Moon landing navigation positioning method based on improved FDAFT matching algorithm

Abstract

The invention relates to a lunar landing navigation positioning method based on an improved FDAFT matching algorithm, and the method comprises the following steps: a feature point matching step: obtaining a lander image, and a digital orthoimage and a digital elevation model of the region, carrying out the down-sampling of the lander image, and then carrying out the blocking, performing FDAFT feature point matching on each block and the digital orthoimage, and obtaining a corresponding relation between image two-dimensional point coordinates and geographic three-dimensional point coordinates in combination with a digital elevation model; and a pose estimation step: solving the pose of the lander by using a pose estimation algorithm according to the corresponding relationship between the image two-dimensional point coordinates and the geographic three-dimensional point coordinates and in combination with the internal reference information of the camera, and optimizing the preliminarily solved pose by using a pose optimization algorithm to obtain the final pose of the lander. Compared with the prior art, the method has the advantages that the number of matching points obtained by the lander image and the digital orthoimage is enough, the matching points are accurate and uniformly distributed, and the accuracy of pose estimation can be improved.

Description

Technical Field

[0001] The invention relates to the technical field of lunar landing navigation and positioning, and in particular to a lunar landing navigation and positioning method based on an improved FDAFT matching algorithm. Background Art

[0002] At present, various countries are conducting active lunar exploration activities. In lunar landing navigation and positioning, the relationship between image points and ground three-dimensional points is obtained through matching methods, which is the basis of positioning. In this field, matching methods can be roughly divided into three categories: template matching, pattern matching, and feature descriptor matching. Template matching uses the generated generalized surface features or part of the pixel matrix around the virtual control point as a template to match the landing image. This method is simple to understand and easy to implement, and has good robustness to simple geometric transformations such as translation. However, it is sensitive to rotation and scale changes and has a large amount of calculation; at the same time, the presence of illumination changes and noise may lead to poor matching results. Pattern matching focuses on using specific features in the scene, such as using craters on the moon for location recognition. Modern pattern recognition methods mostly use machine learning technology. These methods show strong robustness in the face of complex natural environments, but increase the complexity of implementation. In addition, pattern matching depends on the features of objects in the scene and is prone to failure in areas lacking significant ground features. Feature descriptor matching methods have strong robustness in dealing with complex scenes and have gradually become the focus of research. In recent years, various feature point matching technologies have developed rapidly. In addition to improving existing feature descriptors, many scholars have also proposed new feature descriptors to overcome complex scene differences. With the development of computer technology, feature matching methods based on deep learning have gradually emerged. Different feature descriptor matching methods have certain adaptability to different scenes.

[0003] In the lunar exploration mission, the acquisition conditions of the lander images are extremely complex. The changes in illumination and the unclear texture caused by the low contrast of the acquired images have brought severe challenges to the traditional feature matching methods. These factors make it difficult for traditional matching methods to obtain accurate and evenly distributed matching points, thus affecting the accuracy and stability of navigation positioning. Summary of the invention

[0004] The purpose of the present invention is to provide a lunar landing navigation and positioning method based on an improved FDAFT (Fast double-channel aggregated feature transform) matching algorithm.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A lunar landing navigation and positioning method based on an improved FDAFT matching algorithm comprises the following steps:

[0007] Feature point matching steps: obtain the lander image and the digital orthophoto (DOM) and digital elevation model (DEM) of the area, downsample the lander image and divide it into blocks, perform FDAFT feature point matching on each block with the digital orthophoto, and combine the digital elevation model to obtain the correspondence between the image 2D point coordinates and the geographic 3D point coordinates;

[0008] Pose estimation steps: Based on the correspondence between the two-dimensional image point coordinates and the three-dimensional geographic point coordinates, and combined with the camera intrinsic parameter information, the pose estimation algorithm is used to solve the lander's pose, and the pose optimization algorithm is used to optimize the preliminary solution to obtain the final pose of the lander.

[0009] The downsampling is to sample the high-resolution lander image into a low-resolution lander image, and the calculation method is as follows:

[0010]

[0011] Among them, r new With c new represents the row and column number of the low-resolution lander image, r and c represent the row and column number of the original resolution lander image, and n represents the downsampling ratio.

