8 
CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
data, to establish three-dimensional centerlines. Those elevation 
differences of multi-layer areas were marked with an additional 
attribute. 
Other researches based on the mapping concepts to regard road 
surfaces as some parts of terrain. Thus, some filtering 
techniques were developed to extract ground information from 
airborne laser scanning data for DEM (digital elevation model) 
generation. The performances of those filtering methods had 
been compared by Sithole and Vosselman (2004). In some cases, 
single-layer road network could be regarded as a part of bare 
earth. Hu (2003) assumed that road profiles could be piecewise 
continuous and extract road points with the elevation threshold 
from discrete point clouds. Vosselman (2003) used laser 
scanning data to reconstruct single-layer road models 
referencing cadastral maps. This process derived road points 
within road areas first and generated models with triangular 
irregular network (TIN) surface. The refinement step assumed 
that road surfaces without slope, curvature, or torsion and 
smoothed them with the second order constrained polynomial 
functions. Additionally, Sithole and Vosselman (2006) handled 
the multi-layer condition which point clouds of overpasses were 
marked with the analyses of slope and elevation difference. 
They regarded those marked areas as the extended parts of 
terrain so that there was at least one side should connect to the 
ground. Oude Elberink and Vosselman (2006) paid attention to 
multi-layer interchanges using laser scanning data and 
topographic maps. Those roads were TIN-based models, and the 
multi-layer parts were separated into different elevations. Chen 
and Lo (2009) proposed a scheme to fuse airborne laser 
scanning data and topographic maps. The planimetric geometry 
and elevation of each road segment were established. The road 
models were represented as vector-based ribbons. 
As a summary, the integration of heterogeneous datasets seems 
to be a popular way to reconstruct three-dimensional road 
models, especially topomaps and laser scanning data. Most 
studies focused on the modelling processes for single-layer road 
systems, and few of them discussed about multi-layer parts. The 
reconstruction of road systems using a robust method for the 
large coverage is still an ongoing topic. Although the proposed 
scheme (Chen and Lo, 2009) reconstructed multi-layer models, 
this sequential modelling process was a local approach to 
smooth the model surfaces. In a rigorous way, we may need to 
consider a method to handle complete road networks and 
preserve the capacity for model updating. 
This investigation proposes an approach to model three- 
dimensional road networks using laser scanning data and 
topographic maps. Because some countries may have complete 
information of road boundaries and centerlines, others may use 
CAD data to describe roads using piecewise polylines in 
planimetric domain without geometric topology. Therefore, this 
investigation needs to compute topology of road networks and 
derive road elevations from discrete point clouds. In this 
planimetric part, each road segment would be generated its 
centerline and connect to others for network topology with 
conjunction points. The successive processes then include laser 
scanning data to derive road surfaces of each segment and refine 
all conjunction points to maintain the continuities in elevation 
and slope. When road systems encounter changes over time, 
new roads for example, they are needed to rebuild according to 
the latest dataset. Those new parts are digitized from aerial 
photos in this modelling process and refined their elevations 
with existed models to keep the system coincidence. The results 
are to be represented as three-dimensional ribbons. 
2. METHODOLOGY 
Based on the viewpoint of surface modelling, we integrate 
multi-source datasets to reconstruct complete surface modelling. 
In this investigation, we assume that the vertical and horizontal 
alignments of each road segment are continuous within a local 
area. Moreover, a global approach implements B-spline surface 
fitting refines the elevations of network conjunctions by keeping 
the continuities. The local approach sequentially modifies the 
elevation of each road segment. The proposed scheme has also 
considered the multi-layer condition. The processes have three 
parts: (1) registration, (2) planimetric networking, (3) model 
surfacing. The first part is to register all datasets, i.e. topomaps, 
laser scanning data, and three-dimensional boundaries. The next 
step then produces the networks using roadsides from 
topographic maps. The third part computes the model surface of 
each road segment and combines all roads from different 
sources to refine their vertical and horizontal profiles to keep 
the continuities in elevations and slope. The workflow shows in 
Figure 1. 
Figure 1. Workflow 
2.1 Planimetric networking 
In traditional CAD-based topographic maps, there are several 
road levels like local streets, arterial streets, expressways, etc. 
This kind of data records those boundaries using piecewise 
polylines. In addition, the topomaps may lack some information, 
e.g. attributes, topology, and centerlines. To directly use the 
topomaps for centerline generation is still difficult if those 
boundaries are independent without pair relationship. Therefore, 
this step uses those existed boundaries to compute centerlines 
for the reconstruction of network topology. 
First of this part, the planimetric process separates those 
boundaries into many simple straight lines. Those pieces then 
connect to each other according to the empirical thresholds of 
distance and angle for the development of complete boundary 
lines. The second step pairs those produced edges to position 
centerlines. All the planar conjunctions, i.e. crossroads, are 
automatically added a node point to split those centerlines and 
establish the topology, besides overpasses. The networking 
procedure would detect those multi-layer parts with boundary 
analysis (Chen and Lo, 2009) and mark which centerlines go 
through those areas. 
2.2 Three-dimensional surfacing 
After planimetric networking, laser scanning data is employed 
for road surfacing. The airborne LIDAR data records plenty 
discrete points with accurate elevation information. This 
surfacing step, basing on the planimetric geometry, extracts