Satellite three-dimensional reconstruction method in complex illumination environment

Abstract

The invention relates to the technical field of three-dimensional modeling, and provides a satellite three-dimensional reconstruction method in a complex illumination environment, which comprises the following steps: acquiring images under different illumination conditions, and extracting image features by combining multi-view collaborative observation; performing feature point and multi-view data registration, generating a three-dimensional model and analyzing illumination change; performing system initialization by using the initially matched feature points, and establishing an initial three-dimensional point cloud model; sparse point cloud is recovered through multi-view stereoscopic vision, and motion recovery is carried out; a 2D diffusion model is used to complement the shielded area, and laser radar data is combined to carry out densification processing on the sparse point cloud; a three-dimensional grid model is generated, and texture mapping is carried out by a multi-view image; light compensation is adopted, a polarized light separation technology is used, and ambient light and specular reflection light are separated; according to the method, through multi-view-angle collaborative observation, the shielding problem under a single view angle is reduced, the 2D diffusion model is used for complementing the shielded area, and the completeness and precision of three-dimensional reconstruction are ensured.

Description

Technical Field

[0001] The invention relates to the technical field of spacecraft navigation and three-dimensional modeling, and in particular to a three-dimensional reconstruction method of a satellite under a complex lighting environment. Background Art

[0002] In modern satellite remote sensing technology, 3D reconstruction methods are increasingly used, especially in scene reconstruction under complex lighting environments. Traditional 3D reconstruction techniques usually rely on good lighting conditions to ensure the acquisition of high-quality image data. However, in practical applications, satellites often face problems such as image quality degradation and difficulty in feature extraction when imaging under different lighting conditions (such as shadows, reflections, scattering, etc.). These problems seriously affect the accuracy and reliability of 3D reconstruction;

[0003] Traditional methods rely on single spectrum or visible light images for feature extraction, but these features are often unstable under complex lighting conditions (such as shadows, strong light, reflections, etc.), resulting in reduced accuracy and robustness of feature extraction; and in multi-view images, the target object may be blocked by other objects, resulting in the inability to obtain effective image information in some areas; at the same time, the sparse point cloud generated by the traditional method is difficult to perform high-precision densification processing under complex lighting conditions, resulting in low accuracy of the final 3D model; and when performing texture mapping, the traditional method is difficult to handle the changes in lighting under different viewing angles, resulting in obvious lighting inconsistency in the texture mapping results, affecting the realism of the model;

[0004] To this end, technicians in this field have proposed a satellite 3D reconstruction method in a complex lighting environment, aiming to improve the robustness and accuracy of feature extraction, ensure the integrity and accuracy of 3D reconstruction, and ensure the lighting consistency and realism of texture mapping. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a satellite three-dimensional reconstruction method in a complex lighting environment to solve the problems raised in the background technology.

[0006] A satellite 3D reconstruction method in a complex lighting environment specifically comprises the following steps:

[0007] S1. Feature extraction: Use multispectral imaging to obtain images under different lighting conditions, combine multi-view collaborative observation, and extract SURF features in the image;

[0008] S2, feature matching: the extracted SURF features are registered between feature points of different images. The multi-view data is registered and fused to generate a three-dimensional model, and combined with the illumination stereo technology, the illumination changes are analyzed;

[0009] S3, system initialization: use the initial matched feature points to initialize the system, build a preliminary 3D point cloud model, and use the BA algorithm to optimize the initial model;

[0010] S4, motion recovery of sparse point cloud: restore sparse point cloud through multi-view stereo vision and use SLAM for motion recovery;

[0011] S5, dense point cloud reconstruction: use the 2D diffusion model to complete the occluded area, and combine the lidar data to densify the sparse point cloud;

[0012] S6, Network modeling and texture mapping: Generate a 3D mesh model and perform texture mapping through multi-view images;

[0013] S7. Lighting processing: Lighting compensation is used to reduce the impact of complex light environment on reconstruction results, and polarized light separation technology is used to separate ambient light and specular reflection light.

