A Landing Point Localization Method Based on Small Samples of Extraterrestrial Topographic Images

Abstract

A high-generalization landing point localization method based on small sample data from extraterrestrial topographic images is proposed. The method acquires the DOM (Domain Image of Landing Area) of the spacecraft landing zone and the original descent sequence images. The DOM is segmented into small blocks, denoted as sub-base maps. Adaptive frame extraction is performed on the descent sequence images to calculate the homography matrix between adjacent frames and recover the relative pose of the cameras. The estimated altitude of each frame is obtained, and the resolution of each frame in the descent sequence is calculated. The descent image closest to the physical resolution of the DOM is selected as the descent image keyframe. Exhaustive matching is performed between the descent image keyframe and all sub-base maps to select the optimal sub-base map. Matching point pairs between adjacent descent sequence images are obtained. Combining the matching point pairs between the descent image keyframe and the optimal sub-base map, the coarse latitude and longitude coordinates of the landing point are obtained. The optimal sub-base map is added to the descent sequence for clustering adjustment to obtain the fine latitude and longitude coordinates of the landing point.

Description

Technical Field

[0001] This invention relates to a landing point localization method with high generalization based on small samples of extraterrestrial topographic images, belonging to the field of visual navigation technology. Background Technology

[0002] Precise landing point coordinates play a crucial role in missions such as landing and return, geological analysis, and mineral localization. For various probes landing on extraterrestrial bodies, precise landing point coordinates are of great value for engineering implementation and scientific research, and are also significant for analyzing spacecraft maneuvers during the landing process.

[0003] Given the unique nature of the landing sites for extraterrestrial probes and the short duration of their missions, the data required for landing point localization is scarce and timeliness is extremely important (e.g., lunar far side, Mars surface sample return missions). Traditional landing point localization algorithms often have limitations in matching accuracy and versatility. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a landing point localization method with high generalization based on small sample topographic images of extraterrestrial objects, thereby solving the need for high generalization and rapid realization of extraterrestrial object landing localization under small sample conditions.

[0005] The technical solution of this invention is: a high-generalization landing point localization method based on small samples of extraterrestrial topographic images, wherein:

[0006] Acquire images of the large DOM (Domain Image) within the pre-defined landing area of ​​the spacecraft and the original descent sequence of the spacecraft landing;

[0007] The large DOM region is divided into small blocks, which are denoted as sub-base maps; adaptive frame extraction is performed on the spacecraft landing descent sequence images to obtain the frame-extracted descent sequence images, which are denoted as descent image sequences and used for subsequent processing.

[0008] Calculate the homography matrix between adjacent frames in the descent image sequence, and use the homography matrix to recover the relative pose of the cameras between frames; obtain the estimated altitude of each frame by constraining the physical trajectory of the spacecraft landing process, and calculate the resolution corresponding to each frame in the descent image sequence using camera parameters;

[0009] Based on the physical resolution corresponding to each frame of the descent image sequence, the descent image closest to the physical resolution of the DOM is selected as the descent image keyframe; the descent image keyframe is exhaustively matched with all sub-base images to select the best sub-base image.

[0010] Inter-frame matching is performed on the descent image sequence to obtain matching point pairs between adjacent descent sequence images. Combining the matching point pairs between the descent image keyframe and the best sub-base map, the inter-frame homography matrix is ​​obtained through the image matching algorithm. The landing point is transferred from the last frame of the descent image to the descent image keyframe and then to the best sub-base map through homography transformation to obtain the coarse latitude and longitude coordinates of the landing point.

[0011] The optimal sub-base map is added to the descent image sequence for clustering adjustment. The pose of the optimal sub-base map relative to the descent sequence is obtained by continuously optimizing the reprojection error of the corresponding points. The landing point is then directly projected onto the optimal sub-base map according to the pose, thereby obtaining the latitude and longitude coordinates of the precise positioning point.

[0012] Preferably, the initial acquired large-area DOM has a length and width of tens of thousands of meters, and the original descent sequence images of the spacecraft landing are no less than 500 frames.

