Two-Stage CNN-Based Pose Estimation for Uncooperative Axisymmetric Spacecraft

Pose estimation of uncooperative spacecraft, which is essential for on-orbit servicing, is challenging for highly axisymmetric targets such as rocket bodies. For these targets, accuracy is degraded because the rotation around the symmetric axis is difficult to observe. To address this challenge, this paper proposes a new CNN-based estimation framework that divides the problem into two stages. In the first stage, the framework estimates the rotation around the camera’s line-of-sight to normalize the input image. In the second stage, it estimates the remaining pose components from this simplified image, thereby reducing the overall estimation complexity. Through evaluation experiments using custom datasets of both axisymmetric (H-IIA R/B) and nonaxisymmetric (SOHO satellite) spacecraft, we demonstrate that the proposed method significantly outperforms the accuracy of conventional single-stage estimation. Especially, for the axisymmetric target, the proposed method achieved a median absolute angular error of 0.59°, compared to the best-performing single-stage method result of 1.89°. These results indicate that an approach that progressively simplifies the estimation process is highly effective for the pose estimation of axisymmetric spacecraft, which includes the rotation about the symmetric axis.