Abstract

We apply multiple-object correlation filters with limited distortion invariance to the recognition of road signs. The aspects of the signs in a real scene are dependent on the angle and distance of observation. Therefore, the scale and off-plane rotation invariance are necessary for correct detection of the signs in the scene. If a road sign is enough distant from a camera, the rotation and blur of the sign images in a camera frame are negligible. In that case, the variations of scale and aspect ratio of the sign have to be considered. We use the technique of one dimensional Mellin harmonics decomposition¹ to obtain the horizontal scale invariance and apply the method of synthetic discriminant function² to the adaptation for the scale change. By limiting the range of the invariances, it is possible to calculate the correlation filter which has expected correlation response for pattern classification.

© 1995 IEEE