200 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
Thirdly, the inside edge points and outside edge points are 
back-projected to the image since the image have been 
orthorectified. The edge buffer areas are created around the 
edge points, which can be used to locate the real edge points. 
4.2 Detection of initial edges in image 
The initial roof patch edges can be detected from images by the 
edge buffer area. Because it is sure that there is at lease one 
edges along the buffer area, the non-maximal suppression and 
edge tracking processes are carried out along the buffer area. 
Firstly, the non-maximal suppression is performed in the buffer 
area. A pixel is judged as the candidate edge if it has local 
maximal gradient value by comparing along the gradient 
direction. That process is to make sure of only one response to 
the great extent along the buffer area, and suppress the 
responses of false edge. 
Secondly, the edge tracking is implemented based on the 
candidate edges determined by the non-maximal suppression. 
Traditional Canny operator carries on the edge tracking 
controlled by two thresholds, namely a high threshold and a low 
threshold. The tracking of one edge begin at a pixel whose 
gradient is larger than the high threshold, and tracking 
continues in both directions out from that pixel until no more 
pixel whose gradient is larger than the low threshold. The 
process is called hysteresis. It is usually difficult to set the two 
thresholds properly, especially for remote sensing image. The 
illumination and contrast of different portions of remote sensing 
image are often non-uniform. Even the gradient changes for the 
different parts of an individual building are very different. 
In this paper, the edge tracking is carried out by inside edge 
points. The long edges in the neighbourhood of the inside edge 
points are accepted as building edges. The process is 
implemented without thresholds because there are not many 
candidate edges by NMS in buffer area. 
4.3 Edge extraction by fusion based on morphology 
Practically, it is difficult to detect continuous and stable edges 
solely from the images. There are still some broken lines and 
noise existing in the edges processed by previous steps. The 
point clouds actually provide the initial closed edges around the 
roof patches. So the two kinds of edges from the two sources 
can be fused to form the complete edges. 
Firstly, the morphological closing operation is employed, which 
is produced by the combination of dilation and erosion 
operations. During the process, the edges detected from images 
in buffer areas are integrated into the individual roof patch. 
That is, a new patch is formed, which contain not only the 
original patch but also the edges detected from images. 
Secondly, the edge extraction from the new patch is carried out 
by mathematical morphology as following formula (Gonzalez et 
al., 2003): 
pjA) - A — (A ® B) (1) 
where A = the new patch 
B = structuring element with 3x3 size 
P(A) = the edge of the new patch 
A 0 B = A is eroded with structuring element B 
The ultimate edges are determined using the morphological 
method of edge extraction. 
5. EXPERIMENTAL RESULT AND DISCUSSION 
The LIDAR data used in this research covers an area in Toronto, 
Canada. The data is obtained by Optech ALTM 3100 system. 
The average density of the point clouds is 0.8 point/m 2 . The 
ground sampling distance of the aerial image is lm. 
Figure 5. Aerial image 
The edge points inside and outside each patch are detected from 
LIDAR data as shown in Figure 6. The horizontal range of 
building edges can be determined by the buffer area formed 
from edge points as shown in Figure 7. 
Figure 6. Edge points detected from LIDAR data. The orange 
denotes the edge points inside roof patches, and the blue 
denotes the edge points outside roof patches.