Star map registration method based on remote optical image

Abstract

The invention discloses a star map registration method based on a remote optical image, relates to the technical field of image processing, and solves the problems that the existing method is multi-oriented to a single telescope, most of the single telescope is aligned at a star point level, and the robustness is insufficient when parallax, distortion, optical difference and background pollution exist in remote multi-station imaging. The method comprises the following steps: acquiring a remote star map and extracting a star point centroid; angular distance calculation and triangle construction; registering the centroids of the star points in the whole image; performing similarity transformation and homography estimation; and carrying out sub-pixel refinement and full image registration. According to the invention, dual registration of star point centroids and pixel coordinates can be realized for star maps shot by a plurality of telescopes in different places, and meanwhile, a high-quality registration reference can be provided for subsequent three-dimensional information acquisition of a space target and identification and positioning of the space target.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image registration, in particular to a star map registration method based on off-site optical images. BACKGROUND

[0002] Star map registration is the basic link of multi-site astronomical observation, photometry, orbit inversion and splicing imaging, and space target tracking. Unlike natural scenes, star map targets are point-like, lack local texture, and are disturbed by strong noise and false star points. Traditional matching based on local descriptors (such as SIFT / ORB) is prone to problems such as feature scarcity and high mismatch rate. It is difficult to resist the differences in viewing axes or slight nonlinear projection errors by relying solely on the overall affine / similar model.

[0003] Due to the particularity of astronomical observation, traditional star map registration methods are mostly for single telescopes or star maps taken by the same telescope at the same time. In this case, the viewing axis difference of the images is small, and the geometric relationship of the star points and the registration error are relatively easy to handle. However, with the increasing demand for multi-telescope off-site collaborative observation, off-site multi-station star map registration faces more complex challenges. There are significant parallax, optical system differences, atmospheric disturbances, and imaging differences caused by different shooting times and angles between different telescopes, which makes the traditional registration method based on a single viewing axis or the same device conditions cannot be directly applied. Especially in the registration process of off-site star maps, the spatial position relationship of star points is more complex, and it cannot rely on the pre-calibrated WCS or other unified reference coordinate system, resulting in a significant increase in registration error.

[0004] The present application proposes a star map registration method based on multiple off-site telescopes, which breaks through the limitations of traditional single-telescope registration. It addresses the problems of parallax, optical property differences, and inconsistent shooting conditions between different telescopes. It effectively realizes the accurate registration of the centroids and pixel points of star maps taken by multiple telescopes by using geometric invariants such as triangle angular distance and normalized area, and homography matrix estimation method. SUMMARY

[0005] The present application provides a star map registration method based on multiple off-site telescopes to solve the problem that existing methods are mostly oriented towards single telescopes and most of them only align at the star point level. The method has insufficient robustness when there are parallax, distortion, optical differences, and background pollution in off-site multi-station imaging.

[0006] The star map registration method based on multiple off-site telescopes is implemented by the following steps:

[0007] Step one, multiple astronomical telescopes are used to observe the same sky area from different positions, obtain multiple star maps of the same time period, and extract the star point centroids in the corresponding star maps to determine the star point positions.

[0008] Step two, using the triangle method combined with angle distance calculation and triangle geometric characteristics, the same star points in two star maps are identified to form a matched triangle pair, and are used as seed triangles; preliminary registration is realized;

[0009] Step three, using least square estimation of similarity transformation, the star point centroids of the seed triangles obtained in step two are projected to realize the registration of the star point centroids of the whole star map;

[0010] Step four, the homography matrix of the projection of the first star map to the second star map is calculated, and the re-projection error is measured to find the whole overlapping area of the two star maps;

[0011] Step five, the overlapping area obtained in step four is sub-pixel refined to realize the one-to-one alignment registration of the pixel coordinates.

[0012] The star map registration method has the advantages that: the star map registration method firstly extracts high-quality centroids through the local maximum algorithm and star order sorting method, then finds the matching star point triangle by using the angle distance calculation and triangle geometric characteristics, then realizes the alignment of a large number of star points by using a plurality of seed triangles for similarity transformation, finally finds the overlapping area of the two maps by calculating the homography matrix of the projection, and realizes the registration of the pixel coordinates of the star maps taken in different places, which can provide a good foundation for subsequent spatial target three-dimensional information acquisition and target identification and tracking.

