Object Extraction from Terrestrial Laser Scanning Data

SUMMARY Terrestrial laser scanning emerges as a leading technology for direct 3D documentation of natural scenes irrespective of their complexity. The detailed level of description comes however at the cost of huge volume of data in form of unorganized, unevenly spaced, threedimensional points. The cloud of points provides a geometric description of the scanned scene but carries no semantic information regarding the objects within. Consequently, direct extraction of objects turns a challenging task. So far, research has focused on the extraction of well-defined objects with clear geometric characterization (e.g., plane, cylinders). Objects were extracted using segmentation algorithms which led to heavy computational efforts and were sensitive to scanning resolution and to artifacts. Effective working schemes for the extraction of objects require efficient and more general point-cloud processing methodologies. Such schemes are instrumental if aiming towards turning laser scanners into actual 3D mapping tools, and not only as means for characterization of surface geometry. We present in this paper a model for the extraction of objects in natural and cluttered scenes. The proposed approach is predominantly based on using a panoramic representation of the individual laser scans. We discuss the advantages of this representation for a direct scene interpretation and a clear definition of point connectivity it provides. Our focus is on means for identifying and detecting objects appearing in different shapes, sizes, depths and locations within the scene and relatively to the scanner. Results show that the proposed model is applicable for complex 3D point clouds that depict natural scenes without heavy computational effort.