SEGMENTATION of ROAD SIGN SYMBOLS using OPPONENT-COLOR FILTERS

A novel algorithm for the recognition of Japanese road sign symbols with red annular boundaries is proposed. After detecting a red annular object using HSV color coordinates, a biologically inspired opponent-color filter is applied to extract the symbol parts of road signs. By applying discriminant analysis to the filter output, the inside region circumscribed by red annular boundary can be extracted. By applying the discriminant analysis again to the inside region, the symbol part of road sign can be segmented. The extracted symbol is classified by a decision tree and recognition rate of about 96% was achieved for real road video images. It is shown that the opponent-color filter can lead to accurate recognition even for small and degraded road signs in video images. INTRODUCTION Automatic recognition of road signs for the assistance of safety driving is one of the most important components in the intelligent transportation system (ITS). Although many vision systems for the automatic recognition of road signs have been proposed in the last decades[5],[3], they are still at the experimental level because of many difficulties in the task: Variations in lighting conditions, weather conditions, motion and/or focus blurring, degeneration of paints, etc. One way to challenge these difficulties is to learn how our visual system solves these problems, since our vision is the most reliable and robust system in the world. In this paper, as the first step towards a biologically inspired road sign recognition system, opponent-color filters found in our visual systems are applied to the segmentation of symbols in road signs and it is shown that the filters can extract symbols reliably. This paper first describes algorithms for the segmentation, extraction and classification of road signs. Then, the performance of the system measured by real video images will be discussed and analysed. 1 Presented at ITSWC2004 Nagoya 18-22, October 2004 VISION SYSTEM In this section, video camera and PC, algorithms for the segmentation, extraction and classification of road sign will be described. VIDEO CAMERA AND THE SYSTEM The traffic images analyzed in this paper have been acquired by a video camera (SONY DCRTRV900 NTSC) mounted on a tripod inside a car. Road signs not only stand on the left or the right side of road, but also hang over lane. Overhead signs are common and even hang over opposite lane as shown in Figure 1(a). Since both overhead and side signs are our targets, the focus of video camera is set at the widest angle (f=4.3mm), so that each image of road sign becomes inevitably small (from 20 to 60 pixel wide). Video images are captured through an IEEE 1394 DV terminal on PC (Dell Precision 340, 2.8GHz, 2GB memory), and off-line but video-rate processing has been done on the PC. (a) (b) Figure 1: (a) A scene of overhead and side signs. (b) Nine road signs used in this paper. From left to right and from top to bottom row: Speed limits to “30Km”, “40Km” and “50Km”, “No Parking”, “No Parking and Standing”, “No Passing”, “Close to Traffic”, “No Throughfare” and “No U-turn”. DETECTION OF ROAD SIGNS In this paper we focus on nine road signs shown in Figure 1(b) whose boundaries are composed of red annuli, and try to thoroughly investigate difficulties in this task. Extension to other types of road signs may be less difficult than this task, since most of the difficulties encountered in this task will be common to the other signs. The process of road sign recognition consists of three stages: (1) Detection of red circular objects, (2) segmentation of symbols in the objects and (3) classification of the symbols. Detection of red objects Detection of red circular objects is carried out first by segmenting red parts in a scene according to HSV color coordinates. When a color is defined by RGB color coordinates as (R,G,B), where R, G and B are between 0.0 and 1.0, and let MAX = max{R,G,B} and MIN = min{R,G,B}, the transformation from RGB to HSV coordinates is given by