Scalable real-time visionbased SLAM for planetary rovers

Simultaneous localization and mapping is an important problem for autonomous planetary rovers and other space vehicles. While many authors have addressed the SLAM problem, few have done so in the context of producing largescale maps in real time using vision. This paper is concerned primarily with the issues presented by the large numbers of candidate features obtained from vision sensors, and the implications for data association. We present a Rao-Blackwellised particle filter (RBPF) SLAM implementation that employs a stereo camera and the SIFT feature detector, and demonstrate that we can build maps of unknown environments with an extensive number of visual landmarks.