Seeing Stars: Learned Star Localization for Narrow-Field Astrometry

Violet Felt, Justin FletcherWACV 2024

TL;DR

This paper tackles the challenge of accurately identifying stars in narrow-field, high-speed satellite tracking imagery, a crucial step for precise satellite tracking. They introduce "StarNet," a novel dataset of real-world imagery and ground truth, and demonstrate that learned deep learning models significantly outperform traditional methods in star detection, leading to more reliable astrometric fits.

What Problem Does This Paper Address?

Accurate tracking of satellites is becoming increasingly vital for space safety and operations, especially with the proliferation of large satellite constellations. A fundamental step in this tracking process is astrometry, which involves precisely determining the position of objects in the sky. This typically relies on identifying stars within the telescope's field of view and matching them to known star catalogs. However, traditional astronomical star detection algorithms struggle with the unique characteristics of satellite tracking imagery: narrow fields of view and "rate-track" observational modes, where the telescope moves rapidly to follow a satellite. These conditions can cause stars to appear as streaks rather than points, leading to reduced precision and recall in star detection, ultimately causing astrometric fitting failures and hindering satellite tracking. This paper aims to bridge this gap by developing and evaluating deep learning approaches for robust star detection in this challenging domain.

Methods

The authors propose a learned approach to star detection, framing it as a task of localizing an unknown number of stars, which can manifest as streaks of varying lengths and orientations. They explore three main deep learning paradigms:

  • Object Detection: Models like Deformable DETR, Faster RCNN, and YOLOX are used to predict bounding boxes around stars.
  • Instance Segmentation: Models like HTC, Mask RCNN, and QueryInst predict pixel-wise masks for each star.
  • Line Segment Detection: Models like LETR are employed to detect stars as line segments, directly addressing the streak nature of stars in rate-track imagery.

These learned models are designed to work directly with raw image pixels, eliminating the need for traditional pre-processing steps like dark and flat-field corrections. They are trained on their newly created StarNet dataset and evaluated using object-wise, pixel-wise, and astrometric fit rate metrics.

Datasets

The core contribution in terms of data is the StarNet dataset. This dataset consists of:

  • 64,668 real-world, narrow-field, rate-tracked satellite tracking images.
  • Corresponding ground truth annotations for approximately 4.5 million stars, derived from a combination of classical star detection methods, an astrometric engine (astrometry.net), and the SSTRC7 star catalog.

This dataset is notable for its use of real, rate-tracked imagery, which is a significant departure from many existing datasets that often rely on simulated or wide-field-of-view sidereal imagery.

Key Insights and Ideas

  • Learned models excel in rate-track imagery: Deep learning models, particularly those designed for object detection and line segment detection, significantly outperform traditional methods like SExtractor and AstroGraph in detecting star streaks characteristic of rate-track imagery.
  • Astrometric fit rate is strongly correlated with star recall: The ability to achieve a successful astrometric fit is more dependent on detecting a sufficient number of stars (recall) than on the absolute precision of individual detections. This is because astrometric engines require a minimum number of stars to establish a celestial reference.
  • Line segment detection is well-suited for star streaks: The LETR model, designed for line segment detection, shows competitive performance, particularly in recall, suggesting that explicitly modeling stars as lines is beneficial for rate-track data.
  • No metadata required: A key advantage of the proposed learned approach is its ability to perform star localization and contribute to astrometric fitting using only raw image pixels, bypassing the need for traditional pre-processing and metadata.

Key Results

  • Doubled astrometric fit rates: The proposed learned star detection methods achieve astrometric fit rates more than double that of classical star detection algorithms.
  • Improved dim star recall: The models demonstrate better recall for fainter stars compared to baseline methods.
  • Comparable star localization residuals: The learned models achieve star localization residuals comparable to or better than baseline methods.
  • High precision and recall: Deformable DETR achieves a strong balance of precision (0.89) and recall (0.70) with a high F1 score (0.78) in object-based metrics. HTC also performs well with high precision (0.90) and a good F1 score (0.71). LETR shows excellent recall (0.75) and a high F1 score (0.77).
  • Low false fit rates: The deep learning models achieve very low false astrometric fit rates (0.00), indicating high reliability.

Limitations and Future Work

  • Computational Complexity: While inference time is negligible for real-time applications, training the Deformable DETR and LETR models is computationally intensive due to the large number of queries.
  • Model Scaling: Further exploration of query scaling in models like Deformable DETR and LETR might improve recall in very dense star fields, but this needs to be balanced against model parameter growth.
  • Segmentation Mask Limitations: For instance segmentation models like HTC, the small pixel footprint of point source stars (0.02% of an image) can hinder performance, even with padding.
  • Future Work: The authors suggest exploring end-to-end learned astrometric processing, increasing recall in dense star fields, improving localization precision to sub-pixel levels, and potentially developing learned astrometric engines to achieve O(1) complexity for the entire pipeline.

Summary updated

BibTeX
@inproceedings{felt2024seeing,
  author = {Violet Felt and Justin Fletcher},
  title = {{Seeing Stars: Learned Star Localization for Narrow-Field Astrometry}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure