Incremental error compensation continuous pose estimation method for lunar probe landing process
By using an incremental adjustment method, combined with feature matching and coplanar/collinear equation constraints, a sliding window incremental bundle adjustment model was constructed, which solved the efficiency and accuracy problems of pose estimation during lunar probe landing and achieved efficient and accurate pose parameter recovery.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2022-07-06
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for lunar probe landing rely on pose estimation methods based on descent camera images, which suffer from low computational efficiency and error accumulation. This is especially problematic when dealing with large datasets, making it difficult to meet memory and accuracy requirements.
An incremental adjustment method is adopted, which obtains multiple connection point matching information through feature matching. Combined with sliding window coplanar or collinear equation constraints, Shur complement marginalization and weighted pseudo-observation constraints, an incremental collinear adjustment model is constructed to optimize pose parameters.
It improves the accuracy and computational efficiency of pose estimation, reduces errors, and achieves high-precision pose parameter recovery.
Smart Images
Figure CN115272464B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a pose estimation method, and more particularly to an incremental adjustment continuous pose estimation method for the landing process of a lunar probe. Background Technology
[0002] In recent years, Chang'e 3, 4 and 5 have all successfully completed lunar landing and exploration missions. Continuous attitude information during the landing process is crucial for the safe landing of the lander. The sequence of images acquired by the landing camera carried by the lander can be used to reconstruct the position and attitude parameters of the landing process.
[0003] Current research on pose estimation based on landing camera imagery largely focuses on two approaches: neighbor-transfer solutions and global bundle adjustment. However, as the data size increases, global adjustment methods cannot meet memory and efficiency requirements. In practical engineering applications, we also need to carefully control the scale of the adjustment and computational efficiency. On the other hand, neighbor-transfer solutions suffer from the problem of error accumulation, leading to low estimation accuracy.
[0004] Unlike global bundle adjustment, incremental bundle adjustment first selects key data from a subset of cameras for local adjustment, then gradually adds new camera data while discarding some older data with weak correlation. Specifically, for landing camera images in deep space exploration, the construction of incremental adjustment models requires further research. Currently, there is no research on incremental adjustment for continuous pose estimation during landing, and there is a lack of specific descriptions and improvements to the adjustment models. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an incremental adjustment continuous pose estimation method for the landing process of a lunar probe.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] According to one aspect of the present invention, an incremental adjustment continuous pose estimation method for the landing process of a lunar probe is provided, the method comprising the following steps:
[0008] Step 1) Select the landing camera sequence images from the fast adjustment phase to the hovering phase, obtain multiple connection point matching information through feature matching, and use "dual-model relative orientation - three-image model connection" for initial pose estimation;
[0009] Step 2) Using sliding window coplanar or collinear equation constraints, combined with Shur complement marginalization and weighted pseudo-observation constraints, the pose parameters of the initial pose information are refined by incremental adjustment.
[0010] Step 3) Construct an incremental collinear adjustment model with additional external control points to optimize the results of step 2), thereby reducing pose estimation and positioning errors.
[0011] As a preferred technical solution, step 1) involves obtaining multi-connection point matching information through feature matching and using "dual-model relative orientation - three-image model connection" for initial pose estimation, specifically as follows:
[0012] First, the selected landing camera sequence images are preprocessed, and the mast area on the images is masked.
[0013] Then, based on the SIFT-RANSAC image feature matching strategy, multiple connection point matching is performed to establish a multi-degree overlapping image network.
[0014] The relative positional relationship between two adjacent landing camera images is then calculated using relative orientation.
[0015] Finally, based on the object-side consistency constraint of the triple feature matching connection points, adjacent models are connected.
[0016] As a preferred technical solution, the model connection process will generate error accumulation between model transfers, and the accumulated error will be eliminated by correcting the initial pose parameters.
[0017] As a preferred technical solution, step 2) involves refining the pose parameters of the initial pose information through incremental adjustment using sliding window coplanar or collinear equation constraints, combined with Shur complement marginalization and weighted pseudo-observation constraints.
