Semantic 3D Octree Maps based on Conditional Random Fields

In this paper we present a 3D semantic outdoor mapping system with multi-label and resolution octree maps based on the OctoMap mapping framework. The semantic labeling of point clouds uses conditional random fields. Speeding up the conditional random field, we use an adaptive graph downsampling method based on voxel grids and the histogram-of-oriented-residuals operator to describe the local point cloud distribution. We validate the proposed classification and map representation approach on real-world 3D point cloud data. The presented classification approach achieves an overall precision about 96%. The integration of the classification results into the map data structure offers the opportunity to solve complex task settings. Furthermore, the runtime of the presented approach allows an integration of the classification into a real-time 3D semantic outdoor mapping system.