[0013] Preferably, when using the adaptive frame extraction algorithm to adaptively extract frames from the original descent sequence image of the spacecraft landing, the input parameters include the frame interval and the similarity threshold, and the output is a frame sequence for trajectory reconstruction; wherein: the frame interval is used to accelerate the algorithm and specifies the number of intermediate frames skipped between adjacent extracted frames; the similarity threshold is a parameter used to adaptively adjust the frame extraction interval;

[0014] The Structural Similarity Index (SSIM) is used as a metric to measure the similarity between adjacent landing images: frame extraction ends when the SSIM between adjacent extracted frames meets the metric requirements.

[0015] The Structural Similarity Index (SSIM) is calculated as follows:

[0016]

[0017] Where x and y are two non-negative image signals, C1 and C2 are stability constants used to avoid the case where the denominator is zero; μ x and μ y σ and y are the mean values ​​of the image, representing brightness information; x and σ y These are the standard deviations of the image's x and y values, representing contrast information; σ xy is the covariance of images x and y, representing structural information; l(x,y) is a brightness similarity measure, used to measure the similarity between the average brightness of two images; c(x,y) is a contrast similarity measure, used to measure the similarity between the contrast of two images; s(x,y) is a structural similarity measure, used to measure the structural similarity between two images; SSIM(x,y) is a combined result of brightness, contrast, and structural similarity measures, and is the structural similarity index between two images.

[0018] Preferably, when calculating the homography matrix between adjacent frames of the descending image sequence after frame extraction, a matching algorithm is used and the matching result is optimized by bundle adjustment to obtain an optimized homography matrix, thereby recovering the relative pose of the cameras between frames.

[0019] Preferably, the estimated altitude for each frame is obtained by constraining the physical trajectory of the spacecraft during landing, specifically as follows:

[0020] During the final stage of descent, the lander does not perform any horizontal maneuvers and remains perpendicular to the surface of the extraterrestrial object. At this time, the camera position can be projected onto the vertical axis relative to the surface of the extraterrestrial object, and the estimated altitude for each frame can be obtained by combining the relative pose of the cameras between frames.

[0021] Preferably, the method for calculating the resolution corresponding to each frame of the descending image sequence using camera parameters is as follows:

[0022]

[0023] Among them, R x and R y These are the physical resolutions in the x and y directions, respectively; θ x and θ y represents the field of view angles in the x and y directions, respectively, while pix x and pix y represents the pixel resolution in the camera's x and y directions, respectively; h is the height of the current descending sequence image.

[0024] Preferably, when exhaustively matching the keyframes of the descending image with all sub-base maps to select the best sub-base map:

[0025] The pre-trained deep learning networks SuperPoint and SuperGlue are used. The SuperPoint network extracts key points from image pairs, and then the SuperGlue network performs pairwise matching of key points and evaluates the matching score.

[0026] Using predefined matching scores and a threshold for the number of matched keypoint pairs as evaluation criteria, keypoint pairs exceeding the score threshold are considered as matched keypoint pairs; sub-images with more matched keypoints than the number threshold are considered as potential best sub-images, and potential best sub-images are sorted according to matching scores, with the image with the highest matching score selected as the best sub-image.

[0027] Preferably, during the image matching process:

[0028] The images to be matched are rotated at certain angular intervals to avoid incorrect matching caused by angular differences;

[0029] There is a significant lighting difference between the DOM and the landing image. Using the DOM as a reference image, histogram normalization is applied to the keyframes of the landing image to avoid the significant lighting difference between the DOM and the landing image.

[0030] Preferably, after determining the keyframes of the descent image sequence and the optimal sub-base map, the keyframes of the descent image sequence and their adjacent images contain a sufficient number of connection points projected onto the optimal sub-base map; the attitude of the input image and the projections of several connection points on the corresponding images and their three-dimensional coordinates are calculated using airborne triangulation; considering the attitude of the optimal sub-image and the position of the connection points in the world coordinate system, the projection of the connection points on the optimal sub-base map can be calculated, and finally the latitude and longitude of the landing point are determined according to its physical coordinate system.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] (1) This invention designs an algorithm for matching and adaptive frame extraction by using structural similarity (SSIM) between landing images during the descent of an extraterrestrial probe. The feature point extraction and matching algorithm under small sample conditions is used to obtain the probe pose sequence, thereby improving computational efficiency.

