Advanced Optical Navigation Strategies Based on AI Algorithms

The deployment of neural networks in space has opened new possibilities for achieving high-precision proximity operations, which are crucial for inspection, maintenance, and debris removal in the context of on-orbit servicing (OOS). A core requirement for these missions is accurate pose estimation, defined as the capability of an active spacecraft to estimate the relative position and orientation of a non-cooperative one. It involves substantial technological challenges in sensor architecture selection and algorithm development. For this purpose, visual navigation employs monocular cameras, favoured for their compact form and low resource demands, which serve as ideal sensors for capturing visual data in space environments. By leveraging Convolutional Neural Networks (CNNs), this research aims to enhance pose estimation capabilities, addressing the unique challenges posed by non-cooperative targets under varying lighting conditions and complex orbital backgrounds. This work aims to provide an optimization strategy for neural networks to increase their accuracy, performance, and efficiency for deployment on hardware with limited capacity. Particular emphasis is placed on advanced hyperparameter optimization and compression techniques, such as pruning, to streamline the network while preserving high levels of accuracy. The network model was trained and validated using two synthetic image datasets, each representing distinct, mission-critical phases of proximity navigations. The results demonstrate the effectiveness of hyperparameter optimization (HPO) and pruning techniques in enhancing the performance and efficiency of neural networks for space-based pose estimation tasks. The application of HPO led to a marked improvement in pose accuracy, with optimized networks achieving higher performance while downgrading image resolution, thereby minimizing computational cost and accelerating convergence during solution search and network training. Furthermore, structured pruning techniques were successfully applied, reducing model size without compromising accuracy. These optimizations validate theoretical expectations and establish practical benefits for on-orbit neural network deployment, showing that highly accurate and resource-efficient networks can be realized by reducing resolution and leveraging structured pruning. This work provides a foundation for further exploration into optimization techniques that maintain high performance even on constrained hardware, highlighting efficient strategies for real-time space applications.