Today’s space programs are ambitious and require increased level of onboard autonomy. Various sensing techniques and algorithms were developed over the years to achieve the same. However, vision-based sensing techniques have enabled higher level of autonomy in the navigation of space systems. The major advantage of vison-based sensing is its ability to offer high precision navigation. However, the traditional vision-based sensing techniques translate raw image into data which needs to be processed and can be used to control the spacecraft. The increasingly complex mission requirements motivate the use of vision-based techniques that use artificial intelligence with deep learning. Availability of sufficient onboard processing resources is a major challenge. Though space-based deployment of deep learning is in the experimental phase, but the space industry has already adopted AI on the ground systems.
Deep learning technique for spacecraft navigation in an unknown and unpredictable environment, like Lunar or Martian, is an area of research in space industry. Considering the distance from Earth, real-time ground control is impractical in such space missions. Velocity estimation of a descending spacecraft in Lunar environment is selected for the research work produced in this paper. Precisely estimating object's velocity is a vital component in the trajectory planning of space vehicles, such as landers, designed for descent onto Lunar or Martian terrains. In this paper, an effort is made to investigate the viability of velocity estimates by using images obtained from Lunar Reconnaissance Orbiter Cameras (LROC) that are part of a publicly available dataset released by Arizona State University (ASU) — the dataset contains minimal images. However, this imagery dataset is limited and not sufficient to train a deep learning model. Hence synthetic data is generated in this research. The study investigates usage of Condition-Generative Adversarial Networks(C-GAN) to generate synthetic data for image-based velocity estimates to support the proposed workflow. NVIDIA GPU is used to train the deep learning model. The forecast accuracy of the velocity of atmosphere-less objects is empirically examined, and the final results are reported.