The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
198 
(Chen, et. al., 2004) firstly extract 3D planes from point clouds. 
Then the initial building edges are detected from raster form 
LIDAR data by the Canny Edge Detector. Based on the rough 
edges, the precise building edges are extracted in image space 
through the Hough transform. Obviously, only the straight lines 
can be detected exactly. (Chen, et. al., 2006) target the 
buildings with straight and curvilinear boundaries. The most 
probable radius of a line segment is analyzed to distinguish 
straight lines from the curvilinear ones. There are still two line 
types to consider. 
(Hu, et. al., 2006) extracted interactively outlines of complex 
building shapes from a high-resolution aerial image. Edges are 
detected using Canny edge detector and lines are extracted 
using Hough Transform from the aerial image. Because 
automatic method only works well for simple shapes, a 
primitive based method with user interaction is used to extract 
other outlines. A number of primitives that represent the 
outlines of building shapes are designed to reduce the times of 
user interaction. Obviously, that is a compromise between 
automation and demand of reconstruction building models of 
complex shape. 
As illustrated above, most of researchers put emphases on 
rectifying the edges in LIDAR points using the edges from 
images. During the process of detecting edges from images, 
only the image information is considered, which still have 
disadvantages of detecting edges from images as mentioned 
above. Moreover, it is often only one or two types of lines 
extracted from images that are considered during compensating 
the weakness of edges from LIDAR data. And it is often 
assumed that building outlines are comprised of parallel or 
orthogonal lines to regularize the edges for building modelling. 
Actually, edges in reality are not confined to only several 
certain shapes. So the results of edge regularization are likely 
not in accord with real condition, which obviously lead to the 
reduction of edge accuracy and the trouble of texture mapping. 
1.3 Work flow of adaptive building edge detection by 
fusion 
In order to deal with the weakness of detecting building edges 
solely from point clouds or images, a new adaptive method of 
building edge detection by combining LIDAR data and aerial 
images is proposed in this paper. The purpose of the method is 
to make full use of the complementary advantages of the two 
data sources to make the edge detection adaptive for all kinds of 
building shapes. The work flow of this method is shown in 
Figure 1. 
Firstly, the objects and ground are separated by a filter based on 
morphological gradient. The non-building objects are removed 
by mathematical morphology and region growing. Secondly, 
the aerial image is smoothed by Gaussian convolution, and the 
gradients of the image are calculated. Finally, the edge buffer 
areas are created in image space by the edge points of the 
individual roof patch. The pixels with local maximal gradient in 
the buffer area are judged as the candidate edge. The edges and 
roof patches are integrated by morphological closing operation. 
The ultimate edges are determined by the edge extraction 
method of mathematical morphology. 
This paper is structured as 6 sections. Section 2 presents 
building detection. Section 3 is the pre-processing of aerial 
image. Section 4 describes integrating the edges in image and 
roof patches to generate the ultimate edges. The experiments 
are described and discussed in section 5. Section 6 gives some 
brief conclusions. 
Figure 1. Work flow of adaptive building edge detection 
2. BUILDING DETECTION 
2.1 Mesh division of point clouds 
The LIDAR data is a discrete three-dimensional point cloud that 
is stored according to the acquiring time. The distribution of 
raw points is irregular, and the data is huge. Therefore one 
highly effective method of data organization is in demand. In 
this paper, point clouds are divided by an index mesh which can 
support effective neighbouring search as well as maintain the 
high resolution potential of raw data. Because the laser pulse of 
LIDAR is emitted according to a certain frequency, the average 
point spacing can be calculated, which can be the mesh interval. 
Every mesh cell may contain no point, one point or more than 
one point. During any neighbouring computation or 
morphologic operation of 3D point clouds, the corresponding 
mesh cell is firstly found out. Then all points in the cell are 
taken out to carry out computation or judging respectively. 
2.2 Separating of object and ground points 
The process of separating object points and ground points is 
called filtering. A new method of filtering based on 
morphological gradient is proposed (LI and WU, 2007). The 
morphological gradient of each point is calculated using the 
method suitable for filtering. Then, some points are chosen 
based on gradients to carry on an improved opening operation 
iteratively. The iterative times are controlled through analyzing 
the gradient histogram. During each time of iteration, a point is 
classified as an object point if its difference between the height 
after opening operation and the original height is more than a 
threshold. The filter based on morphological gradient can 
reduce the nonessential computation as well as the possibility 
that errors happen. DSM generated from LIDAR data of the test 
area is shown in Figure 2. The DEM generated after filtering is 
shown in Figure 3.