[0033] (2) This invention achieves high-robustness map feature point matching in a wider range of scenarios through pre-location and matching algorithms, and realizes high-precision positioning calculation. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention;

[0035] Figure 2 This refers to the software interface of the framework proposed in this invention. Detailed Implementation

[0036] Given the significance of accurate landing point positioning for spacecraft landing missions and the limitations of traditional landing point positioning algorithms in terms of data source dependence and applicability, this invention proposes a method to address the need for high generalization and rapid implementation of extraterrestrial landing positioning under small sample conditions.

[0037] Firstly, an algorithm for selecting keyframes of the landing images is designed to obtain the landing image in the landing sequence whose physical resolution is closest to the base map, thus improving the matching of heterogeneous images. To further improve algorithm efficiency, this invention applies an adaptive frame extraction algorithm in this step. Subsequently, the sequence after frame extraction is used for matching with source data and trajectory reconstruction. Through physical constraints, this method converts the relative height between trajectories into altitude relative to the lunar surface and calculates the physical resolution in conjunction with camera parameters to select keyframes of the landing images.

[0038] Secondly, an optimal sub-base map selection algorithm is designed to determine the coordinates of the landing area within the sub-base map. Based on keyframes of the landing image, a large-scale base map is divided into blocks and subjected to iterative matching. A confidence score is used as a measure of matching accuracy to select the optimal sub-base map. For the matching algorithm, a combination of preprocessing and deep learning is employed to obtain more robust matching feature points.

[0039] Thirdly, a coarse landing point localization algorithm is designed, which mainly includes landing point transfer between landing sequences and landing point transfer from heterogeneous data. The former obtains the landing point coordinates on keyframes of the landing image, while the latter obtains the landing point coordinates with latitude and longitude information on the base map. The matching algorithm remains consistent with the previous one, using a combination of preprocessing and deep learning to obtain matching feature points.

[0040] Fourthly, a precise landing point localization algorithm is designed. The optimal sub-base map is added to the descent image sequence for cluster adjustment. By continuously optimizing the reprojection error of feature matching points in different image frames, the pose of the optimal sub-base map relative to the descent sequence is obtained. The landing point is then directly projected onto the optimal sub-base map according to the pose, thereby obtaining the latitude and longitude coordinates of the precise localization point.

[0041] The technical implementation plan is as follows.

[0042] Step 1: Acquire a large DOM (with dimensions in the tens of thousands of meters) of the spacecraft's pre-defined landing area and a sequence of images of the spacecraft's landing and descent (usually more than 500 frames).

[0043] Step 2: Preprocessing the large-area DOM and the spacecraft landing descent image sequence: This includes segmenting the DOM into smaller blocks and adaptively extracting frames from the spacecraft landing descent image sequence. The result is a small-area DOM (sub-base map) with a physical resolution of kilometers at a certain pixel resolution (e.g., 1080P, 4K). In addition, based on the adaptive frame extraction results, a certain proportion (e.g., 1 / 10, 1 / 5) of the original sequence of descent images is obtained for subsequent trajectory reconstruction tasks.

[0044] The adaptive frame extraction method for image sequences captured by landing cameras is characterized by reducing the number of frames in the landing sequence through a specific frame extraction strategy, thereby accelerating the trajectory reconstruction and landing point localization process. The input parameters of the adaptive frame extraction algorithm include frame interval and similarity threshold. The frame interval is used to accelerate the algorithm and specifies the number of intermediate frames skipped between adjacent extracted frames. The similarity threshold is a parameter used to adaptively adjust the frame extraction interval. The algorithm outputs a frame sequence for trajectory reconstruction. When the similarity is low, the interval is reduced until the interval is 1. The algorithm uses the structural similarity index (SSIM) as an indicator to measure the similarity between adjacent landing images. SSIM is based on a mathematical framework of signal processing and includes three main components: brightness (average intensity of pixels), contrast (dynamic range of pixel intensity), and structure (spatial pattern of pixel intensity). As shown in Equation (1), x and y are two non-negative image signals, and C1 and C2 are stability constants used to avoid the case where the denominator is zero. μ x and μ y σ represents the mean values ​​of the x and y sides of the image, respectively, indicating brightness information. x and σ y σ and y are the standard deviations of the image, representing contrast information. xy is the covariance of images x and y, representing structural information. l(x,y) is a brightness similarity measure, used to measure the similarity between the average brightness of two images. c(x,y) is a contrast similarity measure, used to measure the similarity between the contrast of two images. s(x,y) is a structural similarity measure, used to measure the structural similarity between two images. SSIM(x,y) is a combined result of brightness, contrast, and structural similarity measures, representing the structural similarity index between two images. The adaptive frame extraction algorithm effectively saves computational resources while ensuring the accuracy of trajectory reconstruction and keyframe selection from landing images.

