Adaptive Road Image Segmentation from Ladar-derived Labels

We present an approach to image-based road segmentation for autonomous driving in which an appearance model is adaptively learned from laser range-finder data. By tracking linear configurations of ladar obstacles as putative road edges and backprojecting into the image, a coarse partition of pixels into high-confidence on-road and off-road regions, as well as unlabeled bands of uncertainty between them, is obtained. A model of the current appearance of the road is learned by running a classifier on labeled image features. The immediate effect is a more refined segmentation at the pixel level indicating nonlinear shape features such as curves, dips, and rises; and some inference of the road geometry beyond the ladar range. At a higher level, the proposed image-ladar interaction offers an approach to segmenting novel roads and in changing illumination conditions without manual intervention. Some results using support vector machines and neural networks as the classifiers on a varied set of desert road images are discussed.