Space target accurate relative pose estimation method based on reconstruction model

Abstract

The invention discloses a space target accurate relative pose estimation method based on a reconstruction model, and the method comprises the following steps: space target three-dimensional reconstruction, binocular vision local reconstruction, three-dimensional point cloud matching, real-time estimation of the relative pose between a tracking spacecraft and a target spacecraft, and non-cooperative target three-dimensional reconstruction model scale information. And performing real-time estimation on the accurate relative pose of the non-cooperative target in the short-distance control stage.

Description

Technical Field

[0001] The present invention relates to the technical field of spacecraft navigation, guidance and control methods, and in particular to a method for accurately estimating relative position and posture of a space target based on a reconstruction model. Background Art

[0002] In the early days, due to the limitations of technological development, most of the focus of relative motion state estimation of space targets was on pose estimation in motion estimation, which usually required prior information of the space target, such as deploying known labels on the target or giving a target structure model.

[0003] By using the classic orthogonal projection iterative transformation algorithm and perspective-N-Point (PNP) algorithm in machine vision, the strong geometric features on the target are detected and identified to solve its position and posture relative to the camera coordinate system. By identifying the features of fixed position and known size on the target, the position and posture of the target are solved using the iterative solution of the four-point estimated translation and rotation matrix. The advantage of this method is that it has a small amount of calculation and is conducive to engineering design. The disadvantage is that four circular features need to be detected at the same time when solving the target. In addition, by selecting the solar panel bracket and engine nozzle of the space target as the identification object, a target posture solution method based on the intersection features of straight lines and circles is established. This method can achieve relatively accurate measurement, but it is very dependent on the strong geometric features on the target.

[0004] Model matching based methods are also often used for the pose estimation problem of non-cooperative targets, which usually requires a known three-dimensional external structure model of the target. A three-dimensional model of the target is established based on the actual observed target, and then the edge contour features of the target in the image are detected. These contour features are then matched with the contour obtained by projecting the three-dimensional model of the target onto the two-dimensional image plane. The real-time pose of the target is solved by minimizing the error function. The advantage of the model matching based method is that it achieves higher solution accuracy with lower computational effort, while the disadvantage is that it requires the three-dimensional model of the target as prior knowledge, and in non-cooperative target motion estimation, it is usually impossible to obtain prior information on the three-dimensional model of the spatial target. Summary of the invention

[0005] The main technical problem solved by the present invention is to provide a method for estimating the accurate relative pose of space targets based on a reconstruction model. By flying around a non-cooperative target at close range, the outer contour of the non-cooperative space target is three-dimensionally reconstructed according to the collected visual and laser data, and the accurate three-dimensional structure and size data information of the target star is obtained. On this basis, the real-time relative pose of the two satellites during the close-range control phase is accurately measured and estimated.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is to provide a method for accurately estimating relative pose of a space target based on a reconstruction model, comprising the following steps:

[0007] (1) 3D reconstruction of space targets: Obtain visual images of the target spacecraft from different angles, and reconstruct the 3D dense point cloud of the target spacecraft based on the relative pose corresponding to each frame of the visual image;

[0008] 1) Feature point detection and matching: Find feature points in different scale spaces and calculate the direction of feature points;

[0009] 2) Motion recovery sparse point cloud: The motion recovery method is used to recover the intrinsic and extrinsic parameters of the camera, and the 3D point coordinates are obtained by triangulation. The joint correction method is used for optimization and the sparse 3D point cloud is output.

[0010] 3) Establish dense point cloud: Use image-based methods to obtain dense point cloud, and use multi-eye stereo vision algorithm based on image feature points to directly obtain dense point cloud model through feature extraction and matching;

[0011] (2) Binocular vision local reconstruction: Using cameras at two different locations on the tracking spacecraft and the target spacecraft to image, the same-name pixels in the overlapping three-dimensional space areas of the images are searched and matched, and then the three-dimensional spatial coordinate information of the pixels in the two-dimensional images is restored through the principle of triangulation;

[0012] 1) Matching cost calculation: The cost of two primitives to be matched within the disparity range is calculated by the cost function as their similarity measure value. The smaller the similarity measure value, the greater the probability of being a homonymous point;

[0013] 2) Matching effect optimization: Optimize the similarity measure value obtained in the matching cost calculation to facilitate the calculation of the initial disparity map;

[0014] 3) Disparity map generation: Select the disparity of each pixel that minimizes the matching cost of the point as the disparity of the point, solve the disparity value that minimizes the energy function value as the final disparity of the pixel, and obtain the initial disparity map of the image;

[0015] (3) 3D point cloud matching: Obtain the current binocular vision local reconstruction 3D point cloud and the target spacecraft 3D dense point cloud model data, use the normal distribution transformation algorithm to match the point cloud and the map, and obtain the precise position transformation relationship between the tracking spacecraft and the target spacecraft;

[0016] (4) Real-time estimation of the relative position between the tracking spacecraft and the target spacecraft.