[0045]

[0046] Step 3: Calculation of Height Resolution for Descent Image Sequence: Using the descent image sequence after frame extraction, a matching algorithm is used to obtain the homography matrix between adjacent frames, and the matching results of the image sequence are optimized through bundle adjustment. The relative pose of the cameras between frames is recovered using the optimized homography matrix. Considering that the camera is stationary relative to the spacecraft, this relative pose is the spacecraft pose. In addition, the position of the spacecraft on the vertical lunar surface altitude axis is determined by the physical trajectory constraints of the spacecraft landing process to obtain the estimated altitude of each frame, and the corresponding resolution of each frame is calculated using the camera parameters.

[0047] The height resolution estimation of descent image sequences is characterized by the fact that landing point localization depends on the matching between the descent images and the DOM (Domain of Deposition). Overcoming the physical resolution difference between the descent images and the DOM is one of the main challenges. The physical resolution of the descent image sequence decreases from large to small, so it is necessary to find a keyframe with a resolution close to that of the DOM to bridge the resolution gap. After acquiring the external parameters of the cameras corresponding to each image frame in the early stage, the relative positions can be estimated and the relative trajectories can be reconstructed. After obtaining the relative trajectories, the physical information of the descent images is integrated to estimate the true height of each camera position relative to the surface of the extraterrestrial object. The physical information implies that in the final stage of descent, the lander does not perform any horizontal maneuvers and maintains a vertical orientation relative to the surface of the extraterrestrial object. Therefore, to estimate the true height of each camera position, the camera position can be projected onto the vertical axis relative to the surface of the extraterrestrial object. Specifically, each camera position p is defined as follows: t (t is a subscript representing each frame. For example, the height of the 2nd and 5th frames is taken as shown in formula (2):

[0048]

[0049] Where, p e-2 and p e-5 These are the 3D position vectors of the camera in the 2nd and 5th frames before descent; these frames were chosen to ensure no significant horizontal maneuvering relative to the surface of the extraterrestrial object, making the descent essentially vertical. Vector z represents the altitude axis perpendicular to the surface of the extraterrestrial object in the current coordinate system; h ≤ 0 is the projection of the camera position on the altitude axis in the initial landing frame, and is considered as a relative altitude. y The constant represents the absolute altitude at the start of descent, which is a predetermined, known value.

[0050] After determining the actual altitude, the physical resolution of the descent image is calculated using camera parameters. The formula for calculating the physical resolution is shown in formula (3):

[0051]

[0052] Among them, R x and R y These are the physical resolutions in the x and y directions, respectively. θ x and θ y represents the field of view angles in the x and y directions, respectively, while pix x and pix y represents the pixel resolution in the x and y directions of the camera, respectively; h is the height of the current descent sequence image. Considering the tolerance of the resolution estimation results and the extremely small ratio between the descent height and the radius of the extraterrestrial object, the curvature of the extraterrestrial object was not considered in the height estimation process. After the above calculation, based on the known DOM resolution, keyframes of the descent image can be selected.

[0053] Step 4: Selection of keyframes and selection of the best sub-base image for the descent image sequence: Based on the physical resolution of each frame, the descent image closest to the DOM physical resolution is selected as the keyframe of the descent image sequence. By introducing preprocessing and deep learning feature extraction and matching algorithms, the matching score is obtained by exhaustively matching the descent image keyframes with all sub-base images. The sub-base image with the highest matching score is the best sub-base image.

