Map based segmentation of airborne laser scanner data

Abstract

The task of segmenting point clouds is to group points that belong to the same (part of an) object. In this project we make use of an existing topographic map as a kind of background layer to segment point clouds acquired by airborne laser scanning systems. This map is an object based representation into polygons which are labelled into topographic classes like buildings, roads, terrain, water and vegetation. We implemented a point-in-polygon operation which is succeeded by a relabeling step at locations of roof overhangs. Next, the topographic class of the object is used to correctly process the point cloud into meaningful segments. The type of segmentation, e.g. planar or smooth segmentation or connect component analysis, depends on the corresponding topographic class of the object. The segmentation step is extended with a classification step where points are labelled as actually belonging to the corresponding map class or not. This is helpful when dealing with point clouds of cars on roads, or powerlines above the terrain. We segment for example the points on the individual cars into one segment each, but we label those points as not belonging to the class ‘road’. This is useful information to filter points from the point cloud when used to generate a 3D landscape model. The result of the map based segmentation algorithm is a point cloud enriched with information on the corresponding topographic class, corresponding map object, a segment number, and an indication whether this point actually belongs the map class or not. Results on two national datasets show that the use of map information is beneficial compared to a standard segmentation approach. Improvements are shown at situations where in one class a smooth segmentation is more suitable, whereas in the other class a planar segmentation is better, which can only be achieved if the class is known before the segmentation.