A unified framework for high precision and speed identification and tracking of space debris

Automatic observation has become the development trend of optical observation of space debris, and corresponding automatic target identification without human intervention has become an urgent research topic. This paper studies the real-time instance segmentation and tracking of space debris based on ground-based and space-based observation systems. We provide a unified, flexible and universal high precision and speed real-time target recognition and tracking framework. This framework improves the recognition speed of continuous image sequences from 5fps (frames per second) to 27fps on the premise of ensuring high precision instance segmentation and category recognition. Our contributions are threefold: (i) we added a fast loop correlation detection module Siam-Mask into the deep network framework of Mask R-CNN instance segmentation recognition, and we innovatively divided the time-domain tasks of different modules of the framework in different threads; (ii) we insert the CBAM module into each convolutional layer in the ResNet and FPN network to improve the recognition accuracy of small targets and information loss targets; (iii) we apply the singular value decomposition technique to convolution feature compression to reduce the computational and storage requirements of the model. Experimental results show that this methodology can effectively realize the real-time detection and tracking of multi-scale debris in space, and reduce the computing cost and storage space as much as possible. This technology will promote the maturity of "visual navigation based" (VBN) technology. In-orbit satellites will be able to realize on-board processing of navigation algorithms to achieve near-real-time mapping of the movement trajectory of non-cooperative objects in space, which will form a complete "identification and tracking - motion analysis - capture - off-orbit" method of space debris removal.