Change Detection using Models derived from Point Clouds

Abstract

This thesis examines the detection of geometric changes in 3D data based on models derived from the point clouds. The process chain consists of the registration of point clouds, the derivation of a model and the detection of changes in these models. For the registration AprilTags are used, which, in combination with the trajectories of the respective measuring runs, allow the estimation of a transformation between two local coordinate systems without a preceding assessment of the tag position. The point cloud is sampled down to a uniform distribution for faster processing and the normals and curvatures are calculated for the remaining points. A region growing process uses this additional information to divide the point clouds into planar and nonplanar areas. The former are modeled as planes through simplification and meshing using a modified quadtree structure. The non-planar points are clustered by supervoxels and the extracted clusters are approximated by Axis Aligned Bounding Boxes. Both planar and non-planar modeling is performed for all datasets which are to be compared. The bounding boxes are used for the detection of changes. The box model of a dataset can easily be examined for intersections with a second dataset, due to the axis alignment. The intersections allow the detection of changed areas in the derived models. The methods proposed are suitable for both indoor and outdoor applications, provided that the changed objects are well separated and the compared datasets cover overlapping areas. The accuracy depends on the chosen size of the bounding boxes as well as the size of the changed objects. The evaluation has also shown that the proposed method can be integrated into a realtime-capable system.