Research On Crater Identification Method For Autonomous Landing And Patrol Exploration On The Lunar Surfac

Advances in networked communications and artificial intelligence make autonomous spacecraft navigation particularly important in future deep space exploration missions.However,rough terrains and the complicated light condition lead to the relatively low signa l-to-noise ratios,poor textures,and inconspicuous landmarks.Conventional algorithms based on machine vision lack robustness in processing such data.Deep learning offers a viable solution to such problems,but is constrained by the lack of training data and has not been widely studied.China has used manually assisted positioning and navigation methods for both lunar and Mars explorations.The position and attitude of the spacecraft is determined by finding landmarks on the ground that are identical to the images taken by the spacecraft.However,there were two problems in previous missions.During landing missions,it took several hours to find the area within the landing ellipse captured by the landing camera.The low resolution of the satellite images made it difficult to identify landmarks in areas beyond the descent image,which made it difficult to navigate for long distance during rover missions.In this paper,Terrain Relative Navigation(TRN)with craters as landmarks is used to accomplish autonomous navigation of spacecraft.The following problems need to be solved:(1)crater detection learning without annotated training data;(2)place recognition of the camera captured area at the early stage of landing;(3)landmark tracking and matching of sequential images at the late stage of landing;(4)landmark matching in low resolution of satellite images of the lunar surface in long-range rover navigation.To address these issues,this paper presents a detailed study of autonomous spacecraft navigation in planetary exploration missions.For landmark detection,domain adaptation(DA)-based crater detection is proposed to address the lack of training datasets in the field of deep space exploration research.For the autonomous Precision Planetary Landing(PPL),Visual Place Recognition(VPR)is introduced for the first time to search for the initial camera area within the landing ellipse as the initial location for landing navigation.Meanwhile,the scale and rotation invariant crater matching operator is studied to continuously track and match the subsequent image landmarks.For autonomous rover navigation(Rover Navigation,RN),the rover’s panshot-based visual measurements are considered to automatically match the craters in the pan-shot terrain with the DOM navigation maps.Also,for the low resolution of the lunar surface satellite images,extreme small crater detection based on super-resolution is investigated to better support landmark matching as well as long-distance rover navigation.Finally,the proposed algorithm is tested practically with Chang’e-4 mission.The details of the study are as follows(1)DA-based unsupervised crater detection is investigated to effectively detect unlabeled real data with the help of automatically labeled simulated data samples only,solving the problem of lacking training data sets in the field of deep space exploration research.The feature gap between different domain data is reduced by domain alignment at image level and feature level.Better performance than classical supervised and unsupervised crater detection algorithms is achieved,and is close to the popular deep learning-based supervised object detection.The method provides the basis for the subsequent research in this paper.(2)In the autonomous PPL study,due to the negative impact of sparsity,repetitiveness and variability of lunar surface textures on traditional visual feature point extraction and matching,craters are used as the main landmarks to improve the robustness of the algorithm in matching data from different domains.In the early stage of landing,VPR is introduced to identify the region of the ’first landing image’,which uses the crater heat map as an input to provide the initial position and attitude for subsequent visual navigation.In addition,a crater matching operator is proposed based on the distribution relationship between the central crater and its surroundings,which is designed in terms of scale invariance,rotation invariance,and operator description to improve the robustness of crater matching to position disturbance,size disturbance,and false positive/negativity rate.The proposed method effectively solves the problem of difficult inter-image matching under variable lighting conditions and improves the matching efficiency between landing images and TRN navigation maps.(3)A super-resolution based on real degraded lunar surface images is investigated to provide more detailed reference base maps and matching landmarks for autonomous RN.The method reconstructs more potential small craters compared to artificial degradation.It is shown that the correct estimation of the degenerate kernels and noise of real degenerates can effectively improve the detection results.Meanwhile,two SR composite model frameworks driven by the crater detection task are proposed to improve the performance of very small crater detection by considering the semantic information related to crater detection.(4)Taking Chang’e-4 autonomous landing and ’Yutu-2’ autonomous roving missions as examples,the autonomous PPL and RN strategies proposed in this paper are experimentally validated.Among them,the autonomous PPL can be divided into two stages: at the early stage of landing,VPR identifies the initial shooting location of the landing camera.For subsequent images,visual navigation delivery such as sequential inter-image matching and ground crater tracking is used until the landing site.The recovered landing trajectory and reprojection errors demonstrate the feasibility of the strategy.In the autonomous RN of ’Yutu II’,the rover’s pan-shot-based visual measurements are considered to automatically match the craters in the pan-shot terrain with the super-resolution reconstructed DOM navigation map.The results show that the proposed method increases the number of crater landmarks in TRN and effectively improves the positioning accuracy.