Finally, we get simple unit 
     
Diff. €(1-2,2-33-41(n-1,..,3) . 
S e (Diff. Diff 2. Diff 3} according to Eq.(1), and © is the across 
scales difference operation. 
Diff = I(c,s) = I(c)OI(s) (n =1,...,3) (D 
3) To generate C unit. The Diff from the S unit are interpolated to 
the scale of the image /(1) , and then calculate the complex unit C by 
using Eq.(2). 
C - Max«S) SelDiff. Diff» Diff) (2) 
4) To generate saliency map. To further highlight the saliency areas 
and for better visual effect, we smooth the image C with a Gaussian 
filter he whose size is 9*9 and standard variance is 8 to acquire the 
saliency map S(/) . 
S() =h"C (3) 
5 ) Object detection template is obtained. We use threshold 
segmentation to obtain binary image. 
st) = v S(I) > threshold (4) 
otherwise 
In eq.(4), empirically threshold is set to threshold = E(S(1))*3, 
where E(S(1)) is the average intensity of the saliency map. 
6) To obtain the detection result. The detection result is acquired by 
convoluting the real images and detection template of their own. 
Result - I*S(I) (5) 
3. EXPERIMENTS 
The experimental data are selected from the two data collection 
points in Wuhan City, China. A set of data are manually 
collected near Wuhan university. Another set of data are 
automatically collected by Mobile Mapping System in Hankou 
district of Wuhan (Data sources: Wuhan Leador Spatial 
Information Technology Development CO.LTD). To verify the 
effectiveness of visual attention mechanisms, we only use the 
image data provided by the MMS system without using the 
positioning information. And not one particular image, but a 
more general pattern, at the street. The research area and real 
images are shown in Fig.4. 
Collection 
ata Collection Points 
Points 
WE 
  
  
    
  
   
(d) 
Fig 4. Digital map(a) and real images(b) of the research area 
The difficulty in detecting is: 1) there are many kinds of interference 
exist in scene, e.g. the acquired information in experimental area 
includes traffic signs, roads, building, vehicle, pedestrians, lines, trees, 
shadow etc. The spectral, shape, texture of some objects is similar with 
the traffic sign. 2) The same target may have different appearance 
under different weather and light. 
Fig.5. Shows saliency analysis result of different method. Compared 
with Itti's"*! and Hou's"” method, the experimental result shows our 
method can better highlight the saliency of traffic signs and suppress 
irrelevant information in scene. 
     
  
(a) Original Image (b) Proposed method 
(c) Itti’s saliency map 
(d) Hou’s saliency map 
Fig.5. The results of different saliency method 
China's current standard is GB5768-1999 after revised in 1999 and 
mainly consists of warning signs, prohibition signs, indication signs etc. 
Because of their direct relationship with traffic safety, these three kinds 
of signs are the major studied objects. This research aims to prohibition 
signs, but partial indication signs. are also round, so this method is also 
able to detect partial indication signs. 
1) Image down-sampling 
Firstly, the size of images is 640*480. The length side is more than 640 
pixels, which should be resized to be 640 pixels. To illustrate the 
proposed detection method is effective, image keeps the original 
complexity without any cut. 
2) Target Segmentation 
After down-sampling, images were segmented by using the process 
shown in Fig. 1. 
    
  
        
    
     
   
   
     
   
    
    
    
     
   
   
    
  
     
    
    
   
    
   
   
     
  
d 
b 
b 
p 
n 
d 
n 
d