Why This Matters
Geostationary Earth Orbit (GEO) is a highly valuable but increasingly crowded strip of space. Defunct satellites and discarded rocket debris pose catastrophic collision risks to active telecommunications infrastructure. Tracking these objects using low-cost optical telescopes is highly scalable but introduces extreme imaging difficulties.
In typical telescope images, GEO targets occupy only a few pixels, lacking texture, size, or shape. They are easily lost behind bright background stars, atmospheric distortion, and sensor noise. Traditional single-frame detectors fail when the signal-to-noise ratio (SNR) drops, leading to broken trajectories and false alarms.
The Big Idea
Instead of relying solely on massive, computationally heavy deep learning models to predict targets directly, this paper proposes a hybrid two-stage framework. First, a lightweight neural network (WTNet) enhances high-frequency features (edges, points) using 2D Wavelet Transforms to generate initial single-frame candidates. Second, a highly structured, physics-informed sequential trajectory completer uses the Hungarian matching algorithm to link detections across frames, filter out random noise, and reconstruct missing positions during stellar occlusions.
Single-Frame Wavelet Feature Enhancement
WTNet decomposes the input feature maps into low-frequency ($LL$) and high-frequency ($LH, HL, HH$) components using 2D discrete wavelet transforms. The low-frequency components preserve general object contours, while high-frequency components undergo convolutional enhancement to make faint edges stand out against background noise before reconstruction.
Interactive Wavelet Feature Decomposition Simulator
Simulate how wavelet decomposition separates low-frequency backgrounds (stars) from high-frequency targets (satellites) and suppresses noise.
Optimal Cross-Frame Matching & Trajectory Completion
When a target is temporarily lost (e.g., due to stellar occlusion), single-frame detectors fail. The paper solves this by constructing a cost matrix $C \in \mathbb{R}^{m \times n}$ between consecutive frames $f_1$ and $f_2$, where each element is the Euclidean distance:
The Hungarian algorithm solves this matrix to find the globally optimal one-to-one matches. If a frame gap is detected (e.g., a target was missing in an intermediate frame), linear interpolation fills the gap:
Hungarian Matching & Interpolation Sandbox
Drag the points in Frame 1 (blue) or Frame 2 (green) to see how the optimal matching is computed and how the missing intermediate frame (dashed orange) is interpolated.
| Dist (px) | Green 1 | Green 2 | Green 3 |
|---|
Temporal-Consistency-Based Noise Filtering
Single-frame false positives are usually random and isolated, whereas real satellites follow continuous paths. To prune noise, the authors evaluate a Temporal Support Degree. For each detection $p_f$ in frame $f$, a window of radius $w$ is evaluated:
where $S$ is the number of frames within the window containing at least one detection within a spatial threshold distance $\tau$. If the Support Ratio is below a set threshold, the point is classified as noise and discarded.
Temporal Support & Noise Filtering Simulator
Adjust the sliders to see how temporal consistency filters out random noise (red) while preserving the real trajectory (green).
Results & Benchmarks
Evaluated on the public SpotGEO dataset (released by the ESA), the proposed framework demonstrates a major performance leap. Single-frame WTNet outperforms other baselines, but the addition of the multi-frame trajectory completion module pushes the F1 score to its peak of 90.14%.
Ablation Study: F1 Score Comparison
Comparing different stages of the proposed pipeline against baseline methods.
Limitations & Open Questions
While the proposed sequential trajectory completer shows dramatic F1 score improvements, some challenges remain open:
- Extreme Occlusion Gaps: The interpolation-based completion relies heavily on the frame gap being less than 3 frames. For longer occlusions, the algorithm must fall back to aggressive motion vector extrapolation, which accumulates errors over time if the target maneuvers.
- Computational Complexity: Although the Hungarian algorithm is highly optimized, scaling it to dense star fields with thousands of potential candidate detections per frame can increase processing latency.
- Sensor Dependencies: The adaptive thresholding mechanism relies on online statistics, which may require calibration when switching between different telescope apertures or atmospheric conditions.
Glossary
Hungarian Algorithm
A combinatorial optimization algorithm that solves the assignment problem in polynomial time ($O(n^3)$). In this paper, it is used to uniquely and optimally pair object detections between consecutive frames to form coherent trajectories.
Wavelet Transform
A mathematical tool that decomposes a signal or image into different frequency sub-bands. Unlike Fourier transforms, wavelets preserve spatial localization, making them ideal for isolating tiny, point-like high-frequency targets from smooth low-frequency backgrounds.
SpotGEO Dataset
A public benchmark dataset co-released by the European Space Agency (ESA) and the University of Adelaide, specifically designed for evaluating computer vision algorithms tasked with detecting geostationary satellites in low-cost telescope sequences.
How to Cite
@article{lan2025space,
title={Space Object Detection using Multi-frame Temporal Trajectory Completion Method},
author={Lan, Xiaoqing and Xin, Biqiao and Wang, Bingshu and Zhang, Han and Zhu, Rui and Zhang, Laixian and Lan, Xiaoqing},
journal={ChinaXiv Preprint},
volume={chinaxiv-202510.00105},
year={2025}
}