Illumination Invariant Image Matching for Lunar TRN

Abstract

A major challenge for vision-based applications of planetary exploration is large illumination variance that is caused by a combination of the sun position, terrain morphology, and lack of light scattering due to the absence of an atmosphere. This diminishes the capability to recognize terrain features or landmarks that are critical for autonomous robotic operations, including spacecraft pinpoint landing, UAV, or UGV navigation. The upcoming Mars Sample Return (MSR) mission exemplifies this challenge, potentially landing in the early morning, in stark contrast to the late afternoon landings of all previous Mars missions. This significant shift in lighting conditions, from the afternoon imagery used to create the reference map to early morning terrain appearance in descent imagery, poses a substantial challenge to existing Terrain Relative Navigation (TRN) algorithms. Consistently recognizing the same features or landmarks under such varied lighting conditions has been a long-standing open problem. In response, this paper introduces two approaches for robust image matching under illumination variation: 1) A new correlation-based algorithm — the Lighting Invariant Matching Algorithm (LIMA), and 2) a novel application of deep learning for learning illumination invariant features. Both approaches are supported by experiments on synthetic and real datasets.