[0018] 201) Coplanar equation constraints only involve refining the position and attitude parameters. Based on the coplanar equations, the fundamental observation equations for the connection points can be constructed. After linearizing the observation equations, the error equations are obtained:
[0019]
[0020] Where B is the coefficient matrix, V, l represents the residual matrix, the solution parameter matrix, and the constant term vector.
[0021] 202) The error equation shown in equation (1) is transformed into the following form:
[0022]
[0023] The coefficient matrix B is divided into four blocks: M is a fixed adjustment sliding window, N is a marginalization window, and the coplanar equations only consider the refinement of the pose parameters. The upper left corner Λ N Storing edge-mapped camera extrinsic parameters, bottom right corner Λ M-NThe camera extrinsic parameters that are not marginalized are stored. The dimension of each diagonal block is the same as the dimension of the camera pose parameters, and both are diagonal block matrices. The Λ matrix stores the intersection between the camera extrinsic parameters and the 3D points. These represent the edged camera extrinsic parameters and the non-edged camera extrinsic parameters, respectively; N l M-N These are the constant term vectors related to the marginalization parameter and the constant term vectors related to the non-margined parameter, respectively.
[0024] By using Gaussian elimination, the above equation can be transformed into:
[0025]
[0026] The extrinsic parameters of the image within the unmargined window can be obtained using the above formula. and image extrinsic parameters within the edge window They are respectively:
[0027]
[0028]
[0029] Among them, Λ M-N =B M-N T P M-N B M-N B M-N P M-N These are the coefficient matrix and weight matrix, respectively, related to the unmargined parameters;
[0030] 203) The camera-related parameter variables to be removed from the edge window, namely the camera pose and the object coordinates observed only by it, are constructed with the error equation as shown in Equation (6):
[0031]
[0032] in For the marginalized camera pose and the coefficient matrix related to the object coordinates observed by it, The coefficient matrix relating the unmargined camera pose and the object coordinates observed only by the camera. To solve for the edge-scaled camera pose and the object coordinates observed by it, To solve for the pose of the unmargined camera and the object-space coordinates observed by it; These are constant vectors related to the marginalized and non-margined parameters, respectively; Gaussian elimination is performed on the linear equations, with the goal of eliminating the off-diagonal portion Λ in the lower left corner. T ,have to
[0033]
[0034] After elimination, the second row of the system of equations becomes sum. Irrelevant terms; taken out separately, we obtain the incremental equation for the unmargined portion:
[0035]
[0036]
[0037] After solving this equation, take the solution... Substitute into the original equation and solve.
[0038]
[0039] As a preferred technical solution, unlike the coplanar equation which only optimizes camera extrinsic parameters, the collinear equation needs to optimize both camera extrinsic parameters and object point coordinates.
[0040] Based on the same error equation (1), with the fixed adjustment sliding window as M and the marginalization window as N, the construction form of the coefficient matrix B is different from that of equation (2).
[0041] Assume there are m camera extrinsic parameters and n object points in the window. Since the number of object points is usually much greater than the number of camera extrinsic parameters, n >> m. Store the coordinates of all object points in the top left corner, and arrange them in order according to the object points observed by each camera position. Store the camera extrinsic parameters in the bottom right corner, and arrange them in the same order according to the image sequence.
[0042] As a preferred technical solution, this method requires continuous iteration to calculate the coefficient matrix B and residual v during the marginalization process. During the iteration process, the state variables are continuously updated, and attention should be paid to fixing the linearization point when calculating the coefficient matrix related to marginalization.
[0043] As a preferred technical solution, this method incorporates weighted pseudo-observation equations to improve the adjustment model. Equations are established for highly correlated parameters, and different weights are assigned to them to reduce their correlation. The weights of each parameter are determined as follows: P X =P Z =10 -10 ,
[0044] Where P X P Z P φ P ω P κ These are weight matrices for the line elements X and Z, and the three corner elements, respectively.
[0045] As a preferred technical solution, step 3) of constructing an incremental collinear adjustment model with additional external control points to optimize the results of step 2) specifically involves:
[0046] Based on the HOPC matching algorithm, heterogeneous matching of LRO NAC orbiter orthophotos and lander camera sequence images is performed to achieve automatic acquisition of control points.
[0047] By introducing the concepts of sliding window and marginalization, within a fixed sliding window, we test "marginalizing the image frames to be removed and their observed object coordinates", and introduce new observation frames and object points.
