In the following sections, first, the bridge data used for testing is 
described. Then, the entire workflow of DBM generation based 
on straight and curved bridge shapes is presented and explained 
step by step. The last section provides results and discussion. 
2. TEST DATA 
For this study, high resolution DMC aerial images and LIDAR 
data of highway 161 corridor area in Franklin County, Ohio, 
USA were provided by ODOT (The Ohio Department of 
Transportation), shown in Figure 1. A straight bridge enclosed 
in a black rectangle ROI (Region Of Interest) and a curved 
bridge enclosed in a red rectangle ROI are selected as test 
bridges. 
    
Figure 1. Straight test bridge (1) and curved test bridge (2) 
Using the aerial image footprint, LiDAR intensity image is 
generated from LiDAR data; the image resolution is set to 1m 
GSD (Ground Sample Distance). Figure 2 show the aerial image 
(a) and its corresponding LiDAR intensity image (b). 
  
(a) (b) 
Figure 2. Aerial image (a) and LiDAR intensity image (b) 
3. METHODOLOGY 
The proposed workflow consists of four steps: derivation of 
coarse DBM from LiDAR data, co-registration between aerial 
and LiDAR intensity image pair, extraction of smooth 
boundaries from aerial image and precise DBM generation. 
3.1 Coarse DBM Generation 
For our tests, the bridge ROI is manually selected in the LIDAR 
intensity image which is a rasterized LiDAR point cloud, 
created by using the highest intensity value of points falling in a 
ground cell. 1m sample size is used to generate the LiDAR 
intensity image. The relation between LiDAR intensity image 
space and LiDAR data mapping system is the following 2D 
transformation: 
EE, 
xc ED e 
N-N 
y > Va = GSD 
where (x, y) is the image coordinate and (E, N) is the 
corresponding LiDAR mapping Easting and 
Northing coordinate 
(xo, yo) is the image coordinate system origin, 
which is defined as the upper left corner, (Eo, No) 
is the corresponding LiDAR mapping Easting and 
Northing coordinate 
Once GSD is fixed, the bridge ROI in the LiDAR intensity 
image can be easily transformed to ROI in the LiDAR data 
domain, which is the 3D LiDAR ROI point cloud used to 
generate the coarse DBM. 
First, ground points and non-bridge points should be filtered 
out. Ground points can be easily removed based on elevation 
analysis. For non-bridge points having similar height as the 
bridge surface, the intensity data can be used for filtering. 
Nevertheless, pavement markings and/or vehicles on the bridge 
may have different reflectance characteristics, and thus, those 
points can be also removed which could create void areas in the 
bridge surface. In addition, outlier points may also exist; see the 
isolated point clusters in red ellipses in the Figure 3 (a). In order 
to trim those sparse outlier points, a statistical outlier removal 
filler based on statistical analysis on each  point's 
neighbourhood is also applied to clean the bridge points. For 
each point, the 2D mean distance from it to its closest » points 
is computed. It is assumed that those mean distances obey a 
Gaussian distribution. Then, points with mean distances outside 
the interval, defined by the global distances mean and standard 
deviation, can be regarded as outliers and removed from the 
bridge points; see differences between Figure 3 (a) and (b). 
  
(a) (b) 
Figure 3. Elevation/intensity filtered bridge surface (a) and 
cleaned by a statistical filter (b) 
3D RANSAC (Fischler and Bolles, 1981) plane estimation is 
applied on those cleaned bridge points to find the robust 3D 
plane parameters. Elevation of bridge points is recomputed 
through the 3D plane equation using the estimated 3D plane 
parameters. The newly computed points should be perfectly on 
the bridge surface defined by the estimated plane. Subsequently, 
concave hull boundary estimation is performed on the refined 
bridge surface points, as the bridge surface has a concave shape. 
In addition, points on bridge boundaries are also determined 
based 
neigh 
differ 
non-t 
conne 
bridg 
3.2 | 
In o 
regist 
Chatt 
PDF 
RAN 
estim 
datas 
thus, 
whicl 
inten: 
propc 
Fig 
  
Fig 
First, 
geom 
regist 
the sc 
Carlo 
for a 
origir 
equat 
Thes 
comb 
paran 
orien 
Cross 
be us 
pair 1 
NCC