CNN-Based Visual Navigation: Optimization Strategies for Monocular Pose Estimation in Proximity Operations

Proximity operations are becoming increasingly more important for current and future space missions, particularly On-Orbit-Servicing (OOS) and Active-Debris Removal (ADR) ones. In this framework, a high-accuracy estimation of the relative pose (position and attitude) between spacecraft is required to successfully and safely achieve complex proximity operations like inspection, rendezvous, and docking. Visual navigation has recently become one of the most popular techniques for this purpose, thanks to the availability of increasingly compact, precise, and reliable monocular cameras. Traditional approaches relying on hand-engineered feature matching do not guarantee robustness or sufficient generalization, whereas Convolutional Neural Network (CNN)-based architectures have demonstrated improved robustness, noise rejection, and resilience to unseen scenarios. Despite their potential, these algorithms do not frequently reach the desired accuracy due, among others, to the employment of heuristic approaches in the choice of hyperparameters and the unavailability of an adequate large dataset. This work proposes a CNN-based architecture for non-cooperative spacecraft monocular pose estimation exploiting optimization techniques to overcome these limits, improve performances and reduce the computational effort. This is achieved through the usage of a robust analytical method to select the best set of hyperparameters to minimize the pose loss function and the enhancement of the dataset for better feature learning. Moreover, the relationship between hyperparameters and the objective function (pose loss) is investigated, as well as the impact of different sets of hyperpa- rameters on the CNN performance. A Blender® based synthetic dataset of approximately 25,000 synthetic images of an uncooperative target is generated to train the CNN. Such images are used to emulate representative proximity scenarios to validate the proposed approach. The obtained results show that the proposed algorithm achieves centimeter-level position accuracy and near-degree-level attitude accuracy, maintaining, at the same time, high robustness against changes of illumination conditions and background textures.