Pose Estimation Using Starfield Occlusion

Abstract

This thesis introduces and demonstrates a novel localization method for the unique environment that is found in permanently dark lunar craters. The method makes possible near GPS-quality localization without any artificial navigation infrastructure, performing an order of magnitude better than existing methods. This is inspired by the observation that as one traverses across the bottom of a crater, the view of the starfield changes. The view is obscured by the walls of the crater and as one moves, one wall occludes more of the starfield and the opposing wall reveals more of the starfield, demonstrating the direct relationship between the different views and the locations in the crater. By matching a view taken at a position to map-computed views of the starfield, the method can estimate location. Similar vision-based methods use the information provided by the skyline, the boundary between the viewed sky and the terrain. However, the skyline is not directly visible in dark crater environments and must be inferred from the occlusion of the starfield, degrading the accuracy of these methods. Alternatively, instead of using the skyline, celestial navigation techniques can be used which employ the information in the starfield. This work will demonstrate that position estimates computed from the inferred skyline alone or from the starfield alone are generally far less accurate than the proposed method, which fuses the two methods to take advantage of both information sources. While many localization methods tend to be grid based, this method utilizes a significantly faster pseudo-gradient descent method to search the position space. This is made possible by the simplification of the map, a simplification that is supported by general lunar crater morphology. This work presents an analysis of the method and highlights three significant factors that affect the performance of the localization: the ability to detect the starfield, the geometry of the crater, and the relative position in the crater. It is shown that for most lunar craters, this method should perform well. This method was successfully tested in a simulation environment modelling the mission target, Shackleton Crater, with localization accuracy better than 35 meters.