Abstract
Space objects in Geostationary Earth Orbit (GEO) present significant detection challenges in optical imaging due to weak signals, complex stellar backgrounds, and environmental interference. In this paper, we enhance high-frequency features of GEO targets while suppressing background noise at the single-frame level through wavelet transform. Building on this, we propose a multi-frame temporal trajectory completion scheme centered on the Hungarian algorithm for globally optimal cross-frame matching. To effectively mitigate missing and false detections, a series of key steps including temporal matching and interpolation completion, temporal-consistency-based noise filtering, and progressive trajectory refinement are designed in the post-processing pipeline. Experimental results on the public SpotGEO dataset demonstrate the effectiveness of the proposed method, achieving an $F_1$ score of 90.14%.
Introduction
Detecting objects in Geostationary Earth Orbit (GEO)—such as defunct satellites and space debris—is a classic "needle in a haystack" problem. Due to the extreme distances involved, these objects appear in optical telescope images as tiny, faint clusters of pixels, often lacking discernible shape or texture.
This task is further complicated by "stellar interference" from bright background stars, atmospheric refraction, and inherent sensor noise. Standard computer vision models often struggle in these low signal-to-noise ratio (SNR) environments, resulting in high false-positive rates or missing targets entirely when they are momentarily occluded or too faint to trigger a single-frame detection. This paper addresses these challenges by leveraging both frequency-domain feature enhancement and temporal consistency.
Methodology
The proposed framework consists of a two-stage pipeline: initial single-frame detection followed by a robust multi-frame trajectory completion module.
WTNet: Wavelet Transform-based Network
For the first stage, we developed WTNet. Unlike traditional CNNs that treat all pixel data equally, WTNet utilizes a Discrete Wavelet Transform (DWT) to decompose images into four sub-bands: one low-frequency (LL) component representing the background and contours, and three high-frequency (LH, HL, HH) components representing edges and potential targets.
By applying convolutions specifically to these sub-bands, the model enhances high-frequency features while suppressing low-frequency noise. After processing, an Inverse Wavelet Transform (IWT) reconstructs the feature map. This approach is particularly effective for space objects where the signal is incredibly weak compared to the background. Additionally, a Convolutional Block Attention Module (CBAM) is integrated to further refine the spatial and channel-wise features.
Multi-frame Temporal Trajectory Completion
The second stage leverages the predictable linear motion of GEO objects across frames. The core of this stage is the Hungarian algorithm, which provides a mathematically optimal way to link detections across consecutive frames into trajectories. The post-processing pipeline includes three critical steps:
- Interpolation: If an object is detected in frames $t-1$ and $t+1$ but missed in frame $t$, the system uses linear interpolation to "fill in the blanks," recovering targets that were too faint for the neural network to detect in a single frame.
- Temporal Consistency Filtering: This acts as a noise filter. If a detection lacks "supporting" detections in adjacent frames (a low temporal support degree), it is discarded as random sensor noise or a transient glitch.
- Progressive Refinement: The system utilizes motion vectors from previous frames to predict the expected location of a target, allowing it to maintain tracking even during prolonged occlusions.
Experimental Results
The method was evaluated on the SpotGEO dataset (released by the ESA) and the BUAA-MSOD dataset, which features real-world ground-based observations with complex, multi-target scenarios.
Quantitative Performance
On the SpotGEO dataset, our method achieved an $F_1$ score of 90.14%, significantly outperforming established models such as ResNet18 (75.43%) and YOLOv3 (81.89%). Furthermore, the trajectory completion strategy drastically reduced the Mean Square Error (MSE) to 61958.19, indicating highly precise localization.
We also developed a lightweight variant, WTNet-S, which contains only 0.38 million parameters—a fraction of YOLOv3's 35M parameters. Despite its small size, WTNet-S maintains a high $F_1$ score of 85.33% on SpotGEO and 89.54% on BUAA-MSOD, making it suitable for deployment on edge devices with limited computational resources.
Qualitative Analysis
Visual results demonstrate that while single-frame detection often suffers from false positives (marked in red/orange in Figure 3) and missed detections, the sequence-based post-processing effectively cleans the output. By enforcing temporal logic, the system ensures that only detections forming a consistent trajectory are retained, while missing points in a path are successfully interpolated.
Conclusion
By combining wavelet-based frequency enhancement with temporal trajectory completion, this research provides a robust solution for detecting faint GEO objects in noisy environments. The integration of the Hungarian algorithm for cross-frame matching allows the system to overcome the limitations of single-frame detectors. Future work will focus on improving robustness in extremely dense star fields and exploring more sophisticated non-linear motion models to handle a wider variety of orbital maneuvers.
Summary updated