The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
Figure 2. DSM generated from LIDAR data 
Figure 3. DEM generated after filtering 
2.3 Detection of Building points 
The object points derived by filtering include buildings, 
vehicles, vegetation and so on. The building points are usually 
detected by region growing. But there are often other objects 
that are attached to buildings, for example, vegetation and low 
objects. The building detection method proposed in this paper 
consists of three steps. Firstly, those low object points are 
removed if its difference between its height and the adjacent 
ground height is less than 2 meters. Secondly, a morphological 
opening operation is carried out, which can break the non 
building portions that are attached to buildings. Thirdly, the 
connected regions are detected by the region growing technique, 
and those that are larger than a certain threshold are regarded as 
buildings. The mesh cells that contain object points are marked 
with true, and the mesh cells that do not contain object points 
are marked with false. That is equivalent to form a binary image. 
The morphological opening operation and region growing are 
carried out based on the binary image, which can enhance 
computation efficiency and get the same result with computing 
directly to point clouds. The result is shown in Figure 4. 
There are some laser points on wall, which may influence the 
subsequent extraction of roof patches and determination of 
building edge points. So the wall points need to be detected. A 
building point is judged as wall point if there are a much higher 
point, a much lower point and few points that have approximate 
height in its neighbourhood. 
Figure 4. Building detection result. Red points are buildings, 
green points are other objects, and white points are ground. 
3. IMAGE PREPROCESSING 
The Canny edge detector (Canny, 1986) is a very popular edge 
operator, which is widely used in digital image processing 
including remote sensing image processing. The Canny 
operator works in four stages: Firstly, the image is smoothed by 
Gaussian convolution. Secondly, a 2-D first derivative operator 
is applied to the smoothed image to calculate the gradient 
magnitude and direction. Thirdly, the process of non-maximal 
suppression (NMS) is imposed on the gradient image. Finally, 
the edge tracking process exhibits hysteresis controlled by two 
thresholds. 
The non-maximal suppression and edge tracking are mainly 
responsible for the success of the Canny edge detector. But if 
the two steps are carried out solely using images, the 
determinations of edges are random. The building edge 
detection is easily influenced by other edges. Some building 
edges may be missed as shown in Figure 8. So the first two 
steps, namely Gaussian smoothing and gradient calculation, are 
implemented to the aerial image. The non-maximal suppression 
and edge tracking processes are to implement by combining 
LIDAR points and images in the following steps. 
4. EDGE DETECTION BY FUSION 
4.1 Creating of edge buffer area 
In order to conduct the building edge detection in images, the 
buffer areas are derived by building edge points. Firstly, the 
building roof patches are detected from LIDAR point cloud. 
The point sets with continuous heights are detected. And the 
point set whose amount is larger than a certain threshold forms 
a building roof patch. 
Secondly, edge points of the patch are detected. The edge points 
include not only the ones in the patch, but also the neighbouring 
ones that do not belong to the patch. That is, if there is a point 
out of the patch in the neighbourhood of a point in the patch, 
the point in the patch is marked as inside edge point, and the 
point out of patch is marked as outside edge point as shown in 
Figure 6. 
199