[0014] Preferably, in the feature extraction, a multispectral camera is used to capture spectral information of multiple bands, and the image of each band is used to provide different lighting information. Multiple shots are taken under different solar altitude angles, cloud cover and atmospheric conditions, and the collected multispectral images are preprocessed, including radiation correction, atmospheric correction and geometric correction; at the same time, multiple satellites or drones are used to simultaneously shoot the target area from different angles to perform multi-angle collaborative observation.

[0015] Preferably, in the feature extraction, SURF feature extraction is performed on the preprocessed multispectral image, and a set of feature points is generated by detecting key points in the image and calculating the descriptors of the surrounding areas; specifically, the feature point detection process is accelerated by first generating an integral image, and the integral image expression is:

[0016]

[0017] Where I(i,j) represents the pixel value of the original image at position (i,j);

[0018] And at each pixel (x, y) of the image, the Hessian matrix is ​​calculated, expressed as:

[0019]

[0020] Among them, L xx , L xy and L yy They represent the second-order Gaussian derivatives of the image at scale σ respectively;

[0021] By calculating the determinant of the Hessian matrix, the feature points are detected, which is expressed as:

[0022] Det(H)=L xx L yy -(L xy )2

[0023] When the determinant value is greater than the preset threshold, the pixel is considered as a potential feature point;

[0024] Then describe the detected feature points, generate feature vectors, and use Haar wavelet response to calculate the descriptor in the area around the feature points. The descriptor calculation expression is:

[0025]

[0026] Where dx and dy represent the Haar wavelet responses in the horizontal and vertical directions in the area around the feature point, respectively.

[0027] Preferably, in the feature matching, the FLANN algorithm is used to match feature points between images of different perspectives, and then the matched feature points are registered, and the RANSAC algorithm is used to remove mismatched points. Through multi-perspective collaborative observation, image data from different perspectives are obtained, and registration and fusion are performed to generate a preliminary three-dimensional point cloud model; then combined with photometric stereo technology, multiple light sources are used to illuminate the object from different angles to capture the corresponding images, and the three-dimensional shape of the object surface is restored by analyzing the brightness changes of each pixel under different lighting conditions.

[0028] Preferably, in the system initialization, the matched feature points are used for system initialization, and a preliminary three-dimensional point cloud model is generated by a triangulation method using the matched feature points, which is expressed as:

[0029]

[0030] Among them, X represents a three-dimensional point, P1 -1 and They are the projection matrices under two viewing angles, and x1 and x2 are their corresponding pixel points;

[0031] After the preliminary 3D point cloud model is established, the BA algorithm is used for optimization. The camera parameters and 3D point positions are optimized by minimizing the reprojection error. The reprojection error formula is:

[0032]

[0033] Among them, x ij represents the observation value of the j-th 3D point under the i-th camera, P i ,X j Represents the projection function, which transforms the three-dimensional point X jProject it onto the image plane of the i-th camera; and use the LM algorithm for nonlinear minimization to iteratively optimize the camera parameters and the position of the 3D points.

[0034] Preferably, in the motion recovery sparse point cloud, SLAM technology is used to estimate the camera pose through feature point matching and triangulation results, calculate the intrinsic matrix and basic matrix between the cameras, and describe the rotation and translation relationship of the cameras. The basic matrix is ​​expressed as:

[0035]

[0036] Among them, K1 and K2 represent the intrinsic parameter matrix of the camera, and E represents the intrinsic matrix;

[0037] Then use the epipolar constraint to match the feature points so that the matching points meet the geometric constraints, which can be expressed as:

[0038]

[0039] Among them, x1 and x2 represent the matched image points; the BA algorithm is then used to optimize the sparse point cloud to minimize the reprojection error.

[0040] Preferably, in the dense point cloud reconstruction, a pre-trained 2D diffusion model is used to complete the occluded parts in the multi-view image. The diffusion model generates high-quality image completion results by gradually adding noise and learning the denoising process. A mask is generated through a deep learning model or manual annotation to detect the occluded area in the image, and the Mask of the occluded area is input into the 2D diffusion model to generate a completed image.

