Unknown spacecraft detection during proximity operations using neural networks

Growing interest in space activities ranging from debris removal to in-space servicing engenders the need for robust, autonomous spacecraft proximity operations. An expanding body of work employs convolutional neural networks (CNNs) in relative navigation pipelines to pre-process data for higher level tasks, such as pose estimation and tracking. However, this generally requires that the spacecraft of interest is known a priori, such that the underlying CNN may be trained with adequate amounts of data. This renders most solutions incapable of handling unknown, partially damaged, or poorly characterized spacecraft, and thereby unable to extend to generalized applications. This thesis aims to bridge that gap with the proposed Spacecraft Localization Pipeline (SLP), which detects known and previously unseen spacecraft within a monocular image and classifies them appropriately. The design of the novel Spacecraft Localization Network, the workhorse behind the SLP, is presented with accompanying theory and implementation. Metrics for associating spacecraft in a time series are also proposed and analyzed to support tracking objectives. The SLP is tested against one known and several unknown spacecraft to show the efficacy of the pipeline and association metrics. Limitations are examined, accompanied by identification of fruitful areas for future work.