International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
of phase-difference laser rangers (Fröhlich et al., 2000). These
points, but also points on leaves, which typically have no, or only
a few neighbors on the same leaf, can be eliminated by removing
all those points where the plane fitting accuracy is below a certain
threshold. Choosing a threshold of 2cm reduces the number of
points by a few percent only, removing points that are not suited
for tree modelling anyway.
3.3 Cylinder fitting
Cylinders have 5 parameters: the radius is one parameter and
the axis (a line in 3-dimensional space) has 4 parameters. It
is, however, easier to use more parameters for the cylinder axis
and introduce some conditions between those. The following
parameterization of a cylinder is adopted:
I(P — Qu) xal -r-0 Q)
The symbol x denotes the outer product between two vectors and
the symbol ||... || denotes the euclidian norm of the vector. In
the above equation P — (x. yp, zp)! is a point on the axis, a =
(315,35 za)! isa vector of unit length that points into the direction
of the axis, and r is the radius. All points Qi — (zi, yi zi)
fulfilling this equation lie on the cylinder. Demanding that a has
length 1 and that P lies in a fixed, predefined plane reduces the
parameters from 7 to 5 again.
In order to fit a cylinder to a given number of points Qi;, 2 =
1,...n, the unknowns (i.e. P,a,r) are determined by least
squares adjustment. The residuals v; are the the distances from
the measured points to the ideal cylinder:
u = |(P- Qi xal -r
As the problem is not linear the observation equations of the least
squares problem are linearized and the adjustment is iterated.
Approximation values are required for the parameters PC ad of
the cylinder. The radius enters only linearly into the cylinder
equation, and therefore no approximation is needed. An initial
guess for the axis direction can be obtained from the normals
estimated in each point. The normal vectors lie approximately in
one plane, which is perpendicular to the cylinder axis a9. The
point P? can be estimated from the center of gravity of the points
Qi. A method for cylinder fitting without additional constraint
equations is described in detail in (Lukacs et al., 1998).
Typical accuracies achieved in cylinder fitting are in the order
of +3cm. As for the normal estimation this value is below the
accuracy of the laser scanner, but the rough surface of the tree, as
well as the deviation of a real branch from a circular cross-section
explain this difference. Fig. 2 shows an example of a cylinder fit
to a set of points. The r.m.s.e. of the fit is £2.7cm, the cylinder
radius is 29cm.
Applying cylinder fitting to a point cloud, does not only determine
radius and axis, but by projecting the given points to the axis, two
outermost points can be identified, marking the top and bottom
end (or start and end point) of the cylinder.
3.4 Cylinder following
Assuming that the axis direction and radius of a branch are known
at a certain position, this branch can be tracked (in both axis direc-
tions) in the given point cloud, using the cylinder fitting described
above. The base cylinder is described by the known radius and an
axis segment (i.e. the two axis end points). Shifting this cylinder
forward (or backward) in the axis direction for a certain percent-
age of its length, an approximate position for the next cylinder
Figure 2: Cylinder fitted to 10600 points of a tree trunk in a
perspective view. The cylinder is shown semi-transparent and its
axis is drawn, too. Arrow 1 indicates one place where shadows
of another branch between the scanner and the trunk caused oc-
clusions. Arrows 2 and 3 indicate where systematic deviations
from the circular form can be found (2: points inside cylinder, 3:
outside). Fitting accuracy for this example is £2.7cm.
is found. Points that are close (i.e. within a certain distance) to
this new cylinder can be selected and used for cylinder fitting.
If the parameters of the fitted cylinders are in accordance with
quality criteria, the cylinder is accepted and tracking continues.
If the cylinder does not pass the quality checks, another (e.g. a
smaller) forward shift is applied, or a shorter cylinder can be used
for selecting the points for the cylinder fitting.
The method of cylinder following has many parameters, but given
the domain of tree reconstruction, these values can be set easily.
The measurement accuracy, the tree surface roughness, the curva-
ture of branches, the change in radius along the axis, and finally
the deviation of the circular cross-section guide the selection of
the necessary threshold values. Values have been determined em-
pirically and checked for compatibility to knowledge of forest
experts. To account for gross measurement errors in the forest
environment, these values must not be set too tight.
For trunks, usually growing straight, the length ofthe cylinder can
be set to 50cm. Smaller branches showing stronger curvatures
require shorter lengths. Forward shift has been set to 30% of
the length, which gives enough flexibility to adopt to the curved
branch. If cylinder fitting is not accepted, shorter forward shifts
(16%, ...) are tried first, assuming that a point set similar to
the previously successfully used points will give better results. If
these shifts fail too, longer forward shifts are tried (100%, ...),
assuming that a major irregularity in the branch disables the good
fit of a cylinder. If this fails too, cylinder following is stopped in
this direction.
Points being close to the shifted cylinder are selected for the next
least squares fit. Closeness is determined in radial direction to the
cylinder side surface, and good results have been achieved with a
threshold value of 7cm, both for the inner and the outer distance.
This corresponds to a 3c bound of the achievable fitting accuracy.
Points lying inside this bound are counted, but not used for fitting
of the cylinder. If the forward shifts are enlarged (e.g. 100%), the
selection of points has to be made more generous, too, because
of the unknown curving direction of the branch (e.g. 10cm for
100%).
Internationa
CT CHLOE
Figure 3: N
segments a €
and end poi
segment (do
used to dem
Deciding if |
of the fit (3c1
(5cm-50cm)
secutive cyli
percent (e.g.
within the cy
sible measur
points on the
3.5 Segme
The algorith
the given pc
surface and a
algorithm fo:
points are su
the point cloi
relevant bran
Pfeifer, 2004
into voxel s
with matheir
tree. Each sl
a branch in «
based on dist
3.6 Cluste
The cylinder
precisely a s
of axis poin!
different pos
mentioned al
à severe pro
makes sense
but results on
between two
the combinec
shifted apart.
radius. In o
the cylinder
combined to
collection of
For this, a mq
Given are tw
Start and an «
E, for axis a