Synergetic segmentation and registration of TLS point clouds using geometric and radiometric information

The separation between regions of certain homogeneity of a data set in 2D or 3D is termed segmentation. Registration or referencing, on the other hand, denotes the process of transferring different data sets into a common reference frame or system. Since the appearance of Terrestrial Laser Scanning (TLS) – combined with quasi-area-based acquisition methods and the associated enormous amount of data in the form of so-called point clouds – in engineering geodesy and related fields such as architecture and civil engineering, segmentation as well as registration have been among the most important processing steps with regard to an automated evaluation and further processing of point clouds. Several procedures have been established, however, each of them representing independent and detached strategies in their solution of the individual process steps, whose disadvantages are mostly expressed on the segmentation side in incompleteness, the unilateral use of information or inefficiency as well as on the registration side by the lack of a suitable stochastic model. This thesis first of all emphasizes the importance of segmentation and registration, as well as the necessity to use an adequate stochastic model in the context of Terrestrial Laser Scanning. As a result of the discussed disadvantages of existing methods, the motivation points as well as the prerequisites for the following chapters result. A decisive component in any evaluation with measured quantities, also in association with segmentation and registration procedures, is the choice of a suitable individual weighting of the observations via the knowledge of the respective precision. Up to now, the precision-limiting factors influencing the measuring components in TLS could not be mapped in an all-embracing model. Consequently, a new stochastic model addressing this issue is presented. A recent look on the topic of segmentation with TLS mostly shows the less effective application to a 3D data structure in combination with already referenced point clouds and the monolateral use of solely 3D information. Therefore, a new segmentation method is presented that addresses these problems and guarantees almost complete segmentation results based on the natural data structure of a single scan with proven and very efficient image processing routines. The examples show the versatile applicability of the segmentation algorithm using both 3D and intensity information on urban and natural object structures in the individual point clouds. In the last chapter the loop closes. A unique synergetic segmentation and registration procedure is presented, which uses the information gain from both process steps mutually. Based on extensive segments of the segmentation of single scans, plane sub-segments are detected with a special sub-segmentation procedure, incorporating the derived stochastic model. By means of a plane-based matching procedure, approximate values are initially derived for a subsequent comprehensive registration procedure of the individual scans. Finally, the point clouds of the single scans can be transferred into a common coordinate system using the external transformation parameters determined from the registration step. Simultaneously, the existing matching information of corresponding planes can be used to complement segments that are not completely captured due to different scanning perspectives on the object. Thus, the result of this processing chain is a complete registered and segmented 3D point representation of an object previously acquired with a TLS from different perspectives.

Author dc:creator

• Burger, Mathias

Advisor dc:contributor.advisor

• Neitzel, Frank

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-sa/4.0/

Language dc:language.iso

en

Identifier URI

http://dx.doi.org/10.14279/depositonce-15502

OAI identifier oai:identifier

oai:depositonce.tu-berlin.de:11303/16724