[0048] As a preferred technical solution, regarding the method of control points participating in adjustment, the window parameter W for adding control points is set, and a local adjustment is performed using multiple connection points and control points within the window.
[0049] As a preferred technical solution, the window parameter W = 30, which means that for every 30 additional images...
[0050] Compared with the prior art, the present invention has the following advantages:
[0051] I. The pose estimation method for sliding window incremental bundle adjustment of landing camera sequence images proposed in this invention effectively improves the computational efficiency of pose adjustment refinement by adopting the strategy of Shur complement marginalization, and also improves the accuracy compared with the adjacent transfer solution method.
[0052] II. The pose estimation method proposed in this invention, which uses a sliding window incremental bundle adjustment for landing camera sequence images, effectively utilizes coplanar and collinear equation constraints by combining redundant observation information from multiple connection points, weighted pseudo-observation constraints, and additional external control points when constructing the adjustment model, thus improving the stability and reliability of pose estimation. Attached Figure Description
[0053] Figure 1 This is a flowchart illustrating the technical process of the present invention.
[0054] Figure 2 This is a schematic diagram illustrating the construction of the coefficient matrix of the incremental collinear adjustment model of this invention;
[0055] Figure 3 This is a schematic diagram of the Schul complement marginalization of the incremental collinear adjustment model of the present invention. Detailed Implementation
[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0057] This invention provides a pose estimation method for sliding window incremental bundle adjustment of landing camera sequence images, comprising the following steps:
[0058] 1) Select the landing camera sequence images from the fast adjustment phase to the hovering phase, obtain multiple connection point matching information through feature matching, and use "dual-model relative orientation-three-image model connection" for initial pose estimation;
[0059] 2) Using sliding window coplanar / collinear equation constraints, combined with Shur complement marginalization and weighted pseudo-observation constraints, the pose parameters of the initial pose information are refined through incremental adjustment;
[0060] 3) Construct an incremental collinear adjustment model with additional external control points to further reduce pose estimation and positioning errors.
[0061] In step 1), multiple connection point matching information is obtained through feature matching, and initial pose estimation is performed using a "dual-model relative orientation - three-image model connection" approach. Specifically:
[0062] First, the selected landing camera sequence images are preprocessed, and the mast region on the images is masked. Then, a multi-connection point matching strategy based on SIFT-RANSAC is used to establish a multi-degree overlapping image network. Next, the relative positional relationship between two adjacent landing camera images is calculated through relative orientation. Finally, based on the object-space consistency constraint of the triple feature matching connection points, adjacent models are connected. During the model connection process, errors accumulate between model transfers, and the initial pose parameters need to be corrected subsequently to eliminate these accumulated errors.
[0063] In step 2), an incremental coplanar / collinear adjustment model integrating Shur complement marginalization and weighted pseudo-observation constraints is established to refine the pose parameters, specifically as follows:
[0064] Two incremental bundle adjustment models based on sliding window coplanar / collinear constraints are established. By fusing Schur complement marginalization and weighted pseudo-observation constraints, the pose parameters of the initial landing trajectory are refined.
[0065] Coplanar equation constraints only involve refining the position and attitude parameters. Based on the coplanar equations, the fundamental observation equations for the connection points can be constructed. Linearizing the observation equations yields the error equations:
[0066]
[0067] The coefficient matrix B exhibits sparsity, so it is divided into four blocks. The adjustment sliding window is fixed at M, and the marginalization window is N. The coplanar equations only consider the refinement of the pose parameters. The upper left corner Λ... N Storing edge-mapped camera extrinsic parameters, bottom right corner Λ M-NThe camera extrinsic parameters that are not marginalized are stored. The dimension of each diagonal block is the same as the dimension of the camera pose parameters, and both are diagonal block matrices. The structure of the off-diagonal blocks is related to the specific observation data. Therefore, the error equation shown in equation (1) can be transformed into the following form:
[0068]
[0069] By using Gaussian elimination, the above equation can be transformed into:
[0070]
[0071] The extrinsic parameters of the image within the unmargined window can be obtained using the above formula. and image extrinsic parameters within the edge window They are respectively:
[0072]
[0073]
[0074] Among them, Λ M-N =B M-N T P M-N B M-N .
