The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
a2 . = compet(n2),si2'” = 
\,m = m* 
0,m m* 
(9) 
Where, m* is the serial number of maximal value in vector a2 ; . 
3. EXPERIMENTS AND CONCLUSIONS 
To verify the validity of author’s method, it is realized by 
Visual C#.Net and Matlab program language. Many 
experiments are also completed with natural scene images in 
Nanjing. These images, with size of 1392^1040 pixels, are 
taken by the vehicle-borne mobile photogrammetry system at 
different time and in different lighting conditions. The natural 
scene image taken by its CCD camera is shown in Figure 4. 
Figure 4. The experimental data-natural scene image 
Traffic signs detected from image of Figure 4 are shown in 
Figure 5(a), corresponding binary inner images of these traffic 
signs are shown in Figure 5(b), and recognition results of these 
traffic signs are shown in Figure 5(c). 
(a) Detected traffic signs 
+ X *1 
(b) Binary inner images 
Acfossroad ^ 
Cross road A\ Caution P«destrian| 
crossing 
(c) Recognition results 
1 Keep Right 
Figure 5. Traffic sign recognition results in image of Figure 4 
Experimental result shows that author’s traffic sign recognition 
method obtains good effect. The run time of author’s method is 
about 0.4 second under the condition of serial compiling. If the 
special image processing unit and technique of parallel 
compiling are used, the recognition speed will be faster. 
Besides the above experimental results, the paper totally selects 
221 natural scene images taken by the vehicle-borne system at 
different time and different locations to test the proposed 
method. There are totally 500 different kinds of traffic signs in 
these images. The number of detected signs is 480. To compare 
with other recognition methods based on invariant moments, the 
paper selects such three kinds of invariant moment as Hu 
moment (Hu, 1962), Tchebichef moment (Li, et al., 2006), and 
Zemike moment (Fleyeh, et al., 2007). Compared statistical 
results are shown in table 2. In this table, H.M, T.M and Z.M 
respectively represents Hu moment vector, Tchebichef moment 
vector and Zemike moment vector. 
""^-feature vector 
recogniton residí 
Author’s 
vector 
H.M 
T.M 
Z.M 
yellow 
warning 
signs(105) 
recognized 
rate % 
105 
100 
25 
23.8 
54 
51.4 
31 
29.5 
red 
prohibition 
signs(221) 
recognized 
rate % 
217 
98.2 
65 
29.4 
98 
44.3 
68 
30.8 
blue 
mandatory 
signs(154) 
recognized 
rate % 
154 
100 
69 
44.8 
103 
66.9 
86 
55.8 
Table 2. Compared statistical results between author’s central 
projection vector and invariant moments 
Some compared experimental results between author’s 
recognition method based on central projection vector and other 
methods based on invariant moments are shown in Figure 6. 
From left to right, Figure 6(a) shows natural scene images, 
detected traffic signs and their binary inner images. Figure 6(b) 
shows recognition results based on Hu moment. Figure 6(c) 
shows recognition results based on Tchebichef moment. Figure 
6(d) shows recognition results based on Zemike moment. 
Figure 6(e) shows recognition results based on author’s central 
projection vector. 
Vo*** 
60 
50 
50 
0 
0 
■/it;- 
0 
@) 
# 
# 
• 
A 
A 
A 
A 
A 
A 
60 
60 
m 
0 
m 
60 
<¿j); 
0 
0 
0 
• 
• 
» 
(a) (b) (c) (d) (e) 
Figure 6. Compared experimental results between author’s 
recognition method based on central projection vector and other 
methods based on invariant moments 
From experimental results, we can see that the recognition rate 
of proposed traffic sign recognition method is over 98%. This 
recognition rate is higher than that of other methods based on 
invariant moments, which shows that central projection 
transformation can obtain better effect on feature representation 
of traffic signs than invariant moments. The shape feature