Comparative Analysis of Feature Tracking Challenges in Unstructured Space Environments

The robust visual tracking of celestial surface features is imperative for autonomous spaceflight operations such as terrain relative navigation and entry, descent, and landing. Plagued by computational constraints of radiation-hardened processing, traditional photoclinometry-based pipelines require extensive a priori imaging and human-in-the-loop interaction, which inflate mission costs and timelines, lack generalization to unexpected scenarios, and suffer from low throughput and accuracy. While terrestrial paradigms such as simultaneous localization and mapping offer significant advantages in cost, speed, and adaptability, they face substantial operational challenges in complex space environments due to fundamental failures in computer vision and image processing. The unstructured nature of celestial surface terrains, characterized by poor and dynamic illumination, textureless areas, and low-discriminative or self-similar regions, exhibits significant perceptual ambiguity and severely compromises feature reidentifiability; an issue that remains largely understudied. This work seeks to characterize these complexities through detailed feature tracking analysis on real-world space imaging data, correlating modern feature performance across surface, aerial, and orbital viewing modalities of varying visual complexity and in comparison to standardized Earth-based benchmarks. We draw distinctions and examine malfunctions of ten modern handcrafted and learning-based features, investigating properties that quantify the inherent difficulties in space-vision processing and provide foundational insights towards more resilient visual navigation systems for future spacecraft missions.