Isao Miyagawa 
  
If 3D point ( X,, Y,, Z;) is known, 2D points (z;, yj) in another camera view (.X,, Y,, Z,) can be established . In other 
words, this projection allows delta points to be matched to 2D points in each image. 
im 
search domain 
  
Q(Xq, Ya, Zq) 
P (Xp, Yo, Zp) 
  
   
  
   
    
(Cx, Cy) 
  
  
  
  
  
  
  
(Xoc, Ygc, Zgc) ; p Orthometric Height 
Zo |" 77 En 
  
  
  
(Xo, Yo) XY 
  
-W 
  
Figure 2: Delta Points on Orthometirc Height 
Figure 3: Search Domain 
3.2 TRACKING ON HIGH-RESOLUTION IMAGES 
In high-resolution images, it has enough resolution to achieve accurate tracking in simple template matching method. 
Therefore, the complex operations are not needed, we set windows to search feature points of the top surfaces of buildings. 
We assumed that surface size is nearly same for each frames because helicopter have maintained altitude for optical axis. 
In searching process, the center of balance: (x, y) (calculated by the feature points composing the top surface of building) 
is corresponded to the center of window to search feature points. Template matching is done between I(x,y) on this 
window and J (x + [, y + m) in the next window. The texture data in a template window is extracted from shape formed 
several points on top surface of building. Feature tracking is realized using this texture data. To put it concretely, template 
pattern: /(x, y) in one image is matched to texture pattern: J(x + [,y + m) in next image. When this matched value : 
D(l, m) is a minimum value, (/, »n) is displacement between the two images. As the high-resolution image have adequate 
resolution, simple operation is enough to match pattern. 
1 dj d; 
Dil, m)y= Qd 1y NT E n I(x +dx,y + dy) + J(x +1+ dx,y +m + dy) (1) 
The size of search window: d; is determined by the distance from a center of balance to the most far point in all feature 
points. (/, m) is parameter in search domain. If delta points can be oriented on the image, the search domain in searching 
feature points except delta points, becomes (-W < ! < W),(-W < m < W) to (xdelta — dr, € 1 € zat 
dxs), (Ydetta — dÿ1 € m € Yaelta + dy2). 
33 TRACKING ON VIDEO IMAGES AND TRANSFORMATION COEFFICIENTS 
The displacements in video images are identified using the Lucas-Tomasi-Kanade method (Tomasi & Kanade, 1991)(Jianb 
Shi & Tomasi, 1994). This algorithm is robust and simple method. Feature tracking is done in video image sequence 
frame No.m, m+1, - - -, m+n in Figure 22. 
A helicopter carries two cameras parallel. All 2D feature points: (2p, ym) on each calibrated high-resolution image 
are transformed to 2D feature points: (r,, y,) on the calibrated video image as follows. The transformation modd 
is simple; we consider translation, scaling, and rotation. The transformation coefficients are calculated as follows. 
(toy Mui im 10-:-N are measured in the video image, (2g (i), Yui), t = 1---N are measured in high-resolution 
image linked to one video image. p = x, y, j means a Moore-Penrose inverse matrix. 
1 
Tul)  Mw(1) I PH(1) 
XH ay be to De p T2) 442); 1 PH(2) ) 
- l Ti D UAE C 
UH dy dy Yv y D : : 
P 
So(N) Yv(N) l PH(N) 
  
610 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
The | 
featu 
in the 
More 
resol 
video 
4 SI 
We h 
gawa, 
regist 
matri 
methc 
More: 
4.1 
It is i 
the di 
acquit 
in Fig 
level. 
delta | 
side)). 
groun 
|