rived. The process of deriving these surface structures from a set
of sample points is traditionally done using computational geom-
etry based methods such Delauney triangulation or the Voronoi
diagram (Bolitho et al., 2009). The data is assumed to be free
from noise and dense enough to allow a realistic surface to be
derived (Zhou et al. (2010) in Lim and Haron (2012)). When
the point cloud is sparse or noisy the resulting surface is often
jagged rather than smooth. The surface reconstruction process
interpolates heights between sample points (Bolitho et al., 2009).
Each point is considered a moment of height change and between
points terrain height change is assumed to be linear or is solved by
interpolating a least squares fit. An alternative to computational
geometry is function fitting, these approaches define a function
for determining a surface at a given location by global and/or lo-
cal fitting (Bolitho et al., 2009). Kazhdan et al. (2006) developed
a Poisson Surface Reconstruction technique that combines both
global and local function fitting expressed as a solution to a Pois-
son equation (Bolitho et al., 2009). The Poisson approach uses
the orientation of the point normal to create a surface that changes
gradient according to the change in point orientation (Figure 1).
The algorithm obtains a reconstructed surface with greater detail
than previously achievable (Jazayeri et al., 2010).
Figure 1: A TIN versus a Poisson DSM.
This paper evaluates the UAV-MVS generated point cloud and
surface representations of a natural land form by qualitatively
comparing these to a reference dataset generated using close range
terrestrial photography based MVS techniques (T-MVS).
2 METHODS
2.1 Study Area
A dynamic 100 m section of sheltered estuarine coastline in south
eastern Tasmania will be monitored for fine scale change (Fig-
ure 2). The vegetation on the site is grasses and scrub bush along
an erosion scarp with salt marsh at the southern end of the study
site. For this study a section of the erosion scarp was chosen
as the focus area for comparing the close range terrestrial MVS
point cloud to the UAV-MVS point cloud (Figure 3).
2.2 Hardware
The camera chosen to capture photography at sufficient resolu-
tion for UAV-MVS point cloud generation is the Canon 550D
digital SLR camera. This camera has a light weight camera body
and provides control over ISO, aperture and shutter speed set-
tings. The settings are carefully chosen to reduce motion blur
when acquiring images at 1 Hz (one photo per second). The re-
sulting image dataset contains around 300 photographs per UAV
flight with 70-95% overlap. The OktoKopter micro UAV plat-
form (Mikrokopter, 2012) is the basis for the TerraLuma UAV
used for this study. The aircraft is an electric multi-rotor system
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
476
Legend
3 — Profiles
: "uu e [ Study Area
= a x Focus Area
0 2.000 4,000 6,000 8000 10.000 eins
km CC) Inset Boundaries
0 A
RL
£ nn
\ ^ .
N
0 306090120 - / 0153 6 9 12
Sm mio ENEEEM km
0 125 25 50 75 100 |
Eee Metres
Figure 2: Coastal monitoring site.
Figure 3: Images of the focus site (the first is taken looking east,
the second is taken looking west).
(eight rotors) capable of carrying a ~2.5 kg payload for approx-
imately six minutes. The system has an on-board GPS (5-10 m
positional accuracy) and other sensors that allow it to do way-
point navigation. The camera is attached to a stabilised camera
