The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
not be disclosed by the device manufacturers. Also estimating
these parameters requires knowledge of the system design.
So far, the relation of emitted to received optical power was
discussed. The received signal is usually converted to an
electrical signal and amplified by an APD (avalanche photo
diode) and then digitized. This amplification is not necessarily
linear, and frequently a logarithmic behaviour is assumed.
Furthermore, for round-trip-time systems the emitted pulse is
not a Dirac impulse but has a duration in the order of a few ns,
e.g. 5ns, corresponding to a pulse length of 1.5m. The
“intensity” may refer to the maximum, the average below the
pulse shape, or the energy level after a certain amount of time.
It does stand to reason to evaluate alternative approaches,
relating intensity values, ranges, and target reflectivity to each
other without considering the physical foundation.
3.2 Data driven approach
In the previous section it has been shown that a number of
unknown system parameters may exist. This complicates using
a model driven approach, because the system parameters need
to be known for estimating object properties, especially object
cross section a or object reflectivity p-cos(a) for Lambertian
targets, from the observed range r and intensity I.
In a data driven approach the function / is determined. In a
general case, but restricted to Lambertian targets, it can be
written as f[rj,a,p)=0. From a set of observations and
corresponding known object properties the parameters of/ can
be estimated. For typical data from laser scanners r is
determined more precisely, with a coefficient of variation 3
below 1%, whereas the intensity value has a coefficient of
variation of 3% to 10% (Pfeifer et al., 2007, Hofle and Pfeifer,
2007). It therefore stands to reason to choose r, p, and a as
function parameters, because they are determined more
precisely, than the function value I. This allows formulating
residuals in an observed quantity.
I + e, = g(r, p, a) (4)
The function g should fulfil the following requirements. Firstly,
it should explain the data well. The / should have small
residuals: ejj = g (/}, py, aj) - Ij , for sets of data quadruples with
y'=l,... n, (n the number of points). The standard deviation ct 0 of
the residuals 4 , obtained from a function g with u parameters can
be used to judge the explanatory power of g.
a 0 = ( £ ejf /( n-u )) 0 5 (5)
Secondly, the function g should not contradict basic physical
principles. As details of the device model are unknown, this is
limited. However, the parameters p and a of the object have
similar effects on the backscattered power, namely decreasing it
for smaller values of p and larger values of a. The beam
diameter is small, in the order of a few mm, whereas the pulse
length for a round-trip-time measurement system is in the order
of meter. Thus, the stretching of the pulse may be neglected,
too 5 . The effects of p and a may therefore be bundled to k =
P’cos(a), which is in line with the basic physical principles.
3 The coefficient of variation is standard deviation / mean.
4 N.B.: a now refers to a quality measure, and not to the
scattering target cross section, introduced in section 2.
5 Assuming a pulse with constant distribution of power across
and along the beam, a 5mm spot size, a target inclination of 45°,
and a 4ns pulse duration (1.2m), this stretch is below 1%.
Thirdly, the function should avoid effects not observed in the
data. Properties like monotonic behaviour or smoothness, which
are visually apparent in the data, should also be properties of
the function g. A decreasing series of values strictly
interpolated by a polynomial may give rise to an oscillatory
function, which cannot be justified by the data. With an
increasing number of parameters и the function g may also start
modelling the noise. The noise would be different in a
repetition of the experiment and should not be represented by g.
Finally, the function should be invertible. This is necessary for
uniquely finding the reflectivity given observations of r and I.
The following types of functions are possible, of which
especially the nested approach will be investigated.
• Separation approach: 1 = gfr )*g 2 ( к) + g3
•Nested approach: I = g 4 ( k, g 5 ( r))
• Surface fitting approach: 1 = g 6 (r, k)
Next to ct 0 also a r will be used to judge the distribution of
residuals. The term a r refers to the root mean square value of
the residuals, and is different from a 0 by the denominator under
the root.
4. EXPERIMENT
Reference targets with known reflectivity behaviour were used.
These targets are made of Spectralon®, which reflects
according to a Lambertian scatterer (cosine-law). The targets
are quadratic with an edge length of approximately 13cm, and
mounted onto a metal frame, holding all six targets with
reflectivities of about 5%, 20%, 40%, 60%, 80%, and 99%. The
frame was placed in different distances to the laser scanners and
at different aspects. Points measured on one target were
selected, and for the point set of each target at each range and
aspect, a number of parameters was determined. Also, a plane
was fitted to the points minimizing the orthogonal distances.
This allows computing further parameters. Finally, for each
target the following parameters were used (Table 2).
p[]
Target reflectivity
on
Mean angle of incidence (to the plane)
r [m]
Mean range
JU
Mean observed intensity value
Table 2. Parameters determined for each target in the
experiments. Symbols are given with their units.
Scanning was performed with a Riegl LMS-Z420i (termed
Riegl) and an Optech ILRIS 3D (termed Optech). For each
device, two series of measurements were performed. In the first
series the distance of the target frame to the scanner was
changed: up to 15m in lm-steps, thereafter in 5m steps, up to
the maximum length of the laboratory (50m). In the second
series the target frame was placed at a distance of
approximately 15m to the scanner, and rotated in 9° steps from
0° to 72°.
The Riegl scanner operates at a wavelength of approximately
1550nm. At this wavelength the reflectivity of the targets are
0.986, 0.828, 0.653, 0.433, 0.233, and 0.081.
For the distance series the minimum and the maximum mean
range were 2.02m and 50.04m, and mean intensities varied
from 0.1207 to 0.3085. The maximum angle of incidence was
11.4° for the target on the outer end of the frame at the shortest