[0041] Preferably, in the reconstruction of the dense point cloud, densification processing is also performed in combination with lidar data. The point cloud data of the target area is obtained by using a lidar device. The lidar measures the distance between the target and the radar by emitting laser pulses and receiving their reflected signals, thereby generating high-precision point cloud data. The lidar point cloud data is preprocessed, including removing noise and filtering out invalid data. The sparse point cloud generated by the completed 2D image is fused with the lidar point cloud data. Through a registration algorithm, the point cloud data from different perspectives are registered to the same coordinate system to generate a dense point cloud.

[0042] Preferably, in the network modeling and texture mapping step, starting from dense point cloud data, the point cloud is preprocessed, including denoising and filtering; then the PSR algorithm is used to generate a three-dimensional mesh model, and a smooth three-dimensional surface is generated from the point cloud data by solving the Poisson equation, which is expressed as:

[0043] ▽·▽φ=▽·V

[0044] Among them, φ represents the implicit function to be solved, and V represents the vector field generated by the point cloud normal vector; and the generated mesh is simplified and refined to reduce the number of polygons;

[0045] Texture mapping is to obtain high-resolution images of the target area from different perspectives, covering the entire 3D model, aligning the multi-perspective images with the 3D mesh model to align the geometry of the image and the model, and then mapping the texture information of the multi-perspective images to the 3D mesh model, generating UV coordinates for each mesh vertex, which represent the position in the image; using weighted averaging, texture fusion is performed between multiple perspective images to eliminate gaps and overlapping areas.

[0046] Preferably, in the illumination processing step, an illumination model is established, illumination changes under different illumination conditions are analyzed, and histogram equalization is used to process the image to reduce the impact of illumination changes on image feature extraction and matching, and then multi-view image fusion is used to compensate using illumination information under different perspectives; at the same time, a polarization camera is used to capture a polarization image of the target area, light information under different polarization angles is obtained, and the polarization characteristics of polarized light and natural light are analyzed using Stokes vectors and Mueller matrices; and mirror reflections are filtered and analyzed based on Stokes vectors and Mueller matrices to calculate different polarization azimuths and separate ambient light and mirror reflection light.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention uses multispectral imaging technology to acquire images under different lighting conditions, and combines a deep learning model to extract key feature points in the image, thereby improving the robustness and accuracy of feature extraction with more spectral information.

[0049] 2. The present invention reduces the occlusion problem under a single perspective through multi-perspective collaborative observation, and uses a 2D diffusion model to complete the occluded area to ensure the integrity and accuracy of three-dimensional reconstruction.

[0050] 3. The present invention densifies sparse point clouds using lidar data to further improve the density and accuracy of point clouds and generate high-quality three-dimensional models. At the same time, polarized light separation technology is used to separate ambient light and specular reflection light to ensure the lighting consistency and realism of texture mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The figure is a flow chart of the satellite 3D reconstruction method under complex lighting environment of the present invention. DETAILED DESCRIPTION

[0052] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0053] As attached Figure 1 As shown:

[0054] Embodiment 1: The present invention provides a satellite 3D reconstruction method in a complex lighting environment, which specifically includes the following steps:

[0055] S1. Feature extraction: Use multispectral imaging to obtain images under different lighting conditions, combine multi-view collaborative observation, and extract SURF features in the image;

[0056] S2, feature matching: the extracted SURF features are registered between feature points of different images. The multi-view data is registered and fused to generate a three-dimensional model, and combined with the illumination stereo technology, the illumination changes are analyzed;

[0057] S3, system initialization: use the initial matched feature points to initialize the system, build a preliminary 3D point cloud model, and use the BA algorithm to optimize the initial model;

[0058] S4, motion recovery of sparse point cloud: restore sparse point cloud through multi-view stereo vision and use SLAM for motion recovery;

[0059] S5, dense point cloud reconstruction: use the 2D diffusion model to complete the occluded area, and combine the lidar data to densify the sparse point cloud;

[0060] S6, Network modeling and texture mapping: Generate a 3D mesh model and perform texture mapping through multi-view images;

[0061] S7. Lighting processing: Lighting compensation is used to reduce the impact of complex light environment on reconstruction results, and polarized light separation technology is used to separate ambient light and specular reflection light.