[0017] In a specific embodiment of the present invention, the feature point detection and matching further includes the following steps:

[0018] a) Extract key points: Search image locations in all scale spaces and identify potential scale- and rotation-invariant interest points through Gaussian derivative functions;

[0019] b) Locate key points and determine feature directions: At each candidate location, determine the position and scale by fitting a sophisticated model, and assign one or more directions to each key point location based on the local gradient direction of the image;

[0020] c) Through the feature vectors of each key point, compare them two by two to find several pairs of feature points that match each other and establish the corresponding relationship between the scenes.

[0021] In a specific embodiment of the present invention, the generation of the disparity map also includes: a process of performing disparity correction on the initial disparity map to obtain an accurate disparity map, and obtaining the accurate disparity map is as follows: using a left-right consistency check algorithm to eliminate erroneous disparities, eliminating small connected areas to remove abnormal points, using a median filter or a bilateral filter algorithm to filter and smooth the initial disparity map, and using quadratic curve fitting to achieve sub-pixel refinement to obtain an accurate disparity map.

[0022] In a specific embodiment of the present invention, in step (3) three-dimensional point cloud matching, the point cloud used for matching is:

[0023] x i =[x i ,y i ,z i ] T ,i={0,1,2,…,M-1}

[0024] Among them, x ki is the i-th 3D point, x i ,y i ,z i is the three-axis coordinate value corresponding to the point, M is the number of three-dimensional points, and these point clouds are distributed to the three-dimensional voxel grid of the standard normal distribution. For each voxel grid, there are M k 3D points, the mean and variance of the NDT distribution of this grid are:

[0025]

[0026]

[0027] After obtaining the mean and variance of each standard normal distribution 3D voxel grid, get the estimated value for each grid:

[0028]

[0029] By finding suitable coordinate transformation parameters to maximize the sum of the NDT probability density corresponding to the point cloud of the current frame, the corresponding posture transformation matrix is ​​solved to achieve point cloud matching.

[0030] In a specific embodiment of the present invention, step (4) of real-time estimation of relative posture further includes the following steps:

[0031] The variables such as the target spacecraft mass center position, velocity, rotation center position, attitude quaternion and rotation angular velocity are combined into a state vector. The state vectors of the target translation and rotation motion are X T =[d T v T r T ] T , X R =[q T ω T ] T , let the state vector Then the target kinematic state equation is:

[0032] X M,k =f M (X M,k-1 )+w k

[0033] Among them, f M (X M,k-1 ) is the dynamic equation, w k is the system noise;

[0034] Assume t k By matching the three-dimensional point cloud at all times, the relative pose estimation of the two spacecraft is obtained as Z k , X k is the real relative motion state variable at that moment, and the nonlinear observation function can be expressed as h(X k ), then the system observation equation can be organized into the following expression:

[0035] Z k =h(X M,k ,X S,k )+n k ;

[0036] Based on the above state equation and observation equation, real-time estimation of the relative motion state between the tracking spacecraft and the target spacecraft can be achieved.

[0037] In a specific embodiment of the present invention, a binocular vision camera is used to obtain visual images of the tracking spacecraft and the target spacecraft. The image acquisition frame rate of the binocular vision camera is 14 frames per second and the image resolution is 5 million pixels.

[0038] The beneficial effects of the present invention are as follows: the present invention provides a method for estimating the precise relative pose of a space target based on a reconstruction model, by constructing a dynamic state space model of the relative position and relative pose of two spacecraft, constructing an extended Kalman filter algorithm for relative state estimation of a space target, using inertial measurement unit data to realize recursive solution of the relative pose, using pose data obtained by point cloud matching to update the state, and after completing the three-dimensional reconstruction of the non-cooperative target, the accurate external structure parameters of the target star can be obtained; according to the reconstructed non-cooperative target model parameters, the three-dimensional information of the non-cooperative spacecraft is reconstructed by using a binocular camera, and then the measurement results of the visual system are matched with the three-dimensional model of the target through the algorithm, thereby realizing the estimation of the relative pose of the target, and the relative pose of the two spacecraft can be accurately calculated by the point cloud registration method, the point cloud registration algorithm is combined with the EKF, the pose estimation result obtained by the point cloud registration algorithm is used as the input of the EKF algorithm, and then the filter estimation result of the EKF algorithm is used as the initial value of the point cloud registration algorithm to improve the convergence characteristics of the point cloud registration algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0040] Figure 1 It is a flow chart of feature extraction and matching in the present invention;

