A high-precision satellite attitude correction method, system and device for extremely large initial positioning errors

Abstract

The application discloses a high-precision satellite attitude correction method, system and equipment for a maximum initial positioning error, belongs to the technical field of machine vision image matching, and selects a proper reference area to prepare a first reference map and a second reference map, and extracts reference map feature descriptors offline; the remote sensing image collected by a satellite remote sensing camera is down-sampled, and the down-sampled real-time map is matched with the first reference map; the best window position of the reference area ground calibration point in the real-time map is calculated according to the matching result; the real-time map is cut into pieces according to the window position to obtain a high-resolution calibration point real-time map; the high-resolution calibration point real-time map is matched with the second reference map; and the longitude and latitude coordinates of the image center of the real-time map are calculated according to the matching result and used for satellite attitude correction. The application gradually reduces the target area position error through a two-stage remote sensing image matching mode, accurately positions the longitude and latitude coordinates of the real-time map center, and realizes high-precision satellite attitude correction.

Description

Technical Field

[0001] This invention belongs to the field of machine vision image matching technology, and more specifically, relates to a high-precision satellite attitude correction method, system and device for extreme initial positioning errors. Background Technology

[0002] Geostationary orbit satellites are located approximately 36,000 km above the Earth's equator, with their orbital plane making zero angle with the equatorial plane. Furthermore, their angular velocity around the Earth is the same as the Earth's rotational angular velocity, thus making them stationary relative to the ground. Geostationary orbit satellites offer advantages such as relatively fixed Earth observation positions, high temporal resolution, and wide observation range, making them ideal for long-term continuous Earth monitoring and rapid access. They complement low-Earth orbit remote sensing satellites. Geostationary orbit satellites aim to acquire multispectral image information of relevant regions across the country with extremely high temporal resolution and medium spatial resolution, meeting the diverse needs of users in disaster reduction, forestry, meteorology, and other fields.

[0003] The imaging accuracy and positioning error of the space camera on a remote sensing satellite directly affect the normal operation of the satellite's various functions. Typically, before launch, the camera parameters of a remote sensing satellite undergo rigorous geometric calibration in a laboratory. However, factors such as satellite flight vibrations and temperature changes during long-term orbital flight can alter the rigidity of the payloads and the geometric parameters of the space camera. In particular, during camera scanning, accumulated errors in mechanical transmission directly affect the camera's attitude accuracy, thus impacting the satellite's image positioning accuracy. Therefore, remote sensing satellites need to achieve high-precision positioning through on-orbit geometric attitude correction based on laboratory calibration parameters.

[0004] Due to their extremely high altitude, geostationary satellites are susceptible to ground positioning errors exceeding hundreds of kilometers even with minor pointing errors. Remote sensing satellite attitude correction methods based on measured control data are costly and data acquisition is difficult. To improve the geometric positioning accuracy of satellite remote sensing imagery when measured control points are insufficient and no additional measured control data is required, a high-precision satellite attitude correction method for large initial positioning errors is urgently needed. This method addresses the problem of image matching difficulties caused by large initial positioning errors in target areas of large-format real-image remote sensing images, thereby achieving high-precision satellite attitude correction. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a high-precision satellite attitude correction method, system, and device for large initial positioning errors, thereby solving the technical problem of image matching difficulties caused by large initial positioning errors in target areas in large-format real-shot remote sensing images.

[0006] To achieve the above objectives, according to one aspect of the present invention, a high-precision satellite attitude correction method for large initial positioning errors is provided, comprising the following steps: S1, using the ground calibration field as the target area, select a reference area to prepare a first reference map and a second reference map respectively; the resolution and area of ​​the first reference map are lower and larger than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively; the resolution and area of ​​the second reference map are not lower and smaller than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. S2, downsample the images acquired in real time by the satellite remote sensing camera and generate a low-resolution real-time image with a resolution lower than that of the real-time image; perform image matching between the low-resolution real-time image and the first reference image, calculate the optimal window position of the ground calibration point in the reference area in the low-resolution real-time image based on the matching result, and cut the real-time image into blocks based on the optimal window position to obtain a high-resolution calibration point real-time image. S3, perform image matching between the high-resolution calibration point real-time image and the second reference image, establish a matching mapping mathematical model based on the matching result, and calculate the latitude and longitude coordinates of the real-time image center through bilinear interpolation to achieve satellite attitude correction.