[0075] Unlike the coplanar equation, which only optimizes camera extrinsic parameters, the collinear equation requires optimizing both camera extrinsic parameters and object point coordinates. Also based on the error equation (1), with a fixed adjustment sliding window of M and a marginalization window of N, the construction of the coefficient matrix B differs from equation (2). Assume there are m camera extrinsic parameters and n object points within the window. Since the number of object points is usually much greater than the number of camera extrinsic parameters, n >> m. The upper left corner stores all object point coordinates, arranged sequentially according to the object points observed at each camera position. The lower right corner stores camera extrinsic parameters, also arranged in image order. Both matrices are diagonal block matrices; the results for the off-diagonal blocks are related to the specific observation data. Considering that the number of object points is much greater than the number of camera variables, and that inverting a diagonal block matrix is much easier than inverting a general matrix, only the diagonal matrices need to be inverted separately.
[0076] The specific implementation of the Schul complement marginalization idea in the error equation solution process is as follows: the camera-related parameter variables to be removed from the marginalization window, namely the camera pose and the object coordinates of the object points observed only by it (hereinafter referred to as the object coordinates of the connected points within the marginalization window), are constructed as shown in equation (6):
[0077]
[0078] Gaussian elimination is performed on the system of linear equations, with the goal of eliminating the off-diagonal part Λ in the lower left corner.T ,have to
[0079]
[0080] After elimination, the second row of the system of equations becomes sum. Irrelevant terms. Taken out separately, we obtain the incremental equation for the unmargined portion:
[0081]
[0082]
[0083] After solving this equation, take the solution... Substitute into the original equation and solve.
[0084]
[0085] It's important to note that during the marginalization process, the B matrix and residual v need to be calculated iteratively. During this iteration, the state variables are continuously updated. When calculating the coefficient matrix related to marginalization, it's crucial to fix the linearization point; that is, the value of the derivative variable must be fixed when calculating the coefficient matrix, rather than using the value updated in each iteration. Calculate the coefficient matrix; otherwise, the conditionally independent variables in the marginalization variable will become correlated due to the fixed marginalization variable.
[0086] Considering the significant correlation among a set of exterior orientation parameters, the regional network has high degrees of freedom, making the normal equations prone to ill-conditioned conditions during adjustment, leading to unstable solutions. Therefore, a weighted pseudo-observation equation is incorporated to improve the adjustment model. This method establishes equations for highly correlated parameters and assigns different weights to them to reduce their correlation. The weights of each parameter are determined as follows: P X =P Z =10 -10 ,
[0087] In step 3), the incremental collinear adjustment model with additional external control points is constructed to further improve the pose estimation and positioning accuracy, specifically as follows:
[0088] Building upon the aforementioned incremental collinear adjustment, this paper considers the effectiveness of introducing external control points in reducing positioning errors. Heterogeneous matching of LRO NAC orthophotos and lander camera sequence images is performed using the HOPC matching algorithm to automatically acquire control points. A sliding window and marginalization approach are introduced. Within a fixed sliding window, the process of "marginalizing out image frames to be removed and their observed object-space coordinates" is tested, and new observation frames and object-space points are introduced. Regarding the method of control point participation in adjustment, a window parameter W is set for control point inclusion. For example, if W = 30, a local adjustment is performed every 30 images using multiple connection points and control points within the window. Specific Implementation
[0090] like Figure 1 As shown, this invention provides a pose estimation method for sliding window incremental bundle adjustment of landing camera sequence images. This method mainly consists of the following three steps:
[0091] (1) Initial pose estimation is performed using a “dual-model relative orientation-three-image model connection”:
[0092] First, the selected landing camera sequence images are preprocessed, and the mast region on the images is masked. Then, a multi-connection point matching strategy based on SIFT-RANSAC is used to establish a multi-degree overlapping image network. Next, the relative positional relationship between two adjacent landing camera images is calculated through relative orientation. Finally, based on the object-space consistency constraint of the triple feature matching connection points, adjacent models are connected. During the model connection process, errors accumulate between model transfers, and the initial pose parameters need to be corrected subsequently to eliminate these accumulated errors.
