68  Yubin Xin, Jonathan Li, and Qiuming Cheng 
  
edge detection is a binary edge image but it may contain wide ridges 
around the local maxima. The required final position of the edge lies ap- 
proximately in the middle of this wider edge (Habib et al. 2003). To ex- 
tract the central position of the edge, a thinning (or skeleton) filter is ap- 
plied. We have implemented the Kreifelts (1977) algorithm to filter out 
non-skeleton pixels. 
3.2 Corner detection 
Since a GCP is always located near a corner feature, we can capture corner 
structures in patterns of intensities (Watson 2006). Using the gradient im- 
ages, we immediately calculate the eigenvalues of a sum-gradient matrix 
C. Comparing the minimal of the two eigenvalues with a pre-defined 
threshold, we may decide the candidate corner position. 
Azimuthal characterization matching is an algorithm that was first pro- 
posed by Motrena and Rebordao (1998). The approach is to solve the rota- 
tion-variant problem existing in area-based correlation method. It is based 
on the autocorrelation function of one azimuthal projection around a can- 
didate GCP and has rotation and contrast invariances. If an orientation 
variance (rotation) exists, the azimuthal projections of the sub-orthoimage 
and the GCP chip image about the corresponding central pixels will differ 
by a phase factor: the sets are the same but the initial elements are differ- 
ent. We may apply the autocorrelation function to determine the rotation 
angle. 
3.3 Linear feature extraction and comparison 
To improve the robustness, we add line comparison in the neibourhoods of 
the candidate point and of the GCP position in the chip. The Hough Trans- 
form line detection is performed in the thinned edge image surrounding 
those two positions. Usually there are two to four line sections in each 
patch that can be found. We can directly compare their slopes and/or the 
lengths and the brightness between the conjugate features to filter out 
mismatching. 
3.4 Pyramid cross-correlation 
Based on the rotation angle and the initial camera/sensor model, we may 
need to re-calculate the resampling to generate the sub-orthoimage for 
each GCP chip, then perform coarse correlation and fine correlation in a 
hierarchical matching process. The matched position is back projected 
onto the raw image (Xin and Parent 2004). Fig. 3 shows raw/orthorectified 
image patches and a GCP chip. 
IS e ^ ^ Ar A 
— 
AY rm [o o9 m