Research On The Key Technologies Of Space Debris 3D Reconstruction

The acquisition of 3D structure of space objects has become more and more important as there are increasing number of spacecrafts swarmed into the space in recent years.Accurate 3D information provides prior for space services such as spaces objects cataloguing,space debris catching.etc.During the process of space object image acquisition,both the hardware limitation and complicated imaging condition reduce the final image quality and dataset size,which brings challenges to space object 3D reconstruction.In this thesis,we focus on the 3D reconstruction on space object using both geometry and deep learning methods,and propose new methods and algorithms.The main contributions are listed as follows:(1)To address the problem of too many false correspondences in space object image pairs,we propose an adaptively ranked sample consensus algorithm.The quality of feature points are frequently evaluated and updated as a guidance for non-uniform sampling,which boost both the efficiency and accuracy in epipolar geometry estimation.At the mean time,we use geometric constraint to alleviate the degeneration during fundamental matrix estimation to improve the estimation accuracy.(2)In order to strike a balance between efficiency and accuracy in Sf M-based space debris3 D reconstruction,we propose an algorithm which combines the strength of both global and incremental Sf M paradigm.We design a loop-consistency based subgraph extraction algorithm to extract the robust part of the whole scene graph,and conduct global estimation on the robust graph.For the remaining images,we propose a local incremental algorithm which improves local reconstruction accuracy.Experiments show that the reconstruction accuracy is improved with comparatively less time consumption.(3)To solve the problem of lacking in data for space object reconstruction,we design a deep learning based 3D reconstruction method.We use a RNN-based network to estimate the 3D structure with a single frame depthmap as input.Given that the spatial relationship of different components of the spacecraft follows some specific pattern,we represent the 3D structure using 3D geometric primitives(3D cuboids).The method achieves space object 3D reconstruction with only a single frame input depthmap.(4)To study how the space object image degeneration influence the final 3D reconstruction result,we add different types of degenerations to the input data of both Sf M-based methods and learning based method,then analysis the variation tendency of every method’s result.The experiments provide reference for data requirements of space object 3D reconstruction.