Classification and Recognition of Space Debris and Its Pose Estimation Based on Deep Learning of CNNs

Abstract

The increasing population of orbital debris is considered as a growing threat to space missions. During the recent decades, many enabling space debris capturing and removal methods were investigated. Thus, estimating automated recognition and on-board pose in an uncooperative target spacecraft by implementing using passive sensors such as monocular cameras is considered as a main task for the removal. However, these tasks are challenging since there is a semantic gap between visual features of target, as well as the lack of scalable, descriptive features and reliable visual features because of illumination conditions in space environment. For this purpose, Convolutional Neural Network was implemented based on transfer learning and data augmentation in order to conduct satellite classification and pose regression. Transfer learning method is performed by using popular pre-trained CNNs, which is limited to the small size of dataset in space environment. Then, just the last fully connected layers in the proposed structure were trained by the BUAA-SID dataset. In particular, augmenting synthetic BUAA-SID dataset was used with Keras Data Generator Tool through a number of random transformations like rotating, shifting, rescaling, and zooming for the purpose of enhancing the classification accuracy. In addition, the effects of un-centered and noisy images as well as different illumination conditions were analyzed by implementing different pre-trained networks. Based on the results, the present method could identify satellites and evaluate their poses against different space conditions effectively.