| With the increasingly frequent human exploration activities in space,the outer space of the Earth is closely occupied by various space targets such as failed satellites and micrometeoroids,so the processing of these space targets is the key to the success of subsequent space missions.With the development of deep learning discipline as well as computer vision technology,the application of artificial intelligence in spacecraft space target processing tasks is gradually becoming a hot research topic nowadays.In orbital space missions,space targets often need to be orbited and tracked,docked and repaired,and ablatively captured,etc.Therefore,correctly identifying space targets and obtaining their positional attitude and 3D models are the key elements of in-orbit services,and this paper investigates the recognition,positional estimation and 3D reconstruction of space targets based on deep learning technology for processing space target images.Firstly,this paper constructs a binocular vision system model and a spatial target dynamics model,deeply investigates the camera model and binocular calibration principle,and completes the calibration of the binocular camera in the vision system by using the Zhang Zhengyou calibration method.Secondly,this paper proposes a spatial target recognition and localization method based on YOLO algorithm,using Mosaic data enhancement and adaptive anchor frame to process input images to improve detection speed,constructing YOLOv5 s network,using GIOU＿Loss as loss function,and combining with binocular ranging principle to achieve accurate recognition and 3D spatial target in binocular vision simulation system.localization.Again,a visual SLAM-based target pose estimation method is designed for spatial targets containing chapter-motion spin.The method uses the relative motion principle to transform the motion relationship between the target and the camera,extracts ORB feature points from the target image and matches them,then uses the EPn P method to obtain the camera coordinates of each frame,then constructs ICP point cloud matching to solve the camera pose,and obtains a local map by iterative solution of the RANSAC model.Spatial circle fitting of the camera trajectory in the map can achieve the determination of the spatial target motion pose.Finally,this paper realizes the 3D reconstruction of the spatial target based on the motion recovery structure,constructs the scale space for the noisy images existing in the practical application and matches the SIFT feature points based on the kd-tree data structure,compares and filters the reasonable filtering algorithms,and finally completes the optimization of the 3D reconstruction of the spatial target. |
