| dc.description.abstract | Aerial imagery and lidar point clouds are complementary in terms of geometric information
contents. Aerial imagery has good definition of object edges but the information of object
elevation is implicit. Lidar point clouds, on the other hand, explicitly record the 3D
information of scanned object points but object edges are less clear than that in an aerial
imagery. This study proposes a method that integrates single imagery and lidar point clouds to
reconstruct 3D building models. Most buildings have multi-facet shapes, so that this paper
mainly focuses on the reconstruction of polyhedral buildings.
The proposed scheme is composed of five major parts, (1) Segmentation of Roof Patches, (2)
Roof Patch Classification, (3) Determination of Structure Lines, (4) 2D Line Segmentation,
and (5) 3D Building Model Reconstruction. The roof patches were segmented via the
coplanirity analysis with the lidar points on the roofs. Then, the roof patches were classified
into flat, multi-pitched and mono-pitched roof patches. Considering the errors of Lidar data,
the localization of low-pitched roofs from multiple slope ones could be difficult. Thus, we
analyzed the point clouds to find the optimal roof patches progressively. Once the patches
were found, we determined the initial building boundaries and the zones of roof structure lines
in the third step. The initial boundaries and the zones of roof structure lines were then
projected to the image space for the determination of a work area. Next, we detected the edges
in the working areas to find the edges and vectorized the edges to form the roof structure line
iii
segments and candidate boundaries segments. The refined boundaries were extracted and then,
the line segments were projected to object space to reconstruct 3D building models.
The accuracy of the results was validated by examining the discrepancy between the manually
measured building models and generated ones. The test data included (1) DMC aerial imagery
with a spatial resolution of 16 cm, and (2) Lidar point clouds from Leica ALS 50. Experiment
results indicate that the accuracies are ±0.242m in X-dir, ±0.246m in Y-dir, and ±0.260m in
Z-dir.
| en_US |