[0041] Figure 2 Schematic diagram of binocular stereo vision in the present invention;

[0042] Figure 3 is the parallax space image matrix in the present invention;

[0043] Figure 4 It is a schematic diagram of the appearance of the dense point cloud of the three-dimensional reconstruction of the target spacecraft in the present invention. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The embodiment of the present invention is: a method for accurately estimating relative position and pose of a space target based on a reconstruction model, comprising the following steps:

[0046] (1) 3D reconstruction of space targets: Obtain visual images of the target spacecraft from different angles, and reconstruct the 3D dense point cloud of the target spacecraft based on the relative pose corresponding to each frame of the visual image;

[0047] 1) Feature point detection and matching: Find feature points in different scale spaces and calculate the direction of the feature points. SIFT features are local features of the image. They remain invariant to rotation, scaling, and brightness changes. They also maintain a certain degree of stability to perspective changes, affine transformations, and noise. Therefore, the SIFT features of the image are extracted here. By aligning the obtained SIFT features, the spatial mapping relationship between the pixels of one image and the pixels of another image is found.

[0048] SIFT feature extraction and matching is used. SIFT, or scale-invariant feature transform, essentially searches for key points in different scale spaces and calculates the direction of the key points. The feature extraction and matching process is as follows: Figure 1 shown.

[0049] The working process is as follows:

[0050] a) Extract key points: Key points are some very prominent points that will not disappear due to factors such as lighting, scale, rotation, etc., such as corner points, edge points, bright spots in dark areas, and dark spots in bright areas. This step searches for image positions in all scale spaces and identifies potential scale- and rotation-invariant interest points through Gaussian derivative functions.

[0051] b) Locate key points and determine feature directions: At each candidate position, a finely fitted model is used to determine the position and scale. Key points are selected based on their stability, and then one or more directions are assigned to each key point position based on the local gradient direction of the image. All subsequent operations on the image data are transformed relative to the direction, scale, and position of the key points, thereby providing invariance to these transformations.

[0052] c) Through the feature vectors of each key point, compare them two by two to find several pairs of feature points that match each other and establish the corresponding relationship between the scenes.

[0053] 2) Motion recovery sparse point cloud: An iterative and robust motion recovery method is used to recover the camera's intrinsic and extrinsic parameters. The three-dimensional point coordinates are obtained by triangulation, and then optimized using a joint correction method to output a sparse three-dimensional point cloud.

[0054] The algorithm process mainly includes:

[0055] a) Camera calibration, obtaining camera internal parameters;

[0056] b) Use a calibrated camera to take pictures of the same scene from multiple angles and save them in sequence;

[0057] c) Calculate matching feature points for adjacent images pairwise;

[0058] d) Calculate the basic matrix F using the corresponding point pairs calculated previously;

[0059] e) calculating the eigenmatrix through the fundamental matrix;

[0060] f) Calculate the motion between the two views, i.e. R, T, through the intrinsic matrix;

[0061] g) After calculating the [R|T] matrix, all feature points can be reconstructed using optical triangulation.

[0062] 3) Establishing dense point cloud: The image-based method is used to obtain dense point cloud. The Multi-View Stereo (MVS) algorithm based on image feature points directly obtains the dense point cloud model through feature extraction and matching. Intuitively, the point of the spatial point on the left view is p 1 , if there is a corresponding p in the right view 2 , then p 2 must be on the corresponding epipolar line. This relationship is called the epipolar geometry constraint. Under this constraint, s is known as p 1 , want to find p in another picture 2 , can be directly on the polar line L 2 MVS mainly searches for the best matching point in different images.

[0063] In MVS, pixel-by-pixel matching is performed and searches are performed on the epipolar lines to determine whether the points between images correspond to the same point in the scene. Here, a consistency discriminant function between image points is proposed:

[0064] G ij (p) = ρ(I i (Ω(π i (p))), I j (Ω(π j (p))))

[0065] Among them, π(p) is the function that projects point p to the image, Ω(p) function defines the area of ​​​​a point p, I(p) function represents the intensity characteristics of the photo area, ρ(f, g) is used to compare the similarity between two vectors, and the specific choice of ρ function and Ω function determines the accuracy of this "consistency judgment".