[0007] Preferably, the first reference image has a field of view of 100 kilometers, and the second reference image has a field of view of kilometers.

[0008] Preferably, in step S2, a pixel block center aggregation method is used to downsample the images acquired in real time by the satellite remote sensing camera, and the specific formula is as follows:

[0009] in The image pixel grayscale values ​​after downsampling. The original image pixel grayscale values, This represents the downsampling ratio.

[0010] Preferably, step S2, which involves image matching between the low-resolution real-time image and the first reference image, includes the following steps: Offline extraction of point feature descriptors of scale-invariant feature transformation of the first reference image, and construction of a point feature descriptor library of the first reference image; The low-resolution real-time image is divided into regions to obtain multiple window image blocks, and point feature descriptor extraction is performed sequentially on each independent window image block. The point feature descriptors of a single window image block are compared one by one with the point feature descriptor library of the first reference image. The feature similarity is calculated by the feature similarity evaluation algorithm, and the feature point pairs with the highest matching degree are initially selected based on the similarity score. After completing feature matching for all window image blocks, the feature point pairs output by each window image block are integrated to construct a first initial set of feature point pairs covering the entire area of ​​the low-resolution real-time image. The first initial feature pair set is processed by a random sampling consensus algorithm to select the first effective feature point pair set and the first perspective transformation matrix that satisfy the preset constraints.

[0011] Preferably, the image matching between the high-resolution calibration point real-time image and the second reference image in step S3 includes the following steps: Offline extraction of point feature descriptors of scale-invariant feature transformation of the second reference image, and construction of a point feature descriptor library of the second reference image; The high-resolution calibration point real-time image is divided into multiple window image blocks, and the point feature descriptor extraction operation is performed sequentially on each independent window image block. The point feature descriptors of a single window image block are compared one by one with the point feature descriptor library of the second reference image. The feature similarity is calculated by the feature similarity evaluation algorithm, and the feature point pairs with the highest matching degree are initially selected based on the similarity score. After completing feature matching for all window image blocks, the feature point pairs output by each window image block are integrated to construct a second initial feature point pair set covering the entire area of ​​the high-resolution calibration point real-time image. The second initial feature pair set is processed by a random sampling consensus algorithm to select the second effective feature point pair set and the second perspective transformation matrix that meet the preset constraints.

[0012] Preferably, Euclidean distance is used to measure feature similarity during image matching, and the specific formula is as follows:

[0013] in For a low-resolution real-time image or a high-resolution calibration point real-time image, the point feature descriptor is a point feature descriptor for a feature point in the image block point feature descriptor library; A feature descriptor for a feature point in the feature descriptor library of the first or second reference image.

[0014] Preferably, step S3, calculating the latitude and longitude coordinates of the real-time image center, includes the following steps: Based on the second set of effective feature point pairs, a matching mapping mathematical model is established; according to the matching mapping mathematical model, the pixel coordinates of the center point of the real-time image are calculated by matrix transformation to determine the corresponding pixel position on the second reference image. Based on the pre-established correspondence between pixel coordinates of the second reference map and geographic latitude and longitude coordinates, the bilinear interpolation method is used to calculate the latitude and longitude coordinates corresponding to the pixel positions on the second reference map.

[0015] According to another aspect of the present invention, a high-precision satellite attitude correction system for large initial positioning errors is provided, comprising: The reference image creation module is used to prepare a first reference image and a second reference image. The resolution and area of ​​the first reference image are lower and larger than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. The resolution and area of ​​the second reference image are not lower and are smaller than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. The real-time image processing module is used to downsample images acquired in real time by satellite remote sensing cameras and generate low-resolution real-time images with a resolution lower than that of the real-time images; it is also used to slice the real-time images according to the optimal window position to obtain high-resolution calibration point real-time images. The matching module is used to perform image matching between the low-resolution real-time image and the first reference image and image matching between the high-resolution calibration point real-time image and the second reference image, and to obtain the first perspective transformation matrix and the second perspective transformation matrix respectively. The coordinate calculation module is used to calculate the latitude and longitude coordinates of the real-time map center based on the image matching results; An attitude correction module is used to adjust the satellite attitude based on the latitude and longitude coordinates.

