Unknown Low-Earth-Orbital Satellite Detection by Learning-Based Methods

With the development of convolutional neural networks, the application of learning-based methods in aerospace attracts much attention. Compared with sufficient datasets in the generic scene, most of the datasets in aerospace lack diversity in model structure, which leads to the low generalization ability in on-orbital tasks, such as satellite detection. In this article, this drawback is relieved by a synthetic dataset whose structure diversity is extended by 21 models in Yale-CMU-Berkely Object and Model set (YCB dataset) and two realistic satellites. This dataset is rendered in a low-earth-orbit scene and is utilized in target detection tasks. Experiment shows unknown satellites can be detected in our synthetic images. The generalization is also evaluated by combining our dataset with other public satellite datasets. The aforementioned experiments are implemented by YOLOv5s and YOLOv5m, we also propose a light-weighted refinement pipeline that improves the performance for an average of 2.57% in the F1-score in the generalization experiments.

Automated summary

With the development of convolutional neural networks, learning-based methods in aerospace are gaining attention. The lack of diversity in model structure of aerospace datasets hampers generalization ability in on-orbital tasks, such as satellite detection. This article addresses this drawback by creating a synthetic dataset with extended structure diversity from the YCB dataset and realistic satellites, enabling the detection of unknown satellites in synthetic images. The generalization is evaluated by combining this dataset with other public satellite datasets, and a light-weighted refinement pipeline is proposed to improve performance.