Vision-based pose estimation of space resident objects using deep learning: Design and testing

Reliable 6-DoF pose estimation of uncooperative Resident Space Objects (RSOs) is a key enabler for autonomous rendezvous, on-orbit servicing, and debris removal, yet the scarcity of labeled in-orbit imagery makes visual deep learning solutions difficult to validate and deploy. This paper presents a scalable, end-to-end framework for real-time monocular pose estimation of a known RSO, spanning synthetic data generation, training, geometric pose recovery and dynamics-aware refinement. A physics-based rendering pipeline in Blender is used to generate a large labeled dataset with photometric distortions and domain-agnostic appearance randomization, while a lightweight YOLO11n-pose CNN is trained to regress 2D keypoints and their visibility. Camera-relative pose is recovered via keypoints confidence thresholding and RANSAC-robust EPnP, then filtered by an Extended Kalman Filter that propagates translation and torque-free rotational dynamics using quaternion kinematics and an inertia-parameterized model, with innovation gating to reject inconsistent measurements. The framework is tested on the SPEED benchmark through repeated holdout testing, on eight controlled synthetic rendezvous sequences combining distance, tumbling rate, and target damage, and on real laboratory imagery acquired with a 1:1 manufactured target. Results show that synthetic-only training, when paired with robust geometric estimation and dynamics filtering, can achieve accurate and temporally consistent pose estimates, while retaining computational efficiency suitable for on-board execution on both GPUs and optimized CPU deployments. (c) 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.