Road Sign Recognition by Means of the Behavioral Model of Vision

Algorithms and procedures to solve the task of road sign detection and recognition invariant of viewing conditions and results of testing during computer simulation with British and Russian signs are presented. After preliminary colour segmentation of initial real world images and classification according to road sign colours and external forms, biologically plausible Behavoiral Model of Vision (BMV) [1,2] which was modified under task, identified correctly about 80% potential traffic sign images for various weather conditions, shading, and other transformations. Possible ways to increase the model performance are considered. Introduction. An important task in developing intelligent systems of driver support and traffic safety is road sign detection and recognition [3-5]. At present, the problem of traffic signs recognition invariantly to their possible transformations in real world conditions has no effective solution in the frameworks of standard computer vision approaches. Evidently, some aspects of this problem may be solved by means of biologically plausible approach in image recognition [1,2,4]. In the present paper, modified algorithms of the BMV [1,2], based on imitation of foveal vision properties in the real visual system, and results of their testing for traffic sign recognition under various viewing and weather conditions are presented. 1. Algorithms and procedures. Colour Segmentation. Images from the standard databases are firstly utilised to find the range of colour vectors for the colours used in the signs, mainly red, blue, black and white. The ranges for each colour vector, e.g., (red, lightness, chroma) and (blue, lightness, chroma), are found by calculating the values by use of the CIECAM97 model [6] and are plotted on a u’v’ chromaticity diagram after conversion. During the study, images of British road signs have been taken using a digital camera, Olympus Digital Camera C-3030, Russian real-world images were got by Fuji photo camera and then scanned (450 dpi). These images are then classified visually according to the viewing and weather conditions, such as cloudy, sunny, etc. Based on the images in each group, the parameters for each viewing condition are found from [7] (e.g., direct sun light with colour temperature 5335K and light from overcast sky with colour temperature 6500K) for application of the colour appearance model. Images taken under real viewing conditions are then transformed from RGB space to CIE XYZ values and then to LCH (Lightness, Chroma, Hue) by use of the model of CIECAM97. Based on the range of sign colours, traffic-signs-to-be are segmented from the rest of scenes for further identification. Classification of traffic signs according to their external form. For all signs, both from standard databases and from real world images, preliminary classification according to the colour, their external form (circle, rectangle, or triangle) can be determined by means of histograms of orientations detected at resolution level 3 (RL 3). RL 3 is emulated by Gaussian convolution (kernel size is equal 9). Each sign with a certain external form (in spite of its inner content) has characteristic relationship of horizontally, vertically, and obliquely oriented segments at RL 3. In particular, all oriented elements had nearly equal representation for circle signs contrary to rectangle signs (Fig. 1) that had preferable horizontal and vertical orientations (in sum, more than 50% of all oriented segments). For each external form, quantitative estimations were obtained for classification into particular groups of signs from British and Russian standard databases and from real world images.