Environmental Engineering Reference
In-Depth Information
LiDAR point-cloud
Original aerial image
registration
registered aerial image
Off-terrain points
Building points
Simplification of building boundary
On-terrain points
On-terrain points
Refinement of building boundary
Tree points
filtering
Detection of roof surface
Other features
classification
Reconstruction of 3-D building model
Building reconstruction
Operation
Data
FIGURE 6.1
The workflow of our strategy from building detection to building reconstruction.
returns, provided by lidar point cloud data, are all used as the
clues for classification. The homogeneous regions of the aerial
image can be attained by the watersheds transform segmentation
algorithm and boundary tracking which can be treated as the
object of classification. In each homogeneous region, both spec-
tral and geometric information can be comprehensively used for
the classification of lidar data in accordance with the experiments
to determine the conditions and rules of classification.
At the building reconstruction step, first, the coarse building
roof boundary (a polygon) is determined based on the classified
building regions. Boundary reconstruction is formulated as a
regularization or simplification problem. Because lidar points
are randomly collected, the traced boundary cannot be directly
used as the final building boundary due to its irregular shape and
possible artifacts introduced in the previous steps. Second, the
well-known Douglas-Peucker algorithm (Douglas and Peucker,
1973; its improved algorithm by Hershberger and Snoeyink,
1992) is employed for simplification of building region polygons.
Third, the least-squares template matching (Ackermann, 1983)
with and the right-angle constraint are used for extraction of
simple right-angle shaped buildings. Due to disadvantages of
lidar data mentioned earlier, some bias exists in the building
boundaries. Therefore, the simplified building boundaries are
projected onto the registered aerial image for further refinement.
Fourth, the random sample consensus (RANSAC) algorithm
(Fischler and Bolles, 1981) is used to detect the roof surfaces and
then the ridge points of a gable building can be reconstructed
through building the neighbor correlation of the roof surfaces,
from which the ridge lines of the roof can be attained. The most
challenging process is the modeling of the individual roof faces
representing the roof. We define a roof surface as the closed,
polygonal boundary of a roof segment.
6.2.2
Building detection
Separation of individual buildings from lidar point cloud data
is the key to an accurate reconstruction. Like buildings, trees
are also one of the dominant features in urban areas. Thus, a
new object-oriented supervised classificationmethod is proposed
to detect individual buildings and differentiate trees from lidar
point clouds at the same time.
6.2.2.1
Region-based segmentation
Segmentation, a process of partitioning an image space into some
non-overlapping meaningful homogeneous regions (polygons),
is crucial to the classification result. Segmentation of color aerial
image contributes to the quality of classification because they
can provide more additional information than gray level images.
On the other hand, abundant spectrum information of color
image may lead to increase of the segmentation difficulty to
some extent which can be discriminated based on the height
and textural information from lidar data as additional channels.
The color aerial image segmentation includes two issues: (1)
the choices of color space, and (2) the choice of segmentation
methods. The RGB color space is suitable for color expression,
but not good for color image segmentation and analysis because
of the high correlation among the red, green, and blue bands. The
Search WWH ::
Custom Search