3D Reconstruction And Pose Estimation For Lunar Images

3D reconstruction and pose estimation are important tasks for visual navigation of lunar rovers.Due to the lack of atmospheric scattering,the lunar environment has a large variation in lighting,and there are many shadow areas.At the same time,there are many weak and repetitive textures on the lunar surface,which pose many difficulties for reliable 3D reconstruction of images at the same site.For the wide baseline image pairs collected between different sites,it further superimposes the characteristics of large-scale and large angle of view changes,making the process of image feature matching and pose estimation very difficult,and prone to excessive incorrect matching,directly leading to pose estimation failure.Existing methods require manual assistance to complete feature matching between the front and back stations.Therefore,conducting research on 3D reconstruction of images from the same site on the lunar surface and pose estimation techniques for images from different sites is of great significance for improving the autonomous navigation capability of lunar rovers.To solve the problem of poor 3D reconstruction effect of the same site on the moon surface,this paper proposes an improved scheme based on Markov random field and Gwc Net disparity estimation network.First,dense disparity map is generated from SIFT sparse disparity map using Markov random field,and then the dense disparity map was used as a pseudo truth value to supervise the training of the Gwc Net network and improve the disparity estimation accuracy of lunar images.At the same time,in order to further improve the 3D reconstruction accuracy of key areas such as meteorites,this article uses the Yolov5 object detection network to detect the key areas.Then,in the loss function design of the disparity network Gwc Net,the key areas are given higher weights,which improves the accuracy of disparity estimation in the key areas and provides favorable assistance for obstacle avoidance during lunar navigation.To address the difficulty of feature matching and pose estimation in wide baseline images of different lunar sites,this paper proposes a feature matching and pose estimation scheme based on synthetic views and ASpan Former self-attention network.Based on the obtained dense disparity map of the same site and prior pose information between sites,the image of the latter site is converted to the previous site to obtain a composite image.The composite view is used as a mediator for feature matching between the front and rear sites,and the matching point pairs are obtained using ASpan Former and projected back to the original image.Finally,based on the obtained matching pairs,an optimization algorithm is used to estimate the pose of different sites.This algorithm can effectively overcome the problem of weak lunar texture and high matching error rate in scenes with duplicate textures,and improve the robustness of the algorithm in complex scenes.