[0012] The standard of the block division is that the sum of the standard deviations of the pixels in the block after the block division is minimized.

[0013] The calculation formula of the standard deviation of pixels within the block is:

[0014]

[0015] Among them, σ i is the pixel standard deviation of the i-th block, N is the total number of pixels in the block, pix j is the pixel value of the jth point, is the average value of all pixels in the block.

[0016] The calculation formula for the sum of the standard deviations of pixels within the block is:

[0017]

[0018] Among them, M represents the number of blocks, and E represents the sum of the standard deviations of pixels in all blocks.

[0019] In the block process, the minimum block size S is set min As a constraint, the size of each block is limited to not less than a given threshold S min *S min .

[0020] The FDAFT feature point matching comprises the following steps:

[0021] The structural features of the image are extracted through ECM, and the phase features of the image are obtained by calculating the maximum moment mapping through weighted phase consistency;

[0022] The structural feature space and phase feature space are constructed by Gaussian filtering, and the FAST algorithm is used to detect corner points in the structural feature space, and the KAZE algorithm is used to detect corner points in the phase feature space;

[0023] Based on the detected corner points, the direction of the local block is divided into 12 parts through GLOH to construct the feature vector;

[0024] The initial matching of low-frequency and high-frequency feature points is achieved by calculating the SSD between feature vectors, and the FSC algorithm is used to eliminate potential mismatching points.

[0025] The pose estimation algorithm performs the following steps:

[0026] Based on the correspondence between the acquired 2D image points of the lander image and the 3D geographic points, four non-coplanar 3D points are selected as virtual control points, and all other 3D points are represented by these four virtual control points:

[0027]

[0028] Where P i w is a geographic 3D point, is a virtual control point, a ij is the coefficient;

[0029] According to the camera internal parameters and image point coordinates, solve the virtual control point coordinates

[0030]

[0031] Among them, f x ,f y ,u0,v0 are the camera internal parameters, u i 、v i is the image point coordinate, is the coordinate of the virtual control point in the camera coordinate system;

[0032] According to the coordinates of the virtual control point in the camera coordinate system, the three-dimensional coordinates of the geographic three-dimensional point in the camera coordinate system and the corresponding position and posture are solved:

[0033]

[0034] Among them, P i cis the coordinate of the geographic 3D point in the camera coordinate system, R cw is the rotation matrix from the world coordinate system to the camera coordinate system, t is the translation vector from the world geographic coordinate system to the camera coordinate system, is the position of the camera in the world coordinate system.

[0035] The pose optimization algorithm performs the following steps:

[0036] Note that the pose (R cw ,t) is p0, and the optimization is performed with the goal of minimizing the reprojection error, where the reprojection error vector r is expressed as:

[0037]

[0038] Among them, R cw is the rotation matrix from the world coordinate system to the camera coordinate system, t is the translation vector from the world geographic coordinate system to the camera coordinate system, P i c is the coordinate of the geographic 3D point in the camera coordinate system, P i w is a geographic 3D point, n is the number of geographic 3D points;

[0039] Calculate the first-order partial derivative of the reprojection error vector r with respect to the rotation matrix and translation vector Denoted as the Jacobian matrix, the damping factor λ is dynamically adjusted in each iteration according to the current reprojection error vector and the Jacobian matrix:

[0040]

[0041] Among them, Δp is the increment of the pose p0 after iterative calculation;

[0042] The initial iteration result U0 is The maximum value among the diagonal elements, when λ is greater than the preset threshold, increase U k , otherwise it decreases, and the following equation is solved to obtain Δp to implement iteration:

[0043]

[0044] Where I is the identity matrix, U k is the result of the kth iteration.

