Pose estimation of non-cooperative spacecraft based on Convelutional Neural Network

On-orbit proximity operations such as rendezvous need to obtain high-precision pose information. However, the pose estimation of space target is difficult, because the space target is greatly affected by the illumination and the earth background. In order to solve the above problems, a Convolutional Neural Networks (CNN) based pose estimation method for known non-cooperative spacecraft is proposed. Three branches CNNs are designed to estimate orientation, position and spacecraft category respectively, and the loss function of three tasks is balanced by the method of automatic learning coefficient. Because CNN needs a lot of training data, the manual annotation method will bring huge workload. In this paper, we use 3D Max to generate spacecraft rendering data quickly without manual annotation. The experimental results show that the method can accurately predict the position and orientation of spacecraft, and classify spacecraft types at the same time.