[0054] The lunar surface image matching and DOM filtering algorithm is characterized by the following: After determining the keyframes of the landing image, a small-scale optimal sub-base map can be obtained using the keyframes of the descent image sequence to facilitate subsequent localization; a pre-trained deep learning network SuperPoint and SuperGlue are used to identify sub-images containing landing points by evaluating the matching score between the keyframes of the descent image sequence and each sub-base image. The sub-image containing the landing point is referred to as the optimal sub-base map; the SuperPoint network extracts keypoints from the matched image pairs, and then the SuperGlue network performs pairwise matching of the keypoints and evaluates the matching score; a predefined matching score and a threshold for the number of matched keypoint pairs are used as evaluation criteria, and keypoint pairs exceeding the score threshold are considered as matched point pairs; sub-images with more than one matching keypoint are considered as potential optimal sub-images, and the potential optimal sub-images are ranked according to the average matching score to select the best sub-image;

[0055] During the matching process: To address mismatches caused by angle differences, each sub-image is rotated from 0 degrees to 360 degrees at certain angular intervals; Due to the significant lighting differences between the DOM and the descent image, histogram normalization is applied to the keyframes of the descent image using the DOM as the reference image.

[0056] Step 5: Coarse Localization of Landing Point. In this framework, coarse localization relies on the transfer of homography matrix sequences between image frames, including the homography matrix between descent image sequences and the homography matrix between descent image keyframes and the optimal DOM. The process involves using the same preprocessing and deep learning-based feature point extraction and matching algorithm as in Step 4 to perform inter-frame matching on the descent image sequence. This yields matching point pairs between descent images. Combining these with the matching point pairs between the descent image keyframes and the optimal DOM, homography transformation transfers the landing point from the last descent image frame to the keyframe, and then to the optimal DOM, thereby obtaining the coarse latitude and longitude coordinates of the landing point.

[0057] The landing point coarse localization algorithm is characterized by the fact that after determining the key frames of the descent image and the best sub-base map, it can obtain the inter-frame homography matrix through the image matching algorithm and transfer the camera pose.

[0058] The camera pose transfer algorithm is characterized by two main stages. The first is the matching of source images from the center point of the last frame of the descent image to the key frame of the descent image, which is the landing point. The second is the matching of source images between the key frame of the descent image and the best sub-base image. Since there are significant lighting differences between the latter image pairs, histogram normalization is applied to the key frame of the descent image with the DOM as the reference image.

[0059] Step 6: Fine localization of the landing point. Compared with coarse localization, fine localization adds an optimization method of inter-sequence clustering adjustment. The best sub-base map is added to the descent image sequence for clustering adjustment. By continuously optimizing the reprojection error of corresponding points, the pose of the best sub-base map relative to the descent sequence is obtained. The landing point is then directly projected onto the best sub-base map according to its pose, thereby obtaining the latitude and longitude coordinates of the finely localized point.

[0060] The landing point precision localization algorithm is characterized by the following: after determining the keyframes of the descent image and the optimal sub-base map, the keyframes of the descent image sequence and their adjacent images contain a sufficient number of connection points projected onto the optimal sub-base map, i.e., matching point pairs; the attitude of the input image (descent sequence image) and the projections of several connection points on the corresponding images and their three-dimensional coordinates are calculated using airborne triangulation; considering the attitude of the optimal sub-image and the position of the connection points in the world coordinate system, the projection of the connection points on the optimal sub-base map can be calculated, and finally the latitude and longitude of the landing point are determined according to its physical coordinate system.

[0061] Example:

[0062] The specific application of this invention is the landing of a lunar probe in the South Pole-Aitken Basin on the far side of the moon. The entire detailed process can be found in [link to documentation]. Figure 1 .

[0063] Image data preprocessing is performed. After acquiring the lunar probe's descent images and the base map with latitude and longitude coordinates of the landing site, the base map is segmented to a certain pixel resolution. An adaptive descent image frame extraction algorithm is used on the descent image sequence. Redundant data is removed by manually setting the start and end frames and the frame extraction interval. The frame extraction interval is flexibly adjusted according to the similarity index SSIM between descent images to achieve the adaptive frame extraction algorithm.

