The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3h. Beijing 2008 
be solved by an iterative bundle adjustment with equation (10) 
and (11) simultaneously. To ensure the stability of adjustment, 
more than two overlapping images are expected. However, one 
should be careful that the parameter a and b in equation (9) 
have to be known for opening curves, while it is not necessary 
to be known for closed curves, such as circles and ellipses. 
Degenerate cases must be avoided to reconstruct the curves. For 
example, the perspective center falls in the plane where a 3D 
planar curve lies. In this case, the image feature will be an edge 
instead of a curved feature. 
4. EXPERIMENT 
Our test data is a set of big overlap aerial imagery, and the 
curve features of interest are building curve edges. Snakes 
algorithm requires that the seed points of object curve must be 
near to the true edge. Otherwise, it can not converge to the ideal 
contour. In the first place, one image with a good object 
imaging is selected, and some seed points are located manually 
(as shown in figure la, the black cross points), then object 
curves are extracted automatically by snakes (as shown in 
figure lb, the red curve), and curve is expressed by B-splines. 
Secondly, the seed points are transferred to other images by 
corresponding points matching. The maximum correlation 
algorithm is adopted in our experiment (Zuxun Zhang, Jianqing 
Zhang, 2005). The result of corresponding seed points on 
another images are shown in Figure 2a (the red cross points). In 
this way all seed points are determined, and the corresponding 
curves can be extracted automatically on other images (as 
shown in figure 2b, 2c and 2d, the red curves. The black cross 
points are corresponding seed points by matching.). In addition, 
the interior and exterior orientation elements can be acquired by 
triangulation or other methods. Furthermore, the space 3D 
curve model can be achieved by traditional photogrammetry 
adjustment and generalized point photogrammetry adjustment. 
Figure2. Semiautomatic curve extraction on a signal image, (a) Initial seed points of curve feature, (b) Curve extraction result. 
(c) (d) 
Figure3. Multiple images corresponding curve extraction result, (a) The corresponding points of seed points on another image. 
(b)(c)(d) Automatic curve extraction result on other images. 
5. CONCLUSION 
In this paper we have presented a new approach for curve 
feature extraction. The method of active contour models 
(Snakes) constrained by both photometric and geometric 
conditions extracts curve feature accurately on a single 
image. Through corresponding point matching, corre 
sponding curve extraction by Snakes is implemented 
automatically on other images. Instead of a set of points on 
the feature, a B-spline representation of the curve feature is 
733