2D vision-based tracking algorithm for general space non-cooperative objects

Space non-cooperative object visual tracking (SNCOVT) with monocular or binocular camera is one of the most significant tasks for on-orbit services. In recent years, most of vision-based trackers utilized in spacecraft are based on hand-crafted features (e.g. point, edge and color), which are vulnerable to the dynamic and harsh environment in space. Although various correlation filter and deep learning based generic object trackers with high performance have been proposed in tracking community. To verify the availability of those generic trackers in aerospace domain, we present a moderate and simulated space non-cooperative object visual tracking dataset, which contains 60 binocular video sequences with manual annotations, 54742 frames in total. Meanwhile, a new evaluation protocol that is more appropriate to the practical scenario and some novel metrics are adopted into our SNCOVT evaluation toolkit. We implement extensive monocular and binocular experiments with state-of-the-art trackers on our dataset. Finally, we make a comprehensive analysis to those algorithms and experiment results for future SNCOVT research.