The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
466 
2D *&&&#№ 
Figure 8. The produced accumulator array using the LiDAR 
points in Figure 7. 
Points contributing to the peaks are clustered while 
simultaneously considering their attribute similarity and the 
spatial neighbourhood among the points. In other words, the 
clustering procedure is implemented while globally assessing 
the attributes in the parameter space together with the local 
proximity of the points in the object space at the same time. 
This procedure provides a robust and accurate segmentation 
solution. Moreover, it is more efficient compared to the existing 
methods in terms of computation load due to the utilization of 
only two attributes in the procedure. Figure 9 displays the 
segmentation results produced from the LiDAR point cloud in 
Figure 7. In the Figure 9, the points in green and blue are 
clustered and recorded from two highest peaks in the 
accumulator array in Figure 8. 
Figure 9. The produced segmentation results from the LiDAR 
points in Figure 7. 
Figure 10 displays the results produced from the segmentation 
procedure applied to the generated building hypotheses in 
Figure 5. As before, the points in different colors belong to the 
different planar patches. One can compare these segmentation 
results, in particular those enclosed by the white solid ellipses, 
with the building hypothesis results in Figure 5. The points 
belong to a single building hypothesis have been separated into 
different clusters. As an additional output from the 
segmentation procedure, we use a least squares adjustment to 
derive an estimate of the plane of best fit through each cluster. 
Figure 10. Clusters produced from the segmentation procedure. 
The modified convex hull approach (Jarvis, 1977) is adopted to 
determine the boundary for each of the segmented clusters. The 
produced intermediate boundaries are displayed in Figure 11. 
Figure 11. Intermediate boundaries produced by using the 
modified convex hull algorithm. 
These boundaries will be used as initial approximations of the 
planar surfaces comprising the building model of a given 
hypothesis. 
2.4 Boundary Refinement and 3D Wire Frame Generation 
The last step of the proposed methodology utilizes the initial 
boundaries to come up with a refined set of boundaries, which 
are connected to produce a wire frame representing the DBM. 
The refinement process is based on several steps. The first one 
inspects the boundaries of the segmented patches to detect the 
presence of neighbouring planar patches which can be 
intersected (i.e., checks for the presence of ridge lines along 
gable roofs). After detecting the parts of the boundaries 
corresponding to the ridge lines the remaining boundaries of the 
sloping planar patches are further investigated. Next, horizontal 
planes are constructed by using the horizontal parts of the 
remaining boundaries. More specifically, height frequency of 
the boundaries is investigated to construct horizontal planes. 
The horizontal lines along the eaves of the sloping planar 
patches are acquired through the intersection between the 
constructed horizontal planes and corresponding planar patches. 
Then, the other remaining boundaries are regularized through 
Douglas-Peuker method and line fitting algorithms. After the 
refined lines are acquired by three different boundary 
regularization procedures, the proximity and collinearity in 2D 
space between the refined lines are investigated to figure out if 
the planar patches to which these lines belong are physically