Research On Pose Estimation For Non-cooperative Targets Based On Multi-view Reconstruction

With the increasing demand for autonomous on-orbit servicing such as space debris cleaning and faulty spacecraft maintenance,visual measurement solutions for positioning and grabbing non-cooperative target have become the research emphasis of space robotics technology.Different from cooperative target,non-cooperative target usually does not have recognizable cooperative marks and the priori three-dimensional structure,which brings severe challenges to the stable identification and robust localization of spacecraft during the rendezvous.3d reconstruction technology can effectively restore the spatial structure of the target.With the improvement of its accuracy and efficiency,pose measurement schemes combined with multi-view reconstruction have been paid more and more attention by researchers.In this paper,a new non-cooperative target pose measurement scheme based on multi-view reconstruction is proposed on the basis of feature extraction,3d reconstruction and pose estimation.At the same time,considering that virtual reality technology can provide immersive perception and operation experience to users,this paper also explores the interactive mode of space robots and a 6-dof telexistence drone controlled by a head-mounted display is developed.The specific research contents are as follows:First,we propose the docking ring and apogee thruster center extraction algorithm based on fast ellipse detection and contour vertices of main body extraction algorithm based on rectangle detection.Different from the feature extraction methods based on key points detection,our method can quickly and robustly extract the typical component features of non-cooperative target and provide stable and effective image feature information for subsequent pose estimation.Secondly,we combine the sequential image feature tracking algorithm based on SIFT points,the incremental structure from motion algorithm and the multi-view stereo algorithm based on depth map fusion to reconstruct the target densely.Based on the detailed description of the multi-view reconstruction pipeline,a new feature matching algorithm combined with RANSAC and an improved depth map estimation algorithm are proposed.Experiments show that the multi-view reconstruction scheme used in this paper can effectively restore the dense point cloud of the spacecraft and provide accurate three-dimensional information of typical components for pose estimation.Based on the above research,this paper treats the non-cooperative target pose measurement as the perspective-n-point problem and solves it by an improved non-iterative pose estimation algorithm,which can not only deal with the cases that feature points are coplanar or non-coplanar,but also improve the computational efficiency and accuracy.A ground semi-physical test platform is built and a software simulation system for visual interaction based on Qt and OpenGL is developed,which further verifies the effectiveness of our pose measurement scheme.Finally,an improved rate-control method with an adaptive origin update strategy and an omni-directional six degree of freedom viewpoint control method are proposed,so that the user can control the drone more flexibly without any auxiliary equipment and obtain immersive perception experience.We build a real prototype which consists of data transmission module,real-time rendering module and motion mapping module.The user study in Unity simulation environment shows that our motion mapping method is superior to existing methods in terms of subjective and objective evaluation indicators and can provide an intuitive and natural interactive experience to users.