627 
AUTOMATIC RECOGNITION OF TRAFFIC SIGNS IN NATURAL SCENE IMAGE 
BASED ON CENTRAL PROJECTION TRANSFORMATION 
Ka Zhang 3 , Yehua Sheng 3 , Peifang Wang b , Lin Luo c , Chun Ye a , Zhijun Gong d 
J Key Laboratory of Virtual Geographic Environment, MOE, Nanjing Normal University, No. 1, Wenyuan Road, 
Nanjing, P. R. China, 210046, - zhangka81@126.com, -(shengyehua, yechun)@njnu.edu.cn 
b Liuji Secondary School of Wuhe County, Liuji Town, Wuhe County, Anhui Province, P. R. China, 233300, 
- wangpfwh@126.com 
c Wuxi City Communication Engineering Ltd., No. 888, Taihu East Avenue, Wuxi, Jiangsu Province, P. R. China, 
214026, -luolin8398@163.com 
d Nanjing Lear Xindi Automotive Interiors System Co, Ltd., No. 376, Hebancun, Maigaoqiao, Nanjing, P. R. China, 
210028,-gzj4573@126.com 
KEY WORDS: Vision sciences, Photogrammetry, Recognition, Image, Transformation, Neural, Networks 
ABSTRACT: 
Considering the problem of automatic traffic signs recognition in natural scene image (mainly including three kinds of traffic signs: 
yellow warning signs, red prohibition signs and blue mandatory signs), a new method for traffic signs recognition based on central 
projection transformation is proposed in this paper. In this method, self-adaptive image segmentation is firstly used to extract binary 
inner images of detected traffic signs after they are detected from natural scene images. Secondly, one-dimensional feature vectors 
of inner images are computed by central projection transformation. Lastly, these vectors are input to the trained probabilistic neural 
networks (PNN) for exact classification, the output of PNN is final recognition result. The new method is applied to 221 natural 
scene images taken by the vehicle-bome mobile photogrammetry system in Nanjing at different time. Experimental results show a 
recognition rate of over 98%. Especially, the problem of confirming optimal projection number in central projection transformation 
is solved by the information entropy in this paper. Moreover, the proposed recognition method is compared with other recognition 
methods based on three kinds of invariant moments. Results of contrastive experiments also show that the method proposed in this 
paper is effective and reliable. 
1. INTRODUCTION 
With the development of society and economy in recent years, 
the problem of traffic safety and traffic jam is becoming more 
and more serious. Thus the development and application of 
intelligent transportation system (ITS) have been attached great 
importance by governments and academia. Traffic sign 
automatic recognition system, which is an important sub-system 
in intelligent transportation system, has become one of the hot 
spots in the field of ITS, and it’s also a very difficult problem in 
real scene image recognition. How to detect traffic signs from 
complex real scene images rapidly and efficiently is the key 
step for traffic sign automatic recognition. Colour and shape 
features are usually used to detect traffic signs. The method of 
edge extraction and shape analysis is a common one for traffic 
sign detection (Loy, et al, 2004; Alefs, et al, 2007). In addition, 
the method based on combination of colour and shape features 
is another common one for traffic sign detection. In this kind of 
method, colour segmentation is firstly used to eliminate 
background-objects and obtain binary image. Then shape 
analysis is used to detect traffic sign regions in the segmented 
binary image (Gao, et al, 2006; Filipe, et al, 2007; Zhang, et al, 
Once candidate regions of traffic signs are detected from the 
image by colour segmentation or edge detection, traffic signs 
can be recognized according to their shape features. Hsu firstly 
used template images in database to train matching pursuit (MP) 
filters to find a set of best MP filter bases of each road sign, 
then used trained MP filter bases to find the best match between 
detected road sign and template sign generated by MP filter 
bases(Hsu, et al, 2001). Perez firstly carried out Fourier 
transformation to input image and template image, then 
computed similarity measure between them by nonlinear 
correlator for road sign recognition (Perez, et al, 2002). 
Escalera carried out normalization processing and binary 
processing on detected road signs first, then input binary image 
to the adaptive resonance theory paradigm neural network for 
road sign recognition (Escalera, et al, 2003). Fang presented an 
automatic road sign detection and recognition system based on 
a computational model of human visual recognition processing 
(Fang, et al, 2004). Paclik proposed a novel concept of a 
trainable similarity measure, which can alleviate some 
shortcomings of traditional cross correlation similarity. He used 
this new similarity measure to recognize prohibition signs 
(Paclik, et al, 2006). Maldonado-Bascon used DtB, which is 
distance vector composed of distances between edges of road 
signs and borders of sign candidate regions, as the shape feature 
of road sign, then input the feature to the trained support vector 
machines for road sign recognition (Maldonado-Bascon, et al, 
2007). Apart from the above methods, many methods based on 
template matching, neural networks and genetic algorithm are 
used for traffic sign recognition (Miura, et al, 2000; Liu, et al, 
2002; Gil-Jimenez, et al, 2007; Kuo, et al, 2007; etc.). 
However, the main shortcoming of above methods is that these 
methods need the same size between detected real sign and 
reference sign. Detected sign regions must be normalized to the 
uniform size, which will damage the original information of 
detected signs and influence the recognition rate. In this paper,