Traffic sign detection system via multiple feature cooperation

We present a system to tackle challenges in traffic sign detection due to individual or multiple
difficulties: chromatic aberration, geometric distortions and scale variations. Our multiple
feature cooperation include two schemes, one for single images and the other for stereoscopic image
sequences. The first scheme is adopted to extract regions of interest (ROIs) of possible traffic signs
by integrating chromatic and scale-normalized geometric features. ROIs of potential traffic signs are
then refined using estimated depth-size ratios by the second scheme. Furthermore, via our multiple
feature cooperation schemes, traffic signs are detectable at different distance ranges as they
approach the vehicle, requiring no offline preprocessing, nor 3D reconstruction of potential traffic
signs. Experiments are conducted on three public datasets, two of which are regrouped based on the
appearance of traffic signs into five subsets: normal situation, chromatic aberrations, geometric
distortions, scale variations and entire situations. Our results demonstrate that the first scheme
gives high detection completeness, and the second scheme improves detection exactness significantly,
compared to ground-truth traffic signs.