[0045] The posture optimization algorithm determines whether the optimization is completed by setting a preset number of iterations or a threshold. When Δp is less than the threshold or the iteration reaches the preset number of iterations, the posture optimization is considered to be completed.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention proposes a lunar landing navigation and positioning method based on an improved FDAFT matching algorithm. First, the lander image is pre-processed by downsampling to improve the computational efficiency. The pre-processed image is divided into blocks based on the criterion of the minimum sum of the standard deviations of the pixels in the blocks. The divided images are matched with the digital orthophotos by FDAFT, and the relationship between the image points and the geographic three-dimensional points is obtained by combining the digital elevation model. Combined with the camera internal parameters, the accurate posture of the lander can be obtained through posture estimation and posture optimization algorithms. Compared with the existing technology, the present invention can obtain a sufficient number of matching points with high accuracy and more uniform distribution in the feature point matching process, and can solve a more accurate lander posture. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0049] Figure 2 A dataset for verifying the effectiveness of the feature matching method of the present invention.

[0050] Figure 3 The performance of the present invention and the existing feature matching algorithm in three data sets is compared.

[0051] Figure 4 This is an indicator diagram for evaluating the quality of matching from different dimensions.

[0052] Figure 5 A comparison diagram of the lander positioning position and the lander design position for the present invention.

[0053] Figure 6 This is an object space accuracy evaluation diagram for the lander positioning result of the present invention.

[0054] Figure 7 This is an image accuracy evaluation diagram for the lander positioning result of the present invention. DETAILED DESCRIPTION

[0055] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0056] The present invention relates to a lunar landing navigation and positioning method based on an improved FDAFT (Fourier Descriptor-based AffineTransformation Feature) feature matching algorithm, including using a landing image obtained by a lander to match feature points with a digital orthophoto (DOM), dividing the image into blocks by taking the minimum sum of pixel standard deviations within the blocks as a criterion, performing FDAFT matching on each block of the image with the digital orthophoto, and obtaining the relationship between the two-dimensional points of the lander image and the geographic plane coordinates; then, combining a digital elevation model (DEM), obtaining the elevation of the coordinates of the geographic two-dimensional point, and obtaining the corresponding relationship between the coordinates of the image point and the coordinates of the geographic three-dimensional point. According to the corresponding relationship and the camera internal parameters, the pose of the lander is calculated by a pose estimation algorithm, and the pose of the initial solution is optimized with the minimum reprojection error as the goal, so as to obtain the pose of the final lander.

[0057] Specifically, Figure 1 As shown, the following steps are included:

[0058] S1, feature point matching step: obtain the lander image and the digital orthophoto (DOM) and digital elevation model (DEM) of the area, downsample the lander image and divide it into blocks, perform FDAFT feature point matching on each block with the digital orthophoto, and combine the digital elevation model to obtain the correspondence between the image 2D point coordinates and the geographic 3D point coordinates.

[0059] S11, downsampling

[0060] Downsampling is to downsample the high-resolution lander image to a low-resolution lander image. The calculation method is as follows:

[0061]

[0062] Among them, r new With c new represents the row and column number of the low-resolution lander image, r and c represent the row and column number of the original resolution lander image, and n represents the downsampling ratio.

[0063] The low-resolution image pixel values ​​are obtained through bilinear interpolation, which is calculated as follows:

[0064]

[0065] Where q is the coordinate of a point on the low-resolution image in the original high-resolution image, the coordinate value is (x, y), p(q) is the pixel value of the point in the low-resolution image, p(Q 11 )、p(Q 12 )、p(Q 21 )、p(Q 22) is the pixel value of the four corner points mapped by q in the original high-resolution image, whose coordinates are (x1, y1), (x1, y2), (x2, y1), and (x2, y2).

[0066] S12, Block

[0067] The criterion for block division is that the sum of the standard deviations of the pixels within the block is minimized after block division.

[0068] The calculation formula of the standard deviation of pixels within a block is:

[0069]

[0070] Among them, σ i is the pixel standard deviation of the i-th block, N is the total number of pixels in the block, pix j is the pixel value of the jth point, is the average value of all pixels in the block.

[0071] The calculation formula for the sum of the standard deviations of pixels within a block is:

[0072]

[0073] Among them, M represents the number of blocks, and E represents the sum of the standard deviations of pixels in all blocks.

[0074] In addition, in order to avoid information dilution caused by over-refinement, the minimum block size S is set during the block process. min As a constraint, the size of each block is limited to not less than a given threshold S min *S min .