[0013] The star map registration method firstly extracts high-quality centroids by using the local maximum detection of the DAOStarFinder algorithm and intensity weighted centroid based on ROI and background estimation, and preferentially takes star points with high magnitudes. In the scene of stray light and weak signal, sub-pixel level star point positions can still be stably obtained, which lays a high-precision starting point for subsequent registration.

[0014] The star map registration method uses triangle matching in the registration of star maps taken by different telescopes, and realizes the registration of a large number of star point centroids in the case of parallax by using the angle distance between star points and triangle geometric characteristics.

[0015] The star map registration method finds the overlapping area between the two star maps by calculating the homography matrix between the two star maps taken in different places, which provides a basis for the alignment of the whole image pixels in the next step.

[0016] The star map registration method realizes the registration of the pixel coordinates between the star maps taken in different places by using the homography matrix calculation, and performs sub-pixel refinement to output the one-to-one corresponding diagram of the pixel coordinates between the two maps.

[0017] The star map registration method described in this invention achieves dual registration of star maps taken by telescopes at different locations in terms of star centroid and pixel coordinates, providing accurate basic data for subsequent applications such as space target tracking and 3D reconstruction. Attached Figure Description

[0018] Figure 1 This is a flowchart of the star map registration method based on off-site optical images according to the present invention;

[0019] Figure 2 Flowchart for star centroid registration and pixel registration;

[0020] Figure 3 This is a schematic diagram of star points extracted from two star images taken at different locations for registration; where (a) is... Figure 1 (a) is the effect image of the extracted 300 star points; (b) is Figure 2 The result image of the extracted 300 star points;

[0021] Figure 4 A schematic diagram of triangles formed by matching the same star points for triangle matching; (a) and (b) are schematic diagrams of star point triangles matched for the entire star map; (c) and (d) are schematic diagrams of star point triangles matched in the lower left 1 / 4 region.

[0022] Figure 5 Registration diagram of the centroids of all stars in the map;

[0023] Figure 6 To calculate the pseudo-color image of the overlapping region of two star images taken at different locations using the homography matrix;

[0024] Figure 7 The image is a pixel registration diagram of a star map taken at a different location. (a) is a line diagram showing the one-to-one correspondence of all pixel coordinates, and (b) is a line diagram showing the pixel connections extracted at uniform intervals after adjusting the step size. Detailed Implementation

[0025] Specific Implementation Method 1: Combination Figure 1 and Figure 2 This embodiment describes a star image registration method based on multiple telescopes located at different sites. This method processes two or more sets of star images captured by multiple telescopes at different locations. Specifically, one telescope remains stationary while the others are moved to different positions to capture images of the same celestial region. Due to the baseline distance between the telescopes, parallax exists between the star images captured by different telescopes. To achieve more accurate star image registration, this embodiment can also optimize the overlapping area and parallax of the star images by adjusting the baseline distance, thereby providing support for subsequent star image registration and tracking of space targets.

[0026] In the star map registration method described in this embodiment, after data acquisition, star centroids are extracted. The DAOStarFinder algorithm is used for star candidate detection, eliminating oversaturated or low SNR stars. The brightness of the stars is then sorted, and stars with higher brightness are selected to improve the efficiency of subsequent registration. Based on this, the star centroids are calculated with sub-pixel precision to obtain accurate star positions.

[0027] After the star points are extracted, angular distance calculation is used to match the same star points in the overlapping parts of the two star images. Triangle matching combined with angular distance calculation is used to analyze the geometric relationship of the matched star points, extract features such as triangle side length and angle, and further verify and solidify the matching results.

[0028] After completing triangle matching, the least squares method is used to estimate the similarity transformation, projecting the centroids of all stars in the first star image onto the second star image. This expands the triangle matching result to full-image star point matching, achieving a large number of reliable centroid-level registrations. After centroid-level registration, this invention uses the homography matrix to register the pixel coordinates of the entire image. First, the homography matrix between the two star images is calculated through robust estimation. Then, the projection error is measured for consistency, invalid regions are eliminated, and the overlapping parts of the two star images are determined. Based on the calculated homography matrix, sub-pixel-level thinning is performed to generate accurate one-to-one pixel coordinates, and coordinates without overlap or with large errors are eliminated to ensure the uniqueness of pixel coordinates.

[0029] Finally, by generating a line graph connecting the corresponding star points and pixel coordinates of the two star images, the overlapping area and the coordinates covered by the two star images are visually displayed. Furthermore, a residual graph is generated to evaluate the registration accuracy. Through these steps, this invention achieves dual registration of star images captured by a remote telescope in terms of star centroids and pixel coordinates, providing accurate basic data for subsequent applications such as space target recognition and 3D information acquisition.

[0030] like Figure 1 As shown, the star map registration method based on off-site optical images described in this embodiment includes the following steps:

[0031] Step 1: Remote Star Image Acquisition and Star Centroid Extraction; Star images taken by multiple telescopes at different locations are acquired, and then a star centroid extraction algorithm is used to extract stars with magnitude height and no contamination; the specific process is as follows:

[0032] In this embodiment, star images are acquired by two or more separate telescopes, with multiple telescopes simultaneously photographing the same celestial region. The entire image is extracted from two or more star images of the same celestial region, or identical, low-contamination, uniformly backgrounded regions are selected.

[0033] First, estimate the background mean of the star map. and noise and through threshold Background subtraction is performed to remove highlighted and contaminated areas. Here, m is a noise factor used to determine the boundary between the background and the star points. After background removal, the DAOStarFinder algorithm is used to extract candidate star points through neighborhood local maxima detection. By setting constraints such as connected component area and aspect ratio, false stars and trailing lines are eliminated to ensure high reliability of the extracted star points.

[0034] For each candidate star point, calculate the star point centroid with sub-pixel precision within a local window. , The calculation formula is as follows:

[0035]

[0036] in, The magnitude of the star is ( Using the coordinates of the star point, the precise position of the centroid is obtained. The accuracy of the star point position can be further improved through two-dimensional Gaussian fitting, and the full width at half maximum (FWHM) and centroid uncertainty can be calculated for weighted estimation in subsequent steps.

[0037] Then, based on the star point's SNR, flux, FWHM, and ellipticity, the extracted star points are sorted by quality, and duplicate star points or redundant data are eliminated by nearest neighbor deduplication to ensure the subsequent registration effect. Based on the filtered star point set, a Top-N strategy is used to sort the star points from highest to lowest magnitude, selecting stars with higher brightness, and prioritizing stars closer to the principal point to further limit the number of matching star points and control the data size and matching coverage.

[0038] Finally, the pixel coordinate system was used to locate the star points in both star images, resulting in two sets of star points. and :

[0039]

[0040] in, and These are the pixel coordinates of star points in the two star images. and These represent two star maps respectively; these methods ensure the acquisition of high-quality star points, and input metadata such as principal point coordinates, pixel size a×b, focal length f, and star point resolution as high-quality input for subsequent angular distance calculation, triangle construction, and full map registration.

[0041] Step 2: Angular Distance Calculation and Triangle Construction; Extract star point triangles from both star images, and use angular distance calculation and other geometric features to find triangles formed by identical star points in both star images, achieving initial alignment; these triangles serve as seed triangles; the specific process is as follows:

[0042] By combining the triangle method with angular distance calculation and triangle geometric features, the star points extracted from two or more star images taken at different locations are identified to form matching triangle pairs, thus achieving registration.

[0043] Because there are too many stars in both star maps, there will be too many triangles to match, which consumes computational resources and is inefficient. To limit the number of triangles and ensure matching efficiency, the ROI module first selects a portion of the same observation area in both star maps, such as extracting stars and triangle groups by quadrant, and only retains candidate stars in the selected area that meet the minimum separation and minimum area thresholds. This module significantly suppresses the formation of triangle combinations by using a spatially confined center region priority and Top-N bright star extraction method. To address the complexity issue of exponential growth in the number of triangles with the number of star points, we employ a three-dimensional algorithm. On the other hand, we ensure the distribution coverage of candidate points within the ROI through grid equalization / non-maximum suppression, making subsequent triangle matching more efficient without causing local congestion or biased edge distortion areas. For star points registered in certain regions, they are merged into a fully aligned star map after registration is complete.

