Vjacheslav Listsyn 
  
3D-OBJECT DETECTION METHOD BASED ON THE STEREO IMAGE TRANSFORMATION 
TO THE COMMON OBSERVATION POINT 
V. M. Lisitsyn', S. V. Tikhonova" 
State Research Institute of Aviation Systems, Moscow, Russia 
"Ivm  gosniias.msk.ru 
Hk 
svetlana-t ? mtu-net.ru 
KEY WORDS: algorithms, stereoimages processing, 3D object detection, computer vision. 
ABSTRACT 
The method of 3D-object detection by stereo image processing is considered where object is situated on the smooth 
surface. Stereo image is obtained by passive sensor such as TV-camera, IR-camera and so on. It is supposed that some 
long landmarks are present on the underlying surface and position of those landmarks is determined for booth left and 
right images of the stereo-pair. 
To detect object both images of a stereo-pair are transformed to a common observation point. For images’ tying, 
equations are used that describe position of long landmarks on the left and right images. After the considered 
transformations, the fragments of the stereoscopic picture are subtracted. On the difference image, the patterns 
corresponding to figures on the underlying surface are close to zero, while contrast points related to three-dimensional 
objects are marked as face-down triangles. The presence of such patterns becomes a sign that 3D-object is detected. To 
search for such patterns, the contour selection on the difference picture and matched filtering of contour preparation is 
performed. 
1 INTRODUCTION 
The method of 3D-object detection by stereo image processing is considered where object is situated on the smooth 
surface. Stereo image is obtained by passive sensor such as TV-camera, IR-camera and so on. 
As a smooth surface the plane or cylindrical surface, i.e. surface obtained by transition of a straight line parallel to itself 
along some curve, can be considered. Here the underlying surface does not shade itself and is oriented along an optical 
axis of cameras. It is supposed that some long landmarks are present on the underlying surface and position of those 
landmarks is determined for booth left and right images of the stereo-pair. The landmarks position is used for images 
tying on both left and right images of the stereoimage. Landmarks should be oriented along an optical axis of cameras. 
Landmarks could be given by lines or curves equations in a camera plane or as a array of points for both left and right 
images of the stereo-pair. 
2 DESCRIPTION OF THE METHOD 
To detect objects both images of a stereo-pair are transformed to a common observation point, i.e. the stereoscopic 
images are transformed in such a way as if the TV-cameras have been displaced in a plane-parallel manner to the 
common observation point. For example it may be the mid-position between the cameras. Such transformation 
corresponds to rotation of the underlying surface around an infinitely remote point. 
In the image plane this leads to such an image transformation when the pixels on the line of horizon (that correspond to 
an infinitely remote point) remains unaltered and pixels below the line of horizon are shifted. And in this case, the lower 
the pixel is (the less is the distance to the observed surface) the greater is the shift. 
Lets define a co-ordinates with the center in the focal plane of a camera and with the axis Y directed upwards, axis X 
along an optical axis of the camera and axis Z to the right. In a such a co-ordinates the shift S in pixels is determined as 
  
  
534 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
whe 
alon 
If th 
then 
whe 
COIT 
tran 
Afte 
both 
the ı 
Afte 
subt 
surf 
The 
anal 
surf 
Sup 
forn 
stra: 
true 
whe 
of: 
con 
lool 
obi 
Dif! 
filtr 
con 
For 
for