[0075] Assume that the input image matrix size is R*C, and the size of each block is not less than the given threshold S min *S min Under the condition of , the image division position is traversed, and the block boundary position is set to (r, c), where r and c represent the boundary position of the row and column respectively, and r∈[S min ,RS min ],c∈[S min ,CS min ], find the block partitioning scheme that minimizes E.

[0076] Calculate the partitioning scheme that minimizes E and partition the lander image into blocks.

[0077] S13, FDAFT feature point matching

[0078] S131, extracting structural features of the image through ECM (Edge Confidence Maps). w(weighted phase congruency) calculates Mmax (Maximum MomentMap) to obtain the phase characteristics of the image.

[0079] For the extraction of structural features, the Canny operator is first used to detect the input image, so that a grayscale image can be obtained, in which each pixel value represents the probability of an edge at that location. In order to improve the effect of subsequent feature extraction, the EdgeConfidence Map is enhanced by using a Gaussian filter to reduce noise, and finally the random decision forest method is used to extract features. The specific extraction is:

[0080] Let the input be Edge Confidence Map E∈[0,1] W*H , where H and W are the height and width of the image respectively, and E represents the probability of an edge at each pixel value. Select a set of local feature functions that describe edge characteristics Considering the features at different scales, it can be expressed as:

[0081]

[0082] where σ i Represents different Gaussian kernel standard deviations, which are used to capture information at different scales. A large number of local regions P are randomly sampled from the ECM. i , each region contains k neighboring points, and records its corresponding edge labels (that is, the existence of edges), so as to train the decision tree. j i ,y j i Denotes the prediction result at the i-th position of the j-th dimension. The predicted results are weighted averaged and a Gaussian filter is applied to reduce the influence of noise. Finally, the structural label graph is further optimized based on the context information to ensure that it conforms to the actual geometric constraints, thereby extracting the structural features of the image.

[0083] For the extraction of phase features, the calculation process of Mmax is as follows:

[0084]

[0085] Where θ is the angle in each direction.

[0086] S132, according to the calculation results of S131, it is divided into 6 directions, and the structure (low-frequency) feature space and phase (high-frequency) feature space are constructed respectively through Gaussian filtering. The FAST algorithm is used to detect corner points in the structure feature space, and the KAZE algorithm is used to detect corner points in the phase feature space.

[0087] S133, based on the detected corner points, the direction of the local block is divided into 12 parts through GLOH (Gradient Location and Orientation Histogram) to construct a feature vector to encode the feature information more efficiently.

[0088] First, the gradient magnitude and direction are calculated. Based on the detected corner points, a neighborhood window is defined for each point, and the gradient magnitude and direction within the window are calculated.

[0089]

[0090] M(x,y) is the gradient magnitude, θ(x,y) is the gradient direction, and I represents the image intensity function.

[0091] The second step is histogram construction. The neighborhood of the key point is divided into several smaller sub-regions, and each sub-region generates an independent directional histogram. For each pixel in each sub-region, its weight is assigned to the corresponding position histogram group according to its gradient direction. Linear interpolation is usually used to smoothly distribute the gradient information between the two nearest groups. A group of 12 directional histograms is constructed for each sub-region to record the gradient accumulation in different directions. This can capture the directional characteristics of the local structure.

[0092] S134, preliminarily matching the features in the low-frequency feature space and the high-frequency feature space by calculating the SSD (Sum of Squared Differences) between the feature vectors. In addition, in order to improve the reliability of the matching, the FSC (Fast Sample Consensus) algorithm is used to eliminate potential mismatching points.

[0093] First, set the internal point ratio ω and the expected success probability P. Use the formula to calculate the minimum number of iterations N required to ensure that the probability of finding a good model is at least P:

[0094]

[0095] Where s is the minimum number of data points required to estimate the model, and ω is the proportion of internal points. Suppose there is a data set D = {x1, x2, x3...., x M}, randomly select s sample points from D as the initial sample set s = {x i1 ,x i1 ,...x in},in And they are different from each other.

[0096] Assume that the model parameter vector is θ, and use the least squares method to fit the model parameter θ according to the selected sample s points.

[0097]

[0098] Where X is the input matrix and Y is the output vector, is the transpose of X.

