Spatio-Temporal Terrain Classification for Mapping and Robot Localization

Abstract

Detection and classification of the surrounding terrain is a fundamental ability of a mobile robot in outdoor navigation to enable safe and efficient path planning. Our robot is equipped with a 3D LiDAR and a color camera, since these sensors complement each other very well. The terrain in front of the robot is divided into a grid, and each grid cell is classified individually using the sensor measurements. A new method for 3D LiDAR-based terrain classification with easy-to-compute and yet discriminative 3D features based on intensity and roughness histograms is presented. To exploit the fact that terrain appears in contiguous areas, spatial dependencies between the individual cells of the terrain grid are taken into account by modeling the grid as a Conditional random field. As the robot moves, we constantly update a terrain map with the current classification result. In this way, we are not only able to exploit temporal dependencies, but we are building whole terrain maps of the environment. Finally, we show how to use these maps for a semantic localization of the robot.