Mobile detection of traffic infrastructure

Abstract

This paper presents a mobile mapping system that aims at automatic detection of traffic infrastructure from image sequences. The visual information will be georeferenced with GPS/INS information to represent a most realistic model of the environment in a GIS database. The integration of innovative methods such as vertical structure segmentation and probabilistic object matching provides a qualified framework for the detection of arbitrary traffic infrastructure. The detection of infrastructure starts with a segmentation of the video frame into regions of interest by exploiting track localization and a-priori knowledge about the visual scene. A further step concerns the extraction of 3D structure and the segmentation of its associated distance map w.r.t. vertically accentuated objects. Color information derived from learned classification filters contributes to a characterization of class-specific support regions for further processing. Eventually, traffic signs and lights are robustly identified by probabilistic matching using a RBF neural network that was trained from real imagery of traffic infrastructure.