Space target docking ring recognition and center point positioning based on Tiny Darknet YOLOv3 fusion CenterNet

Aiming at the visual measurement requirements of space manipulators for grasping non-cooperative targets, a space target docking ring recognition and center point positioning method based on Tiny Darknet YOLOv3 fusion CenterNet is proposed. First, training network model based on open source ImageNet VOC 2007 and self-built spatial non-cooperative target data set and use optimized Tiny Darknet YOLOv3 fusion CenterNet deep learning algorithm to identify space target docking and obtain two-dimensional pixel coordinates of the docking center point; secondly, using the EnsensoN10-408-18 depth camera to obtain the 3*3 neighborhood data of the depth value corresponding to the center point and calculate the neighborhood weighted optimal value to get docking center spatial coordinates in the camera coordinate system. Combined with the hand-eye calibration relationship, the docking center spatial coordinates are converted to the UR5 manipulator base coordinate system. A ground verification system for manipulator to capture the target was built to test the target docking ring identification and center point positioning, and the accuracy error evaluation is completed based on the OptiTrack motion capture global measurement benchmark system. The experimental results show that target positioning accuracy is better than 10mm, and real-time data update rate is better than 2Hz in the dynamic approximation process from 1.5m to 0.2m, which can effectively solve the slow speed and poor accuracy caused by the influence of environmental lighting, target surface material, target attitude scale changes and other factors in traditional feature extraction methods. It lays a foundation for the safe arrival, capture and other manipulator fine operations.