This thesis explores the application and development of visual-based navigation techniques for spacecraft, particularly within the context of close-proximity operations. While these techniques have been successfully applied to terrestrial vehicles and UAVs, their extension to the complex space environment poses unique challenges. The growing necessity for autonomous systems in space missions, especially those involving non-cooperative targets like space debris, necessitates advanced navigation solutions. Vision-based navigation methods are critical in these scenarios, requiring precise estimation of a target’s position and orientation. This thesis investigates the development and evaluation of such methods. The thesis begins with an overview of spacecraft navigation, emphasizing recent advancements in autonomous and vision-based systems. The focus then shifts to proximity operations, a critical aspect of space missions. In this context, deep learning-based approaches to pose estimation are examined, particularly emphasizing their application to space scenarios. The thesis work includes a series of experiments conducted within a proprietary simulation environment developed by Vimotek AB, where the performance of the proposed algorithms is rigorously tested. Results indicate that the object detection algorithm developed in this thesis effectively identifies spacecraft under various conditions, demonstrating satisfactory performance metrics. Using pre-trained weights has proven beneficial in adapting the model to the space domain. However, the algorithm faces challenges distinguishing targets from complex backgrounds, suggesting that further optimization, such as enhanced data augmentation or an expanded dataset, could improve performance. 

The pose estimation network, although generally successful, shows limitations in accuracy when compared to existing methods. Factors such as insucient data augmentation and limited training iterations due to computational constraints likely contribute to this discrepancy. The research highlights the trade-offs in pose estimation accuracy with different model configurations, pointing out that increased parameter complexity does not necessarily translate to better performance. Moreover, the thesis underscores the importance of minimizing the domain gap between training data and real-world conditions, which is crucial for reliable deployment in space. 

The fusion of object detection and pose estimation algorithms into a single pipeline yielded improvements, particularly in correcting positional errors. However, orientation estimation remains problematic, likely due to differences between training and simulation environments. 

In conclusion, while the object detection algorithm shows promise and could be integrated into guidance, navigation, and control (GNC) systems, the pose estimation algorithm proposed in this thesis requires significant refinement before being considered for operational use in space. Future research should focus on enhancing the generalization capabilities of these models, possibly through more sophisticated pre-processing pipelines and the incorporation of sequential data processing techniques like LSTMs. Additionally, integrating sensor fusion, particularly with LiDAR, could address some of the limitations observed. Finally, deploying these algorithms on space-qualified hardware remains an essential step toward their practical application in future space missions.