Identification, Tracking, and Pose Estimation of Non-cooperative Satellites Using Deep Learning Techniques

The identification, servicing, and maintenance of partially damaged satellites is a crucial field in space situational awareness. Failure of one satellite can affect the constellation’s ability to perform the intended operation, requiring identification, trajectory tracking, and pose estimation of the damaged satellite. Standard models of orbital motion have been found to be deficient in analysing the unpredictable orbital motion of these uncooperative satellites. To capture this realistic behavior, authors propose the development of CNN models, a class of neural networks commonly used for image recognition, which lowers the high dimensionality of images. The work uses this idea to space object pose estimation for effective docking in rendezvous missions. The focus is on training deep learning models on the SPEED and SPEED+ datasets, improving their accuracy by combining it with advanced algorithms. The algorithms developed are the first step in implementing an actual deployment of a service satellite.