[0064] Descent trajectory reconstruction and altitude resolution estimation. Airborne triangulation was used to perform feature matching and bundle adjustment on the extracted descent image sequence to obtain the pose of the camera corresponding to the descent image. Based on the relative pose, physical information from the descent images was integrated to estimate the true altitude of each camera attitude relative to the lunar surface and the physical resolution of the image. The descent image keyframes that are closest to the base map resolution were then obtained sequentially.

[0065] Determining the optimal sub-base image: Deep learning feature extraction and matching algorithms are introduced. By exhaustively matching the descent image keyframes with all sub-base images, a matching score is obtained. The sub-base image with the highest matching score is the optimal sub-base image.

[0066] A coarse localization algorithm for the landing point is employed. Using feature point extraction and matching algorithms, the SuperPoint algorithm is first used to extract feature points from the descent image sequence, and the SuperGlue feature matching algorithm is used to find matching point pairs between images. Simultaneously, feature point matching is performed on keyframes of the descent images and the optimal sub-base map to obtain matching point pairs. While ensuring a sufficient number of matching point pairs, points with lower matching scores are filtered out to guarantee the accuracy and robustness of the homography matrix calculation. Next, the calculated homography transformation matrix is ​​used to transfer the landing point from the last frame of the descent image to the keyframes, and then further to the optimal sub-base map. Finally, based on the landing point position transferred to the optimal sub-base map, the coarse latitude and longitude coordinates of the landing point are obtained using the geographic coordinate information of that sub-base map.

[0067] The algorithm for precise localization of the landing point works as follows: First, the optimal sub-base map is added to the descent image sequence, making it an integral part of the entire sequence. Next, cluster adjustment is performed. By continuously optimizing the reprojection error of feature matching points in different image frames, the relative position and pose of the entire image sequence are gradually corrected and adjusted, thereby obtaining the precise pose of the optimal sub-base map relative to the descent sequence. After completing these steps, the landing point is directly projected onto the optimal sub-base map based on the optimized pose. In this way, the latitude and longitude coordinates of the precise localization point are finally obtained.

[0068] Figure 2 The software interface design for the framework proposed in this invention includes modular functional windows such as data management, preprocessing, coarse positioning, and fine positioning.

[0069] This embodiment is a preferred example of the present invention, but it is not limited thereto. The extraterrestrial bodies to which this system is applicable include, but are not limited to, the Moon and Mars. Others skilled in the art can make different extensions and variations based on the structure described in this invention, but as long as they do not depart from the inventive spirit of this invention, they are all within the protection scope of this invention.

[0070] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A high-generalization landing point localization method based on small samples of extraterrestrial topographic images, characterized in that... include: Acquire images of the large DOM (Domain Image) within the pre-defined landing area of ​​the spacecraft and the original descent sequence of the spacecraft landing; Divide a large DOM region into smaller blocks, denoted as a sub-base map; Adaptive frame extraction is performed on the spacecraft landing descent sequence images to obtain the extracted descent sequence images, which are denoted as the descent image sequence and used for subsequent processing. Calculate the homography matrix between adjacent frames in the descent image sequence, and use the homography matrix to recover the relative pose of the cameras between frames; obtain the estimated altitude of each frame by constraining the physical trajectory of the spacecraft landing process, and calculate the resolution corresponding to each frame in the descent image sequence using camera parameters; Based on the physical resolution corresponding to each frame of the falling image sequence, the falling image that is closest to the physical resolution of the DOM is selected as the falling image keyframe. Exhaustive matching of keyframes of the descending image and all sub-base maps is performed to select the best sub-base map; Inter-frame matching is performed on the descent image sequence to obtain matching point pairs between adjacent descent sequence images. Combining the matching point pairs between the descent image keyframe and the best sub-base map, the inter-frame homography matrix is ​​obtained through the image matching algorithm. The landing point is transferred from the last frame of the descent image to the descent image keyframe and then to the best sub-base map through homography transformation to obtain the coarse latitude and longitude coordinates of the landing point. The optimal sub-base map is added to the descent image sequence for clustering adjustment. The pose of the optimal sub-base map relative to the descent sequence is obtained by continuously optimizing the reprojection error of the corresponding points. The landing point is then directly projected onto the optimal sub-base map according to the pose, thereby obtaining the latitude and longitude coordinates of the precise positioning point.

