Star detection for low-altitude atmospheric star trackers

Celestial navigation provides an alternative to global navigation satellite systems (GNSSs) for uncrewed aerial vehicles (UAVs), particularly in environments where GNSS signals are denied. However, traditional star detection methods can be ineffective for low-altitude atmospheric flight due to factors such as motion blur, clouds, and light sources. This paper addresses the challenge of robust star detection under these highly dynamic and noisy conditions. We propose a novel, to our knowledge, two-stage deep learning architecture that fundamentally departs from traditional single-frame spatial thresholding. The first stage utilizes a UNet convolutional neural network for robust star segmentation. The second stage introduces a novel multi-task, inter-frame keypoint detector that processes sequential regions of interest. By explicitly mapping the geometric bisection across the temporal boundary of two consecutive frames, this architecture resolves the directional ambiguity and endpoint fading that impede traditional algorithms in dynamic environments. This inter-frame point is highly salient, allowing for precise temporal alignment between the star sensor and other systems. The models were trained and tested on a large, synthetic dataset generated with a high-fidelity simulator, which includes augmented imagery to replicate motion blur, cloud cover, and lens flare. We compare the performance of our UNet star segmentation model against three baseline adaptive thresholding techniques: adaptive Gaussian, Niblack's, and Bernsen's. The results, evaluated using precision, recall, and F1 scores for star detection, demonstrate that the UNet method consistently outperforms the other techniques across all viewing conditions, achieving an F1 score of 0.85 in clear conditions compared to a maximum of 0.40 for the baseline methods. Furthermore, the keypoint detector demonstrated improved accuracy over standard methods, achieving a mean absolute error of 0.51 pixels under dynamic conditions. Both the segmentation model and the keypoint detector were tested on real data, confirming the validity of the synthetic data and demonstrating the ability of both models to generalize. A star identification success rate of 85.9% was achieved across 1000 frames of dynamic flight data, representing a 23.3% reliability improvement over the baseline peak extraction method. This study's primary contribution is a robust star segmentation model and keypoint detector that form the foundation for a reliable strapdown star tracker for use in challenging low-altitude scenarios. Unlike traditional satellite star trackers that rely on high-precision gyroscopes for motion compensation, our method obviates the need for angular rate sensors altogether, making it highly suitable for size, weight, and power constrained UAV platforms.