[0044] Secondly, the relative distances between the three stars need to be constrained. In a star chart, the celestial coordinates of the stars are difficult to obtain, so we can start from the perspective of pixel coordinates, assuming the pixel coordinates of two stars are... and The pixel distance between star points is calculated using the following formula:

[0045]

[0046] When constructing a triangle, ensure that this distance is as small as possible.

[0047] Next, the angular distance is calculated; angular distance is a unique indicator in astronomy regarding the distance between stars, representing the angle between rays extending from the telescope towards two stars on the telescope's target surface. The formula for calculating angular distance is:

[0048]

[0049] in, denoted as angular distance between two stars, a×b as the pixel size of the star map, and f as the telescope focal length. and These are the pixel coordinates of star points in the two star images. The formula for calculating the angular distance between any two star points using vectors is as follows:

[0050]

[0051] in, , , where a×b is the pixel size of the star map.

[0052] In this embodiment, the preliminary determination of whether the selected triangles in two star images are composed of the same star points is achieved by calculating the angular distance. By calculating the angular distance between each pair of the three star points of the selected triangles in the two star images respectively, the angular distance difference between the corresponding star point pairs is calculated. If the difference between each pair is less than a very small value, i.e., the angular distance tolerance, it can be determined that the two triangles in the two star images are composed of the same three star points. In this way, the same star points in the two star images can be found in advance, and the preliminary registration can be achieved.

[0053] Assume the three star points of triangles A and B in the two star diagrams are respectively The condition for determining whether two triangles are matched is:

[0054]

[0055]

[0056]

[0057] in, Indicates the angular distance between two points. This represents the angular distance tolerance. If the above conditions are met, then the two triangles are considered to be matching triangles, thus achieving the initial registration of the two star maps.

[0058] Furthermore, the matching triangle set can be further solidified and verified by examining its geometric features such as side lengths and angles. The shape characteristics of the triangle can be obtained by calculating the lengths of its three sides. If the three sides of two triangles are the same, it can be further verified that the two triangles are composed of the same star point. The formula for calculating the side length is:

[0059]

[0060]

[0061]

[0062] in, , , These are the side lengths of the three sides of the star-shaped triangle.

[0063] In this embodiment, angles can also be used as a method to verify angular distance registration. The size of each angle can be calculated using the law of cosines. For example, angle... The calculation formula is:

[0064]

[0065] Similarly, angle and It can also be calculated using the law of cosines. If the angles of the two matching star-point triangles selected by the angular distance calculation are also equal, then the triangles selected by the angular distance can be verified.

[0066] Step 3: Centroid Registration of Star Points Across the Entire Image; For the obtained highly consistent seed triangle, a least-squares estimation similarity transformation is used to project the centroids of the star points forming the triangle from the two star images, achieving centroid registration and alignment of a large number of star points across the entire image; the specific implementation process is as follows:

[0067] Based on the highly consistent seed triangles obtained in step two (registration groups consisting of multiple sets of identical star points forming triangles), these triangles are used as starting matching points, and the registration is extended to all star points in the image through similarity transformation.

[0068] Each seed triangle consists of three pairs of matching star points, exhibiting good geometric consistency. Based on these seeds, the least squares method is used to estimate the similarity transformation, projecting the centroids of all star points in the first star image onto the second star image, thus expanding the seed triangles. The transformation matrix for each pair of triangles is set as follows: in As a scaling factor, For rotation matrix ( (for rotation angle) Let x be the translation vector, and let x be the coordinate vector of the two-dimensional star point. In this way, the initially selected triangle seed is extended to the entire graph, resulting in more matching point pairs.

[0069] For the projected star points, nearest neighbor matching is performed in the second star image. By calculating the pixel distance between each projected point and all star points in the second star image, the star point with the smallest distance is selected as a candidate matching pair. For each candidate matching pair, its angular distance difference is calculated. If the angular distance difference is within the allowable range (i.e., less than the preset tolerance), the two points are considered to be matched.

[0070] Among all candidate matching pairs, the most reliable matching point is selected through a voting mechanism. Each star point is voted on based on its frequency of occurrence in multiple similarity transformations, and the matching pair with the most votes is finally selected as the final registration result. To avoid one-to-many or many-to-one situations, a one-to-one uniqueness strategy is used to ensure that each star point is matched only once.

