Research On Motion And Structrue Estimation Methods Of Non-cooperative Targets Based On Stereo Vision

With the advancement of human space exploration, it's essential to clear space junk and strengthen space management. As a powerful tool to deal with non-cooperative targets, space robotics technology has gotten scholars' attentions. However, unknowns and uncertainty of non-cooperative targets pose a great challenge to orbit service of space robots. In order to capture the target accurately, certain measurements must be used to obtain 3D information and motion information of targets in real time. Measurement technique based on vision has become a research hotspot in the field of space robotics.Firstly, the imaging principle of camera is analyzed and the mathematical model of stereo vision system is built to illustrate acquisition methods for 3D coordinates of spatial points. Then calibration of binocular vision system is completed and image preprocessing techniques including image correction and target segmentation are studied.Secondly, problems of feature extraction and matching are studied. For left and right frames, a matching strategy combining rough and fine matching methods is adopted. In rough matching stage, KNN algorithm is used to get the initial feature matching. In fine matching stage, RANSAC algorithm is used to optimize matching results. By comparing matching results of commonly used feature extraction algorithms, ORB is chosen. For front and rear frames, a method based on ORB feature optical flow is adopted to simplify the matching process.Next, EKF is applied to estimate motion and structure of targets. First motion model, structure model and observation model of targets are built. Then a serial estimation method based on Tri-EKF is proposed to estimate the rotation, translation, and structure of targets aiming at solving coupling estimation problems of traditional Joint EKF and Dual EKF. By comparing three algorithms through simulations, it is illustrated Tri-EKF which has used the decoupled estimation method has faster convergence speed, higher estimation precision and better stability for translational motion estimation.Finally, an experimental platform based on a 2-DOF target motion simulation platform and binocular vision system is established to verify the algorithm proposed. Through the experimental study of different motion forms, feasibility and correctness of the algorithm are verified and motion and structure estimation of targets can be achieved.