3D building reconstruction from airborne lidar point clouds fused with aerial imagery - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 83

(a) (b)

FIGURE 6.19 Perspective views of the 3D building models with textured roofs reviewed from two different viewpoints.

FIGURE 6.20 The created 3D city model with textured building models, trees and terrain.

buildings are different in height, lidar point clouds are used to
separate themwith a total of six planar surfaces detected using the
RANSAC algorithm (see Fig. 6.18b). Then the outlines of these
surfaces are generated, simplified and regularized. Finally, the
individual building with a gable roof can be determined by the
roof ridge line from the intersection of the two planar surfaces
(see Fig. 6.18c). Figure 6.18(d) shows the three reconstructed
building models overlaid on the color aerial image.
The color aerial image (orthoimagery) was also used as
texture and draped onto both building rooftops and terrain
surface. From a data-structure point view, orthoimagery is
expressed as 2D raster data, which can be stored or manipu-
lated as a special layer in a 3D geospatial system. This work
focuses on a 3D scene with only three object classes: build-
ings, trees and terrain. The textured building roofs are required
for creation of a photorealistic 3D city model towards fly- or
walk-throughs and simulation (particularly for the planning and
visualization of building models). Figure 6.19 shows the per-
spective reviews of the study area from two different viewpoints.
Figure 6.20 shows the resulting 3D city model with buildings,
trees and textured terrain. It presents useful opportunities for
site visualization, which enhances community participation in
decision-making.

Concluding remarks

In this chapter we have presented a comprehensive approach
for the determination of 3D building models from lidar point
cloud data fused with color aerial imagery. A two-step extraction
strategy, building detection followed by building reconstruction,
has been developed and implemented. Building detection is first
done by filtering to separate on-terrain and off-terrain points,
and further, to integrate information from color aerial imagery
into an object-oriented classification process to catalog three
classes (buildings, trees, and terrain).
Toensure completeness, it is advisable to initialize the veryfirst
step, namely the coarse selection of building regions, interactively.
All subsequent steps (i.e., outline extraction and regulariza-
tion, planar surface detection, and building model generation)
are applied automatically. Possibly erroneous buildings can be
improved by interactive post-processing (i.e., manual editing).
Compared to lidar point cloud data, the advantage of color
aerial imagery is its higher accuracy in sharp linear features, which
can be utilized to refine building boundaries. The rectangular-
shaped buildings can be delineated well by using the least-squares
template matching with orthogonal constraints.