[0071] Step 4: Similarity Transformation and Homography Estimation;

[0072] After centroid registration is completed, robust estimation is used to calculate the homography matrix of the projection from the first star image to the second star image. Then, a consistency measure is performed to eliminate projection errors, invalid projection regions are removed, and all overlapping areas between the two star images are found. The specific process is as follows:

[0073] First, based on the matching point pairs obtained in step three, the homography matrix H between the two images is estimated using the RANSAC algorithm. The homography matrix describes the perspective transformation between the images, i.e., the mapping relationship from image 1 to image 2. RANSAC iteratively calculates by selecting random matching point pairs, eliminating outliers, and finally obtaining the optimal homography matrix. Specifically, at least four pairs of matching points are needed to calculate the homography matrix between two star images, and the perspective transformation matrix is ​​solved using the least squares method. In this process, reprojection error is used to determine whether the matching point pairs are inliers, outliers are eliminated, and the homography matrix H and the inlier mask are finally calculated, playing a crucial role in the image registration process.

[0074] In this embodiment, the homography matrix maps the coordinates in image 1 to image 2 through the following perspective transformation:

[0075]

[0076] Wherein, the homography matrix H is a 3×3 homography matrix. The coordinates of the corresponding point in Image 1, yes Figure 2 Zhongyu Figure 1 The coordinates of the corresponding points. Through homogeneous coordinate transformation, all points in the image are mapped from image 1 to image 2.

[0077] After obtaining the homography matrix H, reprojection and error calculation are performed. The coordinates of the star points in image 1 are then determined using the homography matrix H. Projecting the image back onto image 2, we obtain the predicted coordinates. ;

[0078] The predicted projection point of the second projection Coordinates of the corresponding point in the first projection reprojection error , as a measure of registration error. The calculation formula is:

[0079]

[0080] The reprojection error is used to evaluate the accuracy of the matching. If the error is large, it may be an incorrect match, and RANSAC will remove these matching point pairs with large errors.

[0081] In this embodiment, to avoid the influence of invalid regions, after calculating the homography matrix H, an overlapping region mask can be generated. The unit mask of image 1 is projected onto image 2 through the homography matrix H to obtain the mask of the overlapping region, thereby limiting the registration calculation area. The registration operation will be limited to the overlapping region of the two images, eliminating the influence of invalid regions on the registration result.

[0082] Step 5: Subpixel thinning and pixel coordinate registration; Subpixel thinning is performed using the calculated homography matrix to achieve one-to-one alignment and registration of pixel coordinates, providing a foundation for subsequent spatial target recognition and tracking; the specific process is as follows:

[0083] Step 51: Based on the mapping from image 1 to image 2 in step 4, perform one-to-one registration from centroid level to pixel level for the overlapping area of ​​the two star images. During this process, refine H at the sub-pixel level so that the pixel coordinate correspondence can be accurate to the decimal point, thereby improving the overall accuracy and robustness. Finally, output the result image of the one-to-one correspondence of pixel coordinates in the overlapping area.

[0084] First, match the interior points. To minimize the weighted projection error, the homography matrix is ​​finely refined to reduce the reprojection error as much as possible. The process is described by the following formula:

[0085]

[0086]

[0087] Where H below min indicates that the homography matrix H is the optimization variable. Let [X,Y,W] be the projection operator, where [X,Y,W] is the homogeneous coordinate vector, and X, Y, and W are the three components of the homogeneous coordinates, respectively. It's a scaling factor. Weights The centroid uncertainty or signal-to-noise ratio can be adaptively set to suppress the influence of low-confidence matching pairs. The refined homography matrix serves as a unified model for the overall image coordinate transformation, performing a forward projection on any point (x, y) in image 1:

[0088]

[0089] in, The input coordinates are the homogeneous coordinates of the pixels in image 1. for The homogeneous output coordinates obtained by multiplying by H. It is a scale component, only when Only at a specific point in time can the two-dimensional representation be restored.

[0090] Step 5.2: To establish correspondences only within the truly valid overlapping field of view, first construct an overlapping mask Ω: project the unit mask of image 1 onto image 2 using a homography matrix, and take pixels with values ​​> 0 as the effective domain. If necessary, perform morphological erosion on Ω to avoid boundary numerical instability and extrapolation errors. Then, perform regular sampling on the integer pixel grid of image 1 with a step size s ∈ {1, 2, 4, 8, ...}, and calculate the forward projection for each source pixel. ,Require >0 and the forward projection falls within the imaging boundary of image 2; then perform the nearest integer rounding.

