n 
tA A lA P rci at 
i 
  
HÀ | Ps 
E A A OR te Dome ms mem tene 
vem n V T A f 
  
(b) 
  
Figure 6. (a) (b) — first derivative of vertical intensity 
projection for left and right orthophoto; (c) — 
correlation between projections 
Performing the shift of “feature” images to correlation peak we 
calculate vertical projection of equal edge pixels by 
implementing pixel-by-pixel comparison (Figure 7). 
Left image 
  
  
  
  
  
  
    
  
Right image > | i | 
. \ 
  
Vertical 
projection 
ul 
Figure 7. Counting of equal edge pixels 
  
  
  
Straight-line edges correspond to local peaks on vertical 
projection and can be found by implementing statistical analysis 
of the projection form, which provides more reliable detection 
results compared to the technique based on adaptive 
thresholding. Two statistical hypotheses are tested: Ho — 
projection form is a straight-line (no signal); H; — projection 
form is a “A”-shaped peak. 
The closeness of projection form to the hypothesis H, is 
evaluated as a likelihood ratio (8). 
260.) ® 
o(H,) 
where o(H,)» o(H,) denote standard deviation of 
projection values for Ho, H, hypotheses respectively. 
Straight-line edges with K value exceeding the experimental 
threshold and locating inside the road lane represent the 
obstacle cluster. Such cluster provides description of the 
obstacle: distance, width, position in a road lane and height 
above the road. Kalman filtering of the detected cluster gives 
relative obstacle speed. The final decision about obstacle 
presence is made after the special procedure of separate 
lineament edges analysis. 
5. RESULTS 
This section shows examples of obstacle detection based on the 
developed method. Figure 8 represents left images of real road 
scenes. Detected obstacle is indicated as a white rectangle. At 
the right of the rectangle two obstacle parameters are shown - 
distance to the obstacle and obstacle width in meters. 
  
Figure 8. Examples of obstacle detection 
6. CONCLUSION 
The developed method was tested on extensive database of real 
road scenes and provided reliable obstacle detection in different 
conditions including poor weather and night time. 
In case of using standard video cameras with the basis of stereo 
system about Im the method allows to detect obstacles with 
height of 10cm on distance up to 50m (Zheltov, Sybiryakov, 
2000). The working distance range of the method is from 15m 
to 80m where several obstacles are visible in both cameras. 
Implementation of the method for car collision avoidance 
system operates with frequency of 10 frames per second. 
Further development of the method is connected with 
introducing the global matching based on dynamic 
programming technique for detection of obstacle clusters of 
similar structure. This approach is seen quite perspective in 
sense of increasing the quality of periodical structures detection. 
7. REFERENCES 
Bertozzi M. and Broggi A., 1997, Vision-based vehicle 
guidance, IEEE Computer, vol.30, pp.49-55, July 1997 
Zheltov S., Sybiryakov A., 2000, Method of 3D-object 
detection based on orthophoto difference analysis. IAPRS, 
Vol.XXXIII, Part B3, Amsterdam 2000. 
—128-