SDPENet: A Lightweight Spacecraft Pose Estimation Network With Discrete Euler Angle Probability Distribution

Abstract

Utilizing deep learning techniques for spacecraft pose estimation enables using low-cost sensors like monocular cameras. However, the existing methods have drawbacks, such as complex models or low estimation accuracy. Therefore, this letter proposes the Spacecraft Discrete Pose Estimation Network (SDPENet). Firstly, we design a feature fusion network and a pose estimation head applicable to the spacecraft pose estimation task and devise the Spatial-Semantic Interaction Attention (SSIA) mechanism for feature fusion. Secondly, the discrete Euler angle probability distribution is proposed to represent the spacecraft attitude, significantly reducing the number of parameters while improving the accuracy. Finally, we put forward three data augmentation methods named CropAndPad, DropBlockSafe and Z-axis Rotation Safe to improve the performance of the network for the spacecraft pose estimation task. The experimental results demonstrate that, compared with the existing works, the errors in the spacecraft position and attitude estimated by SDPENet are reduced by 8.7%–83.1% and 31.7%–87.8% respectively, and simultaneously, the number of parameters is decreased by 33.3%–82.4%.