[0091]

[0092] Step 53: Map the subpixel projection points to integer pixel candidates on the target image. If multiple source pixels are mapped to the same target pixel q, then the forward rounding error is considered. Preservation of the best criterion The smallest value ensures uniqueness on the target side, preventing "many-to-one" conflicts from the outset. To further guarantee bidirectional consistency, a reverse mutual check is performed on the retained q.

[0093]

[0094] in, These are candidate integer pixels from the source image after back projection and rounding. This involves taking the nearest integer pixel as the sub-pixel real coordinate. The corresponding pixel is retained only if the projection returns to the source pixel after round trip, thus obtaining a pixel-level one-to-one correspondence set that satisfies bidirectional consistency.

[0095] Specific Implementation Method Two: Combination Figure 2 to Figure 7 This embodiment describes an application of the star map registration method based on off-site optical images described in Specific Embodiment 1. The specific implementation process of this embodiment is as follows:

[0096] 1. Working conditions;

[0097] This experiment uses an Intel Core i7-7700K CPU @ 4.20GHz*8 processor, a PC running Windows 10, a GeForce GTX 1070Ti graphics card, and Python as the programming language.

[0098] 2. Experimental content and results analysis;

[0099] like Figure 2 As shown, Figure 2The flowchart for achieving star centroid registration and pixel coordinate alignment briefly summarizes the detailed process.

[0100] First, image processing is performed on two or more sets of star images taken by multiple telescopes at different locations. One telescope remains stationary while the others move to different positions to photograph the same celestial region. Due to the baseline distance between the telescopes, there is a certain parallax between the star images taken by different telescopes. The overlapping area and parallax between the star images can be adjusted by adjusting the baseline distance, so as to facilitate subsequent registration between star images and tracking of space targets.

[0101] The centroids of the stars are extracted, and a matching filter is performed using a Gaussian kernel that matches the size of the star image. Local bright peaks are then found on the filtered image, and false stars and trailing lines are removed according to a set of morphological indicators. Star candidates are generated, and oversaturated or low SNR stars are removed. The magnitudes of the extracted stars are sorted, and stars with higher brightness are selected to improve the efficiency of subsequent registration. The centroids of the stars are calculated with sub-pixel precision to obtain the star set.

[0102] Construct star point triangles and find the common star points in the overlapping areas between different star maps by calculating the angular distances between each pair of star points. For star points matched using the triangle method, extract their geometric relationship features, including side lengths and angles, to solidify and verify the registration using the triangle method.

[0103] Triangles are registered using angular distance and geometric features. After registering the triangle group, a large number of star points in the entire star map are registered and aligned. The obtained highly consistent "seed triangle" set is used to project the centroids of all star points in image 1 onto image 2 using least squares estimation similarity transformation, thereby expanding the triangle group into a large number of reliable centroid-level correspondences of star points in the entire image, and one-to-one alignment between star points.

[0104] To achieve full-image star point registration, after completing the registration of a large number of star points across the entire image, the pixel coordinates of the entire image are aligned and registered using the homography matrix. First, robust estimation is used to calculate the homography matrix of the projection from image 1 to image 2. Then, consistency measurement is performed to remove projection errors, invalid projection areas are eliminated, and all overlapping parts of the two star images are found.

[0105] Pixel coordinate alignment: Based on the calculated homography matrices of the two star images, sub-pixel refinement is performed to generate one-to-one corresponding pixel coordinates. Coordinates without overlap or with large errors are removed to achieve coordinate uniqueness. Finally, a line graph connecting the corresponding pixel coordinates of the two star images is generated to visually show the overlapping area and the coordinates it covers. A residual map is also generated to achieve dual registration of the centroid of stars and image pixels by a telescope at a different location.

[0106] like Figure 3 As shown,Figure 3 To extract the centroids of the stars, the neighborhood local maximum detection algorithm filters out the brighter stars. By setting constraints such as the area of ​​the connected region, the peak-to-background difference, and the aspect ratio, false stars and stray lines are eliminated and the data is collected evenly. This step extracts a total of 300 stars, which greatly improves the efficiency of subsequent star map registration.

