An Online Approach to Structure from Motion

In this paper, we present an online FastSLAM-based algorithm for extracting three dimensional structure from the motion of a single monocular camera. Nearly all published structure from motion (SFM) algorithms are offline, that is, they require the entire data set for processing. Our algorithm exploits the notion of conditional independence in Rao-Blackwellized particles, allowing us to decouple the complexity of the algorithm from the size of the data set. By doing so, we are able to utilize a particle filter to perform structure from motion as an online algorithm. At the heart of our algorithm is an online least squares technique to estimate the locations of all the landmarks as each new image comes in. Using this algorithm, we successfully extract structure from images taken by a camera constrained to linear motion.