Wednesday 12 March 2025
The quest for accurate space object tracking has long been a challenge in the field of astronomy. With the increasing number of satellites and debris orbiting our planet, it’s crucial to develop reliable methods for identifying and predicting their movements. A team of researchers has made significant progress in this area by developing a novel framework for generating realistic synthetic datasets of Resident Space Object (RSO) imagery.
The project utilizes an innovative approach to create effective point spread functions (PSFs) using real optical systems and convolving them over images rendered in Blender, a popular 3D modeling software. This allows the team to generate high-quality synthetic data that can be used to train machine learning models for image recovery and pose estimation tasks.
One of the key challenges in space object tracking is dealing with blurred or degraded images caused by atmospheric interference or optical system limitations. Deconvolution techniques, which aim to restore original images from noisy or distorted ones, are commonly used to address this issue. However, these methods often rely on complex mathematical formulas and may not produce optimal results.
The researchers have tackled this problem by developing a U-Net architecture that can effectively recover degraded RSO images. The U-Net is a type of convolutional neural network (CNN) designed specifically for image segmentation tasks, but the team has adapted it to perform deconvolution on RSO imagery.
The experiment involves training the U-Net model using synthetic datasets generated by the PSF approach. The results show that the U-Net outperforms traditional deconvolution methods in recovering high-quality images, with a mean squared error (MSE) of 8.36 and peak signal-to-noise ratio (PSNR) of 79.90.
The team has also explored the application of their framework to pose estimation tasks, which involve determining the orientation and position of RSOs in space. They have trained a ResNet50 model on synthetic datasets generated using the PSF approach and achieved impressive results, with an average angular error of 0.414 radians.
One of the key benefits of this research is its potential to greatly expand the availability of labeled training data for machine learning models. Traditionally, collecting large amounts of high-quality labeled data for space object tracking has been a significant challenge due to the complexity and cost of acquiring real-world datasets. The synthetic dataset generation approach offers a more feasible solution, allowing researchers to generate vast amounts of realistic data that can be used to train and test their models.
Cite this article: “Realistic Synthetic Datasets for Space Object Tracking”, The Science Archive, 2025.
Space Object Tracking, Resident Space Objects, Synthetic Datasets, Point Spread Functions, Psfs, Optical Systems, Image Recovery, Pose Estimation, U-Net Architecture, Convolutional Neural Networks