[0107] like Figure 4 As shown, Figure 4 To achieve matching triangles using triangle matching, the matching triangles were found by combining the angular distance between star points with other geometric features. (a) shows the paired triangles extracted from the entire image, and (b) shows the matching star points in the lower left quarter extracted for block registration to improve subsequent registration efficiency. The results show that the two triangles have identical geometric shapes and the same surrounding star points, indicating that they are triangles composed of the same star points. Since the star image data was taken by two telescopes at different locations, there is a baseline distance between the two telescopes, resulting in a certain parallax between the two star images. Therefore, the pixel coordinates of the matching triangles are different in the two images. This achievement successfully registered triangles composed of multiple star points, providing a good foundation for subsequent centroid registration of star points across the entire image.

[0108] like Figure 5 As shown, Figure 5 To register the centroids of stars in optical images from different locations, a large number of matching star points were found. Figure 4 Based on the obtained highly consistent "seed triangles", these triangles are used as starting matching points, and the registration of star points in the whole map is achieved through similarity transformation.

[0109] like Figure 6 As shown, Figure 6 To calculate the homography matrix H between the two graphs, and then... Figure 1 Projected to H Figure 2 In the middle, and projected Figure 1 Rendered with pseudo-color and Figure 2 The grayscale base images are overlaid. The pseudocolor layer is used to highlight the differences in brightness and structure, and the colored areas are used to highlight the overlapping parts of the two images. The black wedge-shaped boundary is the non-overlapping area. Within the overlay area, the peak positions of the pseudocolor stars and grayscale stars basically coincide, indicating that the differences in rotation, scale, and perspective have been effectively compensated by the homography matrix.

[0110] like Figure 7 As shown, Figure 7Line diagrams for pixel registration in optical images from different locations. The overlapping areas of the two images correspond one-to-one with each other. (a) shows the line diagram of all pixels in the overlapping area. Due to the large number of pixels, the lines between corresponding pixels are too dense, making it difficult to distinguish the specific corresponding pixels. (b) shows the corresponding line diagram extracted at uniform intervals after adjusting the step size. The one-to-one pixel correspondence is more clearly visible here.

[0111] The experimental results above demonstrate that for star images captured by multiple telescopes at different locations, this invention, based on a unified ROI, first uses a local maxima-based method to sort star candidates by magnitude, significantly improving the reliability of star identification and localization. Then, using geometric invariants such as triangle angular distance and normalized area as constraints, highly consistent seed triangles are selected from multiple sets of triangles. Furthermore, through similarity transformation projection and voting strategies, local matching is robustly extended to a large number of centroid-level pairs across the entire image. Based on this, RANSAC is used to estimate the homography matrix H between the two images, and reprojection error is used as a consistency metric to eliminate outliers. Pixel-level one-to-one correspondence is achieved within overlapping regions. Comprehensive visualization results show that star points and pixel matching points are one-to-one, matching pairs are evenly distributed across the entire image, residuals are concentrated, and edge regions are stable. Dynamic or non-infinitely distant targets can be automatically eliminated by large residuals. Thus, this invention not only achieves precise registration of star images from multiple telescopes at different locations at both the centroid level and the pixel level, and outputs a coordinate list and error statistics, but also provides a high-quality, reusable registration benchmark for subsequent space target identification and acquisition of three-dimensional information of space targets.

[0112] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for star pattern registration based on off-site optical images, characterized in that: The method is realized by the following steps: Step one, a plurality of astronomical telescopes are used to observe the same sky area from different positions, a plurality of star maps in the same time period are obtained, and the star point centroids in the corresponding star maps are extracted to determine the star point positions; Step two, the star points extracted from the plurality of star maps are identified by using a triangular method combined with angle distance calculation and triangular geometric characteristics to form matched triangular pairs in the two star maps, and the triangular pairs are used as seed triangles; preliminary registration is realized; Step three, the star point centroids of the seed triangles obtained in step two are projected by using least square estimation of similarity transformation to realize the registration of the star point centroids of the whole star map; Step four, a homography matrix of the projection of the first star map to the second star map is calculated, and the consistency is measured by taking the re-projection error to find the whole overlapping area of the two star maps; Step five, the sub-pixel is refined in the overlapping area obtained in step four to realize the one-to-one alignment registration of the pixel coordinates.