[0016] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described high-precision satellite attitude correction method for maximum initial positioning error.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The high-precision satellite attitude correction method proposed in this invention for large initial positioning errors gradually achieves accurate positioning of the target area through a two-stage coarse-to-fine matching strategy. By reducing the resolution of the real-time image through downsampling, a fast matching effect can be achieved. Then, the matching accuracy is further improved through local high-resolution image matching. This can effectively reduce the initial positioning error of the satellite from the hundreds of kilometers level to the kilometers level, thus achieving high-precision satellite attitude correction.

[0018] 2. The high-precision satellite attitude correction method proposed in this invention for large initial positioning errors has efficient feature matching capabilities. It adopts a coarse-to-fine positioning scheme, combined with a sliding window matching strategy, and uses offline preparation of reference image feature descriptors, which can quickly and reliably achieve image matching, effectively ensuring the timeliness and positioning accuracy of satellite attitude correction tasks.

[0019] 3. The high-precision satellite attitude correction system proposed in this invention, designed for extreme initial positioning errors, can be used for on-orbit attitude correction of satellites without ground control points. This invention utilizes open-source remote sensing imagery as a reference, improves satellite positioning accuracy through visual means, and enables fully autonomous processing, completing attitude correction on-orbit without data downlink. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a high-precision satellite attitude correction method for maximum initial positioning error proposed in this invention. Figure 2 This is a schematic diagram of the two-stage remote sensing image matching results of the high-precision satellite attitude correction method for large initial positioning errors of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] This invention proposes a high-precision satellite attitude correction method for large initial positioning errors. It achieves precise satellite positioning of a target area through ground reference map preparation and two-stage remote sensing image matching. First, using a ground calibration field as the target area, a first and second reference map are prepared by selecting a suitable reference area. Feature descriptors are extracted offline from the reference maps. The resolution and area of ​​the first reference map are lower and larger than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. The resolution and area of ​​the second reference map are not lower and are smaller than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. Second, the large-format high-resolution remote sensing image acquired by the satellite remote sensing camera in real-time is downsampled, and the downsampled real-time image is matched with the offline-prepared first reference map. Then, based on the matching results, the optimal window position of the ground calibration point in the reference area is calculated in the real-time image, and the real-time image is segmented according to the window position to obtain a high-resolution calibration point real-time image. Subsequently, the high-resolution calibration point real-time image is matched with the second reference map. Finally, the latitude and longitude coordinates of the center of the real-time image are calculated based on the matching results for satellite attitude correction. This invention addresses the problem of large initial position errors in target areas of large-format remote sensing images, which make positioning errors difficult to correct. It gradually reduces the target area position error through two-stage remote sensing image matching, thereby accurately locating the latitude and longitude coordinates of the real-time image center and achieving high-precision satellite attitude correction. Specifically, it includes the following steps: (1) Preparation of reference diagram Using the ground calibration field as the target area and centering it, a large-format, low-resolution remote sensing image is selected from open-source remote sensing imagery as the first reference image for initial matching. The size of the first reference image is a field of view of hundreds of kilometers, slightly larger than the field of view of the remote sensing satellite imaging. Subsequently, the point feature descriptors of the scale-invariant feature transform (SIFT) of the reference image are calculated offline. Meanwhile, with the ground calibration field as the center and the field of view as kilometer-scale, a small-format high-resolution remote sensing image is selected on the open-source remote sensing image as the second reference image for secondary matching, and the SIFT point feature descriptor of the reference image is calculated offline.

[0023] (2) Real-time image downsampling The satellite focuses on observing the ground calibration field and acquires large-format, high-resolution remote sensing images containing the target area in real time. Then, the captured images are downsampled and converted into low-resolution images, which serve as the low-resolution real-time remote sensing image for the initial matching stage.

[0024] (3) Coarse matching of remote sensing images For the aforementioned low-resolution real-time remote sensing imagery, a block-sliding window strategy is employed for region segmentation. After region segmentation, SIFT point feature descriptor extraction is performed sequentially for each independent window image block.