[0093] (2) Establish an incremental coplanar / collinear adjustment model that integrates Shur complement marginalization and weighted pseudo-observation constraints:
[0094] Two incremental bundle adjustment models based on sliding window coplanar / collinear constraints are established. By fusing Schur complement marginalization and weighted pseudo-observation constraints, the pose parameters of the initial landing trajectory are refined.
[0095] The coplanar equation constraint only involves the refinement of position and attitude parameters. Based on the coplanar equation, the basic observation equation for the connection points can be constructed. The coefficient matrix B is sparse, so it is divided into four blocks. The adjustment sliding window is fixed at M, and the marginalization window is N. The coplanar equation only considers the refinement of pose.
[0096] Unlike coplanar equations, which only optimize camera extrinsic parameters, collinear equations require simultaneous optimization of both camera extrinsic parameters and object point coordinates. For example... Figure 2As shown, the top left corner of the coefficient matrix stores the coordinates of all object points, arranged sequentially according to the object points observed at each camera position. The bottom right corner stores the camera extrinsic parameters, also arranged in image order. Considering that the number of object points is much larger than the number of camera variables, and that inverting a diagonal block matrix is much easier than inverting a general matrix, only the diagonal matrices need to be inverted individually. The specific implementation of the Schur complement marginalization idea in solving the collinearity error equation is as follows: the camera-related parameter variables to be removed are marginalized in the window; the corresponding graphical model is shown in [link to graphical model]. Figure 3 (b) It is important to note that during the marginalization process, the B matrix and residual v need to be calculated iteratively. During this iteration, the state variables are continuously updated. When calculating the coefficient matrix related to marginalization, it is crucial to fix the linearization point; that is, the value of the derivative variable must be fixed when calculating the coefficient matrix, rather than using the value updated in each iteration. Calculate the coefficient matrix; otherwise, the conditionally independent variables in the marginalization variable will become correlated due to the fixed marginalization variable.
[0097] Considering the significant correlation among a set of exterior orientation parameters, the regional network has high degrees of freedom, making the normal equations prone to ill-conditioned conditions during adjustment, leading to unstable solutions. Therefore, a pseudo-observation equation is incorporated to improve the adjustment model. This method establishes equations for highly correlated parameters and assigns different weights to them to reduce their correlation. The weights of each parameter are determined as follows: P X =P Z =10 -10 ,
[0098] (3) Construct an incremental collinear adjustment model with additional external control points:
[0099] Building upon the aforementioned incremental collinear adjustment, this paper considers the effectiveness of introducing external control points in reducing positioning errors. First, the HOPC matching algorithm is used to perform heterogeneous matching between orbiter orthophotos and landing camera sequence images, enabling automatic acquisition of control points. Then, a sliding window and marginalization approach are introduced. Within a fixed sliding window, the process of "marginalizing out image frames to be removed and their observed object-space coordinates" is tested, and new observation frames and object-space points are introduced. Regarding the method of control point participation in adjustment, a window parameter W is set for the control points. For example, if W = 30, a local adjustment is performed every 30 images using multiple connection points and control points within the window.
[0100] Example:
[0101] 1. Experimental Data
[0102] The descent camera images from the Chang'e-3 lander were used as experimental data. During the Chang'e-3 landing process, a total of 4,666 images were received and downloaded frame by frame. In the experiment, one image was selected every ten images, mainly from the 126 descent sequence images from the fast adjustment phase to the approach phase.
[0103] 2. Experimental Results
[0104] The initial pose parameters of 126 landing images were recovered based on the "dual-image relative orientation-three-image model connection". Using the image exterior orientation parameters obtained from the initial pose estimation, the image coordinates of all object-side points were calculated through backprojection from lunar space to image space. The difference between the backprojected image points and the actual image points was used to calculate the 2D image coordinate difference. Table 1 shows the numerical values of the image-side backprojection residuals at the connection points.
[0105] Table 1
[0106]
[0107] Table 2 lists the differences in 3D coordinates between control points and check points calculated using the adjacent transfer solution method. It can be seen that the large positioning error in the XYZ direction reflects the accumulation of errors and inaccuracies in pose recovery under weak intersection conditions.