[0062] In feature extraction, a multispectral camera is used to capture spectral information of multiple bands. The image of each band is used to provide different lighting information. Multiple shots are taken under different solar altitude angles, cloud cover and atmospheric conditions to ensure that the acquired images cover various lighting changes. The collected multispectral images are preprocessed, including radiation correction, atmospheric correction and geometric correction, to eliminate errors caused by lighting changes and atmospheric effects. At the same time, multiple satellites or drones are used to simultaneously shoot the target area from different angles to conduct multi-angle collaborative observations to reduce occlusion problems under a single perspective. SURF feature extraction is performed on the preprocessed multispectral image. By detecting the key points in the image and calculating the descriptors of the surrounding area, a set of feature points is generated. Specifically, the feature point detection process is accelerated by generating an integral image. The integral image expression is:

[0063]

[0064] Where I(i,j) represents the pixel value of the original image at position (i,j);

[0065] And at each pixel (x, y) of the image, the Hessian matrix is ​​calculated, expressed as:

[0066]

[0067] Among them, L xx , L xy and L yy They represent the second-order Gaussian derivatives of the image at scale σ respectively;

[0068] By calculating the determinant of the Hessian matrix, the feature points are detected, which is expressed as:

[0069] Det(H)=L xx L yy -(L xy ) 2

[0070] When the determinant value is greater than the preset threshold, the pixel is considered as a potential feature point;

[0071] Then describe the detected feature points, generate feature vectors, and use Haar wavelet response to calculate the descriptor in the area around the feature points. The descriptor calculation expression is:

[0072]

[0073] Where dx and dy represent the Haar wavelet responses in the horizontal and vertical directions in the area around the feature point, respectively.

[0074] In feature matching, the FLANN algorithm is used to match feature points between images of different perspectives, and then the matched feature points are registered. The RANSAC algorithm is used to remove mismatched points to ensure the accuracy and robustness of the matching. Through multi-perspective collaborative observation, image data from different perspectives are obtained, registered and fused, and a preliminary three-dimensional point cloud model is generated. Combined with photometric stereo technology, multiple light sources are used to illuminate the object from different angles to capture the corresponding images. By analyzing the brightness changes of each pixel under different lighting conditions, the three-dimensional shape of the object surface is restored, thereby enhancing the accuracy of feature matching.

[0075] During system initialization, the matched feature points are used for system initialization. The matched feature points are used to generate a preliminary 3D point cloud model through triangulation method, which is expressed as:

[0076]

[0077] Among them, X represents a three-dimensional point, P1 -1 and They are the projection matrices under two viewing angles, and x1 and x2 are their corresponding pixel points;

[0078] After the preliminary 3D point cloud model is established, the BA algorithm is used for optimization. The camera parameters and 3D point positions are optimized by minimizing the reprojection error. The reprojection error formula is:

[0079]

[0080] Among them, x ij represents the observation value of the j-th 3D point under the i-th camera, P i ,X j Represents the projection function, which transforms the three-dimensional point X j Project it onto the image plane of the i-th camera; use the LM algorithm for nonlinear minimization, iteratively optimize the camera parameters and the position of the 3D points, reduce the reprojection error, and improve the accuracy of the 3D point cloud model. Verify the optimized 3D point cloud model by comparing it with known GIS data or LiDAR data to ensure its accuracy and consistency.