[0066] Through this method, almost every pixel in the image can be matched. With dense matching, the three-dimensional coordinates of almost every pixel can be reconstructed, so that the density of the point cloud obtained can be close to the clarity of the resolution of the image.

[0067] (2) Local reconstruction of binocular vision: After stereo vision matches the same-name pixels between two or more stereo images containing overlapping areas, the obtained parallax can restore the three-dimensional spatial coordinates of the pixels in the image according to the intrinsic and extrinsic parameters of the camera. Stereo images are divided into sequential images, binocular images, and multi-camera images. After calibration, the binocular camera can obtain a relatively stable stereo image pair. This project uses binocular stereo matching to obtain a three-dimensional point cloud. Binocular stereo matching is a process of searching and matching the same-name pixels in the overlapping three-dimensional space area of ​​the image after imaging by two cameras at different positions. Its purpose is to restore the three-dimensional spatial coordinate information of the pixel points in the two-dimensional image through the triangulation principle based on the binocular stereo vision model. The binocular stereo vision model is as follows: Figure 2 shown.

[0068] 1) Matching cost calculation: Matching cost calculation is a step to measure the similarity of matching primitives between the reference image and the target image. The cost function is used to calculate the cost of two matching primitives within the disparity range as their similarity measure. The smaller the value, the greater the probability of being a homonymous point. The disparity with the greatest similarity can be taken as the initial disparity of the pixel, but this will discard a lot of information and generate a large number of mismatched points due to factors such as noise. Another approach is to construct a disparity space image matrix (DSI) with the size of the image width, height and disparity range to store the cost value of each pixel at each disparity, such as Figure 3 As shown;

[0069] 2) Matching effect optimization: The similarity measure value obtained in the matching cost calculation is optimized to facilitate the calculation of the initial disparity map. Specifically, the matching cost calculation result reflects the similarity between the pixel points in the reference image and the pixel points in the target image under different disparities, but this only considers local information. Therefore, if the disparity with the smallest cost function value is directly selected as the actual disparity of the pixel point, the result can be easily affected by noise or weak texture, repeated texture, etc., resulting in mismatch.

[0070] In order to make the cost value effectively express the real correlation between pixels, it is necessary to follow certain constraints, such as the parallax smoothness constraint, for each pixel under a certain parallax, and recalculate the cost value according to the values ​​of the adjacent pixels in the parallax neighborhood to obtain a new DSI matrix, so as to achieve global optimization of the cost value. Its essence is the propagation of parallax. The area with high signal-to-noise ratio can reflect good correlation after matching cost calculation, and it can be propagated to the area with low signal-to-noise ratio by optimization to reduce the generation of mismatched points. The goal of the global stereo matching algorithm when calculating the matching cost is to convert the cost calculation into an energy minimization problem, which is achieved by establishing a global energy function and finding the minimum value. After the energy function is determined, the disparity map is calculated by the global stereo matching algorithm.

[0071] 3) Disparity map generation: Select the disparity of each pixel that minimizes the matching cost of the point as the disparity of the point, and solve the disparity value that minimizes the energy function value as the final disparity of the pixel. This step means that the result of the cost optimization process directly determines the final matching effect, so it must accurately reflect the correlation between pixels.

[0072] The above steps can calculate the initial disparity map of the image pair, but there are still a lot of mismatched points caused by noise, occlusion, etc., and the accuracy is only at the pixel level. Therefore, to correct the disparity of the initial disparity map to obtain a more accurate disparity map, it is necessary to use the left-right consistency check (LRC) algorithm to remove the wrong disparity; remove the abnormal points by removing small connected areas; use median filtering, bilateral filtering and other algorithms to filter and smooth the disparity map; use quadratic curve fitting to achieve sub-pixel refinement, etc. After obtaining the disparity of the pixel points in the stereo image pair, the internal and external parameters of the binocular camera can restore its coordinates in three-dimensional space to obtain point cloud information.

