3D building reconstruction from airborne lidar point clouds fused with aerial imagery - Urban Remote Sensing: Monitoring, Synthesis and Modeling in the Urban Environment - page 74

First-return

Laser pulse

First-return

Laser pulse

Last-return

Last-return

(a) (b)

FIGURE 6.2 Height difference of first and last returns.

of point v j , N is the number of all lidar points, M is the mean
matrix of its neighborhood points.
In this 3 × 3 dispersion matrix, each point v j has three
eigenvalues. An eigenvalue represents the spatial information of a
lidar point because it is a scalar of association with the eigenvector
that reflects the spatial distribution of a lidar point. Three
possibilities are considered. (1) If one of the three eigenvalues is
much larger than the others, a lidar point is labelled as an ''edge''
point. (2) If two of the three eigenvalues are much larger than the
other, the lidar point is labelled as a ''plane'' point. (3) If the three
eigenvalues are all larger than a given threshold, the lidar point
is labelled as a ''discrete'' point. Figure 6.3(a) shows the result
of dispersion matrices of lidar points in Fig. 6.3(b), in which
the white, black, and gray points represent ''discrete'', ''edge,''
and ''plane'' points, respectively. As illustrated in Fig. 6.3, trees
show the divergence property, while bare ground and most of
buildings exhibit the local planarity.

6.2.2.5 Extraction of linear features
fromaerial imagery

In general, a building consists of regular geometric primitives
(lines, corners, etc.). Given the fact that the lines (or edge)
of building boundaries appear clearer that those in lidar point
clouds, we apply the Canny edge operator and 2D Hough trans-
form(Sonka, Hlavac and Boyle, 2002) to extract the line segments
and use them as one of the clues in further classification.

6.2.2.6 Object-oriented supervised
classification

Comparing with an unsupervised approach, a supervised
approach is preferred by most researchers because it generally
gives more accurate class definitions and higher classification
accuracy. As a rule, a statistical supervised classification can be
carried out by the following three steps (Tso and Mather, 2001):

(a) (b)

FIGURE 6.3 Discrete measurements by eigen-analysis.