[0099] For each data point x i , calculate its distance to the model d(x i ,θ), and judge whether it is an interior point. Usually a threshold ζ is set. i ,θ)<ζ, then it is considered that x i is an interior point. Among them:

[0100] d(x i ,θ)=|y i -Ax i +b|

[0101] x i ,y i is the data point, A and b are the fitting model coefficients, and all qualified inliers are recorded. Calculate the residual sum of squares d(x i ,θ) 2 As the score, if the current model score is better than the previous best model, the best model and its corresponding inlier set are updated. If enough inliers are found in a certain iteration (for example, the inlier ratio reaches a predetermined high threshold), the algorithm can be terminated early.

[0102] S2, pose estimation step: According to the correspondence between the coordinates of the two-dimensional image point and the three-dimensional geographic point coordinates, and combined with the camera internal parameter information, the pose estimation algorithm is used to solve the pose of the lander, and the pose optimization algorithm is used to optimize the initially solved pose to obtain the final pose of the lander.

[0103] S21, pose estimation

[0104] Based on the correspondence between the acquired 2D image points of the lander image and the 3D geographic points, four non-coplanar 3D points are selected as virtual control points, and all other 3D points are represented by these four virtual control points:

[0105]

[0106] Where P i w is a geographic 3D point, is a virtual control point, a ij is the coefficient.

[0107] According to the camera internal parameters and image point coordinates, solve the virtual control point coordinates

[0108]

[0109] Among them, f x ,f y ,u0,v0 are the camera internal parameters, u i 、v i is the image point coordinate, is the coordinate of the virtual control point in the camera coordinate system.

[0110] According to the coordinates of the virtual control point in the camera coordinate system, the three-dimensional coordinates of the geographic three-dimensional point in the camera coordinate system and the corresponding position and posture are solved:

[0111]

[0112] Among them, P i c is the coordinate of the geographic 3D point in the camera coordinate system, R cw is the rotation matrix from the world coordinate system to the camera coordinate system, t is the translation vector from the world geographic coordinate system to the camera coordinate system, is the position of the camera in the world coordinate system.

[0113] S22, pose optimization

[0114] Note that the pose (R cw ,t) is p0, and the optimization is performed with the goal of minimizing the reprojection error, where the reprojection error vector r is expressed as:

[0115]

[0116] Among them, R cw is the rotation matrix from the world coordinate system to the camera coordinate system, t is the translation vector from the world geographic coordinate system to the camera coordinate system, P i c is the coordinate of the geographic 3D point in the camera coordinate system, P i w is a geographic 3D point, and n is the number of geographic 3D points.

[0117] Calculate the first-order partial derivative of the reprojection error vector r with respect to the rotation matrix and translation vector Denoted as the Jacobian matrix, the damping factor λ is dynamically adjusted in each iteration according to the current reprojection error vector and the Jacobian matrix:

[0118]

[0119] Among them, Δp is the increment of the pose p0 after iterative calculation.

[0120] The initial iteration result U0 is The maximum value among the diagonal elements, when λ is greater than the preset threshold, increase U k , otherwise it decreases, and the following equation is solved to obtain Δp to implement iteration:

[0121]

[0122] Where I is the identity matrix, U k is the result of the kth iteration.

[0123] In this embodiment, the reprojection error vector r and the Jacobian matrix J are iteratively calculated. k , damping factor λ and Δp, to determine the result of each iteration U k , and judge whether the optimization is completed by setting a preset number of iterations or threshold. When Δp is less than the threshold or the iteration reaches the preset number of iterations, the posture optimization is considered to be completed and the accurate lander posture is obtained.

[0124] Based on the above method, this embodiment provides the following verification experiment.

[0125] I. Data Introduction

[0126] In order to evaluate the effectiveness of the method proposed in this invention, three representative data sets were selected for testing. In data set 1, there is a significant radiation difference between the lander image and the reference image, which is used to test the robustness of the algorithm under the condition of radiation difference; the main feature of data set 2 is the low image contrast resulting in weak image texture, which aims to test the performance of the algorithm when the image lacks obvious texture information; in data set 3, the radiation difference and texture change between the landing image and the reference image are relatively mild, providing a more balanced test environment. Figure 2 Some image pairs from each dataset are shown. Through comprehensive comparison of these diverse datasets, the performance and applicability of the method under various conditions can be comprehensively and deeply analyzed.