[0025] After the SIFT point feature descriptor of a single window image block is extracted, the real-time matching process is immediately started: the feature descriptor of the window image block is compared one by one with the pre-built SIFT point feature descriptor library of the first reference image, the feature similarity is calculated by the feature similarity evaluation algorithm, and the feature point pairs with the highest matching degree are initially selected based on the similarity score.

[0026] After completing feature matching for all window image patches, the feature pairs output by each window are first summarized and integrated to construct an initial feature pair set covering the entire area of ​​the low-resolution real-time remote sensing image. To eliminate mismatched point pairs caused by noise interference, texture similarity, and other factors in the initial set and improve the reliability of the feature matching results, the Random Sample Consensus (RANSAC) algorithm is introduced for post-processing. Finally, the effective feature point pair set and the optimal perspective transformation matrix that meet the preset constraints are selected.

[0027] (4) Calculation of calibration point location and real-time image segmentation After obtaining the perspective transformation matrix based on the above principle, the corresponding pixel position on the real-time map is calculated by using the known pixel coordinates of the ground calibration points on the reference map through matrix transformation.

[0028] Based on the location of the calibration point in the real-time image, the corresponding image area is extracted from the high-resolution real-time image and used as the high-resolution real-time image for the secondary fine matching stage, thereby providing high-resolution data support for subsequent accurate positioning and matching.

[0029] (5) Fine matching of remote sensing images To further improve the positioning accuracy of calibration points, the same matching process as the coarse matching stage is adopted. Using the high-resolution real-time map of calibration points obtained in the previous operation and the pre-prepared second reference map, the feature extraction and matching operation is re-executed to obtain a set of effective feature point pairs that meet the preset constraints, providing a more reliable benchmark for subsequent parameter estimation and precise positioning of the target area. (6) Calculation of latitude and longitude of the center of the real-time map Based on the feature point pair set obtained through secondary fine-matching, the homography matrix is ​​recalculated, and a high-precision position mapping model from the real-time image to the reference image is established. According to this mapping model, the pixel coordinates of the center point in the real-time image are used to calculate the corresponding pixel position in the reference image through matrix transformation.

[0030] After obtaining the corresponding pixel position of the real-time image center point pixel on the reference image, and combining this with the pre-established correspondence between the reference image pixel coordinates and their geographic latitude and longitude coordinates, the bilinear interpolation method is used to calculate the latitude and longitude coordinates corresponding to that pixel position on the reference image. These latitude and longitude coordinates are the latitude and longitude coordinates of the real-time image center point, which can provide a precise spatial position reference for subsequent satellite attitude correction.

[0031] The technical solution of the present invention will be further illustrated below through specific embodiments. The resolution and area of ​​the first reference image are lower and larger than those of the real-time images acquired by the satellite remote sensing camera, respectively. The resolution and area of ​​the second reference image are not lower and are smaller than those of the real-time images acquired by the satellite remote sensing camera, respectively. Furthermore, the area of ​​the first reference image is a field of view of hundreds of kilometers, and the area of ​​the second reference image is a field of view of kilometers. Therefore, in specific embodiments, the first reference image is called a large-format low-resolution remote sensing reference image, and the second reference image is called a small-format high-resolution remote sensing reference image.

[0032] Specifically, this invention provides a high-precision satellite attitude correction method for large initial positioning errors. By preparing ground reference maps and matching them with two-stage remote sensing images, the position error of the target area is gradually reduced, thereby accurately locating the latitude and longitude coordinates of the center point of the real-time map and ultimately achieving satellite attitude correction.

