The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
628 
we present a new shape feature computation method for traffic 
sign recognition based on central projection transformation. 
Extracted shape feature can reflect the global feature of traffic 
sign and stay invariant to object scales and rotations. Research 
object in this paper mainly includes three kinds of traffic signs: 
yellow warning signs, red prohibition signs and blue mandatory 
signs. Experimental results show that author’s method has a 
higher recognition rate for traffic sign recognition. 
2. DESCRIPTION OF THE METHOD 
After traffic signs are detected from natural scene image by the 
method based on combination of colour and shape features 
(Zhang, et al., 2007), Self-adaptive image segmentation is 
firstly used to extract binary inner images of detected traffic 
signs. Then one-dimensional feature vectors of inner images are 
computed by central projection transformation. Lastly, for each 
detected traffic sign, its feature vector is input to the trained 
probabilistic neural networks (PNN) for exact classification, the 
output of PNN is final recognition results. 
pattern recognition. In this paper, we use central projection 
transformation (Tao, et ah, 2001) to compute feature vector for 
inner image of traffic sign. Through feature computation, we 
can transform two-dimensional image to one-dimensional 
vector. For the binary inner image BJmg of traffic sign, m{jc 0 , y 0 ) 
is its centroid, and M denotes the maximal distance between 
each pixel of the inner image and its centroid. Then central 
projection transformation on inner image can be computed as 
following equations. 
M 
f( e k) = Z e k ,^sin e k ) (2) 
r = 0 
M = Max ||B Img (x, y) - m (x 0 , y 0 )|| 
Where, \\BJmg(x,y)-m(xo,yo)\\ represents the Euclidean distance 
between any point and the centroid of inner image. 0 k — 
k*(27dN)e[0,27t\, ¿=0,1, 2N is the number of projection 
rays during central projection transformation, piycosd, ysin#) 
represents the gray value of the pixel at coordinates (ycos#, 
ysin<9) in the Cartesian coordinates frame. 
2.1 Extract binary inner image of traffic sign 
Each type of traffic signs consists of special outline and inner 
image with specific pattern. If the inner image of traffic sign is 
effectively extracted, it can provide a stable basis for traffic 
sign recognition. Within detected traffic sign region in natural 
scene image, the binary inner image of traffic sign B Jmg can be 
effectively extracted by self-adaptive image segmentation. The 
formulas of image segmentation for extraction of binary inner 
image are described as follows. 
0, if (R(i + Hh, j + Lh) < T) 
[1 ,if(R(i+Hh,j + Lh) > T) 
T = a x pjRhd + b x MinRhd 
BJmg (i,j) = 
(1) 
Where, Hh, Lh respectively represents row and column 
coordinates of top and left comer of the traffic sign region. 
R(i+Hh, j+Lh) shows pixel’s R channel gray value on location 
(i+Hh, j+Lh) of original natural scene image. pjRhd, MinRhd 
respectively represents average value and minimal value of all 
pixels’ R channel gray value in traffic sign region. Coefficient a, 
b are definite value in the interval [0, 1]. 
Figure 1 shows some examples of extracted binary inner image 
of traffic sign from different natural scene images. Figure 1(a) 
shows some detected traffic signs from natural scene images, 
Figure 1(b) shows corresponding binary inner images of traffic 
signs. 
/MÆ 
1 
r T 
Ï2Ü S 
1 
■AAA 
□ 
lì 
M 
(a) Some detected traffic signs from natural scene images 
-i«** x. + # x Ta IS ? 
80 *1 <£ ^ 
© ^ « *1 *1 t t O A % 
(b) Binary inner images of traffic signs in Fig.l (a) 
Figure 1. Examples of traffic signs and their inner images 
2.2 Compute feature vector of inner image 
In pattern recognition, features are used to distinguish one 
pattern from the other. Feature extraction is the key step for 
The central projection vector (J{0\)J{&i),...,A^n)) is the feature 
vector used for pattern recognition. In practical application, 
each element in feature vector f[0 N )) should be 
normalized by the length of vector. Thus the feature vector used 
to recognize traffic signs is (A#i/’($v)) in this paper. 
/ (0* ) = m )J Jè (m ) x m » (3) 
From the definition of central projection vector, we can see that 
the number of N will influence the quality of central projection 
vector. If the number of N is too small, a lot of pixels in the 
binary inner image can’t be projected, which will lead to 
insufficient statistical information for traffic sign recognition. 
Otherwise, too large N will lead to complex computation and a 
lot of computing time. Thus the optimal projection number can 
acquire balance between projection quality and computing time. 
But how to confirm the optimal projection number can’t be 
solved from theoretical aspect in previous research. In this 
paper, we use the theory of information entropy to solve the 
problem of confirming optimal projection number in central 
projection transformation. 
Our main thought of confirming optimal projection number is 
that the information entropy of central projection vector will 
increase with the increment of projection number N. But after 
projection number N increases to a certain large number, the 
ratio of two neighbouring information entropies should 
gradually reach a constant. When the ratio of two information 
entropies approximately reaches a constant, number N is 
considered as the optimal projection number. The steps of 
confirming optimal projection number are described as follows. 
(1) Select three standard traffic signs in national standard as 
experimental data (Shown in Table 1). 
Type 
Binary inner 
image 
Meaning of sign 
Warning sign 
Caution pedestrian crossing 
Prohibition sign 
►sa- 
No honking 
Mandatory sign 
Y 
Turn left and right 
Table 1. Experimental data for confirming optimal number A