shared with other nonparallel lines. For those selected segments, 
after line following, each segment consists of discrete points 
(jc f ,) (/ = 0 ,— ,n). These points are fitted to a straight line with 
two parameters a and b 
y=a+bx (12) 
where 
V 
Suppose two parallel projected lines have parameters (a,b) and 
(a ,b ) respectively, the vanishing point should have coordinates 
A =U -aj)/[bj -b) 
X. =X —ap)l{^ J -b) ( 13 ) 
So another fitted line with parameters (a k ,b k ) is parallel to these 
two lines if and only if 
\cb + xh~y\^ £ (14) 
where £ >0 is relative to the data precisioin. 
Figures 9(a) to (d) are parallel line grouping results through 
perspective geometry. 
(c) (d) 
Figure 9. Extracted parallel lines 
5. ROAD EXTRACTION 
To extract roads in a mobile mapping imagery sequence, a set of 
road line seeds are defined, which have the lowest level of slopes 
and are parallel to each other in the object space. Among the seed 
edges in the image space, those pairs share the same vanishing 
point and have the anti-parallel character (the difference between 
direction angles is less than K but larger than K / 2 , caused by 
gradients of the two opposite road edges) are selected as road 
edge candidates. The nearest end points of the edge candidates 
are then connected with a constraint on direction angles. 
Knowing the camera position in the image space with the help of 
GPS/INS data, the initial curb/lane lines are delineated through 
conditional dilation within the segmented image. Edges nearby 
the lines are selected if their lengths are longer than a specified 
threshold. These edges constitute a geometric constraint used to 
select the preliminary line candidates (Figures 10). 
/ -f 
'X 
(a) (b) 
> -j 
c, * 
(c) (d) 
Figure 10. Preliminary road line candidates 
The candidates are then refined through a parallel line grouping 
process and become those improved in Figures 11. 
X 
¡X 
XX 
(a) (b) 
XX 
XX 
(c) (d) 
Figure 11. Refined road lines 
6. ABOVEGROUND OBJECTS EXTRACTION AND 
CLASSIFICATION 
6.1 Shadow Extraction 
Shadow is an important feature in aerial imagery. In principle, 
using shadows, date, time, and orientation of the image, the sun’s 
position can be determined. Inversely, artificial shadows of 
known objects can be generated for object recognition purposes. 
In photo interpretation, the combination of an object surface (for 
example, a building roof) and its shadow make them 
distinguished from others. After image segmentation, statistic 
properties such as mean and variance are calculated in each 
segmented area. One area (shadow) has a very low intensity mean 
and the adjacent area (its corresponding object) may have a very 
high intensity mean. In addition, the direction of the adjacency 
should be in the same as the sun. 
1A-2-5.