[0033] like Figure 1As shown, firstly, two types of remote sensing image references are prepared on the ground: one is a large-format, low-resolution remote sensing reference image suitable for the coarse matching stage, and the other is a small-format, high-resolution remote sensing reference image adapted for the fine matching stage. Secondly, the high-resolution real-time remote sensing image is downsampled to generate a low-resolution real-time image. Then, a sliding window matching strategy is used to match it with the large-format, low-resolution remote sensing reference image to quickly achieve coarse localization of the target area and obtain its approximate spatial range. Finally, based on the above coarse localization results, a sub-image of the corresponding area in the high-resolution real-time remote sensing image is extracted and subjected to a second fine matching with the small-format, high-resolution remote sensing reference image. The two-stage matching mechanism effectively reduces matching deviation and ultimately achieves accurate correction of the satellite attitude. The specific implementation includes the following steps: (1) Preparation of reference diagram At the ground end, the ground calibration field is used as the target area, and a large-format, low-resolution remote sensing image is selected from open-source remote sensing images, centered on this area. This serves as a reference map for the initial matching. The reference map has a field of view of approximately 100 kilometers, slightly larger than that of remote sensing satellite imaging. Subsequently, the SIFT point feature descriptor sub-base of the reference map is calculated offline. ; Meanwhile, with the ground calibration field as the center and a kilometer-scale field of view, small-format, high-resolution remote sensing images were selected from open-source remote sensing imagery. This serves as a reference image for secondary matching, and the SIFT point feature descriptor sub-library of the reference image is computed offline. .

[0034] Specifically, the SIFT point feature descriptor extraction method is as follows: Input: Image Output: Feature descriptors (a) Constructing the Gaussian image pyramid:

[0035] (b) Constructing the Gaussian difference pyramid:

[0036] (c) Extreme point detection and key point localization:

[0037] (d) Determine the main direction of the feature point: Calculate the gradient magnitude and direction of the neighborhood of the key point, construct a direction histogram, and take the peak value as the main direction. (e) Feature descriptor normalization Within the rotating window that references the main direction at key points, press 4. The statistical gradient direction histograms of the four sub-regions are connected into a vector and normalized to generate descriptors.

[0038] To further explain, after extracting feature descriptors, a large-format, low-resolution remote sensing image feature descriptor library will be created. Small-format high-resolution remote sensing image library They are pre-stored together in a dedicated storage module.

[0039] (2) Real-time image downsampling During the attitude correction phase, the satellite focuses on observing the ground calibration area and acquires large-format, high-resolution remote sensing images containing the target area in real time. Then, the captured images are downsampled and converted into low-resolution images, which serve as the low-resolution real-time remote sensing image for the initial matching phase. .

[0040] Downsampling essentially reduces spatial resolution and the number of pixels by performing "aggregation calculations" on the original pixel set. Let the downsampling ratio be... (i.e., each original pixel is generated) (target pixels), original image pixel grayscale value The grayscale value of the image pixels after downsampling is To simplify the calculation, this invention directly selects the original pixel at the center of the pixel block as the target pixel, as described in formula (11).

[0041] (11) (3) Coarse matching of remote sensing images Regarding the aforementioned low-resolution real-time remote sensing images A segmented sliding window strategy is used to divide the area.

[0042] (12) The sliding window size can be set to After the region is divided, SIFT point feature descriptor extraction is performed sequentially for each independent window image block.

[0043] In the Each window corresponds to an image block. SIFT point feature descriptor sub-library Once extraction is complete, the real-time matching process is initiated immediately: based on The algorithm compares the feature descriptors of the window image patch with a pre-built SIFT point feature descriptor library of large-format low-resolution reference images. One-by-one comparisons are performed, and feature similarity is calculated using feature similarity evaluation methods such as Euclidean distance. Based on the similarity scores, feature point pairs with the highest matching degree are initially selected.

[0044] The Euclidean distance is used to measure feature similarity, and the similarity measurement formula is shown in (13), where A SIFT point feature descriptor sub-library for image patches corresponding to real-time images The feature descriptor of a certain feature point in the data. For reference image SIFT point feature description sub-library A feature descriptor for a specific feature point in the dataset.

[0045] (13) After completing feature matching for all window image patches, the feature pairs output by each window are first summarized and integrated to construct an initial set of m feature pairs covering the entire area of ​​the low-resolution real-time remote sensing image. To eliminate mismatched point pairs in the initial set due to noise interference, texture similarity, and other factors, and to improve the reliability of feature matching results, the RANSAC random sampling consensus algorithm is introduced for post-processing. Finally, a set of n valid feature point pairs that meet preset constraints is selected. With the optimal perspective transformation matrix .

[0046] Specifically, the RANSAC random sampling consensus algorithm method is as follows: Input: Set of matching pairs

[0047] Output: Perspective transformation matrix T (a) Randomly select the minimum sample set from the matching set

[0048] (b) Estimate the parameters of the perspective transformation matrix using the least squares method based on S.

