Research On Trajectory Recovery And Terrain Reconstruction Based On Descent Images Of Lunar/Mars Lander

With the successful completion of the three-step mission of the lunar exploration project "around,landing,return" and the successful implementation of the first Mars exploration mission,China has become the third country in the world to successfully land on the Moon and Mars and complete scientific exploration on their surfaces.In the deep space exploration mission,the high-precision landing site location and landing area terrain information will provide basic data for the rover path planning task,and the lander’s descent trajectory can be used for the high-precision positioning of the landing site,so the trajectory recovery and terrain reconstruction based on the descent images are of great significance.The lack of texture and the richness of morphologically similar features in the descent image reduce the uniformity and robustness of feature point extraction and matching;the flat terrain of the landing area degrades the imaging area of the descent image to an approximate plane,and a robust camera pose solution cannot be obtained;the almost-vertical descent trajectory of the lander makes the angle of the descent image with the same name light too small,and the parallax is not sensitive to the depth change,which reduces the accuracy of visual terrain reconstruction.The accuracy of visual terrain reconstruction is reduced.The above problems bring challenges to the lander trajectory recovery and landing area terrain reconstruction tasks based on descent images.For this reason,this paper fully combines the imaging characteristics of descent images and carries out the research on the trajectory recovery and terrain reconstruction methods based on descent images,and the main research contents include:(1)For the high-frame rate descent images acquired by the lunar lander,the SIFT+KLT strategy is used to extract and match feature points,which improves the computational efficiency and ensures the feature point matching accuracy at the same time.Facing the problem of insufficient number of feature points and uneven distribution due to the lack of texture in Mars descent images,a method of uniform extraction and matching of feature points is proposed,which increases the number of feature points in texture-poor regions by lowering the response value threshold,constructing a multi-scale uniform grid and calculating the number of feature points required for each scale corresponding to the square grid based on the next closest neighbor distance ratio,so that the extracted feature points can be uniformly distributed in the image plane and scale space while having matching points high-quality.To address the problem of a large number of mismatch points in the descent images due to the richness of morphologically similar features,a mismatch point rejection method based on a priori terrain information and image scale monotonicity is studied and proposed: firstly,based on the geometric relationship between plane induced parallax and terrain height difference,the orbiter image information is used to obtain the maximum height difference of terrain in the imaging area of the landing camera and estimate the boundary threshold of plane induced parallax to reject the mismatch points that do not satisfy the threshold.Then,based on the nature of monotonic variation of the scale of the feature points in the descent images with the imaging sequence,the initial set of false matches is rejected by the scale monotonicity to obtain the ideal matching points.After testing the methods using simulated data in the field experiment,it is shown that the SIFT+KLT strategy ensures the matching accuracy and computational efficiency;the uniform extraction and matching method of feature points not only increases the number of feature points extracted,but also ensures the uniformity of feature point distribution,thus improving the robustness of the mapping model solution;the false match point rejection method proposed in this paper can effectively reject the false match points.The proposed strategy and algorithm are applied to the in-orbit data processing of Chang’e-4 and Tianwen-1,and both obtained robust feature tracking sequences.(2)For the high frame rate descent images acquired by the lunar lander,an image key frame selection strategy is proposed to overcome the problem of degenerate plane.For the low frame rate descent images acquired by Mars lander,a plane-induced parallax beam method based on the constraint of epipoles is proposed,and two sets of possible camera positional solutions are obtained by using the homography matrix decomposition as the constraint to find a robust non-degenerate mapping model.The geometric relationship between the two sets of possible camera positional solutions is analyzed,their normal plane anisotropy is proved,and the uncertainty of the camera positional solutions in the plane-induced homography is eliminated by combining the parabolic nature of the Mars lander descent trajectory.The results of field experiment show that the key frame selection strategy proposed in this paper can effectively overcome the problem of degenerate plane;the planar induced parallax beam method based on the constraint of epipoles can obtain a more robust fundamental matrix as a non-degraded mapping model compared with other existing methods;the camera pose uncertainty elimination method based on the parabolic shape constraint of the descent trajectory can effectively reject the wrong camera pose solutions.In the on-orbit missions of Chang’e-4 and Tianwen-1,the method proposed in this paper was used to accurately recover the descent trajectory of the lander and to complete the high-precision positioning of the landing sites of Chang’e-4 and Tianwen-1.(3)The optical axis of the descent image is approximately parallel to the photographic baseline,resulting in a small angle of the corresponding light,which makes the traditional multi-view stereo geometry method not suitable for processing descent images.Meanwhile,the virtual plane constructed by the existing plane-sweep method does not have directional and depth constraints,which cannot fit the local terrain surface well and the recovered inclined terrain slope is "step-like".A reference image selection strategy is established for this model,and a seed point selection and depth propagation strategy is built to recover the depth information of the current spatial triangulated surface elements,and then a dense terrain point cloud and the DEM and DOM of the corresponding area are obtained.The results of field experiment show that the proposed method in this paper improves the terrain reconstruction accuracy.The method has been applied in Chang’e-4 and Tianwen-1 in-orbit missions,and the corresponding terrain products have been generated to provide powerful data support for the rover path planning.