The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
340 
parallax of the other points. Finally, along corresponding 
epipolar lines, the other points are searched and matched. 
After the extraction of massive homogeneous points, these 
points’ coordinates in the object space are computed through 
back-projection by known image orientation parameters and the 
original DSM can be generated. 
5. EXPERIMENTS 
Three ADS40 three-line-array digital images were used in this 
paper. These images ware obtained form the same flight strip. 
The main attributes of the images are shown in Table 1. 
Camer 
a 
Leve 
1 
Location 
Focus 
GSD 
Flight 
height 
ADS4 
0 
1 
Lintong, 
China 
62.7cm 
0.48 
m 
1000 
m 
Table 1. Main attributes of experiment data 
Using the multiple image matching algorithm model, three 
images are matched simultaneously, where the nadir image is 
used as reference image and the backward and forward images 
are used as searching images. During the matching process, for 
every feature points on the reference image, an initial height 
(provided by approximate DSM) is given and gradually 
changed by a certain step. The coordinates of the feature points 
in the object space are then computed and back-projected to get 
the points in searching images. Given a correlation window, 
correlation coefficients between reference image and searching 
images are computed. The relationship between height and 
correlation coefficients are plotted, and some are displayed in 
Figure 2 and Figure 3. 
(a) (b) (c) 
(d) (e) (f) 
Figure2. Feature point matching results based on multi-image 
In Figure 2 , (a), (b), (c) is a part of forward, nadir and 
backward image respectively. The point marked by a red cross 
in the nadir image is feature point. The red line in the backward 
and forward image is the epipolar line and represents the 
searching distance, (d), (e), (f) is the correlation coefficients 
between forward and nadir image, the correlation coefficients 
between backward and nadir image, and the mean coefficients 
of the above respectively. In (d), (e) and (f), the horizontal axis 
represents the height range, which changes 40 times with 0.8m 
step, and the vertical axis represents the correlation coefficients. 
(a) (b) (c) 
Figure3. Feature point matching results based on multi-image 
In Figure 3 , (a), (b), (c) is a part of forward, nadir and 
backward image respectively. The point marked by a red cross 
in the nadir image is feature point. The red line in the backward 
and forward image is the epipolar line and represents the 
searching distance, (d), (e), (f) is the correlation coefficients 
between forward and nadir image, the correlation coefficients 
between backward and nadir image, and the mean coefficients 
of the above respectively. In (d), (e) and (f), the horizontal axis 
represents the height range, which changes 40 times with 0.8m 
step, and the vertical axis represents the correlation coefficients. 
As can be seen from Fig 2(b), there are many similar features 
around the selected feature points. Traditional matching method 
based on two images can not obtain correct matching results, 
see Fig 2(d), (e), while based on the multiple image matching 
algorithm, correct matching results can be obtained, see figure 
2(f). In Fig 3(a), the feature point is occluded in the forward 
image, and traditional matching method once again fails to find 
incorrect results (Fig 3(d)), while this problem can be avoided 
by the method proposed in this paper (Fig 3(f)). 
Feature points and grid points (every 3 pixels of the reference 
image) are combined to generate dense points to match. On 
every image pyramid layer, these points are matched and used 
for the next image pyramid layer. After image matching, an 
initial DSM (Figure 4) can be generated from these points’ 
coordinates obtained through back-projection. The DSM’s 
quality has to be controlled. In our experiment, these points 
with or are reserved, while the other points (around 5%) are 
set as doubt points whose heights are interpolated by bilinear 
interpolation method. After some simple processing, the final 
DSM is shown in Figure 5. 
6. CONCLUSIONS 
From above, a new image matching algorithm model is 
proposed in this paper. This algorithm model can be used for 
aerial three line digital images and can match multiple images