The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Surface matching is a technique used to carry out co
registration of point clouds and has been applied broadly in the
fields of computer vision and geomatics. Its applications can be
characterised as (i) registration of objects or surfaces comprised
of 21/2- or three-dimensional point feature data, (ii) detection of
differences between objects or surfaces, and (iii) integration of
datasets generated from different sources (Mitchell and
Chadwick, 1999). By far the most common algorithms used in
surface matching have been based on some form of least-
squares adjustment, minimising the differences in position
between the surfaces during iterative computation. Once the
matching is finished the transformation parameters are
computed and the surface is re-aligned to match more closely
the reference surface. In addition, one of the by-products of
surface matching is the ability to detect differences, as the
residuals from the least squares calculation are the surface
separations. Examination of these differences may reveal actual
differences that may have occurred between the surfaces
produced due to the use of different techniques (Pilgrim, 1996).
Based on this concept, a surface matching technique is
proposed to determine noise occurring in Mars DTMs.
The reason most of the terrain features can be seen in the
optical image is that their colour digital number (DN) values
vary from neighbouring pixels. That is, for clusters of pixels in
flat area their DN values should be relatively similar in the
absence of any significant albedo changes in the area. Based on
this idea, the noise points being removed from the previous
stage are reviewed and inspected in the corresponding ortho
image. Firstly, the corresponding noise pixel was found in the
ortho-image and the standard deviation of DN values between
the pixel and its surrounding pixels was then computed. If a
small difference was observed, the pixel was deemed as being
located in a flat area. In theory, these points should match well
with MOLA TIN after surface matching. Therefore, if the point
is located in a flat area but had a large disparity from MOLA
after surface matching, the point was confirmed as a noise point.
According to this rationale, all points being considered as noise
after surface matching were reviewed. Points whose standard
deviation of DN values between the surrounding pixels was low
remained as noise, whilst others were re-fed into the non-noise
point cloud representing the Martian surface.
2.3 Workflow
To implement the surface matching algorithm, the “3D Surf’
program was employed. This program was originally developed
at the University of Newcastle, Australia (Pilgrim, 1991;
Mitchell, 1994) and was further modified by Buckley (2003) for
the use of irregular datasets. The detailed algorithm was
described in Mills et al. (2003). In this paper two point clouds,
one from 3D intersection points derived from HRSC stereo
images and the other from a point set from the corresponding
area measured by the Mars Orbiter Laser Altimeter (MOLA),
were input into 3D Surf. The former dataset was computed
using the Video Image Communication and Retrieval (VICAR)
software (Scholten et al., 2005) while the latter point cloud was
collected from the MOLA points whose absolute vertical
accuracy is on the order of 10 m (Ebner et al., 2004). MOLA
was deemed as the “true” terrain model of Mars and was thus
input as the reference surface during the matching. To perform
the comparison, the MOLA points were firstly triangulated
using a Delaunay triangulation and then a least squares
minimisation of vertical differences between the HRSC 3D
intersection point cloud and the corresponding MOLA
triangulated irregular network (TIN) surface was performed.
The workflow of this noise reduction process for HRSC DTMs
is shown in Figure 1. In order to ensure the MOLA points
adopted in the surface matching were correct, it was compared
with a gridded global Mars DTM at about 500m spacing (Smith
et al., 2001) and any points beyond a pre-specified threshold
were removed if there were any. When the gridded MOLA
DTM was created such noisy points in the original MOLA point
cloud were deleted. Moreover, to improve the efficiency of the
surface matching, a median filter was applied to the MOLA and
intersection point clouds to reduce any random noise. It is noted
that once the matching and inspection were finished, the
reference surface could be replaced by non-noise points. Also,
the resolution of the intersection point cloud could be increased.
After updating the two surface models, the matching could be
performed repeatedly.
^ f
Triangulate
z
MOLA (TIN)
Once the matching is finished, the HRSC point cloud is
transformed and the residuals which represent the disparities
between the two point clouds are obtained. The points with
residuals over a pre-specified tolerance value are flagged as
noise and marked for removal from the HRSC point cloud.
Noise Inspection in Ortho-image
Although suspect noise points were removed during the surface
matching, for several reasons (further discussed in Section 3.1)
some points being removed are in fact true terrain features. The
withdrawal of these points could decrease the density and
representativeness of any final DTM. To improve this situation,
a method was developed to further examine the points being
removed.
MOLA points
HRSC
Image matching &
space intersection
Intersection
point cloud
Surface
/ Noise y/'y/ Non-noise y^—
-r-s ^
Noise inspection
" I
Noise
~T
Non-noise
Remove
Figure 1. Workflow of noise reduction.
1056
Hi
The pro!
•mint Cl<
(Figure -
approxia
^•1 No
First oi
extracte