Research On Methods Of Critical Load Identification And Relative Pose Estimation Of Non-cooperative Object

With the development of space exploration,space defense and other fields,observation of target spacecraft from service spacecraft and other corresponding on-orbit service have drawn more attention on a domestic scale.Still,identification of critical loads carried by non-cooperative object as well as relative pose estimation of noncooperative object can be a key foundation of target navigation and take over.Thus,in this paper,methods of non-cooperative target critical loads identification and relative pose estimation has been studied,with multilayer fully connected neural network(MLP)based target loads identification frame been established,and traditional ICP method has been ameliorated in terms of mismatches and locally optimal solutions.In response to mission of neural network based image classification,state-of-the-art frames have been investigated and discussed,with MLP and convolutional neural network(CNN)been lucubrated,and analyzed network structure as well as parameter configuration according to the aim of image classification.Then,an MLP has been established,which was further modified,trained and tested with an existing dataset.Taking non-cooperative object critical loads classification into consideration,a method of image size compression and gray processing has been proposed,along with a procedure of dataset establishment using images of non-cooperative object critical loads.Furthermore,an MLP for non-cooperative object critical loads classification has been set up.Aiming at the task on non-cooperative object relative posse estimation,point cloud of its outer contour should be obtained as a prerequisite.A model of one satellite was built with Solidworks,which then got converted into a URDF file for its usage in Gazebo while obtaining its outer contour point cloud using simulation point cloud sensor.Both point cloud of its outer contour and point cloud observation data were obtained in Gazebo using model satellite and the point cloud sensor.In terms of the task on non-cooperative object relative posse estimation,point cloud ICP algorithm has been testified using point cloud data,after which the ICP alignment results were analyzed and the circumstances under which mismatches and locally optimal solutions will happen were investigated.Taking the high symmetry of the outer contour of non-cooperative object into consideration,along with algorithm principle of ICP,a novel procedure of initial pose acquisition for ICP using point cloud segmentation and point cloud keypoints has been proposed.Then,LCCP point cloud segmentation method as well as Harris3 D keypoint was studied as prerequisite algorithms,and the proposed procedure was carried out using the satellite point cloud model.Furthermore,other ICP pretreatments were selected as comparisons for controlled trails carried out using point cloud data for relative pose estimation,and the experiment results showed that the proposed method of initial pose acquisition for ICP can solve the problem of mismatches or locally optimal solutions brought by high symmetry of the outer contour of noncooperative object to a certain extent.