77?e International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
As a building image is only generated for one building, the 
complexity of image processes is largely reduced. In the 
processes, thresholds are needed for creating a BR and a 
buffering zone determined based on the average point spacing 
of Lidar data, the spatial resolution of imagery and the accuracy 
of data registration. For a building image generation, the top- 
left coordinates of each building image need to be properly 
recorded, so that the line segments extracted from a building 
image can be correctly transformed onto the original image. 
2.3 Line segments extraction 
2.3.1 Principal orientation determination 
Most of buildings are rectangular buildings, whose boundaries 
have perpendicular structures with two rectilinear axes. These 
two perpendicular axes have the most dominant contexture and 
can be treated as two major principal axes. The dominant 
orientations can be determined from the angle values statistics 
of the line segments. Rau and Chen (2003) proposed a straight 
line detecting method using Hough transform with principal 
axis analysis to speed up the extraction of straight lines and 
improve the accuracy of detecting lines, in which the key issue 
is to obtain the principal orientations of a building using Hough 
transformation in an image space before line segments 
extraction. A few limitations still exist in this method. One of 
the limits shows its sensitive to the principal orientations 
determination of a building in an image analysis. In this process, 
wrong principal orientations may be obtained, especially when 
poor or high repeatedly textures are appeared in the image. 
Another limit occurs at threshold selection, which is needed for 
filtering counting values in accumulative array from a Hough 
transformation to construct an angle-count histogram. The third 
limit is hard to process an image with complicated and irregular 
building layouts. 
In this study, an algorithm is proposed for determining the 
principal orientations of a building. The principal orientations 
can be accurately and robustly determined based on the 
building image and rough principal orientations constraints. The 
building image is generated, and the rough principal 
orientations of a building can be obtained by analyzing the 
segmented building Lidar points. The proposed algorithm 
consists of two steps: rough principal orientation estimation and 
principal orientation determination. 
(1) Rough principal orientation estimation 
Based on the segmented building points, rough principal 
orientations of a building can be estimated by analyzing the 
Lidar points belonging to the building. A least square approach 
proposed by Chaudhuri and Samal (2007) is usually used to 
determine the directions of major and minor axes of discrete 
points. The method is used to determine the principal 
orientations of a building in this study. Considering the various 
geometric shapes of the buildings, a value range is constructed 
for a rough principal orientation by a threshold, which replaces 
the deterministic value of rough principal orientation in the 
following processes. In most cases, the threshold value is set as 
5. 
(2) Principal orientation determination 
The principle directions are determined by finding maximum 
values in accumulative array from a Hough transformation 
which fall within estimated ranges of rough principle directions. 
There are only two principle directions for a building. Based on 
the rough principal orientation constraints, principal orientation 
determination consists of the following 7 steps. 
Step a: Selecting a building image; 
Step b: Applying Hough transformation on the image, and 
finding the maximum value M in the whole accumulative array; 
Step c: Setting a threshold T=M*t (the initial value of t is set to 
0.9), and keeping those cells in accumulative array with value 
greater than the threshold T; 
Step d: Selecting the range of one rough principle direction; 
Step e: Adding the counting numbers with the same 0 value in 
the range. If all the accumulative values equal 0, then 
decreasing the value of t and going back to step c, if t is greater 
than 0; if t equals 0, the whole processing failed. If some 
accumulative values are greater than 0, go to next step; 
Step f: Selecting the range of the other rough principle direction 
and go to step e. If both rough principle directions are processed, 
go to step g. 
Step g: Detecting a peak in each of the two ranges. Each peak 
refers to a principal orientation of a building. 
The advantages of the proposed algorithm are in three points. (1) 
Based on the building image, the principal orientation 
determination on any complicated building layouts can be 
easily decomposed into some sub-processes on each individual 
building, which make the principal orientation determination on 
any complex images become possible. (2) To get the principal 
orientations of an image, the peak detection just needs to be 
performed in the specific range of a histogram based on rough 
principal orientations constraint by the proposed algorithm, 
which can improve the robustness of the principal orientation 
determination. (3) The automation degree of principal 
orientations determination is also enhanced because the 
threshold of accumulative array can be determined in self- 
adaptive way. 
2.3.2 Line segments extraction 
Having compared the most existing edge detectors, Edison 
detector is chosen to perform edge detection. Then the line 
segments are extracted using Hough transformation with 
principal orientations constraint. The line segments extraction 
becomes accurate and robust, because peak detection in the 
accumulative array just needs to be performed on the principal 
orientations. Since the number of line segments in an image is 
unknown, it is necessary to specify conditions to terminate an 
algorithm. A dynamic termination condition usually works 
better than a static one. Adaptive Cluster Detection (ACD) 
algorithm is a classical dynamic method to detect straight lines 
based on Hough Transform. In this study, a modified ACD 
algorithm is proposed by setting the searching priority for peak 
detection according to the principal orientations; the other 
processes are same as ACD algorithm. The thresholds of the 
minimum length of line segment and of the minimum distance 
between line segments are 20 and 20, respectively. Based on the 
determined principal orientations, line segments extraction 
becomes more accurate and robust. A few line segments with 
weak signals (but always important) on principal orientation 
can be extracted robustly with the principal orientations 
constraint, which may be missed by traditional ACD algorithm 
because of ambiguity in peak detection. It avoids missing 
boundary details. 
2.4 Boundary segments selection 
In this section, the extracted line segments will be automatically 
separated into two sets: boundary segments and non-boundary 
segments. Although the principal orientations constraint is used 
during the line segments extraction, there still exist many non 
boundary segments, especially the line segments of the building 
695