3D spacecraft position and orientation estimation model based on monocular image

For spacecraft position and orientation estimation missions in mid-to-long-range space environments, there is a lack of ideal 3D sensors. Under these circumstances, although monocular cameras can hardly provide accurate depth information, it is widely applied. Some studies have explored monocular target pose estimation in general conditions. However, the degree of freedom of an on-orbit target that is not constrained by ground and gravity is more complicated. To extend this task to orbit target, this paper first estimated a huge spacecraft dataset with automatically labeled 3D poses, and then a deep neural network was designed following ’endto-end’ supervised learning. A modified Intersection over Union method for position-orientation combined evaluation is developed as well.