Driving environment recognition technologies are very important for driver support systems in Intelligent Transport Systems (ITS). Several approaches have been proposed in order to detect road signs so far. However, they could not cope with some bad lighting conditions. The HSL transformation considering brightness or an adaptive limitation of color range has been proposed for several bad lighting conditions. Nevertheless, they were insufficient for night images or for images with strong contrast. Moreover, the previous reports were aimed at extracting only regulatory and warning signs whose shape is regulated and aspect ratio is 1:1. However, the shape of the guide signs targeted in this report is rectangular and the aspect ratio depends on the sign, because the number of characters on the guide sign varies with each sign. A detection method for green guide signs was aimed at selecting the video frame that included the sign with the largest size among past successive frames. Therefore, the method was not effective to extract signs with a small size from a distance for driver support. A real-time guide sign recognition system using multiple binarization threshold determined in advance for various lighting conditions was proposed. However, the performance under bad lighting conditions was poor. The authors are studying a method of extracting blue guide signs in various lighting conditions. The target guide sign has three characteristics, that is, color (background color is blue and character is white), shape (rectangle) and location (overhead and side of the road). However, it is difficult to detect road region reliably because of a lot of traffic. In this paper, therefore, they examine guide sign extraction using color and shape information. Their method consists of color segmentation and sign extraction procedures. The color segmentation procedure uses a histogram on a color space without defining a color range peculiar to the sign in advance, because the sign colors are changed significantly by sunshine. The sign extraction procedure uses a new parameter that represents the likelihood of a rectangle. Lastly, they show some experimental results and consider some problems.
Abstract