Thijs Verhaeghe, Royal Military Academy; Ben Lauwens, Royal Military Academy; Bart De Clerck, Royal Military Academy; Dirk Vandepitte, Ku Leuven; Bart Vandenbussche, Ku Leuven; Tjorven Delabie, arcsec Space
https://doi.org/10.64861/GMBQ6980
Keywords: Space Domain Awareness (SDA), Star Trackers, Space Debris, Onboard Processing, Deep Learning, STar-DETR, Poisson Labeled Multi-Bernoulli (PLMB), Multi-Object Tracking, Random Finite Sets (RFS), Edge Intelligence, Low-SNR Detection, Space-Based Optical
Abstract:
Given the increasing amount of space debris in Earth’s orbit, comprehensive and precise space domain awareness (SDA) is required to ensure the safety of space operations. One cornerstone in achieving such a recognised space picture is the ability to persistently detect and track resident space objects (RSOs). While vast networks of ground-based radars and optical telescopes monitor the space environment, coverage gaps remain due to weather, day-night cycles, and geographical constraints. To address these limitations, there is a growing interest in leveraging space-based sensors to enhance SDA capabilities. Given their scalability, cost-effectiveness, and potential for global coverage, star trackers have emerged as a promising solution for space-based debris detection, as demonstrated by the arcsec DeDUST project and Starlink’s Stargaze initiative.
Real-time detection and tracking of low-SNR objects in star tracker images remains a computational challenge for resource-constrained platforms. In this work, we present ASTRA-Net: Autonomous Space Target Recognition and Association Network, a hybrid onboard architecture that tightly couples a deep learning detector with a Bayesian multi-target tracking filter operating directly in the detector plane.
Using a high-fidelity, physically modelled image generator, we generate a comprehensive dataset of short-exposure images while retaining exact ground-truth positions, magnitudes and trajectories of all stars and resident space objects (RSOs) in the field of view of the star tracker. The simulator models the complete photometric and electronic signal chain, including point spread function rendering, vignetting, and sensor noise sources such as shot noise, dark current and read noise, and accounts for motion blur induced by both spacecraft attitude dynamics and relative target motion.
For detection, we employ a U-Net CNN architecture that ingests a temporal sequence of frame differences together with a synchronised stellar distance map derived from the onboard star catalogue. This spatial prior biases the network’s attention mechanism toward regions away from known stars, allowing the detector to prioritise unknown, potentially uncatalogued objects while remaining robust under attitude motion. A dedicated loss formulation addresses the severe class imbalance between background, star and RSO pixels. The resulting per-detection confidence is propagated directly into the tracking stage as a measurement-noise weighting term, rather than being discarded through a hard detection threshold.
For tracking, we adopt a two-layer Poisson Labelled Multi-Bernoulli (PLMB) filter architecture. This Random Finite Set (RFS) based approach provides a mathematically rigorous framework for multi-object detection, tracking and state estimation under missed detections, clutter and an unknown, time-varying number of targets, while avoiding the combinatorial cost of explicit data association. The first, onboard layer operates directly in the two-dimensional detector plane using a computationally lightweight linear Kalman formulation, producing short, high-confidence tracklets suitable for low-bandwidth downlink. A second, ground- or cloud-based layer subsequently fuses these tracklets into full six-dimensional orbit solutions.
The complete pipeline was validated end-to-end on synthetic star tracker sequences generated for a mixed-inclination Low Earth Orbit (LEO) observer fleet, covering velocity, anti-velocity and zenith pointing geometries to exercise a range of relative angular rates and background conditions. The detector achieves a recall of 0.982 and an F1-score of 0.990, together with a sub-pixel centroid accuracy of 0.201 pixels (9.13 arcseconds). The onboard PLMB tracking layer forms tracklets with a mean purity of 0.981, no false tracks, a Multi-Object Tracking Accuracy (MOTA) of 0.976, and a Multi-Object Tracking Precision (MOTP) of 0.148 pixels (6.70 arcseconds). Tracklets are confirmed from as few as two detections, enabling rapid track initiation without compromising the low false-track rate observed across the full dataset.
These results demonstrate that tightly coupling a confidence-aware deep learning detector with an RFS-based tracking filter yields high-purity, low-latency tracklet formation suitable for onboard, resource-constrained platforms, without sacrificing the detection sensitivity required for faint, low-SNR debris signatures. By replacing the conventional hard-thresholded interface between detection and tracking with a continuous, probabilistic handoff, ASTRA-Net reduces data association ambiguity and clutter before it reaches the off-board tracking stage, while remaining compatible with the onboard processing and downlink constraints of operational star tracker platforms. This approach significantly reduces the data downlink requirements for SDA constellations relative to raw image or full-track downlink, enabling persistent and autonomous monitoring of the space environment and providing a scalable path toward a more comprehensive and accurate recognised space picture.
Date of Conference: September 15-18, 2026
Track: Space-Based Assets