2. The off-site optical image based star pattern registration method of claim 1, wherein: In step one, the process of extracting the star point centroid is as follows: DAOStarFinder algorithm is used for star point candidate detection, and the brightness of the star point is sorted to screen out the star point with high brightness; the sub-pixel accuracy is used to calculate the star point centroid to obtain the accurate star point position.

3. The off-site optical image based star pattern registration method of claim 1, wherein: In step two, the angle distance between two star points is calculated to find the triangular groups that meet the angle distance difference less than the angle distance tolerance between different star maps, and the geometric relationship characteristics of the star point triangle, including the side length and the angle, are extracted to stabilize and verify the registration based on the angle distance triangular method to obtain a plurality of seed triangles.

4. The off-site optical image based star pattern registration method of claim 3, wherein: The angle distance calculation formula between two star points is as follows:

5. Wherein, For the angular distance value between two star points, a x b is the pixel size of the star map, f is the focal length of the telescope, and The pixel coordinates of the two star points; by calculating the angular distance between the three star points of the selected triangle in the two star maps respectively, the angular distance difference value of the corresponding star point pair is calculated, if the difference value between the two is less than the angular distance tolerance , then it is determined that the two triangles in the two star maps are composed of the same three star points.

6. The off-site optical image based star pattern registration method of claim 4, wherein: Let the three stars of triangle A and B in two star maps be The matching condition of two triangles is that ； ； ； In the formula, represents the angular distance between two points, if the above conditions are met, then the two triangles are considered to be matching triangles.

7. The off-site optical image based star pattern registration method of claim 1, wherein: In step two, the matching triangular groups are verified by the geometric characteristics of the side length and the angle of the triangle, the shape characteristics of the triangle are obtained by calculating the lengths of the three sides, and if the lengths of the three sides of the two triangles are the same, it is verified that the two triangles are the triangles composed of the same star points.

8. The off-site optical image based star pattern registration method of claim 1, wherein: In step three, the least square method is used to estimate the similarity transformation to project all the star point centroids of the first star map into the second star map to realize the expansion of the seed triangle; The projected star points are matched by the nearest neighbor in the second map, and the most reliable matching point is selected by the voting mechanism to realize the registration of the star point centroid.

9. The off-site optical image based star pattern registration method of claim 1, wherein: In step four, the coordinates in the first star map are mapped into the second star map by perspective transformation through the homography matrix; After obtaining the homography matrix H, the re-projection and error calculation are performed, the star point coordinates in the first star map are projected into the second star map again through the homography matrix H to obtain the predicted coordinates; The re-projection error between the predicted coordinates and the corresponding point coordinates in the second star map is calculated, and the re-projection error is used as the measurement of the registration error.

10. The off-site optical image based star pattern registration method of claim 8, wherein: In step five, the process of pixel coordinate registration is as follows: Step five, inner point matching pairs To constraint the weighted re-projection error, the homography matrix is refined to make the re-projection error as small as possible. The formula is: ； ； wherein is the projection operator, [X, Y, W] is the homogeneous coordinate vector, X, Y, W are the three components of the homogeneous coordinate, respectively; is the scale factor; is the weight, the forward projection of any point (x, y) of the first star map is as follows: ； wherein is the homogeneous input coordinate of the pixel point in the first star map, is the homogeneous output coordinate of the output, is the scale component; Step five two, an overlapping mask Omega is constructed; the unit mask of the first star map is projected into the second star map through the homography matrix, and the pixel value greater than 0 is taken as the effective area; Regular sampling is performed on the integer pixel grid of the first star image with a step size s∈{1,2,4,8,…}, and the forward projection is calculated for each source pixel. ,Require >0 and the forward projection falls within the boundary of the second star map image; then perform forward projection. The rounding operation; Step five three, the sub-pixel projection point is mapped to the integer pixel candidate on the target map; If multiple source pixels map to the same target pixel q, the forward rounding error is used as a criterion to favor the one that retains the minimum. The inverse projection is performed on the reserved q to obtain the pixel level one-to-one correspondence set satisfying the bidirectional consistency, which is expressed as: ； In the formula, is the integer pixel candidate of the source image after back-projection and rounding, is the projection operator, is the nearest integer pixel of the sub-pixel real coordinate.

Images