Pose and Velocity Estimation of Noncooperative Spacecraft with Deep Landmark Regression and Tracking

The study develops an image-based approach to estimate the pose and velocity of the target spacecraft relative to the servicing spacecraft, from monocular sequential images. Such a problem is crucial in many space proximity operations, such as spacecraft repairing, refueling, and capturing. Specifically, a lightweight neural network model is constructed to regress and track predefined landmarks of the target spacecraft. We adopt EfficientNet-Lite as network backbone and replace traditional convolution with depth-wise separable convolution operations, respectively, to reduce the number of trainable parameters. Subsequently, 6D pose is solved using perspective projection relationship between landmark image coordinate and the target wireframe model. While velocity (angular and linear velocity) is recovered by 2D and 3D landmarks motion equation that is pro-posed to establish 2D landmarks velocity and 6D velocity between two consecutive frames. Numerical simulations are conducted on challenging SHIRT dataset, to validate the performances on pose and velocity estimation, respectively. The results demonstrate the superiority of our method, in terms of estimation accuracy and robustness.