[0073] (3) 3D point cloud matching: Obtain the current binocular vision local reconstruction 3D point cloud and the target spacecraft 3D dense point cloud model data, use the normal distribution transformation algorithm to match the point cloud and the map, and obtain the precise position transformation relationship between the tracking spacecraft and the target spacecraft; the point cloud used for matching is:

[0074] x ki =[x i ,y i ,z i ] T ,i={0,1,2,…,M-1}

[0075] Among them, x ki is the i-th 3D point, x i ,y i ,z iis the three-axis coordinate value corresponding to the point, and M is the number of three-dimensional points. These point clouds are assigned to a three-dimensional voxel grid called a standard normal distribution. For each voxel grid, there are Mk three-dimensional points. The mean and variance of the NDT distribution of the grid are:

[0076]

[0077]

[0078] After obtaining the mean and variance of each standard normal distribution three-dimensional voxel grid, the probability density of each grid can be obtained:

[0079]

[0080] In this way, the corresponding attitude transformation matrix can be solved by finding appropriate coordinate transformation parameters to maximize the sum of the NDT probability density corresponding to the current frame point cloud:

[0081] f(t)=-e(x,t)

[0082] By solving the Gauss-Newton method, the posture transformation matrix can be solved, thereby achieving point cloud matching.

[0083] (4) Real-time estimation of relative pose

[0084] The target spacecraft mass center position d T , speed v T , rotation center position r T 、Attitude quaternion q T and the angular velocity ω T The variables constitute the state vector. The state vectors of the target translation and rotation are X T =[d T v T r T ] T , X R =[q T ω T ] T , let the state vector Then the target kinematic state equation is:

[0085] X M,k =f M (X M,k-1 )+w k

[0086] In the above formula, f M (X M,k-1 ) is the dynamic equation, w k is the system noise, which can be expanded to

[0087]

[0088] The above formula can be expressed as the following matrix form

[0089]

[0090] Assume t k By matching the three-dimensional point cloud at all times, the relative pose estimation of the two spacecraft is obtained as Z k , X k is the real relative motion state variable at that moment, and the nonlinear observation function can be expressed as h(X k ), then the system observation equation can be organized into the following expression

[0091] Z k =h(X M,k ,X S,k )+n k

[0092] According to the above state equation and observation equation, the real-time estimation of the relative motion state of the two spacecraft can be realized. In the state estimation process, the three-dimensional point cloud reconstruction model result of the target spacecraft is used.

[0093] (5) Test Cases

[0094] 1) Test conditions

[0095] The Daheng Camera Mercury series visual camera is used. The image acquisition frame rate of the visual camera is 14 frames per second and the image resolution is 5 million pixels. The baseline length of the binocular camera is 20 cm, and the two cameras are synchronized and triggered. The image data is transmitted through the USB data interface, and the ROS node is used to collect image data. The inertial measurement unit uses XSENS Mti-630. The collected data includes angular velocity and linear acceleration. The data acquisition frequency is 100Hz.

[0096] 2) Target spacecraft reconstruction

[0097] By collecting binocular vision camera and inertial measurement unit data, the target spacecraft is reconstructed in three dimensions by calling the three-dimensional reconstruction algorithm. The dense point cloud result of the three-dimensional reconstruction of the target spacecraft is as follows: Figure 4 As shown, the length, width and height of the reconstructed model were calculated using CloudCompare software and compared with the actual spacecraft model size. The errors in length, width and height of the reconstructed model were 0.63cm, 0.69cm and 0.93cm respectively.

[0098] 3) Relative pose estimation

[0099] Based on the measurement data of visual cameras, inertial measurement units, etc., the relative position and attitude of the two spacecraft are estimated in real time through the extended Kalman filter algorithm. The relative position and relative attitude results of the five points are recorded below and compared with the true values ​​to evaluate the position and attitude accuracy. The results are shown in Tables 1 and 2 respectively.

[0100] Table 1 Relative position estimation accuracy

[0101]

[0102] Table 2 Relative attitude estimation accuracy

[0103]