[0049] (c) For all sets of matching pairs, compute the consistent set.

[0050] (d) Calculate the interior point rate of the current model within the consistency set B. (e) Repeat the above steps until the optimal perspective transformation matrix that satisfies the requirements of the interior point ratio is found. .

[0051] (4) Calculation of calibration point location and real-time image segmentation The perspective transformation matrix is ​​obtained based on the above principle. Then, using the known pixel coordinates of ground calibration points on the reference map... ( , The corresponding pixel position on the real-time image is calculated through matrix transformation. ( , The specific calculation formula is shown in formula (14). Based on this mapping result, points are used in the high-resolution real-time image. Centered on a specific image area, a corresponding region is extracted to serve as the high-resolution real-time image for the secondary fine-matching stage. .

[0052] (14) The processed real-time images fully retain the high-resolution details of the original images; moreover, their coverage is more focused, resulting in a moderate number of features. This not only significantly improves the computational efficiency of feature matching but also further enhances the accuracy of the positioning results, providing reliable high-resolution data support for subsequent high-precision positioning and fine-matching tasks.

[0053] (5) Fine matching of remote sensing images To further improve the positioning accuracy of calibration points, the same matching process as in the coarse matching stage is adopted, utilizing the high-resolution real-time image of calibration points obtained in the aforementioned operation. and pre-prepared small-format high-resolution reference image feature descriptors The feature extraction and matching operations are performed again to obtain a set of k valid feature point pairs that satisfy the preset constraints. This provides a more reliable benchmark for subsequent parameter estimation and precise positioning of the target area.

[0054] (6) Calculation of latitude and longitude of the center of the real-time map The feature point pair set obtained based on the second-order fine matching filter Recalculate the homography matrix A high-precision position mapping model is established from the real-time image to the reference image. Based on this mapping model, the pixel coordinates of the center point of the real-time image are... ( , The corresponding pixel position on the reference image is calculated through matrix transformation. ( , The specific calculation formula is shown in formula (15): (15) To obtain the corresponding pixel position of the center point pixel in the real-time image on the reference image. ( , Then, combining the pre-established correspondence between the pixel coordinates of the reference map and their geographic latitude and longitude coordinates, the bilinear interpolation method is used to calculate the latitude and longitude coordinates corresponding to the pixel position on the reference map. ( , (Latitude and longitude coordinates) This corresponds to the latitude and longitude coordinates of the center point of the real-time image, which can provide a precise spatial position reference for subsequent satellite attitude correction.

[0055] In summary, this invention provides a high-precision satellite attitude correction method for large initial positioning errors. By preparing ground reference maps and matching them with two-stage remote sensing images, the method gradually reduces the position error of the target area and accurately locates the latitude and longitude coordinates of the center point of the real-time image. This effectively solves the problem of difficult satellite positioning caused by large initial position errors in target areas in large-format real-world images.

