AI-ENABLED OPTICAL NAVIGATION AND ANOMALY DETECTION FOR A MOON-ORBITING SPACECRAFT

Autonomous navigation and attitude determination systems are essential in the cis-lunar environment, with optical navigation being one of the key systems, independent of Earth infrastructure, for ensuring safe and resilient mission architectures. Additionally, detecting, localizing and determining the orbits of other space objects around is an important capability for space situational awareness (SSA). This work proposes an integrated methodology that combines optical navigation with anomaly detection to enable resilient spacecraft operations and space situational awareness. Spacecraft position is first determined through a region-matching technique, in which detected lunar surface features are cross-referenced with a catalog generated. Annotated coordinates provide the basis for image-based navigation. Once the spacecraft state is determined, imagery is further analyzed for anomalies indicative of other space objects (natural or artificial). A dual-camera system supplies depth information, and the Herrick-Gibbs method is proposed to reconstruct trajectories from projected surface points observed at multiple epochs. Machine learning methods enhance both navigation and anomaly detection. Convolutional Neural Networks (CNNs) are employed to analyze imagery of lunar regions that are heavily cratered or obscured by shadows, as well as to identify anomalies. A key difficulty in this task is the variability of illumination and the presence of dynamic shadows across the surface. To address this, the CNNs are trained on a region catalog derived from digital elevation models generated in Blender. The training process uses a triplet-based approach with triplet loss minimization to optimize region detection. For accurate matching, a sliding window is applied to align raw images with their corresponding catalog regions. For improved accuracy and contextual awareness, a Shifted Window Transformer (SWIN Transformer) is proposed as a complementary transformer-based approach. For anomaly detection, the Mask Region-based CNN (Mask R-CNN) is employed to perform instance segmentation and classification. A classical instance segmentation approach is applied to identify artificial satellites. Additional disruptive factors, such as natural objects like meteoroids, are detected by analyzing pattern, texture, and color distortions on the lunar surface imagery.