SU-Net: pose estimation network for non-cooperative spacecraft on-orbit

Area of Science:

• Spacecraft Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Spacecraft pose estimation is vital for missions like docking and debris removal.
• Estimating pose in space is difficult due to poor lighting, low resolution, and sparse data.
• Current methods struggle with the unique challenges of space-based imaging.

Purpose of the Study:

• To develop a deep learning model for accurate pose estimation of non-cooperative spacecraft using ISAR images.
• To enhance spacecraft feature extraction under adverse space conditions.
• To improve the robustness and accuracy of pose estimation for on-orbit operations.

Main Methods:

• Proposed a novel deep learning architecture, SU-Net, based on a dense residual U-Net.
• Incorporated dense residual blocks to minimize feature loss during downsampling.
• Utilized a multi-head self-attention block to capture global spatial information.
• Employed transfer learning and image enhancement techniques (contrast improvement, noise reduction) to address data sparsity.

Main Results:

• Achieved state-of-the-art performance in spacecraft pose estimation.
• Demonstrated significant improvements in accuracy with an absolute error range of 0.128° to 0.4491°.
• Reported a mean error of approximately 0.282° and a standard deviation of about 0.065°.

Conclusions:

• The SU-Net architecture effectively extracts features from ISAR images for precise spacecraft pose estimation.
• The model overcomes challenges of low resolution, noise, and data sparsity in space imaging.
• SU-Net offers a robust solution for critical on-orbit spacecraft operations.