Readspeaker menu

3D Point Cloud Analysis

Semantic Segmentation of Large-scale Outdoor 3D Point Clouds

Team: Jhonatan Contreras, Sven Sickert

This ongoing research is a collaboration between the Computer Vision Group at the Friedrich Schiller University Jena and the DLR (German Aerospace Center) Institute for Data Science in Jena. For LiDAR (Light detection and ranging) pulsed beams of light are used to measure distances from a scanner to the surface of objects in a scene to produce 3D point clouds. It is unstructured data composed of a collection of non-uniformly distributed points in a continuous space. In some cases, images are captured simultaneously during LiDAR campaigns to enrich these points with color information. In semantic segmentation we aim to assign one label from a set of pre-defined classes to each point of such a point cloud. For instance, in an urban outdoor scene, the classes could be natural terrain, vegetation, buildings, and cars among other classes. As a result, we get a meaningful (semantic) representation of the input data. This task is an important step in the development of intelligent systems in areas such as autonomous driving, urban planning, as well as disaster prevention and mitigation. A complete semantic understanding of the environment is crucial for automatic decision making.