methodology utilizes rectangular models which are derived 
from LiDAR data and are adjusted during the model-based 
image fitting process. 
The structure of this paper is as follows: after a detailed 
explanation of the proposed methodology (Section 2), 
preliminary results are presented in Section 3. Section 4 
presents concluding remarks. 
2. METHODOLOGY 
As mentioned in the previous section, LIDAR data is a good 
source to detect buildings and generate initial boundaries 
based on the data-driven approaches. However, the quality of 
the derived boundaries is affected by the point density of the 
LiDAR data. Therefore, post-processing to regularise the 
boundary is required. The most popular approaches for the 
regularization include imposing constraints such as regularity, 
parallelism of buildings' sides, connectivity between lines, 
and integrating the boundaries with images (Brenner, 2005; 
Dorninger and Pfeifer, 2008). 
In this research, sets of rectangular models are derived from 
LiDAR data, and the quality of model boundaries are 
improved by the incorporation of images. For this, LiDAR 
data is processed to generate planar building segments 
(Section 2.1), and the initial boundaries are regularised as 
sets of rectangles which will be used as input for the model- 
based image fitting. Recursive Minimum Bounding 
Rectangle (MBR) is introduced to generate initial rectangular 
models (Section 2.2), and then the initial models will be 
adjusted during a sequential MBR adjustment (Section 2.3). 
2.1 Initial LiDAR processing: building detection 
First, to detect the possible buildings which can be used as 
initial building models, this research uses the segmentation 
methodology proposed by Lari et al. (2011). The 
segmentation process identifies individual regions with 
similar attributes and extracts useful features - in this case 
planar rooftops. This methodology starts with organising the 
LiDAR points using kd-tree structure to speed up the process. 
The proposed methodology considers varying point densities 
by calculating local point densities which determines the size 
of the neighbourhood. The points are classified as planar 
points or rough points using the neighbouring points during 
an iterative plane fitting, and these are grouped based on their 
proximity. The clustering procedure will be carried out on the 
grouped neighbouring points that have been classified as 
being part of planar surfaces. After the segmentation 
procedure, ground / non-ground classification is performed to 
filter out the ground points (Lari and Habib, 2012). Planar, 
non-ground group of points whose size and height are larger 
than predefined thresholds, are considered as possible 
buildings. Finally, the Modified Convex Hull algorithm is 
applied to generate boundaries (Sampath and Shan, 2007). As 
mentioned before, the initial traced boundaries from LiDAR 
data show irregular characteristics that need to be regularised. 
In this research, the regularisation will be carried out using 
rectangular models based on a model-based approach. The 
choice of the model parameters and decomposition of the 
complex buildings into rectangular models in an automatic 
way will be discussed in the following sections. 
2.2 Selection of model: Rectangular model 
Traditionally, building models are defined using six pose 
parameters: three of which define the model’s origin using 
coordinates of a reference point while the other three define 
the rotation angles between the model space and the object 
space. Another set of parameters is the relevant shape 
parameters. The most basic model is the one using the box 
primitives. In case of such model, three shape parameters, 
which are the length, width, and height of the box, are 
required. However, in imagery, rooftops and footprints of 
buildings cannot be observed at the same time. On the other 
hand, the vertical accuracy of LiDAR data is higher than the 
horizontal one. Therefore in this research the heights of the 
buildings are determined from LiDAR; this simplifies the 
box model down to a rectangular model. The heights of the 
buildings and the vertical positions of the reference points are 
calculated based on the plane parameters from the 
segmentation. Also, the rotation angles which determine the 
slope and aspect of the building rooftops with respect to the 
object space are derived using the surface normal information 
from the LiDAR data. The final parameters in this research 
then become the three out of six pose parameters (i.e., the 
horizontal positions of the reference point and one rotation 
angle for the orientation of the building) and the two out of 
three shape parameters (the length and width of the building). 
The justification of this choice of the final model parameters 
is explained in detail in Habib et al. (2011). The chosen 
model parameters will be adjusted using edge pixel 
information from available images through a least-square 
adjustment. 
2.3 Initial model parameter generation: Recursive MBR 
This section discusses how to derive the initial model 
parameters automatically as input parameters for the model- 
based image fitting. Since rectangular models are chosen as 
the basic model, the MBR algorithm is applied to regularise 
the initial LiDAR-derived boundaries and decompose them 
into rectangles. MBR is the rectangle with minimum area 
among the rectangles of arbitrary orientation which contain 
all the vertices of a LIDAR boundary (Freeman and Shapira, 
1975; Chaudhuri and Samal, 2007). MBR generation of a 
simple rectangular building is described in Habib et al. 
(2011). For complex buildings which are comprised of 
multiple rectangles, the MBR algorithm can be applied 
recursively. First, the MBR algorithm is applied to the initial 
LiDAR-derived boundary points and the 1* level MBR is 
generated. Then the initial boundary points, which do not 
overlap with the 1* level MBR, are found and then projected 
onto the sides of the 1% level MBR. Using the non- 
overlapping boundary points and their projected counterparts, 
the MBR algorithm is applied again to derive the 24 level 
MBR(s). The same procedure is repeated until there is no 
LiDAR boundary point left. As a result of the recursive MBR, 
different MBR levels are derived and by alternating 
subtraction and addition of each level, the final shape can be 
generated. However, to improve the horizontal accuracy of 
the final product, these MBRs are used as initial models for 
the model-image fitting. Next section describes how these 
different levels of MBRs are adjusted sequentially during the 
image fitting process. 
2.4 Sequential MBR adjustment using imagery 
The MBRs derived from LiDAR give a good approximation 
for the adjustment. The main objective of the model-based 
  
  
  
   
  
  
  
  
   
  
  
   
  
  
   
  
  
  
  
   
  
  
   
  
  
   
  
   
  
  
  
  
   
  
  
  
  
   
   
  
  
   
  
  
   
  
  
  
  
   
  
  
   
  
  
   
  
   
  
  
  
   
    
imag 
initia 
refin 
confi 
simp 
diffe 
the r 
MBF 
boun 
pixel 
actuz 
adju: 
level 
the | 
MBI 
Sect 
Tot 
selec 
selec 
com] 
MBI 
com 
inclu 
loca 
Tect 
airbi 
540 
dista 
poin 
flyin 
  
Fi, 
The 
dist 
are 
resu 
clas 
plar 
pin} 
gro 
wh 
are 
trac