Traffic sign detection in static images using Matlab

In this paper a system for off-line traiiic sign detection is shown. Matlab-Image-processing toolbox is used for this purpose. The vision-based traffic sign detection module developed in this work manages 172x352 color images in RGB (Red, Green, Blue) format. The first step in the algorithm is to obtain the gradient image and its vertical edge projection. In a second step, a color and shape analisys is performed. I . INTRODUCTION Traffic sign detection and recognition have experimented increasing importance in the last times. This is due to the wide range of applications where this system can be used, as for instance, in intelligent vehicles, driver support systems, etc. There are four types of traffic signs in the traffic code: prohibition. warning, obligation and informative. Depending on the shape and color, the warning signs are equilateral triangles with one vertex at the top. Prohibition signs are circular having,a specific figure in each case over a white or blue background and a red' border. To indicate' obligation, signs are circular, with a figure over a blue background. One of the greatest inconveniences of the RGB color space is that it is very sensitive to changes in light [I]. This is the reason why other color spaces are used in computer vision applications, especially the hue, saturation, intensity (HSI) one. This system keeps high immunity to changes in light [2]. The problem with HSI is that its formulas are nonlinear, and the computational cost is prohibitive. Instead, we have used the relation between the RGB components, as this work is proposed for real-time systems and no further processing is needed after digitalization. .. To detect a traffic sign in an image, the algorithm follows I ) Candidate image regions are obtained, by accumulating vertical and horizontal edge projections. 2) Candidate image regions are validated as follows: Red image thresholding. for prohibition sign. Blue image thresholding, for obligation sign. Blob shape analysis. Circular ring template. these steps: 11. BORDERIMAGE The appropriated choice of the color features to use in the process is of crucial importance in order to attain proper and fast detection. Accordingly, only the Red eoniponent is considered as it provides a high capacity for color discrimination in the visual analysis of t rafk signs and no further processing is needed after digitalization. In an attempt to carry out a preattentive strategy, a coarse analysis of vertical edge is performed in a first stage based on differential characteristics computed on the Red component of the image using a gradient filter [3]: as depicted in Fig. I, where two images containing trafiic signs are illustrated together with their associated gradient image applying a vertical edge operator. The gradient image is computed as expressed in (I). Where Gx and Gy denote a Prewitt extended operator, using a 5x5 mask. . . . . . . . _._, f . ' , , , . .. > . . . j : . . ' . . . . . . . . 2,. i . . . . . . . . . . . . . . ." 1. d <, %' " 0-7803-7937-3/'03/'$17.00 02003 IEEE 212 Authorized licensed use limited to: Univ de Alcala. Downloaded on July 20, 2009 at 10:22 from IEEE Xplore. Restrictions apply. 111. CANDIDATE IMAGE REGIONS. Vertical projection of border pixel. One of the most common techniques for trafiic sign segmentation is to use grey-level images, red component in OUT case, and to project pixels at the edges onto the axes. Vertical projections of different types of signs are shown in Fig. 2. As can be observed, a maximum occurs in the area of the image where the sign is placed. 1'9 w Fig. 2. Accumuiated verricoiprojeciion on gradient image. (0) Prohibition (a), rpeciol case, end oJprohibilion (c) ~viirning, id) obligotim As a first step, an adaptive thresholding is performed aiming at removing the common ofl'set component in the projection profile [4]. For this purpose. a threshold U is computed as expressed in (2) where p v stands for the average value of the projection profile, while p.* represents the average of all points in the projection whose value is greater than p. The resulting threshold is depicted in Fig. 2. Finally. the coarse detection phase ends hy removing narrow peaks from the projection profile. This yields a set of candidate image regions that highly reduces and constraints the portions of the image where traffic signs are likely to appear, as depicted in Fig. 3. U I n m I o a m m m ~ ~ m m m m m 1cJ (dl Fiz. 3. Candidole image region aJiw verlicolprojeclion. Horizontal projection of border pixel. In order to restrict the area of interest a bit more, the same method, previously explained, is applied to horizontal projections of edge pixels in the region of interest. In this case the adaptive threshold is obtained as expressed in equation (4)