Publication

Abstract : Satellite pose estimation for uncooperative spacecraft is gaining a lot of interest both by the research community and the space agencies mainly due to the amass of space debris and malfunctioning satellites present in Low Earth Orbit(LEO). These on-orbit proximity operations require accurate estimation of pose of the satellite against highly textured background and under varied lighting conditions. Earlier close-range estimation of pose for space borne applications has typically applied image processing on the basis of features that were handcrafted along with the previous information of the satellite’s pose. However, the major challenges come on our way to efficiently determine the pose of the target satellite, which lies in the ability to extract the significant features as well as to map them. To overcome the challenges posed by the use of manually defined features, CNN architectures are explored for monocular pose estimation of uncooperative space crafts. Therefore, this paper presents two components: first, a deep study on satellite pose estimation using DeepLearning and second an in-depth exploration of the ResNet model. The ResNet model is modified, dropout layer has been added, and the softmax layer is replaced with fully connected layer to estimate satellite pose. The input and output to this modified ResNet model are an image of type greyscale and spacecraft’s estimated pose, respectively. Here the output, an estimated pose of the spacecraft, is a vector of 7D representing position and orientation. The three-dimensional vector, which is nothing but x, y and z coordinates, represents spacecraft’s position of spacecraft, and the four-dimensional vector represents the orientation.