International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 Internat
changes quite dramatically. Fairfield is an urban area, how- Vosselman, 2002). In (Akel et al., 2003), the authors dis- such 2
ever there are some rural-like parts with a heavily treed cuss a method for extracting a DTM in urban areas by ini-
creek area running through the centre of the image. There tially estimating the DTM from the road network present. The pi
are also both residential and industrial urban areas present. of the
The industrial regions have larger buildings with many car f LAST PULSE LIDAR / the cl
parks and private roads whilst the residential regions have INTENSITVAND RANGE DATA the co
a much smaller average building and block size. ria, it
Create Last Pulse effecti
DSM preser
2 BACKGROUND I roads
Morphological
Filtering
Road extraction from High-Resolution Airborne SAR data +
was performed in (Huber and Lang, 2001), by using opera-
tor fusion. À road was characterised by a central homoge- ; A
; ; : x : : z TUM 3)
nous region adjacent to two homogenous regions on either Find all points which are 31 «
side of the road. A method for the automatic extraction within a height tolerance of
9r NRI RE pun SM ES the DTM (Equation 3) Points “on” DTM / To cd
y (Wiedermann and Hinz, e extraction strat-
egy consists of finding the union of lines extracted from Filter LIDAR Points on Sue er eR
all available channels thus exploiting the nature of a multi- intor sity Minimum and d ;
: Maximum (Equation 4) Intensity Filtered €scri
spectral sensor. A graph network is then constructed to List (S2) / will d
derive the best paths and hence the road network itself. Remove LIDAR Points that do (Fiqure 5) by,
not have a minimum point
The extraction of roads by varying the scale space of an im- density (Equation 5) Poinis with Rig
age is another common approach. Extraction of roads from local point
Im-resolution satellite images was preformed by (Lee et : density (S3
: : Create binary image from the
al., 2000), by varying the scale space and applying a wa- mévriist
tershed algorithm before knowledge extraction was per- where
formed based on gray levels and shape cues. Urban road Binary Image 3D coc
networks are extracted from aerial imagery by (Hinz and . pulse s
Baumgartner, 2003). Scale-dependent models are explic- Morphological Filtering lected,
itly formulated at both fine and coarse scales in order to “Cleaned”
extract both detailed and global information. Connected. Component Binary Image
. : E Analysis. Remove small
In (Heipke et al., 1997), three different road extraction regions. (Optional)
techniques were evaluated. The LINE algorithm extraction where
is based on differential geometry, the TUM-G algorithm is ; : |
based on the extraction of lines in an aerial image of re- Check ratio of area against
; ^ area of MBR. (Optional) As des
duced resolution using the approach of (Steger, 1996), and classi
the extraction of edges in a higher resolution image. The € Ó dt
TUM-S algorithm is similar to the TUM-G extraction ex- of Label / 3 last
cept it uses ”snakes” in the form of ribbon-snakes to the Classification p
| 2 created
verify roads. (Figure 6)
scale o
There have been several attempts to extract roads from LI- Figute 2: Ihe Classification Work Flow. S1, Sa and Sa are gressiv
DAR data but most require a form of data fusion to com- “XP lained in section 3.1. femovi
plete the task. In (Hatger and Brenner, 2003), LIDAR data and dis
is used in conjunction with existing database information ^ Road models used by the previously mentioned authors are B
: : ; . : ; ; y mal
to estimate the road geometry parameters. In (Riegeretal, ^ varied. (Heipke et al., 1997) bases his extractions on lines DTM.
1999), roads were extracted from LIDAR data in forested and road edges, (Huber and Lang, 2001) on homogenous tunnels
areas. By detecting the road, breaklines could be generated regions and (Lee et al., 2000) on grey levels and shape ois de
and used to enhance the quality of the digital terrain model ^ cues. The usm.
(DTM) produced. A combination of line and point feature |
extraction was then used to extract the final lines. The presented work flow algorithm detects a road model
based on a continuous network of image pixels. Each road €
Traditionally, the intensity of the returned laser beam is ^ pixel has met a series of criteria, namely, LIDAR points >
registered by most LIDAR systems but this information lie on or near the DTM and have a certain intensity and
has typically not been used for feature extraction. Unfor- normalised local point density. The image pixels appear as where p
tunately, given the footprint size (e.g. 20 - 30 cm) and an visible thick lines which form a road network containing value of
average point density of 1 to 2 metres the intensity imageis ^ all public roads. From the resultant image, existing image maximu
under-sampled and very noisy (Rottensteiner et al., 2003, processing techniques can be used to extract information DTM.