[0081] In motion recovery of sparse point clouds, SLAM technology is used to estimate the camera pose through feature point matching and triangulation results, calculate the intrinsic matrix and basic matrix between cameras, and describe the rotation and translation relationship of the camera. The basic matrix is ​​expressed as:

[0082]

[0083] Among them, K1 and K2 represent the intrinsic parameter matrix of the camera, and E represents the intrinsic matrix;

[0084] Then use the epipolar constraint to match the feature points so that the matching points meet the geometric constraints, which can be expressed as:

[0085]

[0086] Among them, x1 and x2 represent the matched image points; the BA algorithm is then used to optimize the sparse point cloud to minimize the reprojection error. High-precision sparse point cloud recovery and motion estimation are achieved in complex lighting environments, providing reliable data support for subsequent dense point cloud reconstruction and 3D model generation.

[0087] In dense point cloud reconstruction, a pre-trained 2D diffusion model is used to complete the occluded parts in the multi-view image. The diffusion model generates high-quality image completion results by gradually adding noise and learning the denoising process. A mask is generated through a deep learning model or manual annotation to detect the occluded area in the image. The mask of the occluded area is input into the 2D diffusion model to generate a completed image. The completed image remains consistent under different viewing angles to ensure the accuracy of 3D reconstruction. The laser radar data is also combined for densification. The point cloud data of the target area is obtained by using a laser radar device. The laser radar measures the distance between the target and the radar by emitting laser pulses and receiving their reflected signals to generate high-precision point cloud data. The laser radar point cloud data is pre-processed, including removing noise and filtering out invalid data, to improve data quality and accuracy. The sparse point cloud generated by the completed 2D image is fused with the laser radar point cloud data. The point cloud data from different viewing angles is registered to the same coordinate system through the registration algorithm to generate a dense point cloud.

[0088] In the network modeling and texture mapping steps, starting from dense point cloud data, the point cloud is preprocessed, including denoising and filtering; then the PSR algorithm is used to generate a three-dimensional mesh model, and a smooth three-dimensional surface is generated from the point cloud data by solving the Poisson equation, which is expressed as:

[0089] ▽·▽φ=▽·V

[0090] Where φ represents the implicit function to be solved, and V represents the vector field generated by the point cloud normal vector; and the generated mesh is simplified and refined to reduce the number of polygons and improve the smoothness and accuracy of the mesh;

[0091] Texture mapping is to obtain high-resolution images of the target area from different perspectives, cover the entire 3D model, align the multi-perspective images with the 3D mesh model, align the image with the model geometry, and then map the texture information of the multi-perspective images to the 3D mesh model, generate UV coordinates for each mesh vertex, and the UV coordinates represent the position in the image; use weighted averaging to perform texture fusion between multiple perspective images, eliminate gaps and overlapping areas, and ensure the continuity and consistency of the texture. Generate a high-precision 3D mesh model in a complex lighting environment, and perform texture mapping through multi-perspective images to ensure the realism and detail of the model.

[0092] In the illumination processing step, an illumination model is established to analyze illumination changes under different illumination conditions. Histogram equalization is used to process the image to reduce the impact of illumination changes on image feature extraction and matching. Then, through the fusion of multi-view images, illumination information under different perspectives is used for compensation to reduce the illumination impact under a single perspective. At the same time, a polarization camera is used to capture the polarization image of the target area to obtain light information under different polarization angles. The Stokes vector and Mueller matrix are used to analyze the polarization characteristics of polarized light and natural light. The Stokes vector describes the polarization state of light, and the Mueller matrix is ​​used to describe the polarization changes of the interaction between light and space objects, which can be expressed as:

[0093]

[0094] By filtering the specular reflection and analyzing it based on the Stokes vector and Mueller matrix, different polarization azimuths are calculated to separate the ambient light and specular reflection light. Through the above steps, the influence of complex lighting environment on the 3D reconstruction results can be effectively reduced, and the reconstruction accuracy and the realism of the model can be improved.

[0095] Additionally, in order to provide a concise description of exemplary embodiments, all features of an actual embodiment (ie, those features that are not relevant to the best mode presently contemplated for carrying out the invention or those that are not relevant to implementing the invention) may not be described.

