Quadrilateral Pose Estimation for Constrained Spacecraft Guidance and Control Using Deep Learning-Based Keypoint Filtering

The pose estimation is increasingly attracting attention in research fields such as constrained guidance and control, robotics, and communication technology. In the extreme environment of space, existing spacecraft pose estimation methods are not mature. In this regard, this paper introduces a spacecraft quadrilateral pose estimation method based on deep learning and keypoint filtering, specifically designed for spacecraft with coplanar features. A two-stage neural network is employed to detect and extract features from the spacecraft's solar panels, generating a heatmap of 2D keypoints. Geometric constraint equations are formulated based on the homographic relationship between the solar panels and the image plane, yielding the spacecraft's rough pose through the solution of these equations. The predicted confidence of 2D keypoints and rough pose are utilized to construct a pixel error loss function for keypoint filtering. The refined pose is obtained by optimizing this loss function. Extensive experiments are conducted using commonly used spacecraft pose estimation data sets, demonstrating the effectiveness of the proposed method.