1 includes: S20, for each pixel in the original ISAR image, use an exponential weighting function related to the distance between pixels and local statistical characteristics in the filter to perform filtering processing, and the filter output ξ(x, y) is expressed by the following formula; In the above formula, N(x,y) is the neighborhood window centered at the pixel point (x,y), I(i,j) is the pixel value of the input image at position (i,j), d(i,j) is the Euclidean distance between the input image at position (i,j) and the center image (x,y), α is the adaptive damping coefficient, λ is the weight parameter, and the expression of λ is τ is a constant used to control the sensitivity of the filter, σ(x,y) is the local standard deviation of the neighborhood window centered on the pixel point (x,y), and V mean is the global mean variance of the image; S21, redistributing the brightness of the filtered image by contrast limited adaptive histogram equalization, and the pixel output I(x, y) of the redistributed brightness is expressed as: In the above formula, is the histogram density function in the local area of ​​the pixel, and δ is the contrast limit threshold; The Img1 is obtained after two processes of filtering and redistributing the brightness.

4. The non-cooperative spacecraft attitude estimation method based on contour and linear structure as claimed in claim 1, characterized in that: In S3, the acquisition process of Img2 includes: In the original UNet3+ segmentation network model, the attention module CBAM is introduced after each downsampling and upsampling 3x3 convolution operation; Img1 is input into the UNet3+ segmentation network model integrated with CBAM, and the feature map is passed to the attention module. The feature map adjusted by the attention mechanism is then passed to the next layer or fused through jump connections to enhance the feature expression to obtain Img2.

5. The non-cooperative spacecraft attitude estimation method based on contour and linear structure as claimed in claim 1, characterized in that: In S4, the acquisition process of Img4 includes: S40, mark the coordinates of the N contour key points as (x i ,y i ), perform a one-dimensional discrete Fourier transform on each contour key point to obtain the Fourier coefficient Z(k) of the kth frequency component to obtain Img3: In the above formula, (x i +jy i ) is the complex form of each contour key point, j is the imaginary unit S41, spindle direction data θ' main By calculating the phase angle of the lowest frequency component Z(1) of the Fourier transform, we get: θ' main =arg(From(1)) Secondary axis direction data θ' sub By calculating the phase angle of the second lowest frequency component Z(2) of the Fourier transform, we get: θ sub =arg(Z(2)) S42, record the data of the major and minor axis directions (θ' main ,θ' sub ), and the line vectors are marked with lines on the image to obtain Img4, where the line vectors of the major axis and the minor axis in the linear structure are the farthest distances between the intersection of the major axis direction and the contour edge and the intersection of the minor axis direction and the contour edge, respectively.

6. The non-cooperative spacecraft attitude estimation method based on contour and linear structure as claimed in claim 5, characterized in that: In S5, the similarity comparison is achieved by adding a local direction information to constrain the local features based on the shape comparison, so the measurement matrix C of the total similarity is A for: C A =(1-γ)D+γC In the above formula, D is the shape similarity matrix, C is the direction similarity matrix, γ is the adaptive weight parameter, and the expression of γ is μ is a tuning parameter used to adjust the steepness of the function, and t is the mean of the shape similarity matrix D.

7. The non-cooperative spacecraft attitude estimation method based on contour and linear structure as claimed in claim 6, characterized in that: In S5, after the similarity comparison, the images are arranged in descending order according to the total similarity, and the top N images are used as the image sequence to be selected; The major and minor axis direction data (θ main ,θ sub ) and the major and minor axis direction data (θ' main ,θ' sub ) to compare and calculate the angle difference Diff all The smallest image is used as the target pose estimation result: Diff all =Diff main +Diff sub Diff main =min(|θ main -θ′ main |,360-|θ main -θ′ main |) Diff sub =min(|θ sub -θ′ sub |,360-|θ sub -θ′ sub | In the above formula, Diff main is the main axis angle difference, Diff sub The angle difference of the secondary axis is returned. all The smallest image can obtain the optimal spacecraft attitude estimation result.

Images