Space Non-cooperative Target Visual Navigation Fused with Object Detection under Complex Backgrounds

Relative visual navigation of the non-cooperative target is the key technology in the on-orbit service. This task becomes especially challenging when the background is complex, for instance, the Earth appears as the background in the on-orbit image in low-orbit service tasks. Traditional visual navigation algorithms for the non-cooperative target may fail because the inconsistency between the target's motivation and the background's motivation may result in an inaccurate estimation of the target's pose. To solve the above problem, this paper proposes a framework for space non-cooperative target visual navigation fused with object detection. Firstly, the bounding box of non-cooperative target detection is defined by YOLOv5. Secondly, in consideration of unstable bounding box estimation and constraints for on-orbit computing sources, a bounding box refinement strategy is designed for the sake of accurate navigation results. Finally, the refined bounding box and ORB features are fused together by a scheduling system based on Redis for the estimation of non-cooperative target poses regardless of complex backgrounds. The experimental results show that, compared to initial ORB-SLAM3, the space non-cooperative target visual navigation fusing with object detection can estimate target poses under complex backgrounds in most cases and can achieve the navigation accuracy of 0.013m.