[0104] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for accurately estimating relative pose of a space target based on a reconstruction model, characterized in that: The following steps are involved: (1) 3D reconstruction of space targets: Obtain visual images of the target spacecraft from different angles, and reconstruct the 3D dense point cloud of the target spacecraft based on the relative pose corresponding to each frame of the visual image; 1) Feature point detection and matching: Find feature points in different scale spaces and calculate the direction of feature points; 2) Motion recovery sparse point cloud: The motion recovery method is used to recover the intrinsic and extrinsic parameters of the camera, and the 3D point coordinates are obtained by triangulation. The joint correction method is used for optimization and the sparse 3D point cloud is output. 3) Establish dense point cloud: Use image-based methods to obtain dense point cloud, and use multi-eye stereo vision algorithm based on image feature points to directly obtain dense point cloud model through feature extraction and matching; (2) Binocular vision local reconstruction: Using cameras at two different locations on the tracking spacecraft and the target spacecraft to image, the same-name pixels in the overlapping three-dimensional space areas of the images are searched and matched, and then the three-dimensional spatial coordinate information of the pixels in the two-dimensional images is restored through the principle of triangulation; 1) Matching cost calculation: The cost of two primitives to be matched within the disparity range is calculated by the cost function as their similarity measure value. The smaller the similarity measure value, the greater the probability of being a homonymous point; 2) Matching effect optimization: Optimize the similarity measure value obtained in the matching cost calculation to facilitate the calculation of the initial disparity map; 3) Disparity map generation: Select the disparity of each pixel that minimizes the matching cost of the point as the disparity of the point, solve the disparity value that minimizes the energy function value as the final disparity of the pixel, and obtain the initial disparity map of the image; (3) 3D point cloud matching: Obtain the current binocular vision local reconstruction 3D point cloud and the target spacecraft 3D dense point cloud model data, use the normal distribution transformation algorithm to match the point cloud and the map, and obtain the precise position transformation relationship between the tracking spacecraft and the target spacecraft; (4) Real-time estimation of the relative position between the tracking spacecraft and the target spacecraft.

2. The method for accurately estimating relative position and pose of a space target based on a reconstruction model according to claim 1, characterized in that: The following steps are also included in feature point detection and matching: a) Extract key points: Search image locations in all scale spaces and identify potential scale- and rotation-invariant interest points through Gaussian derivative functions; b) Locate key points and determine feature directions: At each candidate location, determine the position and scale by fitting a sophisticated model, and assign one or more directions to each key point location based on the local gradient direction of the image; c) Through the feature vectors of each key point, compare them two by two to find several pairs of feature points that match each other and establish the corresponding relationship between the scenes.

3. The method for accurately estimating relative position and pose of a space target based on a reconstruction model according to claim 1, characterized in that: The disparity map generation also includes: a process of performing disparity correction on the initial disparity map to obtain an accurate disparity map, and obtaining the accurate disparity map is as follows: using a left-right consistency check algorithm to eliminate erroneous disparities, eliminating small connected areas to remove abnormal points, using a median filter or a bilateral filter algorithm to filter and smooth the initial disparity map, and using quadratic curve fitting to achieve sub-pixel refinement to obtain an accurate disparity map.

4. The method for accurately estimating relative position and pose of a space target based on a reconstruction model according to claim 1, characterized in that: In step (3) 3D point cloud matching, the point cloud used for matching is: x i =[x i ,y i ,z i ] T ,i={0,1,2,…,M-1} Among them, x ki is the i-th 3D point, x i ,y i ,z i is the three-axis coordinate value corresponding to the point, M is the number of three-dimensional points, and these point clouds are distributed to the three-dimensional voxel grid of the standard normal distribution. For each voxel grid, there are M k 3D points, the mean and variance of the NDT distribution of this grid are: Where: Pk, mean After obtaining the mean and variance of each standard normal distribution 3D voxel grid, obtain the estimated value of each grid: By finding suitable coordinate transformation parameters to maximize the sum of the NDT probability density corresponding to the point cloud of the current frame, the corresponding posture transformation matrix is ​​solved to achieve point cloud matching.

5. The method for accurately estimating relative position and pose of a space target based on a reconstruction model according to claim 1, characterized in that: Step (4) also includes the following steps in the real-time estimation of relative posture: The variables such as the target spacecraft mass center position, velocity, rotation center position, attitude quaternion and rotation angular velocity are combined into a state vector. The state vectors of the target translation and rotation motion are X T =[d T v T r T ] T , X R =[q T ω T ] T , let the state vector Then the target kinematic state equation is: X M,k =f M (X M,k-1 )+w k Among them, f M (X M,k-1 ) is the dynamic equation, w k is the system noise; Assume t k By matching the three-dimensional point cloud at all times, the relative pose estimation of the two spacecraft is obtained as Z k , X k is the real relative motion state variable at that moment, and the nonlinear observation function can be expressed as h(X k ), then the system observation equation can be organized into the following expression: Z k =h(X M,k ,X S,k )+n k ; Based on the above state equation and observation equation, real-time estimation of the relative motion state between the tracking spacecraft and the target spacecraft can be achieved.

6. The method for accurately estimating relative position and pose of a space target based on a reconstruction model according to claim 1, characterized in that: A binocular vision camera is used to obtain visual images of the tracking spacecraft and the target spacecraft. The image acquisition frame rate of the binocular vision camera is 14 frames per second and the image resolution is 5 million pixels.

Images