Research On Relative Pose Measurement For Visual Guidance Of Lunar Surface Sampling

China’s lunar exploration program is reaching the final stage,and Chang’e-5 lunar probe will take on the first lunar robotic sampling and return mission of China.Measuring the relative pose between the sampler and the probe precisely is the indispensable prerequisite for accomplishing the mission.As a momentous part of the telemanipulation system,the vision-based pose measurement module is able to guide the lunar surface sampling robotic arm to successfully performs operations according to the planned tasks,such as sampling,grasping,encapsulating.Compared with the ground scene,the imaging environment on the lunar surface is more complicated and adverse,including sudden light change,radiation effects in space,and high-temperature phenomena.All these factors bring great challenges to the high precision vision-based pose measurement.Therefore,studying the relative pose measurement method for the visual guidance of lunar surface sampling is of great strategic significance and practical engineering value.In order to effectively improve the efficiency and success rate of the lunar surface sampling robotic arm,the relative pose measurement method for the visual guidance of lunar surface sampling was systematically studied in this dissertation.A series of new algorithms for the key techniques in the relative pose measurement were proposed,including image matching,robust estimation and plane-based pose estimation.Then a complete vision-based pose measurement system was designed,which can be applied to the lunar surface sampling robotic arm.In detail,the main results and contributions of this dissertation are listed as follows:(1)To solve the feature-point matching problem,an iterative mismatch removal algorithm based on the shape interaction matrix was proposed.The algorithm mainly exploits the anti-noise capability of the low rank representation to tell the mismatches apart from the corresponding point sets.First,the shape interaction matrix is introduced to represent the candidate matching points.Then the cosine distance is applied to compute the difference between the column vectors of the shape interaction matrixes evolving from two corresponding point sets.Finally,the adaptive threshold method is used to determine the turning point between the mismatches and correct matches.An iterative matching strategy from coarse to fine is designed to weaken the effect of outliers.Experiments show that the proposed algorithm is robust to noise,a large number of mismatches,illumination or view changes,and also has the advantage of high matching precision and efficiency.(2)A robust template matching algorithm was proposed for the area-based image matching problem,which is based on the best-buddies similarity and regional weighting.The algorithm can handle the problem that the conventional template matching methods are prone to failure under complex conditions including uneven illumination,partial occlusions,and non-rigid transformation.The similarity measure between image patches is represented by Manhattan distance instead of Euclidean distance.The confidence map is constructed by sliding a search window to compute the best-buddies similarity between the template image and the matching window,which is then thresholded and averaged.Considering both the candidate matching region itself and the neighborhood points in its search window,the algorithm can effectively improve the robustness to non-uniform illumination,background clutter,image noise and rotation geometry deformation.(3)To improve the automation level of the vision-based measurement,a novel algorithm based on template matching and peaks of gradient histogram was proposed to extract the corners of diagonal markers automatically.First,the positions of candidate markers are obtained by a fast template matching method to search on the whole image.Then,the candidates are screened depending on the properties,which are two lines intersecting at the center of the marker and the intensity gradient of the marker having multiple peaks.Finally,the Gaussian fitting algorithm is applied to detect the corner at subpixel accuracy.Extensive experiments verify the proposed algorithm has the merits of high robustness,accuracy and generality,which shows it has high practical value in engineering.(4)To improve the performance of model estimation when the outlier ratio is high,a two-view geometry estimation algorithm using RANSAC with the locality preserving constraint was proposed.The algorithm integrates the locality preserving constraint into the universal RANSAC framework,which prunes most of the unreliable correspondences before the hypothesize-and-verify loop and guides non-uniform sampling to generate and verify promising models earlier.Unlike other guided sampling strategies,the proposed algorithm is simple to implement and does not require any prior information.Extensive experiments performed on the publicly available datasets reveal that the algorithm can achieve more accurate and stable solutions at much lower computational cost(in milliseconds on standard CPU)than state-of-the-art methods,particularly when handling problems with low inlier ratios.(5)To solve the problem of pose estimation for planar structures,a plane-based pose estimation algorithm by combing point and line features was proposed.The algorithm first estimates the homography associated with the model-to-image transform from point and line correspondences.Next,the estimated homography is substituted into the partial differential equation as an input,leading to an analytic solution to the rotation matrix.Then the translation vector can be trivially estimated in a linear least squares sense.Finally,a cost function for both point and line correspondences is built.The optimal pose parameters can be obtained by using a non-linear minimization method to minimize this function.The proposed algorithm has good versatility and scalability,which can handle the plane-based pose estimation problem by using point,line or a combination of them.Experimental results show that the proposed algorithm greatly improves the accuracy,robustness to the image noise and stability of pose estimation from planar scenes.(6)According to the requirements of practical projects,a vision-based pose measurement system for the lunar surface sampling manipulator was designed.The designed system relying on a monitoring camera and several fiducial markers will not affect the payload of the lunar probe,which can achieve high precision pose measurement.Ground simulation tests reveal that this system can adapt well to the harsh space imaging environment and efficiently and steadily output high precision pose parameters of the sampler with respect to the probe,which significantly improves the efficiency and success rate of the lunar sampling manipulator for key operations.The proposed pose measurement system based on visual guidance has been used in the Chang’e-5 mission successfully,which powerfully validates the efficiency and practicability of the system.