Shadow-based matching for precise and robust absolute self-localization during lunar landings

Abstract

In this work, we use shadows as predictable landmarks for accurate and robust absolute self-localization during lunar landing. Due to the lack of an atmosphere, shadows on the moon are not diffused and therefore sharper than on Earth. This enables easy pre-computation of their appearance. By matching descent images from a lander's camera with reference images, rendered from available Digital Elevation Models (DEM), and the known sun position during the landing, we perform an estimation of the position. This serves as an input to a vision based navigation system. We first segment the shadows from the image content with adaptive binary thresholding, resulting in a binary shadow image. We have developed an enhanced shadow descriptor to depict a shadow's neighborhood by mapping the constellation of adjacent shadow centroids to a multilayer binary grid. This allows a time efficient and robust matching, which is further enhanced by a RANSAC based outlier detection. To test our method with the perturbations expected during the final phase of an actual lunar mission, particularly time and position deviation from the planned trajectory, we used DEMs of three different landing sites. Two of them were shadow abundant and one contained only a small amount of shadows. For each DEM, our method was separately tested with a position drift of up to 400m and a time drift of up to 50 min between the reference images and the descent images. The time drift was simulated by changing the position of the sun proportionally. With a maximum drift in position and time and at distances to the surface between 4km and 1 km, more than 90% of the shadows were matched correctly at the shadow abundant landing sites. It resulted in an accuracy of at least 1m for the position estimate. At least 40% correct matches were found for the landing site poor in shadows at distances to the surface between 36km and 12 km, resulting in an accuracy of the pose estimation with the same relative magnitude as before, but with less reliability. Overall, for both altitude ranges the accuracy of the position estimate was 1% of the distance to surface or better.