2. The method for high-generalization landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: The initial large-area DOM images were all in the tens of thousands of meters in length and width, and the original descent sequence images of the spacecraft landing were no less than 500 frames.

3. The method for high-generalization landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: When using the adaptive frame extraction algorithm to adaptively extract frames from the original descent sequence images of a spacecraft landing, the input parameters include the frame interval and the similarity threshold, and the output is a frame sequence for trajectory reconstruction; where: the frame interval is used to accelerate the algorithm and specifies the number of intermediate frames skipped between adjacent extracted frames; the similarity threshold is a parameter used to adaptively adjust the frame extraction interval; The Structural Similarity Index (SSIM) is used as a metric to measure the similarity between adjacent landing images: frame extraction ends when the SSIM between adjacent extracted frames meets the metric requirements. The Structural Similarity Index (SSIM) is calculated as follows: Where x and y are two non-negative image signals, C1 and C2 are stability constants used to avoid the case where the denominator is zero; μ x and μ y σ and y are the mean values ​​of the image, representing brightness information; x and σ y These are the standard deviations of the image's x and y values, representing contrast information; σ xy is the covariance of images x and y, representing structural information; l(x, y) is a brightness similarity measure, used to measure the similarity between the average brightness of two images; c(x, y) is a contrast similarity measure, used to measure the similarity between the contrast of two images; s(x, y) is a structural similarity measure, used to measure the structural similarity between two images; SSIM(x, y) is a combined result of brightness, contrast, and structural similarity measures, and is the structural similarity index between two images.

4. The method for high-generalization landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: When calculating the homography matrix between adjacent frames of the descending image sequence after frame extraction, a matching algorithm is used and the matching result is optimized by bundle adjustment to obtain the optimized homography matrix, based on which the relative pose of the cameras between frames is recovered.

5. The method for high generalization of landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: By constraining the physical trajectory of the spacecraft during landing, the estimated altitude for each frame is obtained, specifically: During the final stage of descent, the lander does not perform any horizontal maneuvers and remains perpendicular to the surface of the extraterrestrial object. At this time, the camera position can be projected onto the vertical axis relative to the surface of the extraterrestrial object, and the estimated altitude for each frame can be obtained by combining the relative pose of the cameras between frames.

6. The method for high-generalization landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: The method for calculating the resolution corresponding to each frame of the descending image sequence using camera parameters is as follows: Among them, R x and R y These are the physical resolutions in the x and y directions, respectively; θ x and θ y represents the field of view angles in the x and y directions, respectively, while pix x and pix y represents the pixel resolution in the camera's x and y directions, respectively; h is the height of the current descending sequence image.

7. The method for high generalization of landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: When exhaustively matching the keyframes of the descending image and all sub-base maps to select the best sub-base map: The pre-trained deep learning networks SuperPoint and SuperGlue are used. The SuperPoint network extracts key points from image pairs, and then the SuperGlue network performs pairwise matching of key points and evaluates the matching score. Using predefined matching scores and a threshold for the number of matched keypoint pairs as evaluation criteria, keypoint pairs exceeding the score threshold are considered as matched keypoint pairs; sub-images with more matched keypoints than the number threshold are considered as potential best sub-images, and potential best sub-images are sorted according to matching scores, with the image with the highest matching score selected as the best sub-image.

8. The method for high generalization of landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: During image matching: The images to be matched are rotated at certain angular intervals to avoid incorrect matching caused by angular differences; There is a significant lighting difference between the DOM and the landing image. Using the DOM as a reference image, histogram normalization is applied to the keyframes of the landing image to avoid the significant lighting difference between the DOM and the landing image.

9. The method for high-generalization landing point localization based on small samples of extraterrestrial topographic images according to claim 1, characterized in that: After determining the keyframes of the descent image sequence and the optimal sub-base map, the keyframes of the descent image sequence and their adjacent images contain a sufficient number of connection points projected onto the optimal sub-base map. The attitude of the input image and the projections of several connection points onto the corresponding images and their three-dimensional coordinates are calculated using airborne triangulation. Considering the attitude of the optimal sub-image and the position of the connection points in the world coordinate system, the projection of the connection points onto the optimal sub-base map can be calculated. Finally, the latitude and longitude of the landing point are determined based on its physical coordinate system.

Images