[0127] II. Matching Validity Verification and Lander Positioning Analysis

[0128] Figure 3 The performance of different matching methods on three datasets is shown. Figure 4 The results of these algorithms were systematically evaluated on key performance indicators such as the number of correct matching points, correct matching ratio, matching error, and matching point distribution uniformity. The results show that the improved FDAFT algorithm outperforms other methods under most test conditions.

[0129] The improved FDAFT algorithm not only leads in the number of correct matching points, but also obtains the most correct matching points in all three datasets. In contrast, traditional matching algorithms such as RIFT and image-enhanced SIFT encounter difficulties in dealing with significant radiation differences or weak image surface texture information, and fail to achieve effective matching. This shows that the improved FDAFT algorithm has stronger robustness and adaptability.

[0130] Although in Dataset 3, the improved FDAFT algorithm is slightly inferior to algorithms such as HAPCG in terms of the correct match ratio, which means that the algorithm may introduce some matching points with large errors, it still maintains the lead in terms of the overall number of correct matches.

[0131] In terms of matching error, the improved FDAFT algorithm showed a continuous advantage, maintaining a low root mean square error in the three datasets, especially in datasets 1 and 2, where its matching error was lower than that of the original FDAFT algorithm; and in dataset 3, its performance was on par with or even better than other advanced algorithms. The low matching error directly reflects the high accuracy of the matching and also proves the stability of the algorithm when processing complex images.

[0132] In addition, the improved FDAFT algorithm also performs well in terms of the uniformity of matching point distribution. In datasets 1 and 2, the matching points it generates are more evenly distributed, effectively covering most of the image area, and providing a higher stability guarantee for subsequent pose estimation. Although SuperPoint+SuperGlue has a slight advantage in this regard in dataset 3, the improved FDAFT algorithm still maintains a good performance level.

[0133] In general, the improved FDAFT algorithm can cope with diverse and complex scenarios and achieve a high level of matching accuracy. Compared with the original version, the improved algorithm not only expands the distribution range of matching points, but also improves matching accuracy, further enhancing its practicality and applicability.

[0134] The improved FDAFT matching algorithm and posture estimation algorithm are used to locate the lander. The designed landing point is taken as the origin, the east direction is taken as the positive direction of the X axis, the north direction is taken as the positive direction of the Y axis, and the vertical direction to the reference lunar ellipsoid is taken as the positive direction of the Z axis. A local coordinate system is established, and the landing trajectory is designed to be 2200m to 190m away from the landing point. The restored trajectory is compared with the designed trajectory. Figure 5 shown.

[0135] pass Figure 6 and Figure 7 Analyze the position difference between the design trajectory and the recovery trajectory. Figure 6It can be found that the maximum modulus of the difference between the two is 36.11m, the minimum modulus is 0.58m, the average modulus is 8.23m, and the average positioning result is better than 10m. And as the height of the lander decreases, the positioning error gradually decreases. The final point height is 190m, and the positioning difference with the design point is (2.71m, -0.21m, 0.56m), and the positioning accuracy is better than 3m. Figure 7 From the above, the maximum RMS error in the test case is 1.48 pixels, the minimum RMS error is 0.84 pixels, the average RMS error is 1.1 pixels, and the RMS error is relatively stable. The experiment shows that the proposed method can determine the position of the lander with high accuracy.

[0136] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A lunar landing navigation and positioning method based on an improved FDAFT matching algorithm, characterized in that: The following steps are involved: Feature point matching steps: obtain the lander image and the digital orthophoto and digital elevation model of the area, downsample the lander image and divide it into blocks, perform FDAFT feature point matching on each block with the digital orthophoto, and combine the digital elevation model to obtain the correspondence between the image 2D point coordinates and the geographic 3D point coordinates; Pose estimation steps: Based on the correspondence between the two-dimensional image point coordinates and the three-dimensional geographic point coordinates, and combined with the camera intrinsic parameter information, the pose estimation algorithm is used to solve the lander's pose, and the pose optimization algorithm is used to optimize the preliminary solution to obtain the final pose of the lander.

2. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 1, characterized in that: The downsampling is to sample the high-resolution lander image into a low-resolution lander image, and the calculation method is as follows: Among them, r new With c new represents the row and column number of the low-resolution lander image, r and c represent the row and column number of the original resolution lander image, and n represents the downsampling ratio.

3. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 1, characterized in that: The standard of the block division is that the sum of the standard deviations of the pixels in the block after the block division is minimized.

4. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 3 is characterized in that: The calculation formula of the standard deviation of pixels within the block is: Among them, σ i is the pixel standard deviation of the i-th block, N is the total number of pixels in the block, pix j is the pixel value of the jth point, is the average value of all pixels in the block.

5. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 4, characterized in that: The calculation formula for the sum of the standard deviations of pixels within the block is: Among them, M represents the number of blocks, and E represents the sum of the standard deviations of pixels in all blocks.

6. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 1, characterized in that: In the block process, the minimum block size S is set min As a constraint, the size of each block is limited to not less than a given threshold S min *S min .

7. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 1, characterized in that: The FDAFT feature point matching comprises the following steps: The structural features of the image are extracted through ECM, and the phase features of the image are obtained by calculating the maximum moment mapping through weighted phase consistency; The structural feature space and phase feature space are constructed by Gaussian filtering, and the FAST algorithm is used to detect corner points in the structural feature space, and the KAZE algorithm is used to detect corner points in the phase feature space; Based on the detected corner points, the direction of the local block is divided into 12 parts through GLOH to construct the feature vector; The initial matching of low-frequency and high-frequency feature points is achieved by calculating the SSD between feature vectors, and the FSC algorithm is used to eliminate potential mismatching points.

8. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 1, characterized in that: The pose estimation algorithm performs the following steps: Based on the correspondence between the acquired 2D image points of the lander image and the 3D geographic points, four non-coplanar 3D points are selected as virtual control points, and all other 3D points are represented by these four virtual control points: Where P i w is a geographic 3D point, is a virtual control point, a ij is the coefficient; According to the camera internal parameters and image point coordinates, solve the virtual control point coordinates Among them, f x ,f y ,u0,v0 are the camera internal parameters, u i 、v i is the image point coordinate, is the coordinate of the virtual control point in the camera coordinate system; According to the coordinates of the virtual control point in the camera coordinate system, the three-dimensional coordinates of the geographic three-dimensional point in the camera coordinate system and the corresponding position and posture are solved: Among them, P i c is the coordinate of the geographic 3D point in the camera coordinate system, R cw is the rotation matrix from the world coordinate system to the camera coordinate system, t is the translation vector from the world geographic coordinate system to the camera coordinate system, is the position of the camera in the world coordinate system.

9. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 8, characterized in that: The pose optimization algorithm performs the following steps: Note that the pose (R cw ,t) is p0, and the optimization is performed with the goal of minimizing the reprojection error, where the reprojection error vector r is expressed as: Among them, R cw is the rotation matrix from the world coordinate system to the camera coordinate system, t is the translation vector from the world geographic coordinate system to the camera coordinate system, P i c is the coordinate of the geographic 3D point in the camera coordinate system, P i w is a geographic 3D point, n is the number of geographic 3D points; Calculate the first-order partial derivative of the reprojection error vector r with respect to the rotation matrix and translation vector Denoted as the Jacobian matrix, the damping factor λ is dynamically adjusted in each iteration according to the current reprojection error vector and the Jacobian matrix: Among them, Δp is the increment of the pose p0 after iterative calculation; The initial iteration result U0 is The maximum value among the diagonal elements, when λ is greater than the preset threshold, increase U k , otherwise it decreases, and the following equation is solved to obtain Δp to implement iteration: Among them, I is the identity matrix, U k is the result of the kth iteration.

10. The lunar landing navigation and positioning method based on the improved FDAFT matching algorithm according to claim 9, characterized in that: The posture optimization algorithm determines whether the optimization is completed by setting a preset number of iterations or a threshold. When Δp is less than the threshold or the iteration reaches the preset number of iterations, the posture optimization is considered to be completed.

Images