[0056] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-precision satellite attitude correction method for extreme initial positioning errors, characterized in that, Includes the following steps: S1, using the ground calibration field as the target area, select a reference area to prepare a first reference map and a second reference map respectively; the resolution and area of ​​the first reference map are lower and larger than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively; the resolution and area of ​​the second reference map are not lower and smaller than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. S2, downsample the images acquired in real time by the satellite remote sensing camera and generate a low-resolution real-time image with a resolution lower than that of the real-time image; perform image matching between the low-resolution real-time image and the first reference image, calculate the optimal window position of the ground calibration point in the reference area in the low-resolution real-time image based on the matching result, and cut the real-time image into blocks based on the optimal window position to obtain a high-resolution calibration point real-time image. Step S2, which involves image matching between the low-resolution real-time image and the first reference image, includes the following steps: Offline extraction of point feature descriptors of scale-invariant feature transformation of the first reference image, and construction of a point feature descriptor library of the first reference image; The low-resolution real-time image is divided into regions to obtain multiple window image blocks, and point feature descriptor extraction is performed sequentially on each independent window image block. The point feature descriptors of a single window image block are compared one by one with the point feature descriptor library of the first reference image. The feature similarity is calculated by the feature similarity evaluation algorithm, and the feature point pairs with the highest matching degree are initially selected based on the similarity score. After completing feature matching for all window image blocks, the feature point pairs output by each window image block are integrated to construct a first initial set of feature point pairs covering the entire area of ​​the low-resolution real-time image. The first initial feature pair set is processed by the random sampling consensus algorithm to select the first effective feature point pair set and the first perspective transformation matrix that meet the preset constraints. S3, perform image matching between the high-resolution calibration point real-time image and the second reference image, establish a matching mapping mathematical model based on the matching result, and calculate the latitude and longitude coordinates of the real-time image center through bilinear interpolation to achieve satellite attitude correction; Step S3, which involves image matching between the real-time high-resolution calibration point image and the second reference image, includes the following steps: Offline extraction of point feature descriptors of scale-invariant feature transformation of the second reference image, and construction of a point feature descriptor library of the second reference image; The high-resolution calibration point real-time image is divided into multiple window image blocks, and the point feature descriptor extraction operation is performed sequentially on each independent window image block. The point feature descriptors of a single window image block are compared one by one with the point feature descriptor library of the second reference image. The feature similarity is calculated by the feature similarity evaluation algorithm, and the feature point pairs with the highest matching degree are initially selected based on the similarity score. After completing feature matching for all window image blocks, the feature point pairs output by each window image block are integrated to construct a second initial feature point pair set covering the entire area of ​​the high-resolution calibration point real-time image. The second initial feature pair set is processed by a random sampling consensus algorithm to select the second effective feature point pair set and the second perspective transformation matrix that meet the preset constraints.

2. The high-precision satellite attitude correction method for maximum initial positioning error according to claim 1, characterized in that, The first reference image has a field of view of hundreds of kilometers, while the second reference image has a field of view of kilometers.

3. The high-precision satellite attitude correction method for maximum initial positioning error according to claim 1, characterized in that, In step S2, the pixel block center aggregation method is used to downsample the images acquired in real time by the satellite remote sensing camera. The specific formula is as follows: in represents the grayscale value of the image pixels after downsampling, where and The pixel position in the image. These are the original image pixel grayscale values. This represents the downsampling ratio.

4. The high-precision satellite attitude correction method for maximum initial positioning error according to claim 1, characterized in that, Euclidean distance is used to measure feature similarity during image matching. The specific formula is as follows: in For a low-resolution real-time image or a high-resolution calibration point real-time image, the point feature descriptor is a point feature descriptor for a feature point in the image block point feature descriptor library; A feature descriptor for a feature point in the feature descriptor library of the first or second reference image.

5. A high-precision satellite attitude correction method for maximum initial positioning error according to claim 1, characterized in that, Step S3, calculating the latitude and longitude coordinates of the real-time image center, includes the following steps: Based on the second set of effective feature point pairs, a matching mapping mathematical model is established; according to the matching mapping mathematical model, the pixel coordinates of the center point of the real-time image are calculated by matrix transformation to determine the corresponding pixel position on the second reference image. Based on the pre-established correspondence between pixel coordinates of the second reference map and geographic latitude and longitude coordinates, the bilinear interpolation method is used to calculate the latitude and longitude coordinates corresponding to the pixel positions on the second reference map.

6. The system for high-precision satellite attitude correction method for maximum initial positioning error according to any one of claims 1-5, characterized in that, include: The reference image creation module is used to prepare a first reference image and a second reference image. The resolution and area of ​​the first reference image are lower and larger than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. The resolution and area of ​​the second reference image are not lower and are smaller than the resolution and area of ​​the real-time images acquired by the satellite remote sensing camera, respectively. The real-time image processing module is used to downsample images acquired in real time by satellite remote sensing cameras and generate low-resolution real-time images with a resolution lower than that of the real-time images; it is also used to slice the real-time images according to the optimal window position to obtain high-resolution calibration point real-time images. The matching module is used to perform image matching between the low-resolution real-time image and the first reference image and image matching between the high-resolution calibration point real-time image and the second reference image, and to obtain the first perspective transformation matrix and the second perspective transformation matrix respectively. The coordinate calculation module is used to calculate the latitude and longitude coordinates of the real-time map center based on the image matching results; An attitude correction module is used to adjust the satellite attitude based on the latitude and longitude coordinates.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 5.

Images