'. Istanbul 2004 
e authors dis- 
| areas by ini- 
vork present. 
A 
  
st Pulse 
  
ogical 
ing 
  
e 3) 
1” DTM 
Jure 4) 
E 
"iltered 
32) 
e5 
Ig 
int 
S3 
nage 
od" 
age 
qe 
age 
and S3 are 
| authors are 
ons on lines 
omogenous 
| and shape 
road model 
. Each road 
JAR points 
tensity and 
s appear as 
containing 
ting image 
nformation 
International Archives of the Photogrammetry, Remote Sensing and Spatigl Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
such as centreline, edge and width. 
The present paper limits the amount of data to only a subset 
of the original LIDAR points based on the intensity data, 
the closeness to the DTM, the LIDAR point density and 
the continuity of roads. By considering all of these crite- 
ria, it is possible to extract roads within a surveyed area 
effectively. By using the intensity and height information 
present in LIDAR data, a method is proposed for extracting 
roads from stand-alone LIDAR data. 
3 EXTRACTING ROADS 
3.1 Classification Work-Flow 
To extract roads from a LIDAR point cloud, a hierarchi- 
cal classification technique is used to progressively classify 
the LIDAR points into road or non-road. The work-flow is 
described by Figure 2. For the purpose of this paper, we 
will describe any LIDAR data point p; as being defined 
by, 
pi = ([px,lpy,lpz,lpi), (1) 
where lpz, [py, and lpz represent the last pulse laser strike 
JD coordinates and /pi represents the intensity of the last 
pulse strike. Let S represent the set of all laser points col- 
lected, i.e. 
S = {p1,D2,..., Py}; (2) 
where pi, pz, ..., pw are the individual LIDAR points. 
As described in Figure 2, the first step in the hierarchical 
classification method is to sample the last pulse LIDAR 
data into a regular grid with minimal filtering to produce 
a last pulse digital surface model (DSM). A DTM is then 
created from the last pulse DSM by morphological grey 
scale opening using a square structural element. By pro- 
gressively changing the size of the structural element and 
removing non-terrain type objects a DTM was obtained 
and displayed in Figure 3 (Rottensteiner et al., 2003). 
By making the assumption that roads lie on or near the 
DTM, which is true except for elevated roads, bridges and 
tunnels, it is possible to disregard all LIDAR points that lie 
outside a given tolerance of the DTM (Akel et al., 2003). 
The creation of the subset is defined by (3). 
Si = lbi £8: (Dis). DTM| < Ardy (3) 
where p;,,, is the last pulse z coordinate of p;, DTM is the 
value of the smoothed DTM at location p; and Ahmax is the 
maximum allowable difference between the p;,,, and the 
DTM. 
233 
  
Figure 3: The Generated Fairfield DTM. 
To help visualise the effect of the filtering, figure 4 dis- 
plays an intermediate result showing the position of all ex- 
isting points after applying (3). All non-terrain type objects 
(buildings and trees) have been removed and are displayed 
by areas of white whilst the strips of darker areas show 
areas of swath overlap, i.e. where there is a higher point 
density. 
LIDAR points are then selected if their last pulse inten- 
sity values are between the acceptable range for the type of 
road material being detected (in this case bitumen). Even 
though the intensity values returned by the scanning unit 
are noisy, road material is typically uniform along a sec- 
tion of road. By searching for a particular intensity range 
it is possible to extract all LIDAR points that were on the 
road along with some other false positive (non-road) de- 
tections. If more than one type of road material is to be de- 
tected, two separate subsets should be obtained depending 
on the road material characteristics from (4). The union of 
the two resultant subsets should be taken to form the inten- 
sity filtered set Sy. Equation 4 describes how the LIDAR 
points are filtered on their intensity to create a new subset 
of points 
S2 = lbi € 81 : imin < Pá, € ime } , (4) 
where imi, and i,m, are the minimum and maximum ac- 
ceptable LIDAR intensities at point p;. 
From the road model defined in Section 2, roads are de- 
picted as a continuous network of pixels which form thick 
lines. Due to the nature of roads, a circle around an arbi- 
trary road point p will have at least a quarter of the circle 
lie on the road itself, provided the circles diameter is less 
than the road width. This is the worst case scenario and 
typically we would expect between a half and all the circle 
to lie on the road. By testing all points against a chosen