Experimental validation of a Convolutional Neural Network for proximity navigation between uncooperative satellites

Abstract

Convolutional Neural Networks (CNNs) are a popular deep learning architecture that has been successfully applied to various computer vision tasks. In the field of satellite relative operations, CNNs are an effective method for detecting and classifying an uncooperative target spacecraft in images acquired by a chaser satellite that has to ensure the safety of the satellites when flying in close proximity. In this paper, we propose and validate through experimental tests the first part of a pipeline based on computer vision algorithms for proximity navigation between uncooperative satellites. Specifically, the computer vision algorithms employed are the state-of-the-art CNN called You Only Look Once version 7 tiny (YOLOv7-tiny), used to detect the target satellite and reduce the search field of relevant features on its surface, and the feature detector called Oriented FAST and Rotated BRIEF (ORB). The validation of the measurement system and the computer vision algorithms is carried out using a representative laboratory facility, paying particular attention to computing time and performance metrics of the image analysis algorithms devoted to object detection and feature detection.