[0096] It should be understood that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but for those of ordinary skill having the benefit of this disclosure, without undue experimentation, the development effort will be a routine task of design, fabrication, and production.

[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A satellite 3D reconstruction method in a complex lighting environment, characterized in that: The specific steps include: S1. Feature extraction: Use multispectral imaging to obtain images under different lighting conditions, combine multi-view collaborative observation, and extract SURF features in the image; S2, feature matching: the extracted SURF features are registered between feature points of different images. The multi-view data is registered and fused to generate a three-dimensional model, and combined with the illumination stereo technology, the illumination changes are analyzed; S3, system initialization: use the initial matched feature points to initialize the system, build a preliminary 3D point cloud model, and use the BA algorithm to optimize the initial model; S4, motion recovery of sparse point cloud: restore sparse point cloud through multi-view stereo vision and use SLAM for motion recovery; S5, dense point cloud reconstruction: use the 2D diffusion model to complete the occluded area, and combine the lidar data to densify the sparse point cloud; S6, Network modeling and texture mapping: Generate a 3D mesh model and perform texture mapping through multi-view images; S7. Lighting processing: Lighting compensation is used to reduce the impact of complex light environment on reconstruction results, and polarized light separation technology is used to separate ambient light and specular reflection light.

2. The method for three-dimensional reconstruction of a satellite in a complex lighting environment as claimed in claim 1, characterized in that: In the feature extraction, a multispectral camera is used to capture spectral information of multiple bands. The image of each band is used to provide different lighting information. Multiple shots are taken under different solar altitude angles, cloud coverage and atmospheric conditions, and the collected multispectral images are preprocessed, including radiation correction, atmospheric correction and geometric correction. At the same time, multiple satellites or drones are used to simultaneously shoot the target area from different angles to conduct multi-angle collaborative observation.

3. The method for three-dimensional reconstruction of a satellite in a complex lighting environment as claimed in claim 2, characterized in that: In the feature extraction, SURF feature extraction is performed on the preprocessed multispectral image, and a set of feature points is generated by detecting key points in the image and calculating the descriptors of the surrounding areas. Specifically, the feature point detection process is accelerated by generating an integral image, and the integral image expression is: Where I(i,j) represents the pixel value of the original image at position (i,j); And at each pixel (x, y) of the image, the Hessian matrix is ​​calculated, expressed as: Among them, L xx , L xy and L yy They represent the second-order Gaussian derivatives of the image at scale σ respectively; By calculating the determinant of the Hessian matrix, the feature points are detected, which is expressed as: It(H)=L xx L yy -(L xy ) 2 When the determinant value is greater than the preset threshold, the pixel is considered as a potential feature point; Then describe the detected feature points, generate feature vectors, and use Haar wavelet response to calculate the descriptor in the area around the feature points. The descriptor calculation expression is: Where dx and dy represent the Haar wavelet responses in the horizontal and vertical directions in the area around the feature point, respectively.

4. The method for three-dimensional reconstruction of a satellite in a complex lighting environment as claimed in claim 1, characterized in that: In the feature matching, the FLANN algorithm is used to match feature points between images of different perspectives, and then the matched feature points are registered. The RANSAC algorithm is used to remove mismatched points. Through multi-perspective collaborative observation, image data from different perspectives are obtained, registered and fused, and a preliminary three-dimensional point cloud model is generated. Combined with the photometric stereo technology, multiple light sources are used to illuminate the object from different angles to capture the corresponding images. By analyzing the brightness changes of each pixel under different lighting conditions, the three-dimensional shape of the object surface is restored.

5. The method for three-dimensional reconstruction of a satellite in a complex lighting environment as claimed in claim 1, characterized in that: In the system initialization, the matched feature points are used for system initialization, and a preliminary three-dimensional point cloud model is generated by a triangulation method using the matched feature points, which is expressed as: Among them, X represents a three-dimensional point, P1 -1 and They are the projection matrices under two viewing angles, and x1 and x2 are their corresponding pixel points; After the preliminary 3D point cloud model is established, the BA algorithm is used for optimization. The camera parameters and 3D point positions are optimized by minimizing the reprojection error. The reprojection error formula is: Among them, x ij represents the observation value of the jth 3D point under the i-th camera, P i ,X j Represents the projection function, which transforms the three-dimensional point X j Project it onto the image plane of the i-th camera; and use the LM algorithm for nonlinear minimization to iteratively optimize the camera parameters and the position of the 3D points.

