Generalized Traffic Sign Detection Model for Developing a Sign Inventory

Abstract

The hundreds of traffic sign types on the road and their various shapes and colors make it difficult to develop a generalized method of traffic sign detection. Consequently, agencies performing a sign inventory must manually review millions of roadway video log images. This paper proposes an innovative image processing model that automatically detects traffic signs and dramatically reduces the sign inventory workload. In a test of the proposed model using 37,640 images provided by the Louisiana Department of Transportation and Development, 86 percent of the manual review efforts can be effectively saved. Our method is composed of (1) a generalized traffic sign model to represent the entire class of traffic signs; (2) a proposed new statistical traffic sign color model; (3) a traffic sign region of interest detection system using polygon approximation; and (4) traffic sign candidate decision rules based on shape and color distributions.