Robust pose estimation for non-cooperative space objects based on multichannel matching method.

Area of Science:

• Robotics and Space Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Accurate pose estimation of space objects is vital for tasks like satellite navigation and collision avoidance.
• Existing methods often rely on 3D computer-aided design (CAD) models, limiting their applicability.

Purpose of the Study:

• To develop a generalized deep learning-based pose estimation pipeline for space objects that is independent of 3D CAD models.
• To enable both instance- and category-level pose estimation for non-cooperative targets.

Main Methods:

• A three-part deep learning framework: a keypoint extractor, a multichannel matching network with triple loss for keypoint pairing, and pose graph optimization with dynamic keyframes.
• Utilized a custom dataset of 11,565 images featuring nine types of non-cooperative space objects.

Main Results:

• The proposed method demonstrated robust performance across diverse conditions, including varying illumination, object types, rotation rates, and image resolutions.
• Achieved superior estimation results compared to state-of-the-art approaches, with mean Absolute Pose Error (mAPE) of 0.63° and mean Matching Score (mMS) of 0.767.

Conclusions:

• The developed 3D pose estimation pipeline offers a model-independent and generalized solution for space object pose estimation.
• The approach significantly improves accuracy and robustness, addressing limitations of previous model-dependent methods.