|M.Sc Student||Miller Vera|
|Subject||Marker Free Registration of Terrestrial Laser Scans of|
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Sagi Filin|
|Full Thesis text|
Terrestrial laser scanners provide dense and accurate 3D data offering detailed description of surfaces and objects regardless of their complexity. For that reason, terrestrial laser scanners become a standard technology for characterization of natural environments, tracing deformations and for 3D modeling for various applications. Nevertheless, the volume of extracted data and their irregular distribution in 3D space turn their processing into a difficult and time-consuming task. In that context, there is a growing demand for automation of terrestrial laser scanning data processing. A fundamental and essential processing stage is the registration of individual scans into a common reference frame. For scans of man-made scenes, registration benefits from the relatively simple geometric shapes that may be easily detected in the data. Indeed, studies have made use of well-defined features, e.g. points, lines, or planes, or utilize additional data sources such as digital cameras and intensity data, which relate to the object radiometry. Suitable as they are for urban scenes, such approaches are inapplicable for alignment of natural environment scans, which are characterized by natural, undefined forms, geometrical continuity and albedo homogeneity.
This thesis proposes a new methodology for registration of terrestrial laser scans of natural scenes with no prior positional information. The proposed approach involves curvature based descriptors for key-feature extraction. Analysis of common curvature parameters have shown their unsuitability for finding relevant features in characteristic scans. Thus, transition zones between concave and convex areas were developed here as key-features for the registration. Contrasting saliencies, such curves always exist and form reliable features for registration. The proposed approach is capable of registering point-clouds with variable densities, partial overlap and occluded areas. It is also efficient in finding key-features in monotonous areas and ones characterized by lack of dominant entities.
The registration scheme is demonstrated on different data sets, featuring various topographies and point-cloud characteristics. Experiments show registration accuracy of a few centimeters that is comparable to results of manual registration, while exhibiting high robustness to natural-scene complexity. Such level of accuracy enables the next stage of mutual-alignment refinement using iterative methods.