Deep recurrent neural network-based satellite indirect pose tracking with adaptive Huber loss

The relative pose and motion tracking of a satellite is a key technology for autonomous proximity operation missions. Most convolutional neural network (CNN)-based pose estimation methods are evaluated on satellite images with random poses, and thus cannot capture the temporal correlation of inter-frame features when faced with sequential images. To address this problem, this paper presents an innovative keypoint temporal tracking network, which estimates pixel coordinates of a set of pre-defined keypoints from a motion sequence by incorporating the temporal information. It leverages the sequence-modeling capabilities of long short-term memory units to process features extracted by a CNN backbone. Then keypoint sequences are associated with the corresponding 3D keypoints on a priori satellite 3D model to formulate the PnP problem. Subsequently, this paper also proposes an adaptive Huber loss that adaptively updates the switching threshold between mean squared error and mean absolute error based on the prediction residuals of each batch. Compared with original Huber loss, it can achieve a better trade-off between robustness to outliers and estimation accuracy. Finally, extensive simulations on the SPEED and the proposed HEDE sequence datasets validate the effectiveness of the proposed method.