Instance Segmentation for Feature Recognition on Noncooperative Resident Space Objects

Active debris removal and unmanned on-orbit servicing missions have gained interest in the last few years, along with possibility to perform them through use of an autonomous chasing spacecraft. In this work, new resources are proposed aid implementation guidance, navigation, control algorithms for satellites devoted inspection noncooperative targets before any proximity operation is initiated. particular, convolutional neural networks (CNN) performing object detection instance segmentation proposed, its effectiveness recognizing components parts target satellite evaluated. Yet, no reliable training images dataset kind exists date. A tailored publicly available software has been developed overcome limitation by generating synthetic images. Computer-aided design models existing loaded on a three-dimensional animation used programmatically render objects from different points view lighting conditions, together necessary ground truth labels masks each image. The results show how relatively low number iterations sufficient CNN trained such datasets reach mean average precision value line state-of-the-art performances achieved common datasets. An assessment performance network when conditions provided. To conclude, method tested real Mission Extension Vehicle-1 mission, showing that using only artificially generated train model does not compromise learning process.