402 Vision Technologies for Small Body Proximity Operations Adnan Ansar*, Yang Cheng* *Computer Vision Group, Jet Propulsion Laboratory, California Institute of Technology, USA e-mail: {ansar,ycheng}@jpl.nasa.gov Abstract We present a framework for use of computer vision technologies to localize a spacecraft during small body (comets and asteroids) proximity operations. Our approach is to first detect image-based landmarks at stand-off distance during a mission survey phase, then catalog these landmarks into an easily referenced database, and finally use the cataloged data to recognize the landmarks during proximity operations. The catalog includes 3D body relative coordinates for the landmarks, so that image derived bearing angles to the landmarks enable localization of the spacecraft. In this paper, we describe our method of landmark detection and recognition, the details of the landmark catalog including estimation of 3D body-relative landmark locations, and our approach to vision-based estimation of spacecraft pose (position and attitude). We validate our research using real data from the JAXA MUSES-C mission and the NASA Deep Impact and NEAR missions as well as through detailed simulations. 1 Introduction Any mission, such as sample return, requiring operations in close proximity to a small body demands a high degree of autonomy. This imposes a requirement for precise in-situ spacecraft localization with respect to the target body. A vision sensor provides a low cost, low weight, low power, flight proven solution. We present a suite of computer vision technologies to enable vision-based spacecraft localization. Our operational scenario is a mission with one or more survey phases at a stand-off distance from the target body. During this stage, we detect a set of landmarks specifically chosen to be recognizable at subsequent stages of the mission. For this research we use a variant of the Scale Invariant Feature Transform (SIFT) [7]. SIFT detects highly salient points in the image and associates to each a descriptor based on local gradient data. These points are our landmarks. The descriptor is used to recognize previously detected landmarks when encountered again. The descriptors are highly invariant to image scale change and in-plane rotation, thus accommodating changes in spacecraft location and attitude relative to the target body. Illumination invariance is achieved through a combination of image processing techniques and incorporation of sun angle. Details of the landmark detection and recognition are given in Section 2.1. During the survey phase, detected landmarks are tracked across multiple image frames using image correlation techniques. These tracks are use to generate 3D locations for the landmarks in a body-centric coordinate frame using a least-squares optimization approach commonly referred to as Bundle Adjustment. The landmarks are cataloged by descriptor, 3D location and other contextual information. The catalog is designed for robustness, a high degree of discriminability between landmarks, and ease of search. Details of the catalog generation process are given in Section 2.2. While the BA process produces very accurate spacecraft pose and landmark positions, it is a computationally expensive, strictly batch process. Thus, it is ideal for generating the landmark catalog from stand-off distances but not for proximity operations, which require fast computation. During proximity operations, landmarks are detected and their descriptors compared to those recorded in the catalog. The combination of image coordinates and previously extracted 3D catalog locations allows for a complete body-relative 6 DoF solution for the pose of the spacecraft [3][8]. Note that in an actual mission, the 2D-3D correspondence between image coordinates and 3D locations would be incorporated directly into the navigation filter rather than through an intermediate, purely vision-based pose solution. However, the vision-based pose is still critical as both a sanity check on the filter and as an outlier rejection mechanism for catalog match errors prior to handing off data to the filter. In this paper, we focus on the vision derived pose only. A description of how the vision products can be incorporated into a navigation filter is given in [4]. We begin with an overview of the vision technologies used. We then show results on both real and synthetic data to validate the utility of this approach. 2 Algorithm Overview We now describe the vision algorithms developed and used for this research. 2.1 Landmark Detection and Recognition While our overall approach is agnostic to the choice of landmark type, we chose SIFT as a starting point because i-SAIRAS 2010 August 29-September 1, 2010, Sapporo, Japan