[0108] Table 2
[0109] RMSE X_differ Y_differ Z_differ Control Points 4.4580 7.1743 6.5758 Checkpoint 3.3246 4.8250 6.8312
[0110] The results of the coplanar incremental adjustment are shown in Table 3-4. The image-side backprojection residuals and 3D positioning errors were tested under different marginalization windows with a sliding window M=10. Considering that the coplanar constraint is only established between pairwise image points and there are no constraints on multiple connection points, the image-side backprojection residuals of connection points do not change significantly under different sliding windows and different marginalization windows, remaining within a stable range. The 3D positioning error does not show a clear pattern under different sliding windows and marginalization windows, but the positioning error is significantly reduced compared to the adjacent transfer solution method. Because the coplanar equation only involves pose refinement and is not affected by inaccurate object-side coordinates, the incremental bundle adjustment model based on sliding window coplanar constraints effectively improves positioning accuracy without sacrificing internal consistency, demonstrating significant advantages in both stability and time efficiency.
[0111] Table 3
[0112]
[0113] Table 4
[0114]
[0115] Table 5
[0116]
[0117] Table 6-8 shows the image-side backprojection residuals and 3D positioning error results of the incremental collinear adjustment model for "object-space coordinates of connecting points within the marginalized window". Compared with the initial pose estimation and coplanar incremental adjustment model, the connecting point backprojection residuals of the incremental collinear adjustment are significantly smaller. Regarding 3D positioning error, the accuracy of the incremental collinear adjustment model varies greatly under different window combinations; when the marginalized window is greater than 4, the positioning accuracy is significantly better than the adjacent transfer solution method. Furthermore, the positioning accuracy exhibits a pattern where "the larger the marginalized window N, the higher the accuracy and the less time is consumed".
[0118] Table 6
[0119]
[0120] Table 7
[0121]
[0122] Table 8
[0123]
[0124] Compared to the incremental coplanar adjustment model, the collinear incremental adjustment model, while sacrificing some time efficiency, has the potential to further reduce 3D positioning errors. Table 9-10 lists the 3D positioning errors of the incremental collinear adjustment after adding external control points. Substituting the additional control points into the incremental collinear adjustment significantly improves positioning accuracy.
[0125] Table 9
[0126]
[0127] Table 10
[0128]
[0129] In summary, through a series of qualitative and quantitative experimental analyses using image data from the Chang'e-3 lunar probe's landing camera sequence, the proposed pose estimation method using a sliding window incremental bundle adjustment of landing camera sequence images has been verified. This method effectively achieves high-precision and rapid calculation of continuous position and attitude parameters during the lunar probe's landing process, improving accuracy by 30% compared to the commonly used adjacent transfer solution method. The incremental collinear adjustment model with additional external control point constraints can further improve pose estimation and positioning accuracy.
[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for incremental adjustment of continuous pose estimation during the landing process of a lunar probe, characterized in that, The method includes the following steps: Step 1) Select the landing camera sequence images from the fast adjustment phase to the hovering phase, obtain multiple connection point matching information through feature matching, and use "dual-model relative orientation-three-image model connection" for initial pose estimation; Step 2) Using sliding window coplanar or collinear equation constraints, combined with Shur complement marginalization and weighted pseudo-observation constraints, the pose parameters of the initial pose information are refined by incremental adjustment. Step 3) Construct an incremental collinear adjustment model with additional external control points to optimize the results of Step 2), thereby reducing pose estimation and positioning errors; Step 1) involves obtaining multi-connection point matching information through feature matching and using "dual-model relative orientation - three-image model connection" for initial pose estimation, specifically as follows: First, the selected landing camera sequence images are preprocessed, and the mast area on the images is masked. Then, based on the SIFT-RANSAC image feature matching strategy, multiple connection point matching is performed to establish a multi-degree overlapping image network. The relative positional relationship between two adjacent landing camera images is then calculated using relative orientation. Finally, based on the object-space consistency constraint of the triple feature matching connection points, adjacent models are connected. Step 3) optimizes the results of step 2) by constructing an incremental collinear adjustment model with additional external control points. Specifically, this involves: Based on the HOPC matching algorithm, heterogeneous matching of LRO NAC orbiter orthophotos and lander camera sequence images is performed to achieve automatic acquisition of control points. By introducing the concepts of sliding window and marginalization, within a fixed sliding window, we test "marginalizing the image frames to be removed and their observed object coordinates", and introduce new observation frames and object points.
