Online generation of scene descriptions in urban environments

The ability to extract a rich set of semantic workspace labels from sensor data gathered in complex environments is a fundamental prerequisite to any form of semantic reasoning in mobile robotics. In this paper, we present an online system for the augmentation of maps of outdoor urban environments with such higher-order, semantic labels. The system employs a shallow supervised classification hierarchy to classify scene attributes, consisting of a mixture of 2D/3D geometric and visual scene information, into a range of different workspace classes. The union of classifier responses yields a rich, composite description of the local workspace. We present extensive experimental results, using two large urban data sets collected by our research platform.