6. The method for three-dimensional reconstruction of a satellite in a complex lighting environment as claimed in claim 1, characterized in that: In the motion recovery sparse point cloud, SLAM technology is used to estimate the camera pose through feature point matching and triangulation results, calculate the intrinsic matrix and basic matrix between the cameras, and describe the rotation and translation relationship of the camera. The basic matrix is ​​expressed as: Among them, K1 and K2 represent the intrinsic parameter matrix of the camera, and E represents the intrinsic matrix; Then use the epipolar constraint to match the feature points so that the matching points meet the geometric constraints, which can be expressed as: Among them, x1 and x2 represent the matched image points; the BA algorithm is then used to optimize the sparse point cloud to minimize the reprojection error.

7. The method for three-dimensional reconstruction of a satellite in a complex lighting environment as claimed in claim 1, characterized in that: In the dense point cloud reconstruction, a pre-trained 2D diffusion model is used to complete the occluded parts in the multi-view image. The diffusion model generates high-quality image completion results by gradually adding noise and learning the denoising process. A mask is generated through a deep learning model or manual annotation to detect the occluded area in the image, and the Mask of the occluded area is input into the 2D diffusion model to generate a completed image.

8. The method for three-dimensional reconstruction of a satellite in a complex lighting environment as claimed in claim 7, characterized in that: In the dense point cloud reconstruction, the laser radar data is also combined for densification processing. The point cloud data of the target area is obtained by using a laser radar device. The laser radar measures the distance between the target and the radar by emitting laser pulses and receiving their reflected signals, thereby generating high-precision point cloud data. The laser radar point cloud data is preprocessed, including removing noise and filtering out invalid data. The sparse point cloud generated by the completed 2D image is fused with the laser radar point cloud data. Through the registration algorithm, the point cloud data from different perspectives are registered to the same coordinate system to generate a dense point cloud.

9. The satellite 3D reconstruction method in a complex lighting environment as claimed in claim 1, characterized in that: In the network modeling and texture mapping step, starting from dense point cloud data, the point cloud is preprocessed, including denoising and filtering; then the PSR algorithm is used to generate a three-dimensional mesh model, and a smooth three-dimensional surface is generated from the point cloud data by solving the Poisson equation, which is expressed as: Among them, φ represents the implicit function to be solved, and V represents the vector field generated by the point cloud normal vector; and the generated mesh is simplified and refined to reduce the number of polygons; Texture mapping is to obtain high-resolution images of the target area from different perspectives, covering the entire 3D model, aligning the multi-perspective images with the 3D mesh model to align the geometry of the image and the model, and then mapping the texture information of the multi-perspective images to the 3D mesh model, generating UV coordinates for each mesh vertex, which represent the position in the image; using weighted averaging, texture fusion is performed between multiple perspective images to eliminate gaps and overlapping areas.

10. A satellite 3D reconstruction method in a complex lighting environment as claimed in claim 1, characterized in that: In the illumination processing step, an illumination model is established to analyze illumination changes under different illumination conditions, and histogram equalization is used to process the image to reduce the influence of illumination changes on image feature extraction and matching. Then, through the fusion of multi-view images, compensation is performed using illumination information under different viewing angles. At the same time, a polarization camera is used to capture a polarization image of the target area, light information under different polarization angles is obtained, and the polarization characteristics of polarized light and natural light are analyzed using Stokes vectors and Mueller matrices. Mirror reflections are filtered, and then analyzed based on Stokes vectors and Mueller matrices to calculate different polarization azimuths and separate ambient light and mirror reflection light.

Images