2. The incremental adjustment continuous pose estimation method for the landing process of a lunar probe according to claim 1, characterized in that, During the model connection process, errors accumulate between model transfers. These accumulated errors are then eliminated by correcting the initial pose parameters.
3. The incremental adjustment continuous pose estimation method for the landing process of a lunar probe according to claim 1, characterized in that, Step 2) involves refining the pose parameters of the initial pose information through incremental adjustment using sliding window coplanar or collinear equation constraints, combined with Shur complement marginalization and weighted pseudo-observation constraints. 201) Coplanar equation constraints only involve refining the position and attitude parameters. Based on the coplanar equations, the fundamental observation equations for the connection points can be constructed. After linearizing the observation equations, the error equations are obtained: in B The coefficient matrix, , , The residual matrix, the solution parameter matrix, and the constant term vector are given. 202) The error equation shown in equation (1) is transformed into the following form: The coefficient matrix B Divided into four parts, M To fix the adjustment sliding window, N For edge-optimized windows, the coplanar equations only consider the refinement of pose parameters, with the top left corner... Storing edge-mapped camera extrinsic parameters, bottom right corner Store the camera extrinsic parameters that are not marginalized. The dimension of each diagonal block is the same as the dimension of the camera pose parameters, and both are diagonal block matrices. The matrix stores the intersections between 3D points that participate in the camera's external operations. , These are the extrinsic parameters of the edged-out camera and the extrinsic parameters of the non-edged camera, respectively. , These are the constant term vectors related to the marginalization parameter and the constant term vectors related to the non-margined parameter, respectively. By using Gaussian elimination, the above equation can be transformed into: The extrinsic parameters of the image within the unmargined window can be obtained using the above formula. and image extrinsic parameters within the edge window They are respectively: in, , , These are the coefficient matrix and weight matrix, respectively, related to the unmargined parameters; 203) The camera-related parameter variables to be removed from the edge window, namely the camera pose and the object coordinates observed only by it, are constructed with the error equation as shown in Equation (6): in For the marginalized camera pose and the coefficient matrix related to the object coordinates observed by it, The coefficient matrix relating the unmargined camera pose and the object coordinates observed only by the camera. To solve for the edge-scaled camera pose and the object coordinates observed by it, To solve for the pose of the unmargined camera and the object-space coordinates observed by it; , These are constant vectors related to the marginalization and non-margination parameters, respectively; Gaussian elimination is performed on the system of linear equations, with the goal of eliminating the off-diagonal part in the lower left corner. ,have to After elimination, the second row of the system of equations becomes sum. Irrelevant terms; taken out separately, we obtain the incremental equation for the unmargined portion: After solving this equation, take the solution... Substitute into the original equation and solve. ; 。 4. The incremental adjustment continuous pose estimation method for the landing process of a lunar probe according to claim 3, characterized in that, During the marginalization process, this method requires iterative calculation of the coefficient matrix. B and residual v During the iteration process, the state variables are constantly updated, and when calculating the coefficient matrix related to marginalization, it is necessary to pay attention to fixing the linearization point.
5. The incremental adjustment continuous pose estimation method for the landing process of a lunar probe according to claim 3, characterized in that, This method incorporates weighted pseudo-observation equations to improve the adjustment model. Equations are established for highly correlated parameters, and different weights are assigned to them to reduce their correlation. The weights of each parameter are determined as follows: ； in , , , , These are weight matrices for the line elements X and Z, and the three corner elements, respectively.
6. The incremental adjustment continuous pose estimation method for the landing process of a lunar probe according to claim 1, characterized in that, Regarding the method of control points participating in adjustment, set the window parameter W for adding control points, and use the multiple connection points and control points within the window to perform a local adjustment.
7. The incremental adjustment continuous pose estimation method for the landing process of a lunar probe according to claim 6, characterized in that, The window parameter W=30 means that for every 30 additional images...