Classifying Environment Shape from Image Motion for Mobile Robot Path Following

For a high-speed mobile robot following a road, hallway, or other type of path, modern robot motion controllers must know the coarse shape of the path, such as whether it curves left or right, or on which sides it is bounded by dangerous obstacles such as walls. In this paper, towards addressing the perceptual needs of high-speed mobile robot controllers, we classify these coarse shapes directly from the image motion observed by a camera mounted on a robot driving along the paths. We apply approximate Bayesian model selection over a set of learned probabilistic optical flow subspaces to explain the change between adjacent video frames. We perform inference directly from spatial image gradients instead of first computing optical flow. Optical flow subspaces encode a set of basis optical flow fields, which are valid under the assumption that the scene depth field remains constant over time, a conditional that holds approximately for a robot following a path-like environment. We do not require a calibrated camera or known camera motion. Our experimental results support the claims that our method efficiently classifies between multiple coarse path